Big Data and Cloud Computing: Select Proceedings of ICBCC 2022 9819910501, 9789819910502

Citation preview

Lecture Notes in Electrical Engineering 1021

Neelanarayanan Venkataraman Lipo Wang Xavier Fernando Ahmed F. Zobaa   Editors

Big Data and Cloud Computing Select Proceedings of ICBCC 2022

Lecture Notes in Electrical Engineering Volume 1021

Series Editors Leopoldo Angrisani, Department of Electrical and Information Technologies Engineering, University of Napoli Federico II, Napoli, Italy Marco Arteaga, Departamento de Control y Robótica, Universidad Nacional Autónoma de México, Coyoacán, Mexico Samarjit Chakraborty, Fakultät für Elektrotechnik und Informationstechnik, TU München, München, Germany Jiming Chen, Zhejiang University, Hangzhou, Zhejiang, China Shanben Chen, School of Materials Science and Engineering, Shanghai Jiao Tong University, Shanghai, China Tan Kay Chen, Department of Electrical and Computer Engineering, National University of Singapore, Singapore, Singapore Rüdiger Dillmann, Gebäude 07.21, University of Karlsruhe (TH) IAIM, Karlsruhe, Baden-Württemberg, Germany Haibin Duan, Beijing University of Aeronautics and Astronautics, Beijing, China Gianluigi Ferrari, Dipartimento di Ingegneria dell’Informazione, Sede Scientifica Università degli Studi di Parma, Parma, Italy Manuel Ferre, Centre for Automation and Robotics CAR (UPM-CSIC), Universidad Politécnica de Madrid, Madrid, Spain Faryar Jabbari, Department of Mechanical and Aerospace Engineering, University of California, Irvine, CA, USA Limin Jia, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China Janusz Kacprzyk, Intelligent Systems Laboratory, Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland Alaa Khamis, Department of Mechatronics Engineering, German University in Egypt El Tagamoa El Khames, New Cairo City, Egypt Torsten Kroeger, Intrinsic Innovation, Mountain View, CA, USA Yong Li, College of Electrical and Information Engineering, Hunan University, Changsha, Hunan, China Qilian Liang, Department of Electrical Engineering, University of Texas at Arlington, Arlington, TX, USA Ferran Martín, Departament d’Enginyeria Electrònica, Universitat Autònoma de Barcelona, Bellaterra, Barcelona, Spain Tan Cher Ming, College of Engineering, Nanyang Technological University, Singapore, Singapore Wolfgang Minker, Institute of Information Technology, University of Ulm, Ulm, Germany Pradeep Misra, Department of Electrical Engineering, Wright State University, Dayton, OH, USA Subhas Mukhopadhyay, School of Engineering, Macquarie University, NSW, Australia Cun-Zheng Ning, Department of Electrical Engineering, Arizona State University, Tempe, AZ, USA Toyoaki Nishida, Department of Intelligence Science and Technology, Kyoto University, Kyoto, Japan Luca Oneto, Department of Informatics, Bioengineering, Robotics and Systems Engineering, University of Genova, Genova, Genova, Italy Bijaya Ketan Panigrahi, Department of Electrical Engineering, Indian Institute of Technology Delhi, New Delhi, Delhi, India Federica Pascucci, Department di Ingegneria, Università degli Studi Roma Tre, Roma, Italy Yong Qin, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China Gan Woon Seng, School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore, Singapore Joachim Speidel, Institute of Telecommunications, University of Stuttgart, Stuttgart, Germany Germano Veiga, FEUP Campus, INESC Porto, Porto, Portugal Haitao Wu, Academy of Opto-electronics, Chinese Academy of Sciences, Haidian District Beijing, China Walter Zamboni, Department of Computer Engineering, Electrical Engineering and Applied Mathematics, DIEM—Università degli studi di Salerno, Fisciano, Salerno, Italy Junjie James Zhang, Charlotte, NC, USA

The book series Lecture Notes in Electrical Engineering (LNEE) publishes the latest developments in Electrical Engineering—quickly, informally and in high quality. While original research reported in proceedings and monographs has traditionally formed the core of LNEE, we also encourage authors to submit books devoted to supporting student education and professional training in the various fields and applications areas of electrical engineering. The series cover classical and emerging topics concerning: • • • • • • • • • • • •

Communication Engineering, Information Theory and Networks Electronics Engineering and Microelectronics Signal, Image and Speech Processing Wireless and Mobile Communication Circuits and Systems Energy Systems, Power Electronics and Electrical Machines Electro-optical Engineering Instrumentation Engineering Avionics Engineering Control Systems Internet-of-Things and Cybersecurity Biomedical Devices, MEMS and NEMS

For general information about this book series, comments or suggestions, please contact [email protected]. To submit a proposal or request further information, please contact the Publishing Editor in your country: China Jasmine Dou, Editor ([email protected]) India, Japan, Rest of Asia Swati Meherishi, Editorial Director ([email protected]) Southeast Asia, Australia, New Zealand Ramesh Nath Premnath, Editor ([email protected]) USA, Canada Michael Luby, Senior Editor ([email protected]) All other Countries Leontina Di Cecco, Senior Editor ([email protected]) ** This series is indexed by EI Compendex and Scopus databases. **

Neelanarayanan Venkataraman · Lipo Wang · Xavier Fernando · Ahmed F. Zobaa Editors

Big Data and Cloud Computing Select Proceedings of ICBCC 2022

Editors Neelanarayanan Venkataraman School of Computer Science and Engineering Vellore Institute of Technology Chennai, Tamil Nadu, India Xavier Fernando Department of Electrical, Computer, and Biomedical Engineering Toronto Metropolitan University Toronto, ON, Canada

Lipo Wang School of Electrical and Electronic Engineering Nanyang Technological University Singapore, Singapore Ahmed F. Zobaa Electronic and Electrical Engineering, College of Engineering, Design and Physical Sciences Brunel University London Uxbridge, UK

ISSN 1876-1100 ISSN 1876-1119 (electronic) Lecture Notes in Electrical Engineering ISBN 978-981-99-1050-2 ISBN 978-981-99-1051-9 (eBook) https://doi.org/10.1007/978-981-99-1051-9 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore

Contents

Data Security An Integrated Approach for Big Data Classification and Security Using Optimized Random Forest and DSSE Algorithm . . . . . . . . . . . . . . . S. Castro and R. Pushpa Lakshmi Secure Web Gateway on Website in Cloud . . . . . . . . . . . . . . . . . . . . . . . . . . . Daljeet Kaur, Celestine Iwendi, Thaier Hamid, and Pradeep Hewage Design a Quantum Cryptography Algorithm and Evaluate the Risks of Quantum-Based Nano Computing . . . . . . . . . . . . . . . . . . . . . . . Thilagavathy

3 19

31

Machine Learning Human Odor Security Using E-nose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V. Anush Kumar, C. S. Manigandaa, S. Dhanush Hariharan, Challapalli Manikantaa, G. Saranya, and V. D. Ambeth Kumar

51

Automated Road Surveillance System Using Machine Learning . . . . . . . . Ashish Vishnu, S. Sushmitha, Tina Susan Jacob, A. David Maxim Gururaj, and S. Dhanasekar

67

An Artificial Intelligence-Based Technique to Optimize Hybrid Vehicle Using Renewable Energy Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . S. Dhanush Hariharan, Challapalli Manikantaa, C. S. Manigandaa, V. Anush Kumar, Vetri Priya, and V. D. Ambeth Kumar Stock Market Prediction Using Machine Learning Techniques: A Comparative Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . P. Chinthan, Rashmi Mishra, B. Prakash, and B. Saleena

81

93

Comparative Study on Different Intrusion Detection Datasets Using Machine Learning and Deep Learning Algorithms . . . . . . . . . . . . . . 109 G. Aarthi, S. Sharon Priya, and W. Aisha Banu v

vi

Contents

Big Data Analytics Mitigating Postnatal Depression: A Big Data Self-help Therapy . . . . . . . . 123 Asma Usman, Francis Morrissey, Thaier Hamid, Celestine Iwendi, and F. Anchal Garg Improving Learning Effectiveness by Leveraging Spaced Repetition (SR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 Aamir Mazhar Abbas, Thaier Hamid, Celestine Iwendi, Francis Morrissey, and Anchal Garg Snooping for Fake News: A Cascaded Approach Using Stance Detection and Entailment Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 Ebenezer Ajay Williams, M. Karthik, A. Shahina, Harshithaa Murali, N. Safiyyah, and A. Nayeemulla Khan Early Planning of Virtual Machines to Servers in Cloud Server Farms is an Approach for Energy-Efficient Resource Allocation . . . . . . . 179 P. Kumar and S. VinodhKumar Performance Analysis of Distributed Algorithms for Big Data Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 K. Haritha and M. V. Judy IoT Achieving Sustainability by Rectifying Challenges in IoT-Based Smart Cities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211 Neha Bhardwaj, Celestine Iwendi, Thaier Hamid, and Anchal Garg Framework for Implementation of Smart Driver Assistance System Using Augmented Reality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 K. Baskar, S. Muthuraj, S. Sangeetha, K. Vengatesan, D. Aishwarya, and P. S. Yuvaraj IoT-Based Mental Health Monitoring System Using Machine Learning Stress Prediction Algorithm in Real-Time Application . . . . . . . 249 Md Abdul Quadir, Saumya Bhardwaj, Nitika Verma, Arun Kumar Sivaraman, and Kong Fah Tee Fish Feeder System Using Internet of Things . . . . . . . . . . . . . . . . . . . . . . . . . 265 Jyotsna Malla, Preetham Lekkala, Rishi Raghu, J. Jayashree, J. Vijayashree, and Vicente Garcia Diaz An Efficient and Recoverable Symmetric Data Aggregation Approach for Ensuring the Content Privacy of Internet of Things . . . . . . 279 L. Mary Shamala, V. R. Balasaraswathi, M. Shobana, G. Zayaraz, R. Radhika, and Thankaraja Raja Sree

Contents

vii

Wireless Networks Implementation and Comparative Analysis of Various Energy-Efficient Clustering Schemes in AODV . . . . . . . . . . . . . . . . . . . . . . . 297 S. Dhanabal, P. William, K. Vengatesan, R. Harshini, V. D. Ambeth Kumar, and S. Yuvaraj Adaptive End-To-End Network Slicing on 5G Networks . . . . . . . . . . . . . . . 319 P. Sakthi Saravanakumar, E. Mahendran, and V. Suresh Zero Trust Framework in Integrated Cloud Edge IoT Environment . . . . 331 S. Kailash, Yuvaraj, and Saswati Mukherjee Ph.D. Track Paper O2 Q: Deteriorating Inventory Model with Stock-Dependent Demand Under Trade Credit Policy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345 P. Jayashri and S. Umamaheswari Embedding (K 9 − C 9 )n into Certain Necklace Graphs . . . . . . . . . . . . . . . . 359 Syeda Afiya and M. Rajesh Lyrics Generation Using LSTM and RNN . . . . . . . . . . . . . . . . . . . . . . . . . . . 371 Aarthi Dhandapani, N. Ilakiyaselvan, Satyaki Mandal, Sandipta Bhadra, and V. Viswanathan

About the Editors

Neelanarayanan Venkataraman received his Master of Science in Computer Science from Madurai Kamaraj University, India, in 1995 and his Ph.D. from IT University of Copenhagen, Denmark, in 2012. Currently, he is Professor at VIT University, Chennai, India. Before joining VIT University, he worked as Scientist at the Centre for Advanced Computing (CDAC), India, and as Lecturer at Madurai Kamaraj University, India, and its affiliated institutions. His areas of research include distributed computing such as grid and cloud computing, context-aware computing, network management and security, XML-based security technologies, and e-communities. He has initiated several international research collaborations with universities in Europe, Australia, and South Korea as Research Group Coordinator and Chief Investigator at VIT University. He was instrumental in initiating joint research collaboration between VIT University and industries such as CDAC, SET, and DLink. He served as Head of the department in Cyber-Physical Systems between 2019 and 2023. As head of the department, he was responsible for the design and development of curriculum and syllabi for undergraduate and postgraduate specialization programs. He was instrumental in setting cyber-physical systems lab and an industry 4.0 lab at VIT University, Chennai. Lipo Wang received a bachelor’s degree from the National University of Defense Technology (China) and a Ph.D. from Louisiana State University (USA). He is presently on the faculty of the School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore. His research interest is artificial intelligence with applications to image/video processing, biomedical engineering, communications, control, and power systems. He has 350+ publications, a US patent in neural networks, and a patent in systems to his credit. He has co-authored 2 monographs and (co-)edited 15 books. He was Keynote Speaker for 36 international conferences. He is/was Associate Editor/Editorial Board Member of 30 international journals, including 4 IEEE Transactions, and Guest Editor for 15 journal special issues. He was a member of the Board of Governors of the International Neural Network Society, IEEE Computational Intelligence Society (CIS), and the IEEE Biometrics Council. ix

x

About the Editors

Xavier Fernando is a Professor at the Department of Electrical and Computer Engineering, Ryerson University, Toronto, Canada. He has (co-)authored over 200 research articles, and two books (one translated into Mandarin), and holds a few patents and non-disclosure agreements. He is the Director of Ryerson Communications Lab. He was an IEEE Communications Society Distinguished Lecturer and delivered close to over 50 invited talks and keynote presentations all over the world. His research interests are in signal processing for optical/wireless communication systems. He mainly focuses on physical and MAC layer issues. He has a special interest in underground communications systems of cognitive radio systems, visible light communications, and wireless positioning systems. Ahmed F. Zobaa received his B.Sc. (Hons.), M.Sc., and Ph.D. degrees in Electrical Power and Machines from Cairo University, Egypt, in 1992, 1997, and 2002, respectively. He received his Postgraduate Certificate in Academic Practice from the University of Exeter, the UK, in 2010. He received the Doctoral of Science from Brunel University London, the UK, in 2017. Currently, he is Reader in electrical and power engineering. His main areas of expertise include power quality, (marine) renewable energy, smart grids, energy efficiency, and lighting applications.

Data Security

An Integrated Approach for Big Data Classification and Security Using Optimized Random Forest and DSSE Algorithm S. Castro

and R. Pushpa Lakshmi

Abstract The present digital era handles a massive amount of data every day from various sources. These enormous volumes are termed as big data which is heterogeneous, is dynamic, and includes numerous valuable insights. Handling massive raw data is quite complex, and there might be a chance to miss the important aspects. Data processing methods like clustering and classification models reduce the burden and handle the big data effectively. Recently, numerous clustering and classification models are evolved; however, attaining the maximum classification accuracy for better performance is the main objective of every research work. Similarly, big data security gains more attention equal to the classification process. Encryption procedures will enhance data security so that classified data can be stored securely in the cloud environment. In this research work, a big data classification approach is presented using a random forest algorithm and secured the classified results in the cloud using the DSS encryption technique to attain maximum accuracy and security. The features for the classification process are obtained through a whale optimization algorithm which selects the optimal features and enhances the classification accuracy. The proposed model attains enhanced performance in terms of 98.47% accuracy, 96.48% precision, 96.58% recall, and 96.53% F1-score. Also, the proposed DSS encryption attains better encryption and decryption performances in terms of throughput, encryption, and decryption time compared to existing Encrypting File System standard algorithm (EFSSA). Keywords Big data classification · Security · Encryption · Accuracy

S. Castro (B) Department of Information Technology, Karpagam College of Engineering, Coimbatore 641032, India e-mail: [email protected] R. Pushpa Lakshmi Department of Information Technology, PSNA College of Engineering and Technology, Dindigul 624622, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_1

3

4

S. Castro and R. Pushpa Lakshmi

1 Introduction Technology advancement and innovative applications produce massive amounts of data from different sources. This large volume of data is termed as big data not because of the size but also due to its unstructured, heterogeneous nature. The mixed information may include video, unstructured text, numeric data, audio, and other information. Processing such a large amount of data is quite complex so before processing the data in an application this large amount of data is clustered or classified for better utilization. Various clustering algorithms are employed so far to categorize the big data such as k-means clustering and hierarchical clustering [1]. Though the clustering algorithms will categorize the data, it does not provide any decision as like classifiers. The usefulness of data can be enhanced through classification approaches so that processing of unwanted, redundant data processing steps can be eliminated which parallelly increases the computation ability and reduce the computation time. Various machine learning algorithms like support vector machine, decision tree, linear regression, etc., [2] and neural network models like an artificial neural network, backpropagation neural network, and adaptive neuro-fuzzy inference systems are employed in the data classification process [3–5]. However, the accuracy of the classification process depends on its feature selection. Processing suitable features that are relevant to the application will only provide better accuracy; otherwise, the accuracy of the entire application becomes degraded. Due to this optimal feature selection, procedure is introduced which includes various nature-inspired and heuristic optimization algorithms to select the optimal features. Based on these observations, this research work is aimed to provide a better classification model with optimal feature selection using random forest and whale optimization algorithm to enhance the classification accuracy in the big data classification process. Similar to classification, big data security gains more attention. Since the data obtained from different sources may include privacy details and other information, the intruders try to make use of the data for their application. To avoid data security issues, encryption procedures will be the best choice [6, 7]. Moreover, the classified data stored in a cloud environment is vulnerable to various cloud attacks. So, to enhance data protection and security an encryption methodology is adopted in this research work. The contribution of the integrated approach is summarized as follows. • Presented an optimal feature selection procedure for big data classification using whale optimization algorithm and random forest classifier. • Presented a simple and efficient encryption procedure to secure the big data before moving into the cloud environment. • Presented intense experimental analysis of the proposed model with other machine learning algorithms as a comparative analysis. The rest of the article is arranged in the following order. A brief literature analysis is presented in Sect. 2, the proposed integrated classification and encryption approach is presented in Sect. 3, results and discussion are presented in Sect. 4, and the conclusion is presented in Sect. 5.

An Integrated Approach for Big Data Classification and Security Using …

5

2 Related Works A brief literature analysis of existing classification approaches for big data is presented in this section. Big data classification techniques in the healthcare domain improve the diagnosis accuracy and reduce the manual interventions. A radial basis function neural network model presented in [8] for healthcare big data classification includes manifold analysis and nearest neighbor propagation algorithm. Initially, the data is processed through a manifold algorithm and then the similarity index is adjusted using an exponential function. Finally, the clustering is performed to improve the classification accuracy of the neural network model. Improved accuracy and convergence speed are the features of the radial neural network model-based healthcare data classification approach. An improved k-nearest neighbor algorithm for big data classification is reported in [9] to overcome the limitations in the conventional KNN algorithm. The query instances and weights are assigned to each class in the improved classification algorithm. Also, a class distribution is considered for query instances to make sure that the weight functions are not affected by the outliers. Clustering is used to denoise the data and enhance the classification accuracy compared to the conventional classification approach. Various instance selection algorithms are developed to reduce the high computational requirements of the KNN algorithm. The instances are generally selected as a quantitative metric which may confuse the selection results. To overcome this issue, an instance selection algorithm is presented in [10] for KNN classification rules using evidence theory. Considering the evidence provided by the instances an estimation label is created for each instance, and then, it is combined to identify the conflictions. By removing or revising the conflicting instances, the boundary instances which are efficient to solve the optimization problems are utilized in the classification rules. A fuzzy optimized data management technique presented in [11] classifies the information dependency based on the relationship between data attributes. The attributes are segregated based on the similarity index to process the complex big data with minimum computation time. The presented approach is adopted for weather forecasting and attains better similar index and classification performances compared to existing forecasting applications. A similar fuzzy-based framework is presented in [12] to handle big data classification issues. The presented fuzzy inference system has the ability to make decisions based on the inter- and intraclass distances so that high computation efficiency is achieved in the classification process. An extended belief rule-based big data classification model reported in [13] reduces the computational complexity and increases the computing efficiency in the classification process. The presented approach introduces rule weight calculation, reasoning algorithm, and rule reduction methodology for multiclass classification. Better computation efficiency and time complexity are the observed merits of the presented classification model. The fault classification model presented in [13, 14] identifies the faults in industrial big data. The imbalance data classification problems

6

S. Castro and R. Pushpa Lakshmi

considered in the classification process are handled using k-means Bayes algorithm, and a map-reduce approach is employed to identify the faults in the industrial data. Feature selection is an important process in the classification model as the accuracy of the classification approach relies on processing the optimal features. Various feature selection models are introduced in the recent era based on optimization techniques. The oppositional firefly algorithm presented in [15] identifies the optimal features from a large dataset. The selection process of optimal features is related to the optimal solution of the optimization model. Better computation efficiency and accuracy are the observed merits of the optimal feature selection model. A binary crow search algorithm reported in [16–18] improves the classification performance by selecting optimal features. An opposition-based learning strategy has been incorporated to define the flight length parameter which selects the optimal features better than conventional feature selection procedures. From the above literature analysis, it is observed that optimal feature selection will improve classification accuracy. However, optimization models are quickly trapped into local optima which affect the optimal feature selection process and affect the classification accuracy. So, a suitable optimization algorithm should be selected along with a better classifier to handle the big data effectively. Security features should also be considered to protect data security and privacy. Considering this as research motivation, an integrated approach is presented in this research work for big data classification and encryption.

3 Proposed Work The proposed integrated approach for big data classification and security is presented in this section. The overview of the proposed approach is presented in Fig. 1. The procedure starts from the data preprocessing to remove the redundant features. Then, the essential optimal features are selected using the whale optimization algorithm (WOA). Compared to other optimization algorithms, the whale optimization algorithm avoids the local optima efficiently and has the ability to solve constrained and unconstrained problems. The selected optimal features are classified using a random forest algorithm. Further, the classified results are encrypted using DSS encryption algorithm before transferring the data into the database. The initial preprocessing technique is used to remove the redundant and irrelevant data in the dataset. Entropy calculations are employed to obtain the feature cross-correlation. Based on the correlation factors, the relevance between the data and redundancy is measured. The preprocessing step reduces the computation time in the classification process. Optimal feature selection is the next step after preprocessing. Identifying optimal features for the classification process is attained through a whale optimization algorithm. The nature-inspired algorithm is developed based on the prey-hunting characteristics of whales. The optimization model and proposed

An Integrated Approach for Big Data Classification and Security Using …

7

Optimal feature selection using Whale optimization algorithm

Preprocessing

Dataset Classification using Random Forest algorithm

Data encryption using DSS algorithm

Encrypted data to cloud

Classification Results

Fig. 1 Overview of the proposed model

work is related as follows. The features in the dataset are considered as prey where the system identifies optimal prey, i.e., features from the large database. The mathematical model for whale optimization is formulated into two stages as shrinking encircling mechanism and spiral position updating. Whales adopted a unique style of hunting which is termed as bubble net feeding attack. In which, bubbles will be created to hunt the fishes termed as two schemes as upward spiral and double loops. The whales will dive deep into the sea and create bubbles around the prey while returning from deep to the top surface. In the case of double loops, the process is divided into three phases as capture, coral, and lobtail phase. Generally, the bubble net characteristics are formulated as a mathematical model. The position of the whales is initialized randomly, and the best position to attack the prey is considered as the optimal solution. The remaining whales will update the current position toward the best position which is formulated as follows.  − → − → −  → v1 × y ∗ (t) − − k = → y (t)

(1)

− → − → − → → y (t + 1) = y ∗ (t) − − v2 × k

(2)

where the current iteration is represented as t and the coefficient vectors are represented as v1 and v2 , the best solution position is represented as y ∗ and for each iteration y ∗ is updated. The coefficient vectors are formulated as → − → → → v2 = 2 − v ×− r −− v − → v =2−t



2 max ite

(3)

 (4)

8

S. Castro and R. Pushpa Lakshmi

→ − → v1 = 2 × − r

(5)

where the loop counter is represented as t and the maximum number of iterations → v is linearly decreasing function from is represented as max ite , the vector function − → 2 to 0 over the iterations and the random vector is represented as − r and its range is [0, 1]. In the two-dimensional search space, the position of the whales located at (m, n) is updated based on the best position (m ∗ , n ∗ ). The idea encircling the prey is applied to n-dimensional search space, and the concept can be utilized to obtain optimal solution for complex classification problems. The exploitation and exploration phases in bubble net behavior of whales are explained as an illustration in Fig. 2. In the exploitation phase, two methods are followed such as shrinking encircling and spiral position update. In the encircling process, the value of coefficient vector − → v is decreased and the whale positions are updated based on the Eqs. (1)–(5). In the spiral update process, the distance between the current solution yi and best solution y ∗ is initially calculated and then to replicate the helix-shaped movement a spiral equation is formulated as − → − →  − → y (t + 1) = k .ebσ × cos(2π σ ) y ∗ (t)

(6)

where the constant function which defines the logarithmic spiral is represented as b and σ represents the random number in the range [−1,1]. The distance between − → other whales toward the best solution is represented as k which is given as ( , )

(

∗

,

∗

)

Exploration stage Exploitation stage

(a)

(b)

Fig. 2 Illustration of whale optimization exploitation and exploration phases a Shrinking encircling b Spiral position update

An Integrated Approach for Big Data Classification and Security Using …

−  − →  →  → k . =  y ∗ (t) − − y (t)

9

(7)

The probability of selecting circling or spiral is formulated as − → y (t + 1) =



− →∗ − → → y (t) − − i f P < 0.5 v2 × k − → − →∗  bσ k .e × cos(2π σ ) y (t) i f P ≥ 0.5

(8)

where P represents the probability which is given as [0, 1]. In the exploration phase, the position of the whales is updated based on the search agent. In this phase, search agent plays a major role compared to best solution in the exploitation phase. The → v > 1 and the performed global whales move away from the search agent when − search is formulated as  −  − →  →→∗ −  v1 y − → (9) y k . = − − → − → → → v2 × k y (t + 1) = − y rand − −

(10)

where the whale random position selected from the current population is represented → y rand . For each iteration, the search agent position is updated based on the best as − → → v > 1. If − v < 1 then positions are updated based solution satisfying the condition − on the random search agent. The transformation between exploitation and explo→ v vector, and the parameter P is used to switch between ration is controlled by the − spiral and circular movement. Thus, the optimal features are selected using whale optimization and the selected features are classified using random forest classifier. Basically, random forest is a group classification tree and ability to process large data through its functional elements. Several tree prediction performances are combined as an algorithm which is quite suitable to process heterogeneous big data. Based on the key attributes in the dataset, numerous trees are assembled and the error rate is calculated to select the best tree for further operation. An alternate bootstrap test helps to select the tree and the process initially starts by splitting the node. After split, a random subset is selected from the first set and the best split is used. The three major parameters that random forest classifiers consider are node size, number of trees, and number of predictors. In this node, size defines the perceptions of each tree. Generally, the trees should be obtained with minimum bias factors. In case of tree numbers, higher number of trees will provide better decisions. Whereas the predictors define the sampled predictors at each split and it will be the key parameter in the classifier. The trees are constructed based on the following condition. R F err = R F r1 ,r2 ( f (u 1 , u 2 ) < 0)

(11)

where the parameters u 1 and u 2 indicates the random vectors. The function f (u 1 , u 2 ) represents the degree of random vectors and it is given as

10

S. Castro and R. Pushpa Lakshmi

Fig. 3 Random Forest classifier

f (u 1 , u 2 ) = argI (h l (u 1 ) = u 2 ) − maxargI (h l (u 2 ) = j )

(12)

The two parameters in the above function indicate the accuracy measure of individual classifier and its dependencies between relationship and quality. This parameter indicates the aggregate tree numbers which are utilized in the random forest classifier to obtain best classification results. Figure 3 depicts an illustration of random forest classifier. Further to secure the big data, an encryption model is adopted in this research work. Digital Signature Standard (DSS) is employed in this research work which is an efficient non-deterministic cryptography algorithm. The encryption procedure starts by finding large prime numbers x and y such that y|x − 1. Based on the prime number, a generator of order z is obtained as   x g y = 1 mod x

(13)

The above generator function should generate a subgroup in the range 0 ≤ j ≤ y for g j = 1. The public key is selected in the range 0 ≤ p ≤ y − 1, and the private key is computed as   y g x = 1 mod x

(14)

The data size is reduced before to include a sign in the message a function is formulated as m = g k (mod x)(mod y)

(15)

An Integrated Approach for Big Data Classification and Security Using …

n=

s + mx k

11

(16)

where k ∈ z ∗y . The obtained signature can be verified using the modulus function as follows. m

m

y n g n (mod x)(mod y) = m

(17)

The major advantage of this encryption algorithm is its shorter signature which improves the encryption performance compared to other encryption models. The summarized pseudocode for the proposed classification and encryption approach is given as follows. Pseudocode for the proposed classification and encryption model Input:

,

,

,

,

Initialize Begin

Calculate the entropy values and remove redundant features Initialize optimized feature selection update the current position towards the best position Obtain the coefficient vectors Replicate the helix shaped movement

as per Eqn. (6)

Obtain the distance between other whales towards the best solution Update the position based on the probability

as per eqn (8)

Construct tree based on eqn (11) Obtain the parametric feature functions Classify the data Select the data for encryption Select prime features as per Eqn. (13) Compute the private key as per Eqn. (14) Reduce the data size Include signature as per the formulation given in Eqn. (15) Verify the signature as per Eqn. (16) End process End End

12

S. Castro and R. Pushpa Lakshmi

4 Results and Discussion The performance of the proposed integrated classification and encryption approach is measured using simulation analysis performed in MATLAB 14.1 installed in an Intel i5 processor with 8 GB memory. The operating platform is Windows 10, and the dataset used for experimentation analysis is Wholesale Customer Data. The benchmark dataset has 440 instances with 8 attributes. In the dataset, the attributes and bits are same whereas the data size and vectors are different to indicate the heterogeneous characteristics. The experimentation is repeated for multiple times, and the performances are measured in terms of accuracy, precision, recall, and F1-score. The encryption and decryption performances are measured in terms of encryption and decryption time, and encryption and decryption throughput. To validate the superior performance of proposed classifier, conventional machine learning algorithms like support vector machine and decision tree models are compared with proposed model. Similarly, to validate the encryption performance, Encrypting File System standard algorithm (EFSSA) is compared with DSS encryption algorithm. The simulation parameter used in the proposed work is depicted in Table 1. Figure 4 depicts the performance comparative analysis of proposed model with other machine learning models like support vector machine (SVM) and decision tree (DT) in terms of precision, recall, and F1-score. It is observed that the average values obtained for all the algorithms are above 90%. However, the maximum performance is attained by the proposed approach due to the efficient feature selection using whale optimization algorithm. The optimal features enhance the system performances while the other models are employed without any optimal feature selection procedure which results into reduced performances. The maximum precision, recall, and F1-score attained by the proposed model are 96.48%, 96.58%, and 96.53%, respectively. Whereas the precision, recall, and F1-score values of SVM model are 92.42, 91.56, and 91.99%, respectively. For decision tree model, the performances are 93.56%, 92.84%, and 93.20%, respectively, for precision, recall, and F1-score. The average accuracy is presented in the figure which is obtained through repetitive experimentations performed. The values obtained for each time are arranged, and the final average values are used in the comparative analysis. Maximum of 98.47% Table 1 Simulation parameters

S.no.

Parameter

Range/value

1

Number of whales

200

2

Number of iterations

20

3

Number of trees

1000

4

Maximum samples to split

3–15

5

Maximum samples in leaf

8–12

6

Maximum features

3–8

7

Number of estimators

10

An Integrated Approach for Big Data Classification and Security Using …

13

Fig. 4 Performance comparative analysis

accuracy was attained by the proposed model which is 3% higher than the decision tree and 5% higher than the support vector model. The accuracy of the proposed model is validated through comparative analysis performed with existing models in Fig. 5. The average accuracy is presented in the figure which is obtained through repetitive experimentations performed. The values obtained for each time are arranged, and the final average values are used in the comparative analysis. Maximum of 98.47% accuracy was attained by the proposed model which is 3% higher than the decision tree and 5% higher than the support vector model. Further, the encryption model performance is evaluated in terms of encryption and decryption throughput. The comparative analysis given in Fig. 6 depicts the maximum performance of proposed model for encryption and decryption throughput compared to EFSS algorithm. The encryption throughput attained by the DSSE is 350 Mb/min, and the decryption throughput is 330 Mb/min. Whereas the encryption and decryption throughput of EFSS algorithm are 310 Mb/min and 290 Mb/min which is much lesser than the proposed encryption approach (Fig. 7).

14

S. Castro and R. Pushpa Lakshmi

Fig. 5 Classification accuracy analysis

The encryption and decryption time for the proposed model and EFSS algorithm is measured for different file sizes from 56 MB to 2.5 GB. The minimum file size indicates the classified results. The maximum file size indicates the files that we have included to measure the encryption time. Both observations are presented as a combined figure. The encryption time given in Fig. 6 demonstrates the better performance of proposed DSS encryption algorithm. The average encryption time acquired by the proposed model is 2355 ms whereas EFSSA attains average encryption time as 2800 ms which is higher than the proposed approach (Fig. 8). Based on the experimental results, it can be observed that the performance of the proposed approach is better in terms of classification and security. The optimal feature extraction process using whale optimization algorithm enhances the overall classification performance. Similarly, better encryption performance indicates the enhanced data security while securing the data in cloud environment.

An Integrated Approach for Big Data Classification and Security Using …

15

Fig. 6 Throughput analysis

5 Conclusion An integrated big data classification and encryption approach is presented in this research work using optimized random forest algorithm and DSS encryption algorithm. Initially, the optimal features are selected using whale optimization algorithm and the selected features are classified using random forest classifier. To enhance the data security while proceeding data into cloud environment, DSS encryption procedure is adopted that effectively encrypts the classified data. The performance of the proposed model is experimentally verified and compared with other machine learning algorithms like support vector machine and decision tree algorithms and observed that the proposed model attains maximum classification accuracy of 98.47%. Also, to

16

S. Castro and R. Pushpa Lakshmi

Fig. 7 Encryption time analysis

validate the better encryption performance, DSS encryption algorithm is compared with EFSS algorithm in terms of encryption and decryption throughput, and encryption and decryption time. The proposed integrated approach attains better performance in all the aspects and makes use of the big data in a better manner. Further, this research work can be extended toward authentication and encryption procedures to enhance the data security.

An Integrated Approach for Big Data Classification and Security Using …

17

Fig. 8 Decryption time analysis

References 1. Castro S, Pushpalakshmi R (2020) A novel K-means clustering based FPGA parallel processing in big data analysis. Appl Math Inf Sci 14(1):1–6 2. Adadi A (2021) A survey on data-efficient algorithms in big data era. J Big Data 8(24):1–54 3. Thanga Selvi R, Muthulakshmi I (2021) An optimal artificial neural network based big data application for heart disease diagnosis and classification model. J Ambient Intell Hum Comput 12:6129–6139 4. Hernández G, Zamora E, Sossa H, Téllez G, Furlán F (2020) Hybrid neural networks for big data classification. Neurocomputing 390:327–340 5. Zhou H, Sun G, Fu S, Liu J, Zhou X, Zhou J (2019) A big data mining approach of PSO-based BP neural network for financial risk management with IoT. IEEE Access 7:154035–154043 6. Dumindu Samaraweera D, Morris Chang J (2021) Security and privacy implications on database systems in big data era: a survey. IEEE Trans Knowl Data Eng 33(1):239-258 7. Jiang C, Li Y (2019) Health big data classification using improved radial basis function neural network and nearest neighbour propagation algorithm. IEEE Access 7:176782–176789 8. Xing W, Bei Y (2020) Medical health big data classification based on KNN classification algorithm. IEEE Access 8:28808–28819 9. Gong C, Zhi-gang S, You Y (2021) Evidential instance selection for K-nearest neighbor classification of big data. Int J Approx Reason 138:123–144

18

S. Castro and R. Pushpa Lakshmi

10. Gunasekaran Manogaran P, Shakeel M, Baskar S; Hsu C-H, Kadry SN, Sundarasekar R, Kumar PM, Muthu BA (2021) DM: fuzzy-optimized data management technique for improving big data analytics. IEEE Trans Fuzzy Syst 29(1):177–185 11. Xiaowei G, Angelov P, Zhao Z (2021) Self-organizing fuzzy inference ensemble system for big streaming data classification. Knowl-Based Syst 218:1–13 12. Yang LH, Liu J, Wang YM, Martinez L (2021) A micro-extended belief rule-based system for big data multiclass classification problems. IEEE Trans Syst Man Cybern Syst 51(1):420–440 13. Samuel Manoharan J (2019) Study on Hermitian Graph wavelets in feature detection. J Soft Comput Parad 1(2):24–32 14. Chen G, Liu Y, Ge Z (2019) K-means Bayes algorithm for imbalanced fault classification and big data application. J Process Control 81:54–64 15. Krishnaraj N, Krishamoorthy S, Shankar K (2021) Big Data based medical data classification using oppositional Gray Wolf Optimization with kernel ridge regression. In: Applications of Big Data in Healthcare, pp 195–214 16. Al-Thanoon NA, Algamal ZY, Qasim OS (2021) Feature selection based on a crow search algorithm for big data classification. Chemom Intel Lab Syst 212:1–5 17. Gai K, Qiu M, Zhao H (2021) Privacy-preserving data encryption strategy for big data in mobile cloud computing. IEEE Trans Big Data 7(4):678–688 18. Samuel Manoharan J (2020) Population based metaheuristics algorithm for performance improvement of feed forward Neural Network. J Soft Comput Parad 2(1):36–46

Secure Web Gateway on Website in Cloud Daljeet Kaur, Celestine Iwendi , Thaier Hamid, and Pradeep Hewage

1 Introduction The security level data protects customers from net primarily based totally threats similarly to making use of and implementing company desirable use guidelines [1]. Instead of connecting immediately to the internet site, a person accesses the SWG, that is then liable for connecting the person to the preferred internet site and acting as a feature along with URL filtering, net visibility, malicious content material inspection, and net get admission controls and different protection measures. One of the demanding situations of deploying SWG capability is that it is miles’ general installation as a standby myself surroundings without coordinating workflows, reporting or logging with different protection infrastructures inside organization. This can cause multiplied complexity over the years as agencies frequently have a couple of protection factor merchandise that make their protection operations much less green and effective [2]. More recently, a brand new method for protection infrastructure has emerged. As defined through studies and the advisory company Gartner, a stable get admission to provider edge (SASE; pronounced “sassy”) combines networking and community protection offering rights into a single, cloud-introduced solution [3]. This lets in organizations to supply a couple of styles of protection offerings from the cloud, along with SWG, superior danger prevention, firewall as a provider (FWaaS), area D. Kaur (B) · C. Iwendi · T. Hamid · P. Hewage School of Creative Technologies, University of Bolton, Bolton, UK e-mail: [email protected] C. Iwendi e-mail: [email protected] T. Hamid e-mail: [email protected] P. Hewage e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_2

19

20

D. Kaur et al.

call system (DNS) protection, cloud get admission to protection brokers (CASB), information loss prevention (DLP), and others. This way, organizations are capable to managing the net, get admission to; offer customers with stable connectivity; and defend all their traffic customers and packages from opposed websites and content materials, all from one cloud primarily based totally platform. The aim is to examine the latest methods or techniques and approaches that can be used to secure the website in the cloud compared to old methods [1]. The following are the major contributions of this research: • Blocks get admission to irrelevant websites or content material primarily based totally on desirable use guidelines. • Enforce their protection guidelines to make the net get admission to be more secure. • Help defend information in opposition to unauthorized switches [4]. The rest of this paper is organized as such: Section 2 reviews the related literature. In Sect. 3, the methodology of the research is discussed. Section 4 provides the experimental results of the proposed framework. Section 5 concludes the research and provides the future scope.

2 Related Work The authors of [5] location emphasis on the application of SDN in protection research, and additionally shed mind upon the truth that the reputation of SDN will increase quicker among networking specialists than with protection researchers (Fig. 1). Their book brings sturdy arguments sponsored via way of means of current examples from the SDN protection community, displaying the benefits of the usage of digital networks in securing networks in a singular manner [6]. Among the SDN structures that carry novelty to protection, really well worth citing is dynamic manipulate of malicious or suspicious community flows, a centralized Fig. 1 Different aspects of website security

Secure Web Gateway on Website in Cloud

21

tracking gadget for detection of community flooding or community anomalies, or even the improvement of community programming languages for clean deployment. While our method makes use of SDN for protection research, [7] provides the use of SDN for securing an actual community, element which shall show beneficial for destiny work. An exciting method closer to community protection, which is the usage of SDN, is provided in [8]. The authors employ the famous digital community framework mininet [9] so that it will carry out reconnaissance deception. In short, via the way of means of leveraging the energy of the software program described, networking, every actual host in a community provides to the others a bogus photograph of the community, as soon as taking an element in a community scanning activity. While community deception fails with update protection, scanning, the mechanism significantly aids in the growing time wishing for an insider to deduce the format and shape of the actual community. Compared to our solution, [10] indicates an exciting utilization of the current framework, even as we illustrate the advent of a framework for growing digital networks to be used in protection research. [11–14] illustrate the use of other kinds of technology and algorithms in security.

3 Methodology In this chapter, we are defined about the methods and techniques are using to show the data in the quantitative and qualitative methods mean primary and secondary data. We are using the model software development life cycle to show the data after analysis and testing (Fig. 2).

3.1 Data Collection Because much of the data processing on our website is automated, we are not required to physically process the data in any manner. A copy of the message is sent to our website customer in the most typical scenario of data submission, which involves filling out a form. On more contemporary websites, the message is also saved in the website’s database. The information gathered is based on the needs of our website’s customers and often consists of names, emails, phone numbers, and any messages or requests made by website visitors.

Fig. 2 Elaborates the data from methodology research

22

D. Kaur et al.

Through a secure website admin interface, users can access, export, or remove personal data that has been stored on our website. To make updates and other general tasks, we will log into the website’s admin area.

3.2 Data Analysis Qualitative research: an expression that comes to mind when we need to get inside our clients’ heads to grasp their considerations and sentiments. With the use of evidence gathering, this research methodology aims to respond to a research topic. We can pick from a variety of qualitative research techniques depending on our themes. Quantitative research is based on the secondary data. It is used to define the data in practical form and do testing of data collection. There are different patterns to describe the data such as survey questions, interview, and person opinions.

3.3 Testing Wix has been around for a while and has improved their advertising, displaying the different famous and powerful people who utilize the site. Due to its adaptability and limitless integration and plug-ins, WordPress is without doubt the most wellliked and frequently used content management system. The majority of e-commerce organizations and companies whose success depends significantly on conversion use Shopify as their platform of choice. It has a more narrowly defined target demographic and specialty because it was created particularly for this purpose, and as a result, it succeeds where other platforms fall short. With the help of a number of tools that examine various parts of our system, Bitdefender safeguards us against online threats. Web Protection (Online Threat Prevention), one of the functions, is intended to check all web traffic and block any dangerous information, including infected URLs, dubious websites, and phishing links.

4 Implementation and Results The research issue (or implementation challenge) is utilized as the initial point of inquiry in implementation research, which adopts a pragmatic approach. From there, the research techniques and underlying assumptions are determined. Questions about implementation research can be on a wide range of subjects and are typically grouped according to theories of change or the nature of the research purpose.

Secure Web Gateway on Website in Cloud

23

Table 1 Comparison between three plug-ins Wix

WordPress

Shopify

Entry-level plan includes 500 MB of storage and 1 GB of bandwidth, allowing us to connect our domain

Among other CMS-based websites, WordPress holds a market share of close to 60%

In so many respects, this platform shines at understanding the thoughts of those with an interest in e-commerce

Wix is most popular third-tier plan, which includes an unlimited bandwidth

WordPress is the unquestionably the preferred web construction technology, in our opinion

The ability to add goods from other businesses to their shop and earn a commission

Wix promotes itself as a free website builder

For people who do not have Additionally, Shopify shipping the time or interest to dedicate helps save time and money to learning some programming

4.1 Difference Between Wix WordPress and Shopify The three most often utilized systems for creating websites globally are WIX, Shopify, and WordPress. In this post, we compare WordPress, Shopify, and WIX for their usability, features offered, potential costs, and other factors. An e-commerce-specific website builder platform is called Shopify (Table 1).

4.2 Bitdefender Software With the help of a number of tools that examine various parts of our system, Bitdefender safeguards us against online threats. Web Protection (Online Threat Prevention), one of the functions, is intended to check all web traffic and block any dangerous information, including infected URLs, dubious websites, and phishing links. When we attempt to access a website that has been flagged as hazardous, it is blocked and a warning is shown in our browser. The notice includes details such as the page URL and the threat that was discovered (Fig. 3). Bitdefender can block online programmers like games, services, and utilities in the background that access hazardous links or IP addresses in addition to web pages. Every time Bitdefender prevents something, and a pop-up message titled “Threat detected” appears on our screen.

24

D. Kaur et al.

Fig. 3 Use the activation code with Bitdefender to secure the data

1. Method: With the help of a number of tools that examine various parts of our system, Bitdefender safeguards us against online threats. Web Protection (Online Threat Prevention), one of the functions, is intended to check all web traffic and block any dangerous information, including infected URLs, dubious websites, and phishing links. 2. Method: When we attempt to access a website that has been flagged as hazardous, it is blocked and a warning is shown in our browser. The notice includes details such as the page URL and the threat that was discovered. 3. Method: Bitdefender can block online programmers like games, services, and utilities in the background that access hazardous links or IP addresses in addition to web pages. Every time Bitdefender prevents something, and a pop-up message titled “Threat detected” appears on our screen.

4.3 Survey In this survey, we have defined the qualitative research to secure the website on the cloud (Fig. 4). There are asking multiple five questions from the different age groups of people. These all are defined into four categories of age group people such as age 15–20, age 20–25, age 25–30, and age 30–40. Most of the people agree that we should need to use secure and latest plug-ins and should need to use security software so that the virus can be removed. According to the survey experienced people age group 30–40

Secure Web Gateway on Website in Cloud

25

Fig. 4 Graphical representation of the survey on Google Form

agreeing with this statement, we should use the latest methods and techniques to secure our data.

4.4 Problems • Personal data (PII) of employees, clients, and business partners. • Financial data about the company or its clients. • Intellectual property, trade secrets, and other confidential firm records (Fig. 5).

Fig. 5 Issues to hack the data on the website

26

D. Kaur et al.

There are several potential causes of the exposure of sensitive data, including human error, technological failures, and criminal action. However, if private information ends up in the wrong hands, the organization could suffer serious consequences.

4.5 Solution Impose the use of robust passwords (which should never be shared). If we utilize predictable phrases and numbers in our password, it will be relatively simple to guess (e.g., our name and birth date). Make our passwords longer and more secure to keep criminal actors from getting into our accounts. Additionally, we should never ever reveal our password. Include this rule in the password policy for our company (Fig. 6). Use authentication without a password. Get rid of passwords altogether and switch to certificate-based authentication for resources that we only want certain users to have access to. Users must have authentication certificates installed on their devices to use this approach, which authenticates users without the necessity of complex passwords. Never click on links from unidentified sources. Use our browser to access the website if we wish to. Fig. 6 Applications used to protect the website

Secure Web Gateway on Website in Cloud

27

4.6 Discussion While too many plug-ins and add-ons can put our website at risk, they can also improve the experience for our visitors. They might also make our website less responsive and slow it down. Regular add-on and plugin inspection can lower the probability of faulty additions [4]. Examining the reputation specifications and modification history of plug-ins and add-ons is crucial, regardless of whether these additions come from our CMS or a third party [15]. These website add-ons need frequent maintenance and care, just like a CMS. Their development determines how well their bug defenses work. They expose our business to security threats if they become outdated. Additionally, we might benefit from security-focused add-ons that restrict infections.

5 Conclusion For business owners, the security of the information system is a common worry. As a result, they set out to create official policies and processes to safeguard their companies from any mishaps. This paper’s primary goal is to examine information security concerns in cloud. We should use secure-based latest plug-ins or platforms that are useful on secure websites. As well as antivirus software is also helpful to tell us if any dangerous file is downloaded to the system. Bitdefender is the best software to remove the virus and scan the data with secure methods on website.

5.1 Future Work In the future, we should need to improve the database of websites. Backend security is important as compare to Frontend security. Frintend of website is depend on coding or backend. We should improve the security tools and methods to save the data on the website because every business goes on online in the modern era. Therefore, it is important to safe the data by using the latest methods and tools. Acknowledgements I must first express my gratitude to [Dr. Celestine Iwendi], [Dr. Hamid Thaier], and [Dr. Pradeep Hewage], who oversaw my study. This paper would not have been completed without their help and diligent participation in each step of the procedure. I would say special thanks to my University of Bolton who gave me this best opportunity.

28

D. Kaur et al.

References 1. Ali AA, Zamri Murah M (2018) security assessment of Libyan Government Websites. IEEE Xplore.. https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8626862. Accessed 01 Nov 2018 2. Lavrenovs A, Melon FJR (2018) HTTP security headers analysis of top one million websites. In: 2018 10th international conference on cyber conflict (CyCon). https://doi.org/10.23919/ cycon.2018.8405025. Accessed May 2018 3. Nursetyo A, Ignatius Moses Setiadi DR, Rachmawanto EH, Sari CA (2019) Website and network security techniques against brute force attacks using honeypot. IEEE Xplore. https:// ieeexplore.ieee.org/document/8985686. Accessed 1 Oct 2019 4. Altaf S (2021) A review of the security issues in cloud computing and its remedial actions. Inf Technol Ind 9:444–455. 10.177621itii.v9il.150. 5. Kumar R, Raj H, Perianayagam J (2017) Exploring security issues and solutions in cloud computing services—a survey. Cybern Inf Technol 17. https://doi.org/10.1515/cait-2017-0039 6. Md AQ, Varadarajan V, Mandal K (2019) Correction to: efficient algorithm for identification and cache based discovery of Cloud Services. Mobile Netw Appl 24(4):1198–1198 7. Md AQ, Vijayakumar V (2020) Combined preference ranking algorithm for comparing and initial ranking of cloud services. Recent Adv Electric Electron Eng (Former Recent Pat Electric Electron Eng) 13(2):260–275 8. Li D, Hsu S, Purushotham D, Sears RL, Wang T (2019) WashU epigenome browser update 2019. Nucleic Acids Res 47(W1):W158–W165 9. Md AQ, Vijayakumar V (2019) Dynamic ranking of cloud services for web-based cloud communities: efficient algorithm for rating-based discovery and multi-level ranking of cloud services. Int J Web Based Communit 15(3):248–270 10. Mathew SA, Md AQ (2018) Evaluation of blockchain in capital market use-cases. Int J Web Portals 10(1):54–76 11. Sirajuddin M, Iwendi C et al (2021) TBSMR: a trust-based secure multipath routing protocol for enhancing the QoS of the mobile ad hoc network. Security and Communication Networks, 2021, 5521713, S/N 1939-0114. https://doi.org/10.1155/2021/5521713. (Q2). SCI Impact factor 1.288 12. Anajemba JH, Yue T, Iwendi C, Chatterjee P, Ngabo D, Alnumay WS (2021) A secure multiuser privacy technique for wireless IoT networks using stochastic privacy optimization. IEEE Internet of Things J. https://doi.org/10.1109/JIOT.2021.3050755. (Q1). SCI Impact factor 9.471 13. Iwendi C et al (2021) Security of things intrusion detection system for smart healthcare. Electronics 10(12):1375. https://doi.org/10.3390/electronics10121375. (Main author. Journal Impact Factor 2.412) 14. Rubia A, Iwendi C et al (2021) An optimised homomorphic CRT-RSA algorithm for security and efficient communication. Pers Ubiquit Comput. https://doi.org/10.1007/s00779-021-016 07-3. (Q2). SCI Impact factor 2.395 15. Hari Krishna B, Kiran S, Murali G, Reddy RPK (2016) Security issues in service model of cloud computing environment. Proc Comput Sci 87:246–251. https://doi.org/10.1016/j.procs. 2016.05.156 16. Hou C, Shi J, Cui M, Liu M, Yu J (2021) Universal website fingerprinting defense based on adversarial examples. In: 2021 IEEE 20th international conference on trust, security and privacy in computing and communications (TrustCom), Oct 2021. https://doi.org/10.1109/tru stcom53373.2021.00031 17. Almubayedh D, Khalis MA, Alazman G, Alabdali M, Al-Refai R, Nagy N (2018) Security related issues in Saudi Arabia small organizations: a Saudi case study. In: 2018, 21st Saudi computer society national computer conference (NCC), Apr 2018. https://doi.org/10.1109/ncg. 2018.8593058 18. Zhang S, Yin J, Li Z, Yang R, Du M, Li R (2022) Node-imbalance learning on heterogeneous graph for pirated video website detection. In: 2022 IEEE 25th international conference on

Secure Web Gateway on Website in Cloud

29

computer supported cooperative work in design (CSCWD), May 2022. https://doi.org/10.1109/ cscwd54268.2022.9776224 19. Vaikunth PT, Aithal PS (2016) Cloud computing security issues- challenges and opportunities. Int J Manag Technol Soc Sci 1(1):33–42. https://doi.org/10.5281/zenodo.569920 20. Yan F, Wang Y (2017) A security web gateway based on HTTP reverse proxy. DEStech Trans Eng Technol Res. https://doi.org/10.12783/dtetr/iceta2016/7003

Design a Quantum Cryptography Algorithm and Evaluate the Risks of Quantum-Based Nano Computing Thilagavathy

Abstract Nanotechnology is a cutting-edge field that has gained trust in recent years. One area of research within nanotechnology is nanocomputers, which has opened up numerous possibilities for computer scientists and experts to create novel products in the realm of nano-based electronic manufacturing. There are many ways to implement the nano computing, like nanoelectronic computing, nanomechanical computing, nano chemical and DNA computing, and quantum computing which will provide the easy way to complete our task with very fast and accuracy. The importance of study on quantum computing is hypothetically a new technology, whereas quantum computing preserves a system of qubits. Qubits are the base and fundamental unit of quantum computing. There are several obstacles to developing a quantum computerbased teleportation system that could enable long-distance communication through the transmission of data or messages. Quantum-based cryptography is a fantastic mechanism for safeguarding the internetwork communication between sender and receiver. Here, the paper mainly focuses on designing a quantum cryptography distribution algorithm and analysis of risk factors between sender and receiver in quantumbased nano computing. The Qiskit Python library, created by IBM, was utilized to develop and implement the quantum cryptography algorithm. Keywords Nanotechnology · Quantum computing · DNA—Deoxyribonucleic acid · Qiskit—Quantum Information’s Kit

Thilagavathy (B) Department of Computer Applications, SRM Institute of Science and Technology, Kattankulathur, Tamil Nadu, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_3

31

32

Thilagavathy

1 Introduction 1.1 Nanotechnology Nanotechnology involves utilizing particles that are incredibly tiny, at the sub-atomic level, to facilitate computing processes that differ from classical bits. Nanotechnology is a multi and interdisciplinary field of science. Nanotechnology enables the creation of smaller circuits and computers through innovative design. The utilization of quantum circuits will enhance computation and transmission speeds to a degree far beyond what can be achieved with classical computers.

1.2 Quantum To calculate the outputs in the system quantum mechanics, a quantum is used. The very small discrete and atomic particle of physical property in physics is known as quantum. Neutrinos electrons and photons are the references to properties of atomic or subatomic particles.

1.3 Qubit In quantum computing, basic unit of information is said to be a qubit. In classical computing, bits play a vital role; similarly, qubit plays a same role with different characteristics. Classical computer can store the data as 0 and 1, but qubits can store the data as superposition of all the particles in all states. Binary representation can store only a 0 or 1, but qubits can be represented to store a data in all states of superposition of quantum [1].

1.4 Quantum Computing The quantum computers can have different properties of a quantum that is superposition, entanglement, and quantum interference that should be applied in the quantum computers. This makes a new era from the traditional program computing methods [2].

Design a Quantum Cryptography Algorithm and Evaluate the Risks …

33

1.5 Quantum Cryptography or Key Distribution Quantum cryptography is the method to secure the data by using quantum and distributed symmetric encryption key algorithm. The quantum cryptography is more precisely known as quantum key distributions (QKDs). It communicates using subatomic particles such as electrons, photons across an optical link communication channels [3].

2 Related Work 2020-Yoann PIÉTRI—“Quantum Cryptography”, presented other quantum cryptographic tasks. Quantum randomness is already commercialized, while the other tasks do not seem attainable today. 2020-C.H.Ugwuishiwu1 el, “An overview of Quantum Cryptography and Shor’s Algorithm”. The Shor’s algorithm is not used for cryptography, because it will danger for system. The researcher should give more stress on quantum cryptography mechanism. 2020-ManeeshYati—“Quantum Cryptography”. It promises to solve problems which classical computers practically cannot. But, the cost behind quantum computing is too high. The major challenge that stands right now is to reduce the cost so that it is more accessible for experiments. 2015-Matthew Campagna and el, “Quantum Safe Cryptography and Security”. Here, the challenge for quantum network communication is to ensure the network threatening from untrusted network or from a person. So, it must control and manage for improve their product and will a commercial solution. At the same, all network engineers should be trained a well versed in quantum domain. 2014-N. Sasirekha, el. The title of the paper is quantum key distribution and its applications; this paper gives the right direction to move forward to next-generation computing and more sophisticated way to protect the data from hacker.

3 Proposed Work 3.1 Quantum Cryptography Protocol Implementation Alice sends a message to Bob using quantum bit; in that Alice, using mechanism for computing of qubit is random selection of qubit. Once the message reached to Bob, then Bob will decrypt the message which one sent by Alice mechanism. Sometime, Bob will not read correct message because some eavesdropper involved their communication. Here, some steps of implementation are mentioned in the following [3].

34

Thilagavathy

Alice transmitted a qubit in the state |+⟩|+⟩, represented as 0 on the XX-basis. Bob measured the XX-basis, resulting in a definite measurement outcome of 0, as depicted in Figs. 1 and 2.

But before the Bob receives it, Eve attempts to measure this qubit state from X-basis to Z-basis. He will change the qubit’s state from |+⟩|+⟩ to either |0⟩|0⟩ or |1⟩|1⟩, and Bob’s are not any more to measure 0 [4] (Figs. 3 and 4). Here, Bob has now 50% chance of measuring 1, and If Bob’s measurement outcome is not 1, he has a 50% probability of obtaining that result, and if he does not obtain 1, Alice and Bob will be alerted that an error occurred during their communication. Fig. 1 Alice’s quantum circuit with bit is 0

Fig. 2 Bob’s measure quantum with 100% chance

Fig. 3 Alice’s quantum circuit with Eve

Design a Quantum Cryptography Algorithm and Evaluate the Risks …

35

Fig. 4 Bob’s measure quantum with 50% chance

The quantum key distribution algorithm involved many times to transfer the data between the Alice’s and Bob’s, and the eavesdropper has negligible chances to trapping the message. This may be implemented using the following steps [4].

3.2 Random Bits’ Generation Alice plans to send a following string of random bits: 1000101011010100. Each random bit is converted to the following string: ZZXZXXXZXZXXXXXX. Finally, Alice has two pieces of information ready to send to Bob.

3.3 Encoding Bits into String Every bit is encoded by Alice into a string of qubits using the basis of choice made by her; this exactly each qubit is in any one of the states |0⟩|0⟩, |1⟩|1⟩, |+⟩|+⟩ or |−⟩|−⟩, and random choice is made. String of qubits may be in the following form: |1⟩|0⟩|+⟩|0⟩|−⟩|+⟩|−⟩|0⟩|−⟩|1⟩|+⟩|−⟩|+⟩|−⟩|+⟩|+⟩|1⟩|0⟩|+⟩|0⟩|−⟩|+⟩|−⟩|0⟩ |−⟩|1⟩|+⟩|−⟩|+⟩|−⟩|+⟩|+⟩ This was the message which was sent to Bob by Alice [5].

3.4 Measurement of Qubits Randomly, each qubit is measured by Bob, for instance, using the following as base: XZZZXZXZXZXZZZXZ.

36

Thilagavathy

And, this message was kept privately by Bob.

3.5 Symmetrically Share the Qubits Bob and Alice share the techniques to build an each qubit of a message. If Bob measures the qubit on the basis of Alice techniques, then for this mechanism, both share the secret keys. Suppose, if anything mismatches, the bit information will be discarded [6].

3.6 Symmetrically Share the Keys Eventually, random sample keys were shared by Bob and Alice, and if the both the keys were same, then the message can be successfully transferred from both the ends [7].

4 Implementation of Without Interception To generate random keys, we can initiate the process by introducing the necessary components. For implementation, the numpy randint function can be utilized to reproduce the seed value, beginning from 0 [7]. np.random.seed (seed = 0). Alice sends the message using n variable with 100 qubits long. n = 100.

4.1 Random Bits’ Generation Alice generates set of random bits by the following Python code:

The set of random bits created by Alice’s “alice’s bits” is known only to him. The collection of random bits generated by Alice referred to as “Alice’s bits,” is exclusively known to her. Alice’s bit information should remain private to Alice, with

Design a Quantum Cryptography Algorithm and Evaluate the Risks … Table 1 Alice’s_bits

Alice’s channel

Eve’s channel

37 Bob channel

alice_bits

neither Bob nor Eve knowing its content. The communications are in the following Table 1 in sequence.

4.2 Encoding Bits into String Alice made a choice to encode each of the bit on qubit randomly using X- or Z-basis and stores them in alice’s- bases. Here, 0 is taken as Z-basis and 1 is taken as X-basis [8].

Alice holds this knowledge of information privately (Table 2). By using the encode_message function below, creates a quantum circuits, to represent the each single qubit for Alice’s message.

It is shown that the first bit in alice’s_bits is 0, and the basis has been determined, revealing that the initial bit in Alice’s bits is 0 [9].

In the circuit, represent the Alice first qubit message that should be verified and prepared of a qubit in the state of |+⟩|+⟩ (Fig. 5).

messages [0]. draw() The fourth bit in alice_bits is 1, and it might be encoded in Z-basis functions and produces the corresponding qubit states which are |1⟩|1⟩ (Fig. 6).

Table 2 Alice_bases

Alice’s channel alice’s_bases

Eve’s channel

Bob channel

38

Thilagavathy

Fig. 5 Alice quantum first bit without interception

Fig. 6 Alice quantum fourth bit without interception of Z-bases

Table 3 Alice’s_bases

Alice’s channel

Eve’s channel

Bob channel

message

message

alice’s_bits alice’s_bases message

messages [0]. draw() The qubit message transmitted between Alice and Bob via the quantum channel was intercepted by Eve, as outlined in Table 3.

4.3 Measurement of Qubits The qubit measures in the X- or Z-basis by using random methods for message of Bob qubit and stores the message as bob’s_bases which was the choice made by the Bob to measure each qubit [10] (Table 4). The following function measures_message is used to measure the corresponding qubit and stimulate the results [11]. Result measurement is stored in bob’s_results.

Design a Quantum Cryptography Algorithm and Evaluate the Risks … Table 4 Bob’s_bases

Alice’s channel

Eve’s channel

39 Bob channel

alice’s_bits alice’s_bases message

message

message bob’s_bases

Fig. 7 Alice and Bob quantum circuit

Here, the result of the circuit message of 0 that is represented as 0th qubit by measurement of X-basis is added to it by Bob (Fig. 7).

messages[0].draw() Hence, Bob measures the same result of Alice-encoded qubit and is also guaranteed to get the result of 0 so, for example, to verify the sixth qubit, the measurement of Bob is not as same as the Alice so that the Bob result has 50% chance of matching Alice messages [12] (Fig. 8).

messages[6].draw()

Bob holds the result privately (Table 5). Fig. 8 Alice and Bob quantum sixth bit

40 Table 5 Bob’s_results

Thilagavathy Alice’s channel

Eve’s channel

Bob channel

alice’s_bits alice’s_bases message

message

message bob’s_bases bob’s_results

4.4 Encoded Qubits Alice reviews the messages through Eve channel to know what kind qubits were encoded (Table 6). At the same, Bob reveals the messages through Eve channel measure and stimulate each qubit [14] (Table 7). Bob’s results are compared with the corresponding Alice’s bits, and both parties incorporate this information into their key. Both messages are not similar which have to be remove entry away by using removes_garbage function and remaining bits to form their secret keys. Table 6 Alice’s and bob’s_bases

Alice’s channel

Eve’s channel

Bob channel

alice’s_bits alice’s_bases message

message

message bob’s_bases bob’s_results

Table 7 Bob’s_base with alice

Alice’s channel

alice’s_bases

alice’s_bases

Eve’s channel

Bob channel

message

message

alice’s_bits alice’s_bases message

bob’s_bases bob’s_results alice’s_bases bob’s_bases

bob’s_bases

alice’s_bases

Design a Quantum Cryptography Algorithm and Evaluate the Risks …

41

4.5 Comparison of Selection Bits To verify the Bob and Alice messages, comparison has to be made on some random selection bits and their keys to make ensure that the communication was successful between them [15]. Both Alice and Bob make their keys public, rendering them no longer confidential, as depicted in Table 8.

If the protocol transferred the message between two sides without interference, the sample of data must be matching. bob’s_sample = = alice’s_sample. If both the sides sample data probably matching highly, alice’s_keys = = bob’s_keys. The sender and receiver are ready to share the secret key to encrypt the messages also (Table 9).

5 Implementation of with Interception The sample results of Alice and Bob messages do not match, and both identified that there are some intrusions. So, the sender and receiver conclude that Eve tried Table 8 Alice’s and bob’s sample

Alice’s channel

Eve’s channel

Bob channel

message

message

alice’s_bits alice’s_bases message

bob’s_bases bob’s_results alice’s_bases bob’s_bases

alice’s_bases

bob’s_bases

alice’s_key

bob’s_key

bob’s_sample

bob’s_sample

bob’s_sample

alice’s_sample

alice’s_sample

alice’s_sample

42 Table 9 Alice’s and bob’s sample

Thilagavathy Alice’s channel

Eve’s channel

Bob channel

message

message

alice’s_bits alice’s_bases message

bob’s_bases bob’s_results alice’s_bases bob’s_bases

alice’s_bases

bob’s_bases bob’s_key

alice’s_key bob’s_sample

bob’s_sample

bob’s_sample

alice’s_sample

alice’s_sample

alice’s_sample

shared_key

shared_key

to extract some information from their communication. So, they decided to set a different set of seeds used for reproducible of random results [1]. np.random.seed (seed = 3).

5.1 Random Bits’ Generation Alice generates set of random bits for interception.

5.2 Encoding Bits Alice encodes the message by using X- and Z-bases at randomly to send it to Bob through Eve quantum channel. Here, the first qubit message of Alice’s state is |+⟩|+⟩ (Fig. 9). messages[0].draw(). Fig. 9 Alice encoding quantum bit

Design a Quantum Cryptography Algorithm and Evaluate the Risks …

43

Fig. 10 Eve tried to change quantum bit

Fig. 11 Bob measure quantum bit

5.2.1

Interception

Eve tried to intercept and extract the messages which pass through Eve channel, and random selection of bases was used to try the measure of qubit. Later, the same way will be used by Bob [15]. intercepted_message = measure_message(message,eve_bases). print(Intercepted_message). In the case of qubit 0, Eve random selection is not same as a Alice random selection, because of this, quibit selection state is changed from |+⟩|+⟩ to random selection of Z-basis with 50% of |0⟩|0⟩ or |1⟩|1⟩ [1] (Fig. 10). message[0].draw().

5.3 Measurement of Qubits Eve passed the message through Eve channel to Bob; from that, Bob tries to measure the same basis which Alice sent the qubit. Bob measures 0 definitely if there is no interception, but in this scenario Eve tries to read the Alice message so that the result will be changing. Because of this, only 50% chance is there to measure 1 instead of 0 [15] (Fig. 11). bob’s_bases = randint(2, size = n). bob’s_results = measure_message(message, bob_bases). message[0].draw().

5.4 Qubits’ Noise Removals Alice and Bob check the both results if mismatch discards the noise bit.

44

Thilagavathy

bob’s_key = remove_garbage(alice’s_bases, bob’s_bases, bob’s_results). alice’s._key = remove_garbage(alice’s_bases, bob’s_bases, alice’s_bits).

5.5 Comparison of Selection Bits Alice and Bob will test to determine whether their qubit messages match their respective keys by conducting a comparison. The following will be evaluated during this test. bob’s_sample = = alice’s_sample. Bob and Alice keys were mismatched. As we know, this is purely because of Eve who tried to read the message between communications which in turn changed the qubit state. Alice and Bob come to know that some interruption has been happened in their communication channel, so that they must throw away their messages and try to resend the messages. This in turn fails the Eve attempt [1].

6 Result Discussion of Risk Analysis For any interception happened between Alice and Bob, then it should be happened only by the Eve just to make changes in the qubits’ state and there are some quite few chances where mismatches can happen in the Bob and Alice data. Reason is that Alice sends the vulnerable message through Eve channel. Now, calculation can be made to know the chances of change in the qubit state and to know how much risky the quantum key distribution is [2]. • Case I—With random selection of same basis, the sender and receiver produce the same result. If the same random selection basis function is followed by Eve, then the Eve will get successful data without any error. For this, the chances are 50% only. • Case II—If Eve has chosen wrong different random selection basis functions which do not match with Alice and Bob random selection basis functions, there is only 50% of chance to measure the value of Bob which was sent by Alice. Here, it is unable to find the interception by Alice and Bob. • Case III—If Eve has chosen wrong different random selection basis functions which do not match with Alice and Bob random selection basis functions, there is only 50% of chance which will not measure the value of Bob that was sent by Alice. Here, will raise the error into their keys (Fig. 12). When the first bit is compared with their keys, the probability of Alice and Bob both having a match is 0.75, which would not indicate to them that Eve intercepted the transmission.

Design a Quantum Cryptography Algorithm and Evaluate the Risks …

45

Fig. 12 Alice, Bob, and chance to Qubit transfer

The probability of Alice and Bob having a match when comparing the second bit with their keys is 0.752, resulting in a value of 0.5625 for both parties. There is no choice of interpretation to bring the notice to their concern [2]. The probability of Eve working process was unable to detect from their number of bits that should be calculated by the following way: Pro(undetected) = 0.75 × Pro(undetected) = 0.75× By implementing the method described above for 15 bits, the likelihood of Eve remaining undetected is extremely low, at only 0.00006%. So, this is too risky for us. If the above method is applied to 50 bits, there is a 1.3% probability of being surreptitiously intercepted. Suppose change the sample_size which is low and repeatedly running the same process. The Eve can easily intercept the message.

7 Conclusion The quantum key distribution algorithm has been developed for use in nano computing based on quantum technology. The quantum key distributions’ algorithm is implemented and stimulated by using IBM Qiskits Python library and produces the best result for securing data transformation to sender and receiver. At the same, both sides can detect the transaction which was completed, or some eavesdropper tries

46

Thilagavathy

to intercept and extract the information between the transactions; there is a chance 50%. In the event of an intrusion, the sender and receiver will discard the message and attempt to resend it, with a guaranteed chance of successful communication.

References 1. Quiskit, IBM 2. Quantum research, Microsoft 3. Srilatha T (2017) Types of nano computers and challenges of quantum computing. Assistant Professor, Department of Computer Science R.B.V.R.R Women’s College, (Autonomous), Affiliated to Osmania University Narayanaguda, Hyderabad, Telangana 4. Ekert A (1991) Quantum cryptography based on Bell’s Theorem. Phys Rev Lett 67:661. (5 Aug 1991) 5. Ekert A (2000) What is quantum cryptography? In: Conger S, Loch KD (eds) Ethics and computer use. Communications of the ACM, vol 38, p 12 (entire issue). (Centre for Quantum Computation–Oxford University) 6. Petschinka J (2003) European scientists against eavesdropping and espionage. (1 April 2004). (7. Salkever A (2003) A quantum leap in cryptography. BusinessWeek Online. 15 July 2003) 7. Schenker JL (2004) A quantum leap in codes for secure transmissions. The IHT Online. (28 Jan 2004) 8. Mullins J (2003) Quantum cryptography’s reach extended. IEEE Spectrum Online. (1 Aug 2003) 9. MagiQ Technologies Press Release. (23 Nov 2003) 10. Elliott C (2002) Building the quantum network. New J Phys 4:46 11. Pearson D (2004) High!speed QKD reconciliation using forward error correction. In: Quantum communication, measurement and computing, vol 734, no 1. AIP Publishing 12. Shor PW, Preskill J (2000) Simple proof of security of the BB84 quantum key distribution protocol. Phys Rev Lett 85(2):441 13. Bennett C, Brassard G (1984) Quantum cryptography: public key distribution and coin tossing. In: International conference on computers, systems, and signal processing, Bangalore, India 14. Curcic T et al (2004) Quantum networks: from quantum cryptography to quantum architecture. ACM SIGCOMM Comput Commun Rev 34(5):3–8 15. Buttler WT et al (2003) Fast, efficient error reconciliation for quantum cryptography. Phys Rev A 67(5):052303 16. Piétri Y (2020) Quantum cryptography. Imperial College London 17. Ugwuishiwu CH, Orji1 UE, Ugwu CI, Asogwa CN (2020) An overview of quantum cryptography and shor’s algorithm, Department of Computer Science University of Nigeria, Nsukka, Enugu State, Nigeria 18. Yati M (2020) Quantum cryptography thesis. (Nov 2020). https://doi.org/10.13140/RG.2.2. 34447.61601 19. Sasirekha N, Hemalatha M (2014) Quantum cryptography using quantum key distribution and its applications. 20. Colin JR (2002) MagiQ employs quantum technology for secure encryption. EE Times. (6 Nov 2002) 21. Bienfang J et al (2004) Quantum key distribution with 1.25 Gbps clock synchronization. Opt Express 12(9):2011–2016 22. Inoue K, Waks E, Yamamoto Y (2002) Differential phase-shift quantum key distribution. In: Photonics Asia 2002. International society for optics and photonics 23. Barnum H et al (2002) Authentication of quantum messages. In: The 43rd annual IEEE symposium on foundations of computer science. Proceedings. IEEE

Design a Quantum Cryptography Algorithm and Evaluate the Risks …

47

24. Elliott C, Pearson D, Troxel G (2003) Quantum cryptography in practice. In: Proceedings of the 2003 conference on Applications, technologies, architectures, and protocols for computer communications. ACM

Machine Learning

Human Odor Security Using E-nose V. Anush Kumar, C. S. Manigandaa, S. Dhanush Hariharan, Challapalli Manikantaa, G. Saranya, and V. D. Ambeth Kumar

Abstract This paper presents an innovative security system based on E-nose to optimize the edge-to-edge encryption for highly vulnerable to security threats. The idea of this system is to convert human odors into signals. The signals are converted into binary codes in order to provide high security in authentication of the user. In this way, binary codes are converted to alphabets. The alphabets are converted to encrypted data which is checked by the database and validating the user. This is mainly implemented to increase the security of the user and to the data which they use. Keywords Security · AI · E-nose

1 Introduction The building blocks of an organization are information. Information is a form of communication that expresses knowledge or a message. Information can be communicated, stored, refined, and controlled—it is necessary for the majority of what we do. In the world of business, information is the most valuable asset. Organizations and individuals need to protect their information appropriately. Keeping data and operating procedures secure in an organization involves integrating systems, operations, and internal controls. Information should therefore be protected based on the organization’s needs. Organizations and individuals alike can benefit from information, sometimes even need it. Information of this kind can have catastrophic consequences V. Anush Kumar · C. S. Manigandaa · S. Dhanush Hariharan · C. Manikantaa · V. D. Ambeth Kumar (B) Department of AI&DS, Panimalar Engineering College, Chennai 600123, India e-mail: [email protected] G. Saranya Department of CSE, Panimalar Engineering College, Chennai 600123, India V. D. Ambeth Kumar Department of Computer Engineering, Mizoram University, Aizawl 796004, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_4

51

52

V. Anush Kumar et al.

if it is lost or inaccurate. An administrator can use biometrics to solve problems and to increase security by introducing new biometrics. The systems can be used in two ways. Admins whose biometric photos are stored in the database are compared with those that are already stored in the database. In this case, the biometric data is compared with the admin’s stored data. A priori unknown identity of the admin leads to identification. In the database, every record in the database is matched against the biometric dataset of the admin. In order to match their biometric data, the admin needs not be present in any records in the database. The technical challenges and costs associated with identification are evident. As a database grows, identification accuracy decreases. A biometric characteristic that is sufficiently discriminatory is therefore used to categorize large databases. Afterward, a small subset of records is searched for a particular record. (If the discriminating characteristic is properly chosen), this results in fewer relevant records per search and greater accuracy. A biometric authentication system only collects, processes, and stores the admin’s biometric data after he/she registers with it. To create an admin’s master template, several biometric samples are usually used (usually three or five) due to the importance of the quality of the stored biometric data. It is possible to create a unique dataset from fingerprints, retinal scans, and iris scans when they are processed properly. Information can be accessed with accuracy and security with biometric identification. There are many problems with current methods of password verification (people make up easy-tohack passwords, they forget them, people write them down). A minimum of training is required to automate biometric identification for a very rapid and uniform result. Documents that are stolen, lost, or altered cannot be used to verify your identity. PC admin authentication using biometrics is the most economical technique. An intuitive interface makes it easy to use. The biometric template can be stored in just a few minutes. Reducing the number of memory requirements for the database. A retina cannot be replicated standardized. Approximately five seconds are needed to verify the information. A biometric scanner captures an individual’s biometrics for identity verification. By comparing the scans with the saved database, access to the system can be approved or denied. The key to unlocking access is your body, which makes biometric security possible. A sensor on the E-nose senses human odors and converts odors into signals, which is the methodology of this paper. In order to convert the signals into binary, ASCII values are used. In order to turn the binary code into alphabets, ASCII values are used. An alphabet string is formed by concatenating these alphabets. For the password to be strong, the string formed is encrypted using the Caeser cipher technique. Data in the database is compared with the encrypted string to authenticate the admin. The following sections make up the paper: Sect. 2 offers an overview of existing concepts; Sect. 3 is a description of how the system works. Sect. Section. In Sect. 4, the proposed system is implemented, and in Sect. 5, this study comes to a conclusion.

Human Odor Security Using E-nose

53

2 Related Work This work paper [1] machine olfaction is becoming more objective thanks to sensorbased technology, which mimics the biological system. Also, materials with chemoelectronic properties have been discovered, which have made it easier to develop materials. Currently, human and instrumental sensing are both used to detect odors, and presented here is an overview of the current state of the art. The strengths and weaknesses of human odor detection were discussed by a panel of trained experts using olfactometric techniques. Research on odor sensor system is tested against various odorants in this study [2]. A piezoelectric electrode is coated with human olfactory receptor proteins (ORPs), similar to how human olfactory receptors work. PZ crystals were used to transmit signals [2]. N-caproic acid, isoamyl acetate, n-decyl alcohol, and b-ionone induced in this study. The effect lasts for up to three months and is reversible (30 s). This sensor can detect individual odorant fingerprints using the correlation between the olfactory threshold’s value and the sensor’s sensitivity. A phospholipid probe and five fractionated ORPs make up each sensor for detecting individual fingerprints of odorants. In this paper [3], humans react differently to pleasant and unpleasant odors based on a psychophysical detection test. An assortment of different stimulus strengths was used to compare response latencies, a pleasant odorant (amyl acetate) and malodorant (valeric acid). In a concentration range with iso-intensity, you can compare the responses to the two odors using reaction time as a function of detection rate. As part of this study, odorants are compared at the same intensity but at varying concentrations for the first time. In a 50% detection level, amyl acetate, which causes malodors, was detected faster than valeric acid. This is the level at which amyl acetate detection time was 1.74 s, whereas valeric acid detection time was 1.36 s (380 ms, or 22% faster). Electronic noses (E-noses) have emerged as an innovative application area in medical diagnostics. This paper reviews. An E-nose was used to diagnose illnesses at Warwick University [4]. E-noses have been used in particular to identify pathogens in cultures and diagnose diseases based on breath samples because cell metabolism is the chemical oxidation of organic compounds, such as glucose C H O, to yield *. In addition to stomach pains, ATP and secondary metabolites can cause halitosis. Also, some diseases can produce characteristic odors, such as those affecting the lungs, liver, or intestine. Wine quality can be measured with the E-nose (portable electronic nose). A micromachined resistive sensor array is developed for this portable E-nose, with one polysilicon heater and another platinum heater. Malvar, Airen, Garnacha, and Tempranillo are the four wines produced in the Madrid region which were tested to see how they affected the nose. A probabilistic neural network (PNN) achieved 88% classification accuracy with this E-nose design, which is portable, requires complex sample preparation, has expensive fluidic circuits, and can be easily deployed in a field test [5]. The principal component analysis (PCA) plots show that the responses

54

V. Anush Kumar et al.

for the platinum heater array are similar to those of the sensor array, and there is a slight overlap in the responses. Developing a formulation of a product idea for upscaling production requires a food product research technique. Throughout the entire process, sensory evaluation plays an important role. A product’s acceptance is particularly important in the last step of product development. A key aspect of satisfying consumer expectations during that stage is measuring the product’s aroma. Electronic noses (E-noses) can be useful for this purpose [6]. The E-nose generates signals that are analyzed by a system to detect gases. In order to investigate the scent factor in some foodstuffs, we developed low-cost E-noses with a compact design and based on low-cost technology. Despite the suggestion [7] that female mate choice is influenced by body odor, it is not proven that it can indicate the quality of a mate in general. An opposite-sex rating group rated the scent, while another rated their attractiveness. Body asymmetry was assessed by measuring asymmetrical traits on the subject’s body. Using portrait photographs, distance measurements were used to examine facial asymmetry. The correlation between body odor and attractiveness among females was significantly positive, while the correlation between asymmetry and smell was significantly negative. Loveliness was negatively correlated with male body odor raters. In this paper, stir bar sorptive extractions of samples were carried out using gas chromatographymass spectrometry (GC–MS) with thermal desorption, with the chromatographic profiles analyzed using pattern recognition techniques. Four out of five samples per individual had more consistent peaks over time than urine or saliva, more than 373, based on four out of five observations per individual [8]. The GC–MS fingerprints, reproducible gender differences, and chemical structures of these candidate compounds allowed us to identify 44 individual compounds as well as 12 genderspecific volatile compounds. These compounds are capable of identifying a variety of genetically determined odors. This review paper [9] presents application of an E-nose-type measurement instrument to evaluate ambient air in the vicinity of municipal processing plants in order to detect odorants characterized by unpleasant odors. In addition to monitoring networks and remote-controlled robots and drones, the electronic nose instruments were also used on portable devices. As well, this paper presents commercially available sensors for electronic noses that are capable of eliminating odors at a level below 1 ppm v/v, which is close to the threshold odor for most odorants. Detecting counterfeit cigarettes may be difficult due to the small difference between their odors. An approach for improving E-nose performance in cigarette brand identification is presented in this study. Various brands of cigarettes were classified using a portable E-nose [10]. An artificial neural network (ANN) was used to identify cigarettes using raw data and was trained with the E-nose data. In this laboratory experiment, four different types of cigarettes were identified. Compared with neural networks trained with extracted parameters, E-nose results were better at identifying.

Human Odor Security Using E-nose

55

3 Proposed Work The goal of this paper is to propose and implement a security system based on E-nose, which optimizes edge-to-edge encryption for extremely vulnerable systems. Our summary of the article’s main contributions is based on the above challenges: 1. 2. 3. 4. 5. 6.

Artificial olfactory sense and recognition system. Implicit data from signals to binary. Encoding of data using ASCII. Concatenation of character to form string. Encryption of data using Caesar cipher. Recognition and authentication of admin with database (Fig. 1).

3.1 Artificial Olfactory Sense and Recognition System The smell of human armpits can be detected and differentiated by software based on an electronic nose (E-nose). Volatile organic compounds are detected using metal oxide sensors. Each sensor’s sensitivity is measured by a voltage divider resistor in the measurement circuit. A portable USB card was used to control the E-nose, a device developed in-house. We have developed a new method to compensate for the humidity noise in armpit odor detection samples, since the sensitivity of gas sensors is affected by humidity. It detected human body odor despite the humidity correction and was able to distinguish it from the smell of two individuals in a relative sense despite the correction. It can still identify people even when they are wearing deodorant. In general, chemical sensors are made up of two main parts: a transducer and a receptor. Transducers measure the signal generated by the receptor, which is converted into chemical information by the receptor. A gas sensor detects gas molecules by interacting with the sensor-coated or sensing materials. The modulated

Fig. 1 Architecture structure of proposed system

56 Table 1 Metal oxide sensors present in E-nose

V. Anush Kumar et al. Sensor

Detecting range (ppm)

Gas target

TGS 825

5–100

Hydrogen sulfide

TGS 2602

1–30

Air contaminants

TGS 813

500–10,000

TGS 822

50–5,000

Organic solvent vapors

Combustible gases

TGS 880

10–1,000

Cooking vapors

electrical current is then converted to a recordable signal by recording the frequency of the electrical current. The metal oxide sensors present in E-nose are given below in Table 1.

3.2 Artificial Olfactory Sense and Recognition System Electronic noses mirror human olfaction, a system not confined to any one mechanism, such that smell or flavor is perceived as global fingerprints. In order to characterize odors, the instrument is constructed with sensors, pattern reorganization modules, and headspace sampling. It is composed of three major components, which are detecting and processing systems, as well as delivery systems for sampling which are shown in Fig. 2.

Fig. 2 Architecture of olfactory sense and recognition system

Human Odor Security Using E-nose

57

The sample delivery system: By providing a headspace generated by the sample or volatile compound, the sample delivery system provides a sample space. The headspace is an analysis fraction. These headspaces are then transmitted to the electronic nose to be detected. The computing system: It is the reactionary part of the instrument that contains the detection system, which is composed of a group of sensors. When sensors come into contact with volatile compounds, electrical characteristics change causing them to react. Metal oxide semiconductor sensor: It is the reactionary part of the instrument that contains the detection system, which is composed of a group of sensors. Sensors react when they come into contact with volatile compounds due to changes in their electrical characteristics. Metal oxide sensors (MOSs): These sensors change conductivity when gas molecules adsorb on them. Changes in conductivity are used to measure adsorption of volatile organic compounds. Piezoelectric sensors: Sensors with piezoelectric behavior: Gas molecules adsorb on the surface of polymers, changing the mass of the sensor. The crystal’s resonant frequency changes as a result. Quartz crystal microbalance: A quartz crystal microbalance measures mass per unit area by measuring the change in frequency of the crystal resonator. This information can be stored in a database. Conducting polymers: As gases adsorb onto the surface, the electrical properties of the polymer change. Samples obtained from unknown analysts are often compared with samples taken from reference libraries or elements of odor identification from well-known sources. The simplest way to reduce data is through a graphical analysis. It is necessary to use both trained and untrained methodologies for analyzing multivariate data. In cases where samples are unknown, these techniques are used. The most popular and easiest technique of untrained MDA is principle component analysis. When sensors are partially sensitive to individual compounds in a sample mixer, a nascent data analysis MDA is very useful. PCA is most helpful when there are no known samples available.

3.3 Implicit Data from Signals to Binary An analog signal: There is an infinite amount of detail in analog data. Binary data is the only type of data that computers can store digitally. Signals in which certain characteristics increase and decrease along with transmissions. Both time and amplitude components of analog signals are continuous. Sampling: In the first step, analog signals are sampled. Sampling converts analog signals to digital signals. Analog signals are transformed into discrete numbers by sampling. Signals are sampled at certain frequencies (or sampling rates) every second. An interval of regular time is used to take samples. A discrete interval is created

58

V. Anush Kumar et al.

by reducing a continuous time domain to a series of discrete intervals. As computer numbers are not infinitely precise and may be rounded off, we can only store y-values of a certain precision. Quantization: The amplitude domain in the y-axis remains wide after sampling. By discretizing the analog signal into a number of quantization levels, an analog signal can be quantized. It is possible to convert an amplitude sample into a discretetime signal by converting it into a finite set of levels. At each sampled point, the actual signal value is compared with the quantized value in order to quantify any quantization error. The finite nature of a computer’s memory and its numeric precision, however, always require some quantization when storing analog data in digital form. Binary encoding: Lastly, binary encoding is necessary. A computer does not need to store the actual value if there are a limited number of quantized y-values. The quantized y-value can instead be stored as a much smaller value. Computers would need to know how the sequence was encoded and sampled in order to understand the digitized version. Each sample is encoded with four bits. A sample’s bit depth can also be described as its number of bits. 0 and 1 are the only values that can be described by a bit depth of 1. Reconstruction: Typically, analog signals are stored digitally, so they can be reproduced later, like displaying an image or playing back audio files. Devices will attempt to reconstruct continuous signals when converting digitized signals back to analog signals. Reconstructed signals are similar to the originals, but they lack some details. It is possible to bring the reconstructed curve closer to the original signal by decreasing the sampling interval and lowering the quantization error. It is also possible to reconstruct the signal using different strategies. Finally, samples were formed by converting infinite streams into finite sequences. This sequence was quantized by approximating its values. A computing device then stored the values as bits. It may be possible to reconstruct the original infinite stream of continuous values at a later time by interpreting those bits.

3.4 Encoding of Data Using ASCII Earlier on, the signals are converted into binary code in order to provide high security in authentication of the user. Firstly, the binary codes are converted into ASCII and based on the values and the alphabets are given as output in this process. Binary Codes Are Those Seemingly Endless Sets Of Zeroes. A System Consisting Of Zeros And Ones Can Be Described As A Two-Symbol System, As Its Name Suggests. The TwoSymbol Code Can Represent Information Such As Computer Instructions, Text, And Other Sources Of Information. In This Way, Binary Code Became Widely Used In Electronic Devices, Computers, And Many Other Applications. Binary Code Is Used In Many Applications, Such As Computer Monitors, In Which Binary Is Converted To Text. It Would Be Much More Difficult To Enter Text Typing Commands And For The Resulting Text To Appear On The Screen If Modern Computers Did Not

Human Odor Security Using E-nose

59

Have Super-Fast Processing Speeds. A Machine-Readable Binary Code Is Generated From The Letters And Words We Enter, Which Is Then Converted Back To Human-Readable Text For Display. The Following Sections Will Explain Several Key Concepts That Will Give Us A General Sense Of How Binary Codes Are Converted To Text. Several key points will be covered, but not every step will be covered in detail. In Contrast To The Codes Used By Other Programming Languages, Binary Code Cannot Easily Be Converted. It Doesn’t Automatically Translate A Given Set Of Binary Code Into A Prescribed Character, Text, Or Function When The Codes Are Keyed In. It Needs Another The American Standard Code for Information Interchange (ASCII) is a code used for transferring information. So Computer Systems Can Recognize It And Convert It Into Letters And Words. ASCII Could Be Viewed As An Intermediary That Enables A Company To Reach Its Clients. When Converted From Binary Code Into Text Or Vice Versa, It Performs The Same Intermediary Role. The ASCII Character Set Is Therefore An Integral Part Of Any Binary Translator. Computer Systems Can Recognize The Values Assigned By ASCII. In ASCII, A Combined Combination Of Binary And Decimal Values Represents A Specific Sequence Of Letters And Numbers. According To ASCII Standard Conversion, The Computer Interprets The Binary 01,100,101 As The ASCII Decimal Value 101, Which It Interprets As The Letter "E". On The Computer Screen, This Is Represented As Letters "E". Online Code Conversion Programs Such As Binary To Text Converters And Binary To ASCII Converters Also Follow This Procedure. The BuiltIn Capability Of Computers To Convert Binary Codes To Text Must Be Considered In Light Of Possible Applications Beyond Computers. In Order To Store And Share Information With The Outside World, It Is Important To Recreate The Same Conversion System. A Computer Network Is Connected To The Internet Using The Internet Protocol Suite (TCP/IP) For Communication With Different Devices And Networks. It uses a system of machine-readable codes, including binary codes, to facilitate efficient data exchange between multiple devices over a wide area. In the virtual world, conversion systems create text-intensive information such as articles, posts, comments, emails, chats, and other types of communication. The Ascii Code for Converting Binary Code into Alphabets Are Given Above In the Tables 2 and 3.

3.5 Concatenation of Character to Form String Previously, the binary codes are converted into character, and the output produced is list of characters. The list of character is needed to combine together in order to get output as a single string. The method used to convert character a single string is traversal of list. Firstly, initialize an empty string; secondly, the list of character is in traverse as for every index add an character to the initialized string. Lastly, after the completion of the traverse process, print the string formed by the adding of the characters.

60

V. Anush Kumar et al.

Table 2 Binary code to alphabets (uppercase) Alphabet

ASCII code

Binary code

Alphabet

ASCII code

Binary code

A

065

01000001

N

078

01001110

B

066

01000010

O

079

01001111

C

067

01000011

P

080

01010000

D

068

01000100

Q

081

01010001

E

069

01000101

R

082

01010010

F

070

01000110

S

083

01010011

G

071

01000111

T

084

01010100

H

072

01001000

U

085

01010101

I

073

01001001

V

086

01010110

J

074

01001010

W

087

01010111

K

075

01001011

X

088

01011000

L

076

01001100

Y

089

01011001

M

077

01001101

Z

090

01011010

Alphabet

ASCII code

Binary code

Table 3 Binary code to alphabets (lowercase) Alphabet

ASCII code

Binary code

a

097

01100001

N

0110

01101110

b

098

01100010

O

0111

01101111

c

099

01100011

P

0112

01110000

d

0100

01100100

Q

0113

01110001

e

0101

01100101

R

0114

01110010

f

0102

01100110

S

0115

01110011

g

0103

01100111

T

0116

01110100

h

0104

01101000

U

0117

01110101

i

0105

01101001

V

0118

01110110

j

0106

01101010

W

0119

01110111

k

0107

01101011

X

0120

01111000

l

0108

01101100

Y

0121

01111001

m

0109

01101101

Z

0122

01111010

Human Odor Security Using E-nose

61

Pseudo code: def convert(s): new = "" for c in s: new += c return new s=[] print(convert(s)) I/P : ['a','b','c','d','e','f'] O/P : abcde

3.6 Encryption of Data Using Caesar Cipher Strings are created by concatenating alphabets. In order to make the user’s experience more secure, the strings must be encrypted. Data are encrypted with Caesar cipher. Caesar ciphers encode messages simply. For Caesar ciphers, letters in the alphabet are transformed into encoding alphabets by shifting them by a fixed amount of space, which are given in Fig. 3. Steps for designing and using a Caesar cipher • The alphabet will be shifted by the value you choose. • Make a table with the letters in the top row arranged in standard order and the new shifted alphabet in the bottom row. • You encode the message by substituting the shifted letters for each of the letters in the message. • To enable the recipient to decode the message, you must inform them of the shifting scheme used to encode the message. • By subtracting 26 from the shift value, you can decrypt a Caesar-encoded message by shifting it back to its original form. Fig. 3 Encrypting using Caesar cipher

62

V. Anush Kumar et al.

3.7 Recognition and Authentication of user with Database It is crucial to know your users in order to secure your network. Identification is the first step to determining their privileges and access rights, and the process by which auditing their actions is accomplished. Before a database session can be created, users need to be authenticated in one of a number of ways. Users can be defined in database authentication such that both identification and authentication are performed by the database. The operating system or network service can perform external authentication for you when you define users. Additionally, you can configure your users’ accounts, so they are authenticated via Secure Socket Layer (SSL). Using the enterprise directory, enterprise roles allow enterprise users to authorize access to the database. Your middle-tier server allows you to specify users who may connect through it. The middle-tier server authenticates and acts as the identity of the user, and it is able to create specific roles for the user. This site uses proxy authentication.

3.8 Passwords for Authentication It is a basic form of authentication to use a password. To prevent unauthorized access to the database, users must establish a connection with the correct password. An authentication process is used to verify users connecting to a database with the information stored in the database. Information stored in the database is encrypted. An encrypted user password can be saved in the data dictionary of a database. This makes changing passwords a straightforward process. It is very important that passwords in database security systems remain secret at all times. The problem is that passwords can be stolen, forged, and misused. By using user profiles, DBAs and security officers can control password management policy, not only strengthening the basic password feature but also increasing database security. Password complexity standards, such as minimum password length, can be established by the DBA. A good password should not appear in a dictionary. People’s names or birthdates should not be included in passwords. Eventually, passwords expire and are no longer valid. It is recommended that passwords can be changed periodically to avoid this issue. You can prohibit password reuse for a certain number of days. A server can automatically lock a user’s account if their login attempts exceed a certain threshold. An individual’s biometric characteristics are used to verify that they are who they claim to be through biometric authentication. Authentication is conducted through a comparison of stored data with physical or behavioral characteristics. Two samples must match for authentication to be confirmed. Buildings, rooms, and computers are commonly protected with biometric authentication. The security and convenience it provides are indisputable. Biometric verification uses unique characteristics that are hard to replicate.

Human Odor Security Using E-nose

63

3.9 Proxy Authentication and Authorization The security of middle-tier applications is crucial in multitier environments such as transaction processing monitors as clients’ identities and privileges are tracked and audited. You can accomplish this with proxy authentication. This feature, for instance, allows multiple applications (also known as proxies) to pass information about an individual’s identity to the database server. Sometimes, it lets the application pass the credentials to the database server so that it can validate the credentials of a user. Database administrators are able to control which users are allowed to access a given application through the database server. It enables administrators to monitor the actions taken by a given application for a specific user. Data retrieval is included in query processing. First, each user query is translated into a database language such as SQL. To access a file system, the translated information is converted into expressions. The evaluation of queries is then carried out, along with a number of query optimizations. In order to process a query, a computer system must convert the query into a language that humans can understand. A query’s parser performs the same translation process as the query translator. When parsing a query request into an internal query form, the parser takes into account the syntax, the name of the relation in the database, the tuple, and then the attributes. As a result of the parser, a tree representing the query is generated. By doing so, all the views are automatically replaced when used in the query. Database users and non-database users can both authenticate with Oracle databases. All Oracle database users are typically authenticated using the same method, regardless of the number of authentication methods used at a database instance. Special authentication procedures are required for database administrators running special operations on the database. During the transmission of passwords over the network, Oracle database encrypts them and is mentioned in Fig. 4. Fig. 4 Storing and fetching data using oracle

64

V. Anush Kumar et al.

4 Result and Analysis A research was made in the research laboratory for 2 days with two volunteers. Two male volunteers were asked to provide odor samples from their armpits. For five days, volunteers were asked to collect their armpit odors immediately after waking up (7–8 in the morning) and eight hours later (8 in the afternoon). Armpit odors were transferred to the E-nose using cotton pads. For ten minutes, cotton pads must be placed directly in the armpits and stored in bottles with screw-on lids. Glass bottles collected in the morning were transferred to the laboratory for E-nose measurements. A heat-protected container was used to transfer the samples within 30–50 min after collection, in order to minimize the odor change caused by bacteria. Immediately following odor sampling, E-nose measurements can be conducted for afternoon samples. During the experiment, volunteers were expected to carry on their normal daily routine, including showering twice a day before bedtime and once after waking in the morning after collecting sample in the morning. In order to avoid fluctuating for sex or alcohol consumption, they were prohibited from touching the odor samples. Deodorant was only applied to the right arm of volunteers after showering in the morning. Two male volunteers were asked to provide odor samples from their armpits. For five days, volunteers were asked to collect their armpit odors immediately after waking up (7–8 in the morning) and eight hours later (8 in the afternoon). Armpit odors were transferred to the E-nose using cotton pads. For ten minutes, cotton pads must be placed directly in the armpits and stored in bottles with screw-on lids. The concentration of isovaleric acid levels is mentioned below in Table 3. Glass bottles collected in the morning were transferred to the laboratory for E-nose measurements. A heat-protected container was used to transfer the samples within 30–50 min after collection, in order to minimize the odor change caused by bacteria. Immediately following odor sampling, E-nose measurements can be conducted for afternoon samples. Following the morning sample collection, volunteers were asked to shower twice before going to sleep and once after waking up during the experiment period. For fear of having sex or consuming alcohol, they were not permitted to touch the odor samples. The collected data of odor is processed to get as a binary code. The binary code and ASCII values are previously described in Tables 2 and 3. The binary code which is transformed into alphabets is processed below. Volunteer

Alphabets

ASCII

1

a

097

01100001

1

f

0102

01100110

Binary code

1

l

0108

01101100

1

R

0114

01110010

1

j

0106

01101010

Human Odor Security Using E-nose Volunteer

Alphabets

65 ASCII

Binary code

2

X

0120

01,111,000

2

d

0100

01,100,100

2

N

0110

01,101,110

2

i

0105

01,101,001

2

V

0118

01,110,110

The alphabets need to be concatenated now for the both data of volunteer 1 and volunteer 2. The concatenated string is: Volunteer 1: {a,f,l,R,j} = {aFlRj}. Volunteer 2: {X,d,N,i,V} = {XdNiV}. The string which is concatenated is encrypted using Caeser cipher of shift 3. {aFlRj} = xTbed. {XdNiV} = CYgoH. Then, this encrypted string is stored before in the databases, and then, it validates the user. If the string matches with the encrypted string in database and the string produced and give access to the user.

5 Conclusions In this paper, we are found an effective way of increasing security using biometrics. Human order is used as the biometrics. The human order is converted to signals unprocessed as binary code and then converted into alphabets, and these alphabets are concatenated to form a string and encrypted using Caesar cipher. The user data is created and stored in the database and fetches the information of user while accessing. By this method, we can keep our information safe and secure.

References 1. Brattoli M, de Gennaro G, de Pinto V, Loiotile AD, Lovascio S, Penza M (2011) Odour detection methods: olfactometry and chemical sensors. Sensors 11: 5290–5322. ISSN 1424-8220. (open access) 2. Tzong-Zeng W (1998) A piezoelectric biosensor as an olfactory receptor for Odour detection: electronic nose. Biosens Bioelectron 14: 9–18. (Elsevier) 3. Jacob TJC, Wang L (2006) A new method for measuring reaction times for odour detection at is o-intensity: Comparison between an unpleasant and pleasant odour. Physiol Behav 87:500–505. (Elsevior) 4. Gardner JW, Shin HW, Hines EL (2000) An electronicnose system to diagnose illness. Sens Actuators B 19–24. (Elsevier)

66

V. Anush Kumar et al.

5. Aleixandre M, Lozano J, Gutierrez J, Sayago I, Fernandez MJ, Horrillo MC (2008) Portable e-nose to classify different kinds of wine. Sens Actuators B 131:71–76. (Elsevier) 6. Chilo J, Pelegri-Sebastia J, Cupane M, Sogorb T (2016) E-nose application to food industry production. IEEE Instrum Meas Mag 27–33 7. Rikowski A, Grammer K (1998) Human body odour, symmetry and attractiveness. R Soc 266:869–874 8. Penn1 DJ, Oberzaucher E, Grammer K, Fischer G, Soini HA, Wiesler D, Novotny MV, Dixon SJ, Xu Y, Brereton RG (2007) Individual and gender fingerprints in human body odour. R Soc Interface 4:331–340 9. Szulczy´nski B, Wasilewski T, Wojnowski W, Majchrzak T, Dymerski T, Namie´snik J, G˛ebicki J (2017) Different ways to apply a measurement instrument of E-nose type to evaluate ambient air quality with respect to odour nuisance in a Vicinity of municipal processing plants. Sensors 17:2671 10. Luo D, Hosseini HG, Stewart JR (2004) Application of ANN with extracted parameters from an electronic nose in cigarette brand identification. Sens Actuators B 99:253–257. (Elsiever)

Automated Road Surveillance System Using Machine Learning Ashish Vishnu, S. Sushmitha, Tina Susan Jacob, A. David Maxim Gururaj, and S. Dhanasekar

Abstract Road safety requires an understanding of traffic rules. It is also not just the responsibility of oneself but the coordination of every individual on the road to be aware and alert to avoid accidents. The objective of the paper is to analyze the impact of the accident and identify the vehicle which is being prone to accidents using image classification through machine learning. Machine learning provides the system with an ability to automatically learn and improve from the given dataset without human intervention or assistance. It looks for patterns in the data and takes a decision accordingly. The training process involves the following steps: collecting the images, annotating the image, data ingestion, and data processing. This paper follows the convolutional neural network algorithm that takes image inputs, assigns various aspects to images, and differentiates them from one another. The image recognition model automatically determines whether the incident in the given image is an accident with the help of bounding boxes. These bounding boxes surround themselves on vehicles which are prone to accidents. Keywords Road safety · Accidents · Automated surveillance · Machine learning · TensorFlow · Convolutional neural network

1 Introduction During the year 1886, Karl Benz introduced the first motor vehicle [1]. When the number of motor vehicles and their owners increased, it gave rise to issues of handling traffic. As a response to the issue, in 1901, the state of Connecticut created statewide traffic laws [2]. A. Vishnu · S. Sushmitha · T. S. Jacob · A. David Maxim Gururaj (B) · S. Dhanasekar (B) School of Advanced Sciences, Vellore Institute of Technology, Chennai, India e-mail: [email protected] S. Dhanasekar e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_5

67

68

A. Vishnu et al.

The news reports stated that thousands of cars on roads created more dangerous accidents within city limits. As a precaution to avoid accidents, the first surveillance machine was introduced, in Germany in the year 1942 [3]. These monitoring machines are used for live monitoring. This paper includes the usage of Python as a backend language, TensorFlow, and its dependencies for building machine learning models that are compatible with problems, image classification, and algorithms. This system is an alternative to the traditional method of manually monitoring the roads with the help of CCTV. This automated surveillance system detects the actions that occur on the road, and in case of an accident, a bounding box appears around the affected vehicle. “Impatient on the road, patient in the hospital”, styling the above quote in reallife solutions and implementing it gave rise to our research paper on how to build a surveillance machine for road safety.

2 Literature Survey The early experiments with road safety surveillance are as follows: Road safety surveillance system using hybrid CNN-LSTM [4] is an age and fatigue detection model. It captures images of the driver and uses a face recognition system to detect the drive’s age, and the blink frequency is used to detect fatigue. In case the driver is found drowsy, the system will give an alarm alert. Automated vehicle [5] is an IoT-enabled system that has been trained to get the information from the cloud and abide by the traffic rules. This system is trained to control the vehicle speed according to the area’s speed limit, alert the driver when humps ahead, disable the horn in a No Horn Zone, and notify the traffic department when parked in a No Parking Zone. This system is controlled by a cloud database and has a threat of being hacked or wrong information passed. Computer vision-based accident detection [6] system uses CCTV surveillance footage to detect collision. This system continuously tracks the vehicles to detect any overlapping of vehicles and further records the acceleration rate and angle of the collision to perform a mathematical calculation. It successfully detects collision at the rate of 71% and also has a 0.53% false alarm rate. Car accident detection and notification system [7, 8] is a mobile application consisting of Activating Accident Detection Activity as well as Deactivate Accident Detection Activity. It uploads images as well as videos. During an emergency, the driver’s mobile is installed with it. The electronic call system is an automated accident detector [9] and notifier for smartphones. The complexity of this system is that it uses the smartphone’s builtin accelerometer sensor as an accident sensor. In this scenario, the electronic call system is subjected to a high false alarm rate as well as false emergency calls when the user is not present in the vehicle. The license plate detection model [10] captures the images of the vehicle’s number plate that violate safety rules and compares the number code in the Nation’s

Automated Road Surveillance System Using Machine Learning

69

license plate database. The system extracts information about the vehicle’s owner and sends a violation notification. It works only for rule violation. These are few important recent papers published on road safety surveillance systems and we saw that there are few shortcomings to be addressed. This can be rectified with the help of the TensorFlow object detection API. Sreyan Ghosh et al. [11] proposed the concept of taking every frame of a video and running it in a CNN model which has been educated to categorize frames of a video into accident or non-accident. Pallav Borisagar et al. [12] proposed that CNN training takes a long time, a lot of data, and a lot of processing power. To address these concerns, a unique transfer learning technique has been introduced for the accident detection application, which entails retraining the already learned network. Remigiusz Baran et al. [13] proposed that the smart camera components and their capabilities for automatic detection and recognition of specified automotive attributes, as well as various aspects of system efficiency, are explained and discussed in detail. As the system’s key benefits, smart capabilities such as make and model recognition (MMR), license plate recognition (LPR), and color recognition (CR) are highlighted. Busarin Eamthanakul et al. [14] proposed that the framework brings a traffic picture from a CCTV camera to measure in the framework as an information. From that point forward, the framework finds for gridlock and gets the outcomes in three rush hour gridlock conditions as Stream, Weighty, and Stuck. At long last, a client can utilize the framework for a transportation arrangement or a convergence traffic light.

3 Proposed System The proposed system recognizes the vehicle that is being harmed during a mishap. This system trains the images for another step of progress in its working for identifying the action in the image and is capable of identifying accidents when a live video or an image is passed through the pre-trained neural network. This is one step toward reducing surveillance and improving automated surveillance on the road.

4 Methodology The surveillance system is trained to easily detect the accidents through computer vision. This paper involves the following process: A. Data collection. B. Data cleaning. C. Image annotating.

70

A. Vishnu et al.

Fig. 1 Flowchart

D. Data ingestion. E. Data processing. Model Training—Fig. 1 shows a flowchart representation of each step involved.

4.1 Data Cleaning Data cleaning is the process of detecting and correcting (or eliminating) irrelevant or erroneous images from the data collected and alludes to distinguishing deficient, mistaken, incorrect, or superfluous pieces of the information, or erasing the grimy or coarse data. Data cleaning might be performed intelligently with data wrangling tools or batch processing [15]. The irrelevant images in this dataset are images that include normal roads and vehicles that are not prone to accidents, which are removed from the dataset. Then,

Automated Road Surveillance System Using Machine Learning

71

Fig. 2 Various types of accident Source www.google.com

noise from the suitable images is reduced to avoid misleading the system; once both the steps are done, we get a clean dataset that is ready to be annotated.

4.2 Image Annotation LabelImg is the most supported tool for image annotation [16]. Fig. 2 gives some sample images of the accidents. All the images collected from the previous steps are annotated with the desired label. In our case, the label is accident. The annotation is done by creating bounding boxes around the object in the image. This process is repeated for all the images and saved as an .xml file containing the label data for each image as x and y diagonal coordinates in Fig. 3 (Fig. 2 and 3).

4.3 Data Ingestion Layer Data ingestion is the most crucial step in the process of data handling method. The layer converts the .xml files to a .csv file that contains the consolidated essential information from all the .xml files [17]. Figure 4 shows a sample.csv file generated from the .xml files.

72

A. Vishnu et al.

Fig. 3 Sample.xml file

Fig. 4 CSV file

4.4 Data Processing Layer The data processing layer uses Python code that receives the data from the previous steps and trains the model with TensorFlow and convolutional neural network (CNN) [18] algorithms. TensorFlow [19] is an open-source programming library for AI. It tends to be utilized over a scope of assignments however has a specific spotlight on

Automated Road Surveillance System Using Machine Learning

73

preparing and induction of profound neural networks. It is a symbolic math library that depends on dataflow and differentiable programming. It is utilized for both exploration and creation. (a) Convolutional Neural Network A convolutional neural network (CNN) [20, 21] is an algorithm that takes image inputs, assigns importance to various aspects of the images, and differentiates it from one and the other. The preprocessing done on CNN [18] is lower when compared to other classification algorithms. While in ancient methods, filters are hand-engineered, with enough training, CNN [18] can learn these characteristics. The architecture of a CNN [18] is equivalent to that of the connectivity pattern of neurons in the human brain and inspired by the visual cortex organization. Figure 5 explains the working of a convolutional neural network. CNN has two layers [19] i. The convolutional layer is a critical part of CNN [18].. This layer can extract features or feature maps from input images. Each convolutional layer can have different convolution kernels, which are utilized to acquire various feature maps. ii. A pooling layer is commonly sandwiched between two convolutional layers. The fundamental function of this layer is to lessen the feature map dimensions and keep up the scale invariance of the features. There are two primary pooling strategies: mean pooling and max pooling. A pooling impact chart is shown in Fig. 6. The average value in an image is taken as the pooled value in mean pooling

Fig. 5 Convolutional neural network. Source https://www.javatpoint.com

74

A. Vishnu et al.

Fig. 6 Pooling effect. Source https://www.hindawi.com/journals

as it preserves the background of the image. The maximum value of the image is taken as the pooled value in max pooling, and it preserves the texture of the image. (b) Model Formulation Let convolutional layer be “l”, feature maps as the input and has “k” feature maps as output, so the filter size is given by “n*m”. The convolution layer is calculated as [18, 19]: ⎛ x lj = f ⎝



⎞ xil−1 ∗ kil j + blj ⎠,

i∈M j

where: xil−1 is the characteristic map of the output of the previous layer. x lj is the output of the ith channel of the jth convolution of the layer. f (·) is called the activation function. M j is a subset of the input feature maps used to calculate u lj . kil j is a convolution kernel. blj is the corresponding offset. In the fully connected layer, all sources of info units have a divisible load to each yield unit. For “n” sources of input and “m” yields, the number of weights is “n ∗ m.” This layer has the inclination for each yield, so “(n + 1) ∗ m” parameters [22]. The number of parameters learned within the network is calculated as (n ∗ m ∗ k + 1) ∗ f , where “f” is the filter size. The output size of the CNN [18] layer can be calculated as input_si ze − ( f ilter _si ze − 1).

Automated Road Surveillance System Using Machine Learning

75

For example, consider the input image_si ze as (25, 25) and filter as (2, 2), then the output size of the CNN [18] layer is (25−(2−1)) = 24.

4.5 Model Training Once we have decided on the preferred model and method to train it, we will start our model training [23, 24]. This paper describes the training of a model for detecting the images that have met with accidents or any other form of a mishap to the vehicle. Model training is the key element of ML where we are teaching the machine to recognize the given image and nature of it. (a) Algorithm Step 1: All the .xml files in the train and test folder are converted into .csv file. Step 2: After the train/test .csv is created we convert them into a machine understandable format i.e., record format. (It will decrease load from memory and complexity). Step 3: A python program train.py is used to train the model with train record. Step 4: During model training there incur some loss of data. Step 5: The model saves temporary checkpoints as. ckpt files while training. (This will change as we increase the training), the checkpoints are saved at random. Step 6: Once a desired number of checkpoints are created i.e., loss of data in training is less than 0.05(>= 5%) the training is stopped and the model is saved. Step 7: The object detection API is successfully trained to identify accidents with input images. Figure 7 represents the above given steps as a flowchart.

5 Result The below-given figures are the final outcome of the object detection API, that successfully identifies the vehicles that meet a massive mishap [25] by dropping a bounding box around the affected vehicle. These are the output images of different vehicles that meet with a massive mishap (Figs. 8, 9, 10 and 11).

6 Ease of Use A. Edge of precision Due to long waits and pandemic situations, which paves its way to uncertainty, people tend to move out on-road driving speedily. The more people on the roads, the more

76

A. Vishnu et al.

Fig. 7 Algorithm

the congestion will be. The surveillance system can help in traffic measurement and planning accordingly. This leads to encouraging safe driving. B. Maintaining the Specification Integrity The integrity of the data collected and training the collected data is the prime focus of the present SET. Its accuracy and consistency are taken care of throughout by annotating the image collected into XML units and stored separately.

Automated Road Surveillance System Using Machine Learning Fig. 8 Output-1

Fig. 9 Output-2

Fig. 10 Output-3

77

78

A. Vishnu et al.

Fig. 11 Output-4

7 Scope of Future Direction The forthcoming research will focus on how the system identifies theft, traffic rule violation, and harassment. The other add-on feature expected in the future release is notifying the nearest hospital and police station on account of the accident. The nearest police station will be notified regarding the theft, traffic rule violation, and harassment cases. A webpage that can monitor live streaming videos, that will notify the respective departments regarding the issue, will be created in the forthcoming semesters.

8 Conclusion The research paper is based on the image classification model that is trained to identify accidents. TensorFlow object detection API along with convolutional neural network (CNN) is used to develop this image classification model. This paper specifically focuses on accidents. We have attained 70–90% accuracy in detecting the vehicles that are prone to accidents. The accuracy rate depends on the clarity of mishap.

References 1. Iguchi M (2002) Evolution of automobiles. In: Proceedings of conference on intelligent vehicles, 0.1109/IVS.1996.566396, 06 Aug 2002 2. Silvano AP (2016) Advancing traffic safety-an evaluation of speed limits, vehicle-bicycle interactions, and I2V systems. In: Connecticut’s first speed-limit law-law regulating motor vehicles. ISBN 978-91-87353-94-9 3. Olagoke AS, Ibrahim H, Teohi SS (2020) Literature survey on multi-camera system and its application, vol 8. (6 Sept 2020)

Automated Road Surveillance System Using Machine Learning

79

4. Babitha D, Ismail M, Chowdhury S, Govindaraj R, Prakash K B (2020) Automated road safety surveillance system using hybrid CNN-LSTM approach. Int J Adv Trends Comput Sci Eng 9(2). (March–April 2020). 5. Namratha MM, Navya MN, Niharika R, Namitha NV, Sunitha R (2020) Automated vehicle: the prospective of road safety. Int J Eng Res Technol (IJERT). ISSN: 2278-0181. (Published by www.ijert.org, NCCDS-2020 Conference Proceedings) 6. Ijjina EP, Chand D, Gupta S, Goutham K (2019) , Computervision-based accident detection in traffic surveillance, Proc. 10th Int. Conf. Comput., Commun. Netw. Technol. (ICCCNT), Jul. 2019, pp. 1–6 7. Ali HM, Alwan ZS (2015) Car accident detection and notification system using smartphone. IJCSMC 4(4):620. (April 2015) 8. Gupta R, Patel AS Singh, Ojha M (2021), Accident Detection Using Time-Distributed Model in Videos, Proceedings of Fifth International Congress on Information and Communication Technology 2021, pp. 214-223.. 9. Pinart C, Calvo JC, Nicholson L, Villaverde JA (2009) ECall-compliant early crash notification service for portable and nomadic devices 10. Patel C, Shah D, Patel A (2013) Automatic number plate recognition system (ANPR): a survey. Int J Comput Appl (0975–8887) 69(9). (May 2013) 11. Ghosh S, Sunny SJ, Roney R (2019) Accident detection using convolutional neural networks. In: 2019 international conference on data science and communication (IconDSC), 1–2 Mar 2019 12. Borisagar P, Agrawal Y, Parekh R (2018) Efficient vehicle accident detection system using tensorflow and transfer learning. In: 2018 international conference on networking, embedded and wireless systems (ICNEWS), 27–28 Dec 2018 13. Shimizu K, Shigehara N (2002) Image processing system using cameras for vehicle surveillance. In: Second international conference on road traffic monitoring, 1989, 06 Aug 2002 14. Eamthanakul B, Ketcham M, Chumuang N (2017) The traffic congestion investigating system by image processing from CCTV camera. In: 2017 international conference on digital arts, media and technology (ICDAMT), 1–4 Mar 2017 15. Ridzuan F, Zainon WMN (2019) A review on data cleansing methods for big data. In: The fifth information systems international conference (Jan 2019) 16. Petrovai A, Costea AD, Nedevschi S (2017) Semi-automatic image annotation of street scenes. In: 2017 IEEE intelligent vehicles symposium (IV), 31 July 2017 17. Yoshida S, Yahagi H, Odagiri J (2004) CSV compaction to improve data-processing performance for large XML documents. In: Data compression conference, 2004. Proceedings. DCC 2004, 24 Aug 2004 18. Liu Z, Lian T, Farrell J, Wandell BA (2020) Neural network generalization: the impact of camera parameters. IEEE Access 8:10443–10454 19. Agnes Lydia A, Sagayaraj Francis F (2020) Multi-label classification using deep convolutional neural network. In: 2020 international conference on innovative trends in information technology (ICITIIT), 13–14 Feb 2020 20. Kumar P, Dugal U (2020) Tensorflow based image classification using advanced convolutional neural network, IJRTE 8(6):2277–3878 21. Chirodea MC, Novac OC, Novac CM, Bizon N, Oproescu M, Emilia C (2021) Comparison of tensorflow and pytorch in convolutional neural network-based applications. In: 2021 13th international conference on electronics, computers and artificial intelligence (ECAI),1–3 July 2021 22. Sinha T, Verma B, Haidar A (2017) Optimization of convolutional neural network parameters for image classification. In: 2017 IEEE symposium series on computational intelligence (SSCI), 27 Nov–1 Dec 2017 23. Albayrak NE (2020) Object recognition using tensor flow. In: 2020 IEEE integrated STEM education conference (ISEC), 1 Aug 2020 24. Sujeetha R, Mishra V (2019) Object detection and tracking using tensor flow. IJRTE. 8(1). ISSN: 2277-3878. (May 2019)

80

A. Vishnu et al.

25. Singh R, Singh A, Bhattacharya P (2022) A machine learning approach for anomaly detection to secure smart grid systems. In: Research anthology on smart grid and microgrid development, IGI Global, pp 911–923

An Artificial Intelligence-Based Technique to Optimize Hybrid Vehicle Using Renewable Energy Sources S. Dhanush Hariharan, Challapalli Manikantaa, C. S. Manigandaa, V. Anush Kumar, Vetri Priya, and V. D. Ambeth Kumar

Abstract A review of water electrolysis technologies for hydrogen production is presented in this paper, and a sustainable, economical, and eco-friendly method to save the environment yet normally carries out the everyday activity of new age man which is “Transportation”. Current electrolyzers must be reduced in energy consumption, cost, and maintenance in order to facilitate water electrolysis expansion and, on the other hand, made more durable, efficient, and safe. Due to its harmful environmental effects, fossil fuel use in vehicles is a growing concern. A hybrid electric vehicle is among the most promising technologies currently being developed, as it is high-performing, fuel-efficient, low emission, and has an impressive range. The main aim of this paper is to integrate artificial intelligence and transportation machines and engineer a feasible and non-polluting vehicle that contributes to the making of a sustainable environment. The successful implementation of this project will result in the creation of a better and pollution-free environment that contributes to the conservation of energy and fossil fuels. This paper also discusses a better yet safer way to store hydrogen fuel on a small scale and prevent accidents and save the lives of the users with the help Anti-Accident protocol. Keywords Hybrid vehicle · Renewable energy · Artificial intelligent

1 Introduction Among other factors, energy influences a nation’s economy, infrastructure, transportation, and standard of living. The production of energy in every nation is dependent upon fossil fuels, which are unsustainable. Using fuels derived from fossil and nuclear sources extensively poses dangerous environmental threats, such as S. Dhanush Hariharan · C. Manikantaa · C. S. Manigandaa · V. Anush Kumar · V. Priya · V. D. Ambeth Kumar (B) Department of AI&DS, Panimalar Engineering College, Chennai 600123, India e-mail: [email protected] V. D. Ambeth Kumar Department of Computer Engineering, Mizoram University, Aizawl 796004, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_6

81

82

S. Dhanush Hariharan et al.

exhausting resources derived from nature emitting pollutant gases, generating waste, and causing climate change. To provide adequate energy to the world’s rapidly growing population, there is a need for an alternative, sustainable energy source that does not negatively impact the environment. With the use of renewable energybased electric systems, new challenges arise for storing and utilizing surplus energy, managing distributed generation, ensuring energy supply reliability, as well as integrating it with the automotive sector. New hydrogen technologies could contribute to the development of an energy system that addresses these issues. An economy based on hydrogen envisions low-cost, environmentally clean hydrogen production. Hydrogen has the lowest volume of energy, but the highest weight of energy of any molecule. The high energy efficiency of hydrogen makes it an most appropriate fuel for applications such as fuel cells and rockets. In comparison to fossil fuels, hydrogen does not emit harmful emissions, making it one of the most attractive fuels. In addition, hydrogen has a heating value three times greater than petroleum. In a hybrid electric vehicle (HEV), two power sources are used to run the vehicle. An IC engine (hydrogen or diesel-fueled) and chemical batteries with an electric motor drive are currently used as power sources. With optimal operating points and regenerative braking, a small IC engine can save fuel and emit fewer emissions compared to an engine powered solely by an IC. Various applications rely on batteries for power, including portable electronics and electric cars. Electric vehicles can reduce gasoline consumption by as much as 75% today, making EV batteries a popular option for automotive manufacturers. The advantages of hybrid electric vehicles make them the most promising future vehicle alternatives.

2 Related Work This paper presents a model-driven approach for dynamic load control in real time for parallel hybrid vehicles in this paper [1]. Fuel-optimal managing is sought that relies only on current system operation instead of an idea of the driving conditions that will arise in the future (global optimal control). In addition to problems involving hard constraints on the state of a battery, arc costs, fuel consumption, which are not explicitly determined by the state, the developed methodology can also be applied to problems involving arc costs, such as flight and fuel consumption. With an appropriately defined cost function at each time instant, it is possible to find a suboptimal control. The equivalence factors have been redefined so that the two types of energy can be compared. Mercedes A-Class was modified to hybridize with the “Hyper” prototype of DaimlerChrysler. A simulation has demonstrated potential for fuel economy and deviation reduction with the proposed control system from standard operating procedures down to a minimum. This paper reviews the technology for hydrogen production through water electrolysis and analyzes the current state of the electrolysis of water integration with energy sources that are renewable. In the first part of the course, we examine thermodynamics and electrochemistry to gain a better understanding of how electrolysis cells work and

An Artificial Intelligence-Based Technique to Optimize Hybrid Vehicle …

83

how they can be combined into larger electrolysis modules. The following discussion discusses the electrolysis process, its advantages, disadvantages, and challenges. In this article, the main features of electrolyzers currently available on the market are thoroughly discussed [2]. A few relevant demonstration projects are discussed as well as the possible configurations that can be used to integrate a system that combines renewable energy with water electrolysis, both autonomously and on a grid. In a future energy model based on electricity and hydrogen, water electrolysis can play an important role. As renewable energy sources, primarily hydropower, wind, and photovoltaic, can easily produce hydrogen via water electrolysis, they are able to produce clean and sustainable hydrogen. The polymer electrolyte membrane electrolyzer is capable of electrolyzing alkaline and acidic solutions. Since the advent of mobile technology [3], the issue of energy storage has become increasingly important. It is not inconceivable that fossil fuels and natural gas will always be available. Even if they were, their combustion or end products would not be environmentally friendly. Energy sources such as hydrogen can help solve this problem. In the 1970s and 1980s, oil crises and technological advancements led to an increase in interest in hydrogen as a fuel for vehicles. Hydrogen-powered vehicles have been in use since 1800. All sorts of things have been powered by it, from balloon flights to rockets. Among the non-toxic, clean, abundant, and renewable fuels, hydrogen gas ranks first [1–3]. Hydrogen gas releases nothing other than water vapor into the atmosphere when it is combusted. As it dissipates quickly into the atmosphere, there is no spilling or pooling concern [4–6]. In terms of energy content per mass, it contains nearly twice as much as any hydrocarbon fuel (142 MJ). It is a very light fuel with a high weight energy content (three times more than gasohol), whereas hydrocarbons are very heavy and have a very low volume energy content (four times less than gasohol) and burn faster than gasoline. Petrol molecules are 3.2 times less energetically dense than natural gas molecules and 2700 times less energetically dense than gasoline molecules. In this sense, hydrogen serves as a carrier of energy and not a source as it stores and delivers power. Based on the analysis, it cannot be concluded that one technology/fuel is the best option. One of the most solid recommendations is to avoid alternatives which are less energy efficient [4]. At least for electricity as a starting point, electric vehicles are and will likely continue to be a significantly more energy-efficient option than hydrogen. Meanwhile, hydrogen proves to be a more efficient fuel than electricity in terms of range, so hydrogen vehicles are poised to look more like conventional cars in the future. In conclusion, the limitations of electric vehicles are more a result of misconceptions about what an automobile is than physical limitations. It is likely that changes in the way people perceive the automobile will increase the potential for electric vehicles. If there is no change, there is a great risk of failure, either in the purchase of electric vehicles and not using them, or even worse (from an environmental standpoint), using them in addition to the use of automobiles. It is probably possible to construct a framework with fewer automobiles and more use of electric vehicles if shared or rented cars are given a larger role at the expense of privately owned vehicles. The development of electric vehicles has been accelerated in many countries to reduce dependence on oil and pollution [5]. An electric vehicle, especially a

84

S. Dhanush Hariharan et al.

battery electric vehicle, is considered an environmentally friendly solution to the energy crisis. An overview of the emerging technologies powering electric vehicles is provided in this paper. A summary of battery, charging, motors, and charging infrastructure of EVs is presented. A second contribution of this paper is to discuss the technical challenges and emerging technologies that EVs will face in the coming years to improve efficiency, reliability, and safety. The purpose of this paper is to present a conceptual framework that explains how electric vehicles can successfully be integrated into electric power systems. This framework analyzes two domains [6]: grid technology and electricity markets. Both domains describe the individuals involved in the processes, as well as their activities. To provide insight into the potential impacts and benefits of electric vehicle grid integration under the framework referred to, several simulations, including steady-state and dynamical simulations, are presented. EVs that are powered only by batteries or fuel cells have no internal combustion engines, while hybrids have both. A large deployment of this concept will produce considerable impacts on the design and operating of electric power systems due to the batteries’ high energy storage capacity and the charging of electric loads it will require. However, the use of non-polluting energy sources will be enabled and benefited by it also. Batteries (BMS) are one of the most important components of hybrid vehicles. In this way, batteries are guaranteed to work safely and reliably. BMS deals with the monitoring and evaluation functionality to maintain battery and the user’s safety and reliability. The behavior of an electrochemical product varies according to the operating and environmental conditions [7]. Implementing these functions is difficult due to the uncertain performance of batteries. The paper discusses current BMS concerns. It is crucial for a BMS to evaluate the state of a battery, including its charge level, its health, and its life expectancy. This paper discusses the future challenges and possible solutions of BMSs by reviewing the latest methodologies for evaluating battery states. It is urgent to develop a comprehensive view of the EV and HEV markets in light of their rapid growth. In this paper, the current technologies of HEVs, including drivetrain configurations, electric motors, and energy storage, are described. Electric traction has the potential to reduce vehicle pollution, improve vehicle performance, and improve energy efficiency. There is a new technology called hybrid electric vehicle traction that has great potential [8]. Performance, fuel efficiency, and low emissions, as well as a long operating range, are its benefits. In addition, the technologies of each component hardware are fully available and technically advanced. Many notable automobile manufacturers are working on hybrid electric vehicles, and some have already begun marketing them. A growing number of scientists are turning to environmentally friendly energy sources due to the hazards of conventional fuel vehicles [9]. Even though there are various renewable energy sources, hydrogen is the ideal fuel for vehicles. It delivers high energy levels. When designing fuel cell vehicles, it is important to consider onboard hydrogen storage. An assessment of hydrogen fuel cell engines’ feasibility as a major fuel for transportation systems is presented in this study. An electric generator can be made using fuel cells by reacting chemical gases with oxidants. Through the interaction between anodes and electrolytes, the fuel

An Artificial Intelligence-Based Technique to Optimize Hybrid Vehicle …

85

cell produces electricity by splitting the cation and anion in the reactant. During the chemical reaction, fuel cells produce water as a byproduct, which is non-hazardous for the environment. In its fuel cell, hydrogen is used as a source of direct current (DC), which makes it one of the most efficient energy carriers. In order to produce a sustainable mode of transport, hydrogen fuel cells and batteries should be coupled with a control system and strategies.

3 Contribution of Work When a pressure difference (caused by a leak) inside the tank is detected by the sensors or when the data readings from the sensors are of parameters, the AI system engages the Anti-Accident protocol. The walls of the hydrogen storage tank have a thin inner lining which itself is a tank containing nitrogen gas, as when the data from the sensors are off parameter, the amount of hydrogen remaining in the tank is recorded, and the temperature of the tank is raised to 45 °C and the tank walls are compressed internally until 100 atm pressure is reached. When the temperature and pressure of the tank reach the required amounts, nitrogen gas from the tank walls is flushed inside the tank in the ratio of 1:3. This process converts the constituents of the tank into ammonia, which is less reactive and prevents a major accident (Fig. 1).

Fig. 1 Pictorial representation of the processes carried out

86

S. Dhanush Hariharan et al.

4 Proposed System 4.1 Refilling Raw Materials and Electrolysis In order to produce carbon-free hydrogen from renewable resources, electrolysis is a promising method. With the help of electricity, electrolysis breaks water into hydrogen and oxygen as shown in Fig. 2. It is carried out in an electrolyzer. There are a variety of hydrogen electrolyzers, from small appliances that produce hydrogen locally to large-scale units that produce hydrogen using renewable or non-greenhouse gas emitting electricity. Electrolytes dissolve in polar solvents, such as water, and form electrically conducting solutions. Solvents dissolve electrolytes into cations and anions, which disperse evenly. When dissolved in water or heated, electrolytes form charged ions. A common chemical compound is potassium hydroxide. Chemically, potassium hydroxide is known as KOH. Aqueous solutions of KOH or molten KOH react with water to form K+ and OH- ions. Electricity is easily conducted in its aqueous solution since it contains free ions. In water or molten KOH, the ions separate and conduct electricity.

4.2 Electrolysis of Water An alkaline electrolyte solution, typically potassium hydroxide (KOH), is used for alkaline water electrolysis. Sodium hydroxide is a caustic and strong base. Alkaline water conducts electricity, allowing electrolysis to take place. Diaphragms separate electrodes, separating product gases and transporting hydroxide ions between them. Alkaline water electrolysis is one of the most technologically advanced and cheapest methods in electrolysis technology. In terms of efficiency, alkaline water electrolyzers offer a wide range of benefits. Alkaline water electrolysis technologies have become increasingly popular due to the growing interest in alkaline water electrolysis. The commercially available hydrogen generation systems are capable of producing up to tens of megawatts of power, with 100-MW systems in development.

Fig. 2 Simple electrolysis of water

An Artificial Intelligence-Based Technique to Optimize Hybrid Vehicle …

87

Water splitting is primarily achieved through alkaline water electrolysis. Hydrogen isotope-based military applications have contributed to the development of this technology. In Norway, a first plant for electrolysis of D2O (D2 production) was built. These industrial electrolyzers typically produce 5–500 m3 H2 /h of water. Hydrolysis of water with potassium hydroxide is much more efficient than other methods since it produces pure hydrogen at a faster rate and consumes less energy. By electrolyzing 40 g of potassium hydroxide with 500 ml water for 24 min, 100 ml of hydrogen gas is produced per minute.

4.3 Role of Artificial Intelligence Artificial intelligence is the key component in the working of this vehicle. Artificial intelligence oversees all the system processes and engages safety protocols when necessary. The switching of the vehicle’s driving mode between fuel-based and battery-based shown in Fig. 1 is being carried out by the AI in it. The data from all the sensors in the vehicle are sent to AI and are checked and rectified there. Some nonidentifiable problems are informed to the user at the earliest. The AI is mainly used to carry out the Anti-Accident protocol which is a safety measure to prevent accidents caused due to hydrogen leakage. The range of the vehicle, remaining battery capacity, amount of fuel produced, and the parameters of the storage tank are collected by AI and displayed to the user.

4.4 Hydrogen Storage Hydrogen energy holds great promise for the future. Water electrolysis is a sustainable and renewable chemical technology that has been attracting a great deal of attention among its various methods of production. The hydrogen fuel cell can be used to store intermittent energy. Food, metallurgy, and power plants are among the industries that use water electrolysis. The components of water, oxygen, and hydrogen have multiple functions. An electrolyzed hydrogen fuel, for example, can be renewable, clean, and efficient. Gases such as hydrogen and oxygen are primarily obtained from water electrolysis. Water is decomposed into hydrogen and oxygen by passing it through an electric current. Water atoms have weak hydrogen bonds, which can be dissociated into free atoms by heat, electricity, or chemical reactions. Hydrogen fuel can replace gasoline, diesel, kerosene, and CNG. Hydrogen is the lightest element in the periodic table. There are two protons in each hydrogen atom. Molecules contain gases, even if they do not exist in their natural state, such as water, sugar, proteins, and hydrocarbons. The gas hydrogen is much lighter than air, colorless, odorless, and extremely flammable and reacts violently when combined with other chemicals. A number of governments and companies are being urged by the Intergovernmental Panel on Climate Change to implement decarbonization measures. Human

88

S. Dhanush Hariharan et al.

activities are accelerating global warming, and the effects will be irreversible and devastating, according to a sixth assessment report. The use of renewable energy will reduce greenhouse gas emissions. Even so, solar, wind, and water do not provide enough energy in every region. Furthermore, several industries require energy supplies with high energy density that can be stored at the point of use in large quantities. Diversified energy carriers and storage technologies are necessary to bridge the gap between renewable energy sources and applications. Hydrogen can be stored for long periods of time in its gaseous state without losing energy. There is also the possibility of utilizing much of the existing natural gas infrastructure. Hydrogen has a lower volumetric energy density than natural gas or oil. In mobile applications, tank size and weight are significant concerns, while stationary applications can afford large storage tanks at low pressure. The metallic hydrides are capable of storing hydrogen. Because these low-pressure systems operate at atmospheric pressure, hydrogen molecules remain stable and non-hazardous due to chemical bonding. Renewable energy sources need to be transported and stored in a variety of ways to bridge the gap between renewable energy sources and applications. Hydrogen is a promising solution in this context. The engine runs primarily on gaseous hydrogen. Even though hydrogen is a highly efficient fuel, it is difficult to handle and contain. As a result, hydrogen fuel in the forever bike is not refilled like other fuel vehicles, but produced within the vehicle by electrolysis, reducing the possibility of an accident when refilling or storing. Furthermore, hydrogen gas here is produced in two cycles with a time gap between them, which reduces pressure inside the fuel tank. In a highly inert, temporary storage vessel, hydrogen is produced. There are numerous sensors in the storage tank, such as a pressure sensor, temperature sensor, gas detector sensor, and fire sensor. As the walls of the tank are filled with nitrogen gas, the possibility of a leak in the hydrogen storage vessel is almost zero. In case of an accident, the AI system that oversees the vehicle’s activities immediately engages the Anti-Accident protocol.

4.5 Power Management System All electric vehicles require energy storage systems, usually batteries. In general, a battery is designed to supply electrical energy to the electric starting motor, which, in turn, starts the chemical combustion engine that propels the vehicle. The battery here also serves the purpose of running the motor when the fuel in the vehicle runs out. Therefore, here the battery unit is as important as the fuel. The battery here is divided into two units. Each unit has its own specific task. The battery unit is divided in the ratio 1:9 in terms of capacity. The smaller unit is used to power the process of electrolysis and also to power the spark plug present in the engine. The larger unit is used to power the vehicle to run using motor when the vehicle runs out of fuel. The batteries used here are rechargeable and portable.

An Artificial Intelligence-Based Technique to Optimize Hybrid Vehicle …

89

Fig. 3 Conceptual representation of the engine segment.

4.6 Modified Hybrid Vehicle Driving Apparatus An engine is a type of device that converts energy into mechanical force and motion. The engine for this type of vehicle is modified and fitted with a motor which is connected with the engine via belt and a series of gears as shown in Fig. 3. When the vehicle runs using fuel, the engine drives the motor which acts as a generator and recharges the battery. Also when the fuel runs out, the motor detaches from the engine and runs the wheel on its own until the runs out. The following algorithm describes the overall proposed system: Proposed Algorithm. Step 1: START. Step 2: First fuel production cycle initializes. Step 3: Hydrogen fuel produced and stored in the fuel tank. Step 4: Vehicle operates using fuel and battery gets charged. Step 5: When the fuel run out vehicle is switched to the battery mode. Step 6: Vehicle runs using electricity till the battery runs out. Step 7: Second fuel cycle initializes. Step 8: Repeat Step 3 to Step 6. Step 9: STOP. Proposed PSEUDOCODE V: Denote the vehicle mode. H: Denote the vehicle running using hydrogen. B: Denote the vehicle running using battery. (here increment denotes recharging and refueling).

90

S. Dhanush Hariharan et al.

BEGIN FOR H IN RANGE (0,2): FOR B IN RANGE (0,2) V=H B++ V=B H++ END FOR END FOR END

5 Discussion The working of this project consists of four processes, viz. (a) (b) (c) (d)

Electrolysis and production of hydrogen. Vehicle runs using hydrogen fuel. Vehicle runs using battery. Processes b and c repeat one more time.

The hydrogen production through electrolysis is divided into two cycles, viz first 12 min and second 12 min. During the first 12 min of electrolysis, about 1200 ml of hydrogen gas is produced. The produced gas is stored in the temporary storage vessel which is then used to run the engine and drive the motor attached to it for the next 55 km. The motor which is driven by the engine acts as a dynamo and hence recharges the battery. The battery unit is divided into two compartments; one is used to carry out electrolysis, while the other is used to run the vehicle in electric mode. When the first half cycle of fuel is completed (i.e.) when the fuel runs out for the first time in the cycle, the vehicle is switched to electric mode. Now the engine deactivates and motor activates. When the vehicle runs using battery, the fuel required for second half cycle is produced and stored simultaneously. Now when the battery runs out for the first time in the cycle, the vehicle is switched back to fuel mode and uses fuel to run the vehicle for the next 55 km. As the vehicle runs using fuel, the engine drives the motor which in turn recharges the battery. Now when the fuel runs out for the second time in the cycle, the vehicle is again switched to electric mode and the motor shaft detaches from the engine and drives the wheels of the vehicle. As when the whole cycle is completed, the vehicle would have traveled 175–180 km. Now the electrolyte required for carrying out electrolysis is refilled and the vehicle is ready to go (Table 1 and Fig. 4).

An Artificial Intelligence-Based Technique to Optimize Hybrid Vehicle … Table 1 Distance–time data table

Distance traveled (km)

91

Time taken (min)

50

49.98

100

99.96

140

139.92

180

179.88

Fig. 4 Distance–time graph

6 Conclusions Transportation is one of the major activities carried out by human beings, and vehicles are the most used invention of human history. The use of vehicles has significantly affected the environment over time. Due to the burning of fossil fuels for the engine to run, hazardous gases are emitted as exhaust from the vehicle causing major catastrophes to the environment. These vehicles have been affecting the earth since 1886. The best way to prevent this is to switch to an alternating fuel source or an electric vehicle. The successful implementation of this project will help in the making of a clean environment and also result in the conservation of fossil fuels. Also, the hydrogen fuel storage mechanism used here will help in engineering a better yet safer fuel storage tank that can prevent accidents and can be implemented in future projects. Thus, this type of vehicle will contribute to the conservation of energy and create a sustainable and cleaner environment for future generations.

92

S. Dhanush Hariharan et al.

References 1. Sciarretta A, Back M, Guzzella L (2004) Optimal control of parallel hybrid electric vehicles. IEEE Trans Control Syst Technol 12(3):352:362. (May 2004) 2. Ursua A, Gandıa LM, Sanchis P (2012) Hydrogen production from water electrolysis: current status and future trends. Proc IEEE 100(2):410–426. (Feb 2012) 3. Niaz S, Manzoor T, Pandith AH (2014) Hydrogen storage: materials, methods and perspectives, vol 4. Elsevier, pp 457–469. (17 Nov 2014) 4. Jorgensen K (2007) Technologies for electric, hybrid and hydrogen vehicles: electricity from renewable energy sources in transport. Elsevier, pp 72–79. (1 Oct 2007) 5. Sun X, Li Z, Wang X, Li C (2019) Technology development of electric vehicles: a review, vol 13. Energies, pp 2–29. (12 Nov 2019) 6. Lopes JAP, Soares FJ, Rocha Almeida PM (2011) Integration of electric vehicles in the electric power system. Proc IEEE 99(1). (Jan 2011) 7. Xing Y, Ma EWM, Tsui KL, Pecht M (2011) Battery management systems in electric and hybrid vehicles, vol 4. Energies, pp 1840–1857 8. Ehsani M, Gao Y, Miller JM (2007) Hybrid electric vehicles: architecture and motor drives. Proc IEEE 95(4):719–728. (Apr 2007) 9. Manoharan Y, Hosseini SE, Butler B, Alzhahrani H, Senior BTF, Ashuri T, Krohn J (2019) Hydrogen fuel cell vehicles; current status and future prospect. Appl Sci 1(9):1–17. (4 June 2019)

Stock Market Prediction Using Machine Learning Techniques: A Comparative Study P. Chinthan, Rashmi Mishra, B. Prakash, and B. Saleena

Abstract Stock market prediction is one of the most important and time-consuming processes in the financial sector. This is mainly due to the fact that many factors could influence particular stock prices dynamically. In recent days, algorithm-based predictions have become widely popular among researchers because it provides better results in terms of improved accuracy and reduced error rates. That in turn helps the investors to make precise and informed decisions. In this study, a comparative analysis of various machine learning models was performed and the results were evaluated based on performance metrics such as mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and accuracy. The objective of this comparative study is to help investors take better decisions based on the consistent results provided by the algorithmic models. Keywords Stock market prediction · Machine learning · KNN · SVM · Regression · Auto-ARIMA · LSTM · National stock exchange

1 Introduction Most industries have benefited from the advancement of computers and natural language processing. The Financial Markets, particularly the stock markets, is one such industry. The issue that is currently trending and receiving the most interest is stock market trading [1]. According to widespread observations, a majority of people P. Chinthan School of Mechanical Engineering, Vellore Institute of Technology, Chennai, India R. Mishra School of Electronics Engineering, Vellore Institute of Technology, Chennai, India B. Prakash (B) · B. Saleena School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India e-mail: [email protected] B. Saleena e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_7

93

94

P. Chinthan et al.

in today’s society, including the youth, have been analysing the market (stocks) and making prudent investments in it. As a result, several techniques for forecasting stock values by analysing their open and close price trend over the last few years have been demonstrated to be highly effective for stock market trends. This allows investors to maximise their gains and minimise their losses. Typically, they pay less for the stocks and derivatives up front and then sell them for more afterwards. Before making an investment in a company, a venture capitalist frequently utilises two methods of analysis. The first is technical analysis, which projects future stock prices using historical stock prices, such as opening and closing prices as well as highs and lows. The second type of research is qualitative analysis [2], where a venture capitalist assesses the intrinsic value of stocks and makes predictions about their future price based on outside factors such as market circumstances, national political and economic situations, media and news. Over the years, there has been a widespread misconception in the marketplace that purchasing and selling stock in a firm constitutes gambling. Such delusions are the result of people being misinformed about the market and how it operates [3]. The increasing use of machine learning algorithms in the stock market for stock price prediction and other purposes has aided clever individuals in developing a clear understanding of the market and the commercial activities that take place there [4]. In order to determine which machine learning algorithms are more accurate in predicting stock outcomes, we conducted a comparative study of multiple stock prediction algorithms on historical stock price data and compared and assessed their outputs with various performance criteria. Real-world stock traders employ sophisticated predictive machine learning algorithms in conjunction with other techniques to anticipate market trends, which in turn helps them decide when and where to invest money to realise the greatest returns [5]. Since there are always other influences on market patterns that cannot be anticipated in advance, these algorithms are not entirely correct [6]. There is therefore much scope for development in this area. The National Stock Exchange (NSE) and the Bombay Stock Exchange (BSE) are the two major stock exchanges in India [7]. They offer a marketplace where Indian company equities may be bought and sold. The NSE’s NIFTY index measures the health of the equities that are traded on that exchange, which indirectly measures the health of the nation’s economy. We utilised the NIFTY-50 stock data to train our algorithms for the research. The NIFTY-50 is a stock selection of 50 NSE firms chosen for inclusion in a weighted formula that determines the value of the NIFTY on the basis of their reputation, market size and importance.

2 Related Work Hiransha and her team [8] used a deep learning model to predict the stocks featured in National Stock Exchange (NSE) and New York Stock Exchange (NYSE). This study was based on the prediction of 5 stock prices listed in the two indices and concluded that neural network models outperformed the linear models, in specific ARIMA. The

Stock Market Prediction Using Machine Learning Techniques …

95

study conducted by Mehar et al., [9] selected five large capitalization stocks listed on the NYSE for the closing price prediction of stocks. They have employed machine learning techniques such as artificial neural networks (ANN) and random forest for their analysis. The performance of the models was assessed based on evaluation metrics such as root mean square error (RMSE) and mean absolute percentage error (MAPE). The results concluded that ANN performed better as compared with random forest in terms of stock value prediction. The most widely popular deep learning technique long short-term memory (LSTM) was used by Shen and Shafiq for predicting stock prices in the Chinese stock market dataset [10]. As part of pre-processing, this study built a feature engineering procedure by applying recursive feature elimination (RFE) and principal component analysis (PCA). The combination of feature engineering with LSTM performed better than other existing models in terms of accuracy. Payal Soni et al., [11] performed a systematic review of stock price movement by analysing various machine learning approaches. Different categories of methods like traditional machine learning (ML) techniques, deep learning models and neural networks, time series analysis, and graph-based approaches were selected for the comparative analysis. Another comparative analysis of stock market predictions was performed by Rouf et al., based on the studies conducted in the last 10 years [12]. The factors considered for analysis were the types of data implemented as input, various pre-processing approaches, and the machine learning and deep learning models employed for predictions. The results of this study show that the support vector machine (SVM) is a well-established machine learning model for stock market analysis. Also, other techniques like artificial neural network (ANN) and deep neural network (DNN) present more accurate and faster predictions of stock values. A systematic literature review (SLR) has been conducted by Strader et al., [13] in the field of stock market analysis on 4 categories, namely artificial neural networks, support vector machines, genetic algorithms and other hybrid techniques. This study concludes that artificial neural networks are suitable for stock value index predictions, SVM is useful for forecasting an overall trend of indices, genetic algorithms are used for suggesting the suitable stocks in a portfolio, and hybrid approaches are helpful in overcoming the limitations faced by a single method [15]. A framework was proposed by Khan et al., [14] to determine the future trend of a stock market based on certain external contributing factors like news and social media posts. The analysis also finds that the effect of pre-processing steps such as spam tweet reduction and feature selection process positively impacts the accuracy of stock predictions. Another survey was carried out by Subba Rao et al., to analyse the stock movements based on performing a comparative study of various approaches with their advantages and limitations. The comparative study comprised 8 supervised machine learning models for predicting the Nifty 50 index stocks [16]. The analysis was performed based on the past 25 years of historical data. The analysis shows that linear regression performed better than neural network due to the fact that the regression model deals better with linear dependency data followed by SVM and Gradient Descent. The stock price movements in NYSE were analysed by applying a linear regression model with three-month moving averages and exponential smoothing

96

P. Chinthan et al.

predictions [17]. The result proved that exponential smoothing predictions performed better than linear regression and three-month moving averages. A survey has been conducted on the application of various ML algorithms, helpful in predicting the future prices of stocks in the field of financial sectors [18]. Parmar et al., [5] has implemented two models which are Regression and LSTM for predicting stock price movements. It has been concluded that LSTM performed better compared to regression in terms of prediction accuracy. Similarly, Ashwini and Sakshi [19] analysed 4 machine learning models, namely random forest, SVM, KNN and Logistic Regression for the prediction of the stock market. For this study, the evaluation metric considered were accuracy, precision, sensitivity (recall) and Fscore (F1-score) and the outcome shows that random forest outperformed the other algorithms. Lokesh and his team [2] combined sentiment analysis and machine learning algorithms to determine the trend of a particular stock. In addition, based on the derived results, the risk exposure towards the particular company has been determined and notified to the user. Pooja et al., [20] integrated sentiment analysis with deep learning models to enhance the prediction accuracy of the stock market. The results indicate that the combination of deep learning models along with the sentiment analysis has a positive impact on stock price movement predictions.

3 Methodology 3.1 Dataset The dataset utilised in this analysis was obtained from Yahoo Finances. We conducted our study on NIFTY stocks using historical stock market data for the last five years, beginning on January 1, 2016, and ending on December 31, 2021. As per Table 1 below, the features in the dataset were ‘High,’ ‘Low,’ ‘Open,’ ‘Close,’ ‘Volume,’ and ‘Adj Close.’ The observations in the dataset were indexed by date. The term ‘High’ denotes the stock’s highest value on that particular day. The word ‘Low’ represents the stock’s daily low. ‘Open’ denotes the stock price at the start of the day, while ‘Close’ denotes the stock price at its conclusion. The term ‘Volume’ refers to the quantity of shares purchased or sold over a specific time frame. We employed the ‘Date’ and ‘Close’ features in the machine learning models for the study. Our prediction algorithms were run on 1476 data points.

3.2 Machine Learning Models There are two types of machine learning methods: supervised machine learning methods and unsupervised machine learning methods. In this paper, a comparative

Stock Market Prediction Using Machine Learning Techniques … Table 1 Description of dataset

97

Feature

Description

Symbol

Unique identifiers assigned to each listed company

Date

Over a duration of the past 5 years

High

Highest value of the listed company on the date

Low

Lowest value of the listed company on the date

Open

Opening value of the listed company on the date

Close

Closing value of the listed company on the date

Volume

Total traded volume of the listed company on the date

Previous-closing

Closing value of the listed company on the day before

study was performed based on the results from the following models which were used for stock market prediction. The models that were carefully considered for this study are based on the in-depth literature survey conducted by the authors. • • • • •

Long short-term memory (LSTM) [21], K-nearest neighbour (KNN) [22], Linear regression using moving average [23], Auto-ARIMA [24] and Support vector machine (SVM) [25].

3.2.1

Lstm

A kind of Recurrent Neural Network (RNN) known as LSTM is an unsupervised learning model that may be used to categorise, evaluate and forecast time series data as well as learn order dependence in chronology prediction problems. Long short-term memory has a significant advantage over other typical neural networks in that it can handle the complete collection of inputs. LSTMs are created in this way to avoid the issue of long-term reliance. Compared to other neural networks, it has a longer memory and processing capacity for the full dataset. The cells that make up long short-term memory are represented by the forget gate in Fig. 1, the input gate in Fig. 2 and the output gate in Fig. 3. The three gates synchronise the information flow into and out of the cell. The cell makes it easier to keep track of connections between the elements in the input sequence. The forget gate aids the model in separating out unnecessary information and deleting it. The input gate assists in incorporating data into the state of the cell. The information that will be presented is selected and stored by the output gate.

98

P. Chinthan et al.

Fig. 1 Forget gate

Fig. 2 Input gate

Fig. 3 Output gate

In this model, historical stock data was obtained using python modules, and the LSTM model was modified to use this data to predict future NIFTY values (stock) from the Yahoo Finance website. This data collection had 1476 observations, each of which had six parameters. After pre-processing, only the ‘Dates’ and ‘Close’ columns were used as they are the most important to the dataset and our study. We scaled the data and divided it into 80:20 train and test sets. The LSTM model was developed for the period from January 1, 2016, to December 31, 2021, using the training dataset, which comprises 80% of the total dataset. The remaining 20% was utilised for testing. The predicted prices against the initial values were shown to see how accurate the model was.

Stock Market Prediction Using Machine Learning Techniques …

3.2.2

99

Knn

K-nearest neighbour is most commonly used for classification and regression purposes. K-nearest neighbour keeps track of all the instances and categorises new ones using similarity measurements. KNN is a lazy supervised learning model since it does not have a specific training phase and instead trains and classifies using all of the data. KNN is a non-parametric learning method since it makes no assumptions about the nature of the underlying data. In this, the ‘K’ is a very small integer and indicates the number of nearest neighbours to be included in the classification process. The suitable ‘K’ value is chosen using parameter tuning, which helps in better accuracy. The distance between the points on the plane is calculated using the Euclidean distance as denoted in Eq. 1. dist (d) =

√

(x − a)2 + (y − b)2

(1)

x and y are the coordinates on the plane a and b are the edges of the plane Larger values of ‘d’ have a lesser effect on the noise in the classification process. If all features are on the same scale, it would be easier to check the data point in multidimensional space. As an outcome, data modelling or standardisation will be more effective. If there are too many features, KNN may not work well. As a result, processes like feature selection and principal components can be used to reduce dimensionality. When one of the M feature data of a given sample in the training dataset is missing, we would not be able to find or calculate the distance from that spot. Therefore, both deletion and imputation are required. In this model, historical stock data was obtained using python modules, and the KNN model was modified to use this data to predict future NIFTY values (stock) from the Yahoo Finance website. This data collection had 1476 observations, each of which had six parameters. After pre-processing, only the ‘Dates’ and ‘Close’ columns were used because they are the most important to the dataset and our study. We rescaled the data and divided it into 80:20 training and testing datasets. The KNN model was developed for the period from January 1, 2016, to December 31, 2021, using the training dataset, which represents 80% of the total dataset, and the remaining 20% for testing.

3.2.3

Linear Regression

Linear regression is a method for determining the connection between a scalar response and one or more explanatory factors. Linear regression attempts to model the connection between two variables by fitting a linear equation to observable data, namely dependent and independent variables. Linear regression could be determined as represented in Eq. 2. y =m∗x +c

(2)

100

P. Chinthan et al.

where ‘y’—dependent variable, ‘m’—regression coefficient (slope), ‘x’—independent variable, and ‘c’—constant. Implementation of linear regression is done by the following steps: 1. 2. 3. 4. 5. 6.

Acquiring data from the finance website. Exploration of data. Segmentation of data. Training and splitting the data. Predicting the required parameters and generating the model. Evaluating the accuracy of the model.

The idea of moving average is also used in this approach. To forecast future values, a 30-day moving average is employed. In order to anticipate future NIFTY values (stock) from the Yahoo Finance website, we utilised python modules to collect historical stock data and customised the linear regression model to this data. This data collection had 1476 observations, each of which had six parameters. After preprocessing, only the ‘Dates’ and ‘Close’ columns were used because they are the most important to the dataset and our study. For the time period of January 1, 2016, to December 31, 2021, we scaled the data and divided it into training and testing datasets with 80:20 ratio.

3.2.4

Auto-ARIMA

ARIMA model is a stereotype of the Autoregressive Moving Average (ARMA) model. These algorithms are used to examine and forecast future data points in a time series (forecasting). When data shows indications of non-stationarity in the main function, ARIMA models are used, an initial differencing step is done, several times to eradicate the mean function’s non-stationarity. In order to estimate future values, ARIMA models consider historical values. In ARIMA, there are three key parameters: [Autoregressive] p: Number of Autoregressive terms. [Integrated] d: Degree of differencing. [Moving Average] q: Forecast mistakes from the past are being used to predict future values. where variables p, d and q are non-negative integers. The values of ‘p’ and ‘q’ are calculated using the Partial Autocorrelation Function (PACF) plot and Autocorrelation Function (ACF) plot, respectively. Whereas Kwiatkowski–Phillips–Schmidt– Shin (KPSS) Test and Augmented Dickey–Fuller (ADF) Test are used for checking whether the given data series is stationary or not and attain the value for ‘d’.

Stock Market Prediction Using Machine Learning Techniques …

101

When two of the three components are zero, the model can be referred to solely by the non-zero parameter, eliminating the requirement for the acronyms ‘AR’, ‘I’ or ‘MA’. Implementation of Auto-ARIMA is done by the following steps: 1. 2. 3. 4. 5.

Acquiring data from the yfinance website. Pre-processing the data. Applying the Auto-ARIMA model. Predicting values on the validation set. Evaluating the accuracy of the model.

In order to predict future NIFTY values (stock) from the Yahoo Finance website, we modified the Auto-ARIMA model using previous stock data that we obtained using python modules. This data collection had 1476 observations, each of which had six parameters. After pre-processing, only the ‘Dates’ and ‘Close’ columns were used because they are the most important to the dataset and our study. 80 per cent of the whole dataset was utilised for training the Auto-ARIMA model, and the remaining 20 per cent was used for testing, covering the time period from January 1, 2016, to December 31, 2021.

3.2.5

Svm

SVMs are supervised learning models which are used for classification and regression analysis. Each data point is treated as a single point in n-dimensional space (n being the number of attributes), with the value of each feature becoming the SVM algorithm’s value for a specific position. The hyper-plane that unambiguously divides the class labels is then found to complete classification. These features are then plotted on certain coordinates. The hyper-plane is a boundary established across the data set by the SVM algorithm that splits the data into two groups. Support vectors are just the coordinates of each individual observation. The SVM classifier behaves as a dividing line between the two groups (hyper-plane and line). Implementation of SVM is done by the following steps: 1. 2. 3. 4. 5.

Acquiring data from the yfinance website Segmentation of data Applying the SVM model Predicting values on the validation set Plotting the data and evaluating the accuracy of the model.

In order to anticipate future NIFTY values (stock) from the Yahoo Finance website, we utilised python modules to obtain historical stock data and modified the support vector machine model to this data. This data collection had 1476 observations, each of which had six parameters. After pre-processing, only the ‘Dates’ and ‘Close’ columns were used because they are the most important to the dataset and our study. We rescaled the data and divided it into 80:20 training and testing sets.

102

P. Chinthan et al.

The training dataset, which makes up 80% of the total dataset, was used to train the support vector machine model over the period from January 1, 2016, to December 31, 2021, while the remaining 20% was utilised for testing.

4 Results and Discussions Stock market returns are hard to anticipate since stock prices change often and are based on a variety of factors that form intricate patterns. The historical information offered on the company’s website only includes a few factors, like high, low, open and closing values, the number of shares and equities traded, and so on, few of which were used in this work. We recommended using data gathered from the Yahoo Finance website for the comparative study. Various machine learning-based algorithms in order to predict the daily movement of Market stocks have been analysed. The numerous performance measures given below provide proof of the encouraging findings of our investigation.

4.1 Performance Metrics The performance evaluation of the five machine learning algorithms was carried out based on the metrics MSE, RMSE, MAE, MAPE and accuracy.

4.1.1

Mse

MSE is a common loss function that can be used to evaluate the models. It is based on the mean of the difference in squares between the original values and projected values as represented using Eq. 3 above. Its values ideally lie between 0 and ∞, and a smaller MSE value indicates a better model. This metric is highly sensitive to outliers. MSE =

4.1.2

N 1 Σ (yi − yˆi )2 N i=1

(3)

Rmse

RMSE is the radical of MSE. When dealing with large error values, this metric is superior at expressing performance. RMSE could be evaluated based on Eq. 4. It is an excellent indicator of precision when used to compare different model predictions,

Stock Market Prediction Using Machine Learning Techniques …

103

but not between variables as it is scale-dependent. RMSE values in the range of 0.2 and 0.5 indicate that the model is capable of accurately estimating the data. ⎡ | N |1 Σ (yi − yˆi )2 R M S E = √ N i=1 N

4.1.3

(4)

Mae

MAE is the sum of the total difference between the estimated and actual values as represented by Eq. 5 as below. It is a measure of the differences in errors between paired observations of the same data. A good MAE is dependent on the dataset, but the smaller the value the better the model. It is not sensitive to outliers. It is recommended to establish a baseline MAE using a naïve prediction model for the specific dataset, which can then be utilised to assess the effectiveness of the various models. N 1 Σ M AE = (yi − yˆi )2 N i=1

4.1.4

(5)

Mape

MAPE aka mean absolute percentage division (MAPD) indicates the mean difference between the expected value and actual value, as determined by Eq. 6. It is the mean of absolute percentage errors of predicted values and helps to define the accuracy of the model on the specific dataset. Though the ideal range of MAPE is dataset specific, a MAPE of less than 5% indicates that the prediction is reasonably accurate. M AP E =

4.1.5

N 1 Σ |yi − yˆi | X 100% N i=1 |yi |

(6)

Accuracy

Accuracy is a statistic for deciding which model is the most effective at detecting trends and patterns in a dataset based on training data. It is the percentage of right predictions the model produced on the test data. As demonstrated in Eq. 7, accuracy is calculated as the ratio of the number of correct predictions to the total number of predictions. The greater the value of accuracy, better the related model’s performance.

104

P. Chinthan et al.

Accuray =

TP +TN T P + FP + T N + FN

(7)

where, TP: True positive (A positive sample that the model accurately classified as positive). FP: False positive (A negative sample that the model inaccurately classified as positive). TN: True negative (A negative sample that the model accurately classified as negative). FN: False negative (A positive sample that the model inaccurately classified as negative). The performance metric comparison of the five different machine learning algorithms employed in this paper is illustrated in Table 2. The various error metrics like MSE, RMSE, MAE, MAPE and accuracy of the models were evaluated, and it was concluded that for this concerned dataset, the LSTM model outperformed the others in the prediction of stocks with an overall accuracy of 97%. In terms of accuracy, KNN performed better with 89% next to LSTM, followed by SVM (86%), linear regression (79%) and Auto-ARIMA (71%). In comparison with the other error metrics, it has the least value out of the five models which is consistent with the inference drawn from the model performances. The linear regression model shows the worst performance out of the five models. It has the highest error rate when compared to the other models. This indicates that the model is unable to find a linear relation between the data and hence is overfitting on the train set, resulting in higher test error rates. The SVM classifier improves the error in predictions considerably when compared to linear regression and gives an accuracy of 86%. KNN performs better than SVM, and Auto-ARIMA further improves the performance of KNN. Auto-ARIMA yielded a precision (accuracy) value of 71% on the testing data which is the least compared to the other four models. The graphs represented in Figs. 4 and 5 are the performance of the models listed with respect to five different performance metrics. The MSE and RMSE of linear regression are extremely high compared to that of the other models whereas the MAE of linear regression and SVM are on the higher side. LSTM gives the best accuracy Table 2 Result of performance of machine learning models S. NO

Algorithm

MSE

RMSE

MAE

MAPE

Accuracy (%)

1

Linear Regression

1.76

1.32

6.57

5.05

79

2

LSTM

0.12

0.34

2.49

1.63

97

3

SVM

0.56

0.74

5.71

4.11

86

4

KNN

0.48

0.69

3.94

2.89

89

5

Auto-ARIMA

0.33

0.57

3.72

2.23

71

Stock Market Prediction Using Machine Learning Techniques …

105

and Auto-ARIMA gives the worst accuracy among the five models compared in this study.

Fig. 4 Performance of ML algorithms

Fig. 5 Accuracy of ML algorithms

106

P. Chinthan et al.

5 Conclusion In today’s scenario, anticipating the stock’s future prices is considered to be a more challenging task. This study has analysed the effectiveness of five machine learning models on the NIFTY-50 index. In this research study, the past five years’ historical values of the stock prices were considered for the prediction. The outcomes of the machine learning models have been assessed on the basis of MSE, RMSE, MAE, MAPE and accuracy. According to the results achieved, we conclude that LSTM outperformed all other chosen algorithms in terms of both accuracy and error factors. For future work, (i) the wide range of company stocks could be focussed on for prediction and (ii) hybrid algorithms may be considered for stock price prediction to achieve better results.

6 Limitations This work is limited to utilising historic stock price data for modelling and evaluation of the models. For this reason, these models may not be able to make accurate realtime stock predictions as they lack the robustness of incorporating various other factors that may affect the stock prices. Some of these factors include news releases, change of management, supply and demand, public sentiment towards the company, trading volume, etc. Another limitation of this work is that the analysis has been performed on NIFTY50 stocks which are representative of the Indian Stock market. The same methodology might not apply to different stock markets; thus, more evaluation will be necessary to determine whether the implementation suffers overfitting problems. Our work makes use of machine learning models which could be limiting the scope of the model when compared with the advancements in deep learning algorithms. Newer state-of-the-art models like genetic algorithm and multi-layered feed-forward networks may give better results in this case.

7 Future Enhancements Machine learning algorithms incorporating financial news articles as well as financial metrics such as trading volume, closing price, profit and loss accounts and so on may be developed in future as a means of achieving better results. This would strengthen the models. Additionally, we may expand the scope of our dataset and include new variables, such as financial ratios. This methodology can be made more broadly applicable by using data from other stock markets like the US stock market, etc., and assessing for model overfitting and prediction accuracy.

Stock Market Prediction Using Machine Learning Techniques …

107

Complex neural networks may be coupled with other techniques like fuzzy logic and genetic algorithm in the future. Genetic algorithm may be used to discover the appropriate training settings and network design. Some of the uncertainty brought about by neural network predictions can be accounted for with fuzzy logic. They can be used with neural networks to enhance stock market forecasting.

References 1. Daul S, Jaisson T, Nagy A (2022) Performance attribution of machine learning methods for stock returns prediction. J Financ Data Sci 2. Lokesh S, Mitta S, Sethia S, Kalli SR, Sudhir M (2018) Risk analysis and prediction of the stock market using machine learning and NLP. Int J Appl Eng Res 13(22):16036–16041 3. Illa PK, Parvathala B, Sharma AK (2022)Stock price prediction methodology using random forest algorithm and support vector machine. Mater Today: Proc 56(Part 4):1776–1782 4. Vanukuru K (2018) Stock market prediction using machine learning. https://doi.org/10.13140/ RG.2.2.12300.77448 5. Parmar I et al (2018) Stock market prediction using machine learning. In: First international conference on secure cyber computing and communication (ICSCCC) 2018:574–576 6. Subasi A, Amir F, Bagedo K, Shams A, Sarirete A, Learning SMPUM (2021) Proc Comput Sci 194:173–179 7. Bhardwaj A, Narayan Y, Vanraj P, Dutta M (2015) Sentiment analysis for Indian stock market prediction using sensex and nifty. Proc Comput Sci 70:85–91 8. Hiransha M, Gopalakrishnan EA, Menon VK, Soman KP (2018) NSE stock market prediction using deep- learning models. Proc Comput Sci 132:1351–1362. ISSN 1877–0509 9. Vijh M, Chandola D, Tikkiwal VA, Kumar A (2020) Stock closing price prediction using machine learning techniques. Proc Comput Sci 167:599–606. ISSN 1877–0509 10. Shen J, Shafiq MO (2020) Short-term stock market price trend prediction using a comprehensive deep learning system. J Big Data 7:66 11. Soni P, Tewari Y, Krishnan D (2022) Machine learning approaches in stock price prediction: a systematic review. J Phys: Conf Ser 2161:012065. https://doi.org/10.1088/1742-6596/2161/1/ 012065 12. Rouf N, Malik MB, Arif T, Sharma S, Singh S, Aich S, Kim H-C (2021) Stock market prediction using machine learning techniques: a decade survey on methodologies, recent developments, and future directions. Electronics 10:2717 13. Strader TJ, Rozycki JJ, Root TH, Huang YH (John) (2020) Machine learning stock market prediction studies: review and research directions. J Int Technol Inf Manag 28(4), Article 3 14. Khan W, Ghazanfar MA, Azam MA, Karami A, Alyoubi KH, Alfakeeh AS (2020) Stock market prediction using machine learning classifiers and social media, news. J AmbientIntell Humanized Comput https://doi.org/10.1007/s12652-020-01839-w 15. Rao PS, Srinivas K, Mohan AK (2020) A survey on stock market prediction using machine learning techniques. https://doi.org/10.1007/978-981-15-1420-3_101 16. Singh G (2022) Machine learning models in stock market prediction. Int J Innov Technol Explor Eng 11:18–28. https://doi.org/10.35940/ijitee.C9733.0111322 17. Umer M, Awais M, Muzammul M (2019) Stock market prediction using machine learning (ML) algorithms. ADCAIJ: Adv Distrib Comput Artif Intell J 8(4):97–116. https://doi.org/10. 14201/ADCAIJ20198497116 18. Obthong M, Tantisantiwong N, Jeamwatthanachai W, Wills G (2020) A surveyon machine learning for stock price prediction: algorithms and techniques. https://doi.org/10.5220/000934 0700630071

108

P. Chinthan et al.

19. Pathak A (2020) Study of machine learning algorithms for stock market prediction. Int J Eng Res V9. https://doi.org/10.17577/IJERTV9IS060064 20. Mehta P, Pandya S, Kotecha K (2021) Harvesting social media sentiment analysis to enhance stock market prediction using deep learning. Peer J Comput Sci 7:e476 21. Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9(8):1735–1780 22. Altman N (1992) An introduction to kernel and nearest-neighbor nonparametric regression. Am Stat 46(3):175–185 23. Stanton JM (2001) Galton, Pearson, and the Peas: a brief history of linear regression for statistics instructors. J Stat Educ 9:3 24. Masani R (2012) Norbert Wiener 1894–1964, vol. 5. Birkhäuser, Basel, Switzerland 25. Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20:273–297

Comparative Study on Different Intrusion Detection Datasets Using Machine Learning and Deep Learning Algorithms G. Aarthi, S. Sharon Priya, and W. Aisha Banu

Abstract The tremendous growth in the Internet of Things (IoT) creates great potential which provides us with incredible productivity and simplified our daily lives. But, due to resource constraints and computation, IoT networks are vulnerable to a variety of malicious activities. Thus, protecting the network from hostile attacks should be the top priority. This can be done by planning and implementing effective security measures, one of them is an intrusion detection system. It detects harmful activities on the network and monitors network traffic based on that detection. The aim of the intrusion detection system (IDS) is to afford various approaches for detecting the rapidly growing network attacks, as well as to stop the harmful activities that occur in the IoT devices. Various artificial intelligence methods were evaluated and concluded on various datasets, including BoT-IoT, IoT-23, UNSW-NB15, CSE-CIC-IDS2018, and MQTT-IOT-IDS2020, in search of a suitable algorithm that can easily learn the pattern of network attack activities. The feature extraction and pre-processing data were then fed into IDS as data to train the model for future anomaly detection, prediction, and analysis. This study focuses primarily on the various types of cyber-attacks and machine learning algorithms used to identify cyber-attacks. Keywords Intrusion detection system · Cyber-attacks · Internet of things · Cyber security · Machine learning

G. Aarthi (B) · S. Sharon Priya · W. Aisha Banu Computer Science and Engineering„ B. S. Abdur Rahman Crescent Institute of Science and Technology, Chennai, India e-mail: [email protected]; [email protected] S. Sharon Priya e-mail: [email protected] W. Aisha Banu e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_8

109

110

G. Aarthi et al.

1 Introduction The Internet of Things (IoT) is a set of interrelated nodes[1], which gathers and share the data with other nodes through an Internet connection without human involvement [2]. IoT is a globally adopted technology, which grows rapidly in the automated network system. Because of the rapid proliferation of IoT nodes, even simple devices may now communicate and exchange massive amounts of data with one another and with other nodes. Since huge data are being transmitted in the network, the security of data should be taken into consideration. To keep IoT nodes safe and secure, modern security solutions should be used on the network. In Botnet, the attackers infect nodes with malicious code or overload node information and reduce performance, which is among the most serious security challenges in the IoT. One of the most significant threats to IoT nodes is botnet-based attacks [3]. The IoT node should need specialized standards and communication protocols to face these types of challenges. And they are employed to provide safe and dependable data transmission between IoT nodes. Because of its low bandwidth, low memory requirements, and low packet loss, MQTT is the most widely used protocol. MQTT is an OASIS standard messaging Protocol, designed mainly for IoT nodes. It is a published/subscribe messaging protocol, which has four important components, such as Clients, Broker, Topic, and Message. MQTT-Client (IoT nodes) exchange data via Broker (Central node). If the node capability allows, the broker allows IoT nodes to publish and subscribe to topics at the same time. MQTT topics are a type of structured, hierarchical addressing, and identical to the file system’s forward slash (/) delimiter. It contains communications, such as data collected by various IoT nodes as well as the nodes that serve as the source and destination of transmission messages from other networks. In most applications, transmission control protocol and user datagram protocol are the widely used protocols on the transport layer. However, the IoT applications need interoperable standard ways and different message distribution functions depending upon the requirement. Most of the IoT devices use the MQTT communication protocol, which runs on top of TCP and is used for data transmission and reception between the IoT nodes. In accordance to find the network traffic anomalies, faster and more effective security methods should be implemented on the IoT nodes. Anti-virus software, intrusion detection systems (IDSs), and firewalls are examples of cyber security methods. These methods safeguard data against both active and passive attacks. A passive attacker just observes the data and copies them, without disturbing the operation of the system, which mainly consists of eavesdropping, traffic analysis, etc., whereas an active intruder will try to modify the content of the data, some of them are DDoS, DoS, Man in the Middle attack (MitM), etc., Among all of the methods, IDS is one of the important types of detection system that is used to track the anomalies in the states of the hardware and the software running in the network. The experts have begun to focus mainly on intrusion detection systems (IDSs) using artificial intelligence (AI) methodologies. A branch of AI called “Machine

Comparative Study on Different Intrusion Detection Datasets Using …

111

Learning” (ML) is capable of autonomously gleaning useful information from huge databases. These techniques are used in a variety of cyber security activities, including intrusion detection, network traffic analysis, and botnet identification. ML methods are used to identify and investigate both benign and abnormal conditions. By training using efficient data, machine learning is used not only to detect but also to forecast threats.

2 Related Works Discuss the related research pertinent to IDSs in IoT that use machine learning methods to evaluate IoT datasets in this section [4]. The most widely used dataset in IoT is BoT-IoT, UNSW-NB15 dataset, etc., which are used by several researchers. For analyzing and testing with various machine learning methods. The author gives a brief review of the many forms of intrusion detection attacks, and also the various types of machine learning methods, and demonstrates how the dataset may be used to analyze, evaluate, and test the Model’s performance. Table 1 shows the different intrusion detection public dataset and methodologies used by different research experts. Leevy and other co-authors [5] use the publicly available BoT-IoT dataset, to train and test the model by cross-validation technique, to identify the normal and the information theft traffic. The authors used different ensembles techniques and non-ensembles techniques, to evaluate and analyze the outcome of the data, and to determine the best model. Shinly Swarna Sugi and the co-authors [6] have examined and evaluated the capability of the intrusion detection system to identify assaults using the model utilizing the BoT-IoT dataset. To assess its effectiveness of it, the authors used long short-term memory (LSTM) and K-Nearest Neighbor (KNN). To find anomalies in the data in the IoT-23 dataset, R. Thamaraiselvi and other authors [7], used supervised learning methods such as random forest (RF), Naive Bayes (NB), support vector machine (SVM), and decision tree(DT). N.Abdalgawad and co-authors [8] have analyzed the IoT-23 dataset, to detect malicious activities like DoS, DDoS, and several attacks such as Mirai, tori, and okiruk, using the deep structured learning algorithms. Using the UNSW-NB15 dataset, it finds detrimental data attacks. Mustafa Alshamkhany and the co-authors [9], utilized a variety of machine learning methods to assess and analyze the performance of the best fit model. Muhammad Zeeshan and the co-authors [10] have compared the BoT-IoT and UNSW-NB15 dataset based on their packet flow, TCP, and other features to detect the non-anomalous and different attacks, by using deep learning technique. Ammar D Jaism [11] and the other authors have investigated the CSE-CICIDS2018 dataset and detected the normal behavior and attacks using a variety of artificial intelligence techniques.

112

G. Aarthi et al.

Table 1 Various intrusion detection datasets and different methodologies used Author

Title

Leevy et al. [5]

Detecting 2021 information theft attacks in the Bot-IoT dataset

BoT-IoT

Ensembles and non-ensembles techniques

Swarna Sugi and Ratna [6]

Investigation of machine learning techniques in intrusion detection system for IoT network

2020

BoT-IoT

Long short-term memory and K-nearest neighbor

Thamaraiselvi and Mary [7]

Attack and anomaly detection in IoT networks using machine learning

2020

IoT-23

Random forest, Naive Bayes, support vector machine, and decision tree

Abdalgawad and Generative deep Member [8] learning to detect cyberattacks for the IoT-23 dataset

2022

IoT-23

Adversarial autoencoders and bidirectional generative adversarial network

Alshamkhany et al. [9]

Botnet attack detection using machine learning

2020

UNSW-NB15

Naive Bayes, K-nearestneighbor, support vector machine, and decision tree

Zeeshan et al. [10]

Protocol-based deep intrusion detection for DoS and DDoS attacks using UNSW-NB15

2022

UNSW-NB15 and BoT-IoT

Deep learning technique

Farhan and Jaism [11]

Performance analysis 2022 of intrusion detection for deep learning model based on CSE-CIC-IDS2018 dataset

CSE-CIC-IDS2018

Long short-term memory

Dwibedi et al. [12]

A comparative study on contemporary intrusion detection datasets for machine learning research

UNSW-NB15, BoT-IoT, and CSE-CIC-IDS2018

Random forest, support vector machine, Keras deep learning model and XGBoost

Year Dataset details of the paper

2020

Methodology used

(continued)

Comparative Study on Different Intrusion Detection Datasets Using …

113

Table 1 (continued) Author

Title

Year Dataset details of the paper

Methodology used

Ullah and Mahmoud [15]

Design and development of a deep learning based model for anomaly detection in IoT networks

2021

Convolutional neural networks (CNN) in 1D, 2D, and 3D for IoT networks

BoT-IoT, IoT network intrusion, MQTT-IoT-IDS2020 and IoT-23

Smirti Dwibedi and the other co-authors [12] have analyzed the UNSW-NB15 [13], Bot-IoT [6, 7], and CSE-CIC-IDS2018 datasets, analyzing their performance using supervised learning techniques and also with deep learning models [14]. Imtiaz Ullah and other co-authors [15] have looked at various intrusion detection datasets such as BoT-IoT, MQTT-IoT-IDS2020, IoT-23, and IoT Network Intrusion and have constructed a special intrusion detection model for IoT networks utilizing CNN in 1D, 2D, and 3D [15].

3 Intrusion Detection Dataset Details An IDS is a piece of software or hardware that protects a system by sounding an alarm in the case of a security breach and taking action to thwart the attacker [16]. IDS can be arranged in a wide range of ways, and there are many of them [11]. Some of the most popular IDS datasets are discussed below.

3.1 IoT-23 Dataset This dataset contains 23 scenarios in which various IoT devices and network traffic are recorded. There are also about 20 potential network-based cyber security concerns. The cyberattacks in this dataset include those caused by Trojan, Kenjiro, Okiru, Hakai, Mirai, Gagfyt, Torii, Hakai, Hide and Seek, IRCBot, Muhstik, and others [17]. The details of real-time IoT devices and cyber-attack scenarios on the network are recorded in various pcap files. These files contain information on the various application layer protocols, such as HTTP, DNS, DHCP, TELNET, SSL, SSH, IRC, and so on. The dataset is approximately 20 gigabytes in size [18].

114

G. Aarthi et al.

3.2 BoT-IoT Dataset The protocol name (HTTP, TCP, and UDP), attacks name (DoS, DDoS, OS and Service Scan, Data Exfiltration, and Keylogging), flow duration, source, and destination port addresses, and a label that indicates whether the packet or data is affected or not. These data attributes are in pcap file format, which is approximately 69.3 gigabytes in size, and extracted flow traffic is approximately 16.7 gigabytes in size [19].

3.3 UNSW-NB15 Dataset This dataset contains various attributes such as the protocol name, the name of the attacks, the label, which is represented in 0’s and 1’s to indicate whether the data is benign or malicious, the source and destination port addresses, as well as the packet inflow and outflow details, flow duration, and so on. Backdoor, Analysis, Reconnaissance, Shell-code, Worms, DoS, Exploits, Fuzzers, and Generic are the different types of cyber-attacks captured in this dataset. The dataset is approximately 100 GB in pcap file format [20].

3.4 CSE-CIC-IDS2018 Dataset This dataset includes the following information: the destination port address, flow time, protocol, total number of packets in forward and backward directions, the maximum and minimum packet sizes, average, mean, and standard deviation of packets, and more [21]. Web attacks, DoS, Brute force attacks, Infiltration, Botnet, DDoS, and Port Scan attacks are only a few of the numerous assault types described in this report [22]. The dataset, which is in CSV format, is approximately 16.68 gigabytes in size.

3.5 MQTT-IOT-IDS2020 Dataset The collection contains data on TCP flags, TCP time stream, TCP segment length, reserved, acknowledge flags, clean session flags, password flags, QoS level, username flags, connect flags, and more. This dataset includes DoS, MQTT publish flood, corrupted data, SlowITe, and brute force assaults. The dataset is about 9.5 gigabytes in size and comes in pcap and CSV file formats. The protocols and different attacks used in the dataset [23] are described in Table 2.

Comparative Study on Different Intrusion Detection Datasets Using …

115

Table 2 Protocols and different attacks in intrusion detection dataset Intrusion detection dataset

Protocol used

Attacks captured

IoT-23

HTTP, DHCP, Telnet, SSL, and others

Mirai, Torii, Trojan, Kenjiro, Okiru, Gagfyt, IRCBot, Hajime, Muhstik, Hide and Seek, Hakai, and other Cyber Attacks

BoT-IoT

TCP, UDP, ARP, ICMP, and others

DoS, DDoS, OS, and Service Scan, Keylogging and Data Exfiltration

UNSW-NB15

TCP, UDP, ICMP, and others

Backdoor, Analysis, Reconnaissance, Shell-Code, Worms, DoS, Exploits, Fuzzers, and Generic

CSE-CIC-IDS2018

TCP, UDP, and others

Bruteforce, DoS, Web Attack, Infiltration, Botner Attack, DDoS, and PortScan

MQTT-IOT-IDS2020

TCP and MQTT

DoS, MQTT Publish Flood, Malformed Data, SlowITe, and Bruteforce

4 Cyber Attack Details 4.1 Denial of Service (DoS) It prevents the normal use of the target network or the nodes. It will overload the network, by sending more packets to reduce the performance of the system or crash the main node to prevent users from accessing it [24, 25].

4.2 Distributed Denial of Service (DDoS) A DDoS assault occurs when numerous devices are used in a coordinated DoS at-tack against a single targeted node.

4.3 Botnet (BoT) It makes use of the network of Bots, in which nodes infected with malware/viruses that can be activated to perform different attacks like spam e-mail or DoS or DDoS, on the other nodes. The NF-BoT-IoT dataset has DoS and DDoS attacks along with other attacks like theft and reconnaissance.

116

G. Aarthi et al.

4.4 Theft It is a type of group assault in which important data is stolen via weakening the primary node’s security.

4.5 Reconnaissance It is a method for gathering information about a network host [13]. A probe is another name for it. Probing attacks include scanning remote nodes for information about the fatality. These are the various cyber security attacks, which can be analyzed by using different machine learning algorithms [26].

5 Comparative Analysis The area under the curve (AUC) was used in the research paper to evaluate the BoTIoT dataset using ensembles and non-ensemble approaches (AUPRC). The dataset is trained and tested using ensemble methods [27]. For the evaluation of the dataset, nonensemble approaches are utilized [5]. Whereas, according to the expert’s findings, ensemble methods produce good outcomes. In the paper [6], the BoT-IoT dataset was analyzed using the artificial intelligence method. The metrics are detection time, geometric mean, kappa statistic, and sensitivity. According to the author, the LSTM outperforms the KNN. The paper [7] has chosen the IoT-23 dataset, to find the anomaly detection. By using the random forest algorithm, the authors achieved an accuracy of approximately 99.5% when compared to all other algorithms. The paper [8] talked more about the IoT-23 dataset and analyzed it using deep structured learning algorithms like Bidirectional Generative Adversarial Network (BiGAN) and Adversarial Auto Encoders (AAE). The metric used to analyze the attack detection is F1-Score. The author has analyzed and experimented using both the deep learning algorithms and got an efficiency of about 0.99. The paper [9] used UNSW–NB-15 dataset and various machine learning algorithms to detect the attacks in the network. Various metrics like Accuracy, Confusion Matrix, Precision, Recall, and F1-Score are used for evaluation purposes [28]. The author obtained the best results by using the decision tree among these methods. The paper [10] uses protocol-based deep intrusion detection (PB-DID) architecture to analyze and find anomalies in BoT-IoT and UNSW –NB-15 datasets. The authors used deep learning algorithms to achieve a 96.3% accuracy. In the paper [12], the CSE-CIC-IDS2018 dataset was used to evaluate normal behavior and attacks using the deep learning method. The author’s accuracy was approximately 99%.

Comparative Study on Different Intrusion Detection Datasets Using …

117

The paper [13] mainly focuses on the UNSW-NB15 [13], BoT-IoT, and CSECIC-IDS2018 datasets to analyze attacks using machine learning algorithms and Keras deep learning models. The author achieved better results in XGBoost when compared with the entire dataset. In paper [15], MQTT-IoT-IDS2020 and IoT-23 datasets are used to detect attacks using the deep learning method, called CNN [29]. Various metrics are used for evaluation purposes [1]. When compared to existing deep learning methods, the author achieved a maximum accuracy of around 99% [30]. Many experts have already been done in the field of IoT Security; this study headed towards into few research gaps that can be filled or used to create a more efficient future scope of work [31]. Figure 1 depicts the various intrusion detection datasets and research papers. The intrusion detection system uses various artificial intelligence methods to detect any type of abnormality in incoming traffic [27]. Based on the findings of various researchers, the best machine learning algorithm [32] suitable for the various intrusion datasets is depicted in Table 3. Deep learning algorithms such as GAN and CNN perform well on all public intrusion detection datasets, according to an analysis of various research papers. However,

Fig. 1 Taxonomy of artificial intelligence in intrusion detection dataset

Table 3 Comparative analysis of different intrusion detection datasets and their various artificial intelligence methods Intrusion detection dataset

Machine learning algorithms

Deep learning algorithms

DT

RF

LSTM

GAN

CNN

IoT-23

x

99.5

x

99

99.82

BoT-IoT

x

x

97.28

x

99.81

UNSW-NB15

99.89

x

x

x

x

CSE-CIC-IDS2018

x

x

99

x

x

MQTT-IOT-IDS2020

x

x

x

x

99.99

118

G. Aarthi et al.

to predict the algorithm’s efficiency, the dataset must be thoroughly analyzed and feature selection must be performed. In addition, many new types of attacks must be introduced to determine whether the algorithm is performing well.

6 Conclusion and Future Enhancement To achieve our research goal, we analyze five different public intrusion detection datasets, as well as the different techniques and machine learning algorithms used by different experts, and we conclude that deep learning models like CNN and GAN perform well for attack detection when compared to other ensemble methods. In addition, all of these studies make use of publicly available datasets. Since the public dataset has issues such as an imbalance in nature and overfitting, the packets are chosen in the same numbers, from each of the categories. In the future, by incorporating additional well-known and benchmark real-time datasets, and also by adding more attack types and new threats in data to identify the major attacks in the network of IoT devices. In the future, we intend to do further research on the security of the network and also to find the anomalies by implementing various artificial intelligence techniques, and by using data preprocessing techniques, to contrast the outcomes with those attained using current models.

References 1. Ullah I, Mahmoud QH (2021) A framework for anomaly detection in IoT networks using conditional generative adversarial networks. IEEE Access 9:165907–165931. https://doi.org/ 10.1109/ACCESS.2021.3132127 2. Akhtar MS, Feng T (2021) Deep learning-based framework for the detection of cyberattack using feature engineering. Secur Commun Netw 2021:1–12. https://doi.org/10.1155/2021/612 9210 3. Popoola SI, Adebisi B, Ande R et al (2021) SMOTE-DRNN: a deep learning algorithm for botnet detection in the internet-of-things networks. Sensors 21:2985. https://doi.org/10.3390/ s21092985 4. Sengodan T, Murugappan M, Misra S (2021) Advances in electrical and computer technologies. Springer Singapore, Singapore 5. Leevy JL, Hancock J, Khoshgoftaar TM, Peterson J (2021) Detecting information theft attacks in the bot-IoT dataset. In: 2021 20th IEEE international conference on machine learning and applications (ICMLA). IEEE, pp 807–812 6. Swarna Sugi SS, Ratna SR (2020) Investigation of machine learning techniques in intrusion detection system for IoT network. In: 2020 3rd international conference on intelligent sustainable systems (ICISS). IEEE, pp 1164–1167 7. Thamaraiselvi R, Mary SAS (2020) Attack and anomaly detection in IoT networks using machine learning 9:95–103. https://doi.org/10.47760/ijcsmc.2020.v09i10.012 8. Abdalgawad N, Member S (2022) Generative deep learning to detect cyberattacksfortheIoT23dDataset.IEEEAccess10:6430–6441. https://doi.org/10.1109/ACCESS.2021.3140015

Comparative Study on Different Intrusion Detection Datasets Using …

119

9. Alshamkhany M, Alshamkhany W, Mansour M et al (2020) Botnet attack detection using machine learning. In: 2020 14th international conference on innovations in information technology (IIT). IEEE, pp 203–208 10. Zeeshan M, Riaz Q, Bilal MA et al (2022) Protocol-based deep intrusion detection for DoS and DDoS attacks using UNSW-NB15 and Bot-IoT data-sets. IEEE Access 10:2269–2283. https:// doi.org/10.1109/ACCESS.2021.3137201 11. Farhan BI, Jasim AD (2022) Performance analysis of intrusion detection for deep learning model based on Performance analysis of intrusion detection for deep learning model based on CSE - CIC - IDS2018 dataset. 1165–1172. https://doi.org/10.11591/ijeecs.v26.i2.pp11651172 12. Dwibedi S, Pujari M, Sun W (2020) A comparative study on contemporary intrusion detection datasets for machine learning research. In: 2020 IEEE international conference on intelligence and security informatics (ISI). IEEE, pp 1–6 13. Oussous A, Benjelloun F-Z, Ait Lahcen A, Belfkih S (2018) Big Data technologies: A survey. J King Saud Univ - Comput Inf Sci 30:431–448. https://doi.org/10.1016/j.jksuci.2017.06.001 14. Syamsuddin I, Barukab OM (2022) SUKRY: suricata IDS with enhanced kNN algorithm on raspberry Pi for classifying IoT botnet attacks. Electronics 11:737. https://doi.org/10.3390/ele ctronics11050737 15. Ullah I, Mahmoud QH (2021) Design and development of a deep learning-based model for anomaly detection in IoT networks. IEEE Access 9:103906–103926. https://doi.org/10.1109/ ACCESS.2021.3094024 16. Idrissi I, Azizi M, Moussaoui O (2022) A lightweight optimized deep learning-based hostintrusion detection system deployed on the edge for IoT. Int J Comput Digit Syst 11:209–216. https://doi.org/10.12785/ijcds/110117 17. Snehi M, Bhandari A (2022) IoT-based DDoS on cyber physical systems: research challenges, datasets and future prospects. In: 2022 IEEE international IOT, electronics and mechatronics conference (IEMTRONICS). IEEE, pp 1–8 18. Gorbett M, Shirazi H, Ray I (2022) WiP: the intrinsic dimensionality of IoT networks. In: Proceedings of the 27th ACM on symposium on access control models and technologies. ACM, New York, pp 245–250 19. Maurya S, Kumar S, Garg U, Kumar M (2022) An efficient framework for detection and classification of IoT botnet traffic. ECS Sens Plus 1:026401. https://doi.org/10.1149/27542726/ac7abc 20. Souhail et al M (2019) Network based intrusion detection using the UNSWNB15dataset.IntJComputDigitSyst8:477–487. https://doi.org/10.12785/ijcds/080505 21. Zolotukhin M, Hämäläinen T, Kotilainen P (2022) Intelligent solutions for attack mitigation in zero-trust environments, pp 403–417 22. Karatas G, Demir O, Koray Sahingoz O (2018) Deep learning in intrusion detection systems. In: 2018 international congress on big data, deep learning and fighting cyber terrorism (IBIGDELFT). IEEE, pp 113–116 23. Barolli L, Terzo O (2018) Complex, intelligent, and software intensive systems. Springer International Publishing, Cham 24. Thomas L, Bhat S (2021) Machine learning and deep learning techniques for IoT-based intrusion detection systems: a literature review machine learning and deep learning techniques for IoT-based intrusion detection systems: a literature review 6:296–314 25. Al RM, Javeed D, Khan MT et al (2022) Cyber threats detection in smart environments using SDN-enabled DNN-LSTM hybrid framework. IEEE Access 10:53015–53026. https://doi.org/ 10.1109/ACCESS.2022.3172304 26. Sommer R, Paxson V (2010) Outside the closed world: on using machine learning for network intrusion detection. In: 2010 IEEE symposium on security and privacy. IEEE, pp 305–316 27. Sagu A, Gill NS (2020) Machine learning techniques for securing IoT environment, pp 977– 982. https://doi.org/10.35940/ijitee.D1209.029420 28. Abdalgawad N, Sajun A, Kaddoura Y et al (2022) Generative deep learning to detect cyberattacks for the IoT-23 dataset. IEEE Access 10:6430–6441. https://doi.org/10.1109/ACCESS. 2021.3140015

120

G. Aarthi et al.

29. Douiba M, Benkirane S, Guezzaz A, Azrour M (2022) Anomaly detection model based on gradient boosting and decision tree for IoT environments security. J Reliab Intell Environ. https://doi.org/10.1007/s40860-022-00184-3 30. Mohamed TS, Aydin S (2021) IoT-based intrusion detection systems: a review. Smart Sci 1–18. https://doi.org/10.1080/23080477.2021.1972914 31. Sagu A, Gill NS (2020) Machine learning techniques for securing IoT environment. Int J Innov Technol Explor Eng 9:977–982. https://doi.org/10.35940/ijitee.D1209.029420 32. Yang Z, Liu X, Li T et al (2022) A systematic literature review of methods and datasets for anomaly-based network intrusion detection. Comput Secur 116:102675. https://doi.org/10. 1016/j.cose.2022.102675

Big Data Analytics

Mitigating Postnatal Depression: A Big Data Self-help Therapy Asma Usman, Francis Morrissey, Thaier Hamid, Celestine Iwendi, and F. Anchal Garg

Abstract Mother, who gives birth, usually faces a mood disorder called postpartum or postnatal depression. It appears immediately after the third week of the baby’s birth. However, during the first year of delivery, women can suffer anytime with this situation, and it could lead to a couple of years after birth. Few men as a father can face this condition. If it is not monitored immediately, it triggers severe and permanent disorders such as anger issues, isolation, stress, or anxiety. A significant increase has been observed in postpartum depression incidents with harmful consequences on children as well as parents regarding their physical and emotional well-being. This research paper analysed the literature to evaluate the psychotherapies that can be followed as self-help. We also evaluated automated psychotherapy systems and meta-analysed mobile applications that are available online to cope with postpartum depression. We discussed the acceptability of a therapeutic mobile application for reducing depression during parenting and postpartum period for the patients themselves. Finally, a combination of cognitive behavioural therapy and interpersonal psychotherapy; an algorithm, we proposed in this paper as a base to develop the mobile application that can help control and reduce depression during a postpartum situation. Keywords Therapeutic mobile application · Postnatal depression · Postpartum · Digital cognitive behavioural therapy · Computerised interpersonal psychotherapy · Self-help · Self-therapy A. Usman (B) · F. Morrissey · T. Hamid · C. Iwendi · F. Anchal Garg School of Creative Technologies, University of Bolton, Bolton, UK e-mail: [email protected] F. Morrissey e-mail: [email protected] T. Hamid e-mail: [email protected] C. Iwendi e-mail: [email protected] F. Anchal Garg e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_9

123

124

A. Usman et al.

1 Introduction A common topic discussed among mothers is postnatal depression. Significantly, during COVID-19 public has been dealing with mental health problems. Mums with newborn babies and mothers with two or more kids at home in isolation faced challenging situations during the pandemic. Physical effects of the virus have been notified, but negligence has been observed towards mental health issues caused by quarantining the society [1]. During the year 2017, 9.5% in high-income countries, about 20.8% in middleincome countries, and 25.8% in low-income countries cases have been observed depression during the postpartum period [2]. For a variety of reasons, many cases, have not been recorded. A lot of women might feel better after three to six months, while more than 30% of them might still experience depression a year after giving birth. This might result in severe mental, physical, or behavioural disorders [3]. Common symptoms have been observed such as insomnia, loss of appetite, moodiness, continuous worries, and anxiety [4]. As shown in Fig. 1, depression leads to other illnesses. Obesity is more likely to affect depressed people, although both conditions can be addressed using the same treatment [5]. Patients with diabetes who use insulin may experience depression. It has been noted in Europe, the United States, and Asia that chronic heart disease and depression share common triggers. [5]. Patients with chronic heart disease are advised to get screened for depression and receive therapy, according to the American Heart Association and the European Society of Cardiology. Depression is more difficult in women than in males, and it worsens metabolic syndrome. Cholesterol levels and blood pressure may rise because of a depressed condition. According to recent research, depression eventually results in a diagnosis of hypertension five years or later. Depression increases the risk of developing chronic heart disease by 30 to 90%. The idea that depression causes a variety of health problems, including loneliness, anxiety, diabetes, inflammatory bowel disease, alcoholism, schizophrenia, drug usage, and Alzheimer’s, has been investigated [6]. Mobile applications or computerised therapeutic devices are simple and affordable. Effectiveness, accessibility, technical treatment, patient education, management of records, learning setups, and patient preference are eight aspects that support computerised psychotherapy [7]. The goal of this project is to create an algorithm for a mobile application that can offer a therapeutic option for self-help. With the help of this application, anyone may treat and manage mild to moderate postpartum depression on their own. The following are the research’s main contributions: • To conduct a thematic analysis of the literature in search of therapeutic solutions for cure of postpartum depression. • To assess these therapies can achieve by self-help. • To analyse the postpartum depression-related mobile applications that are offered in the Apple and Android app stores. • To create an algorithm that can offer the best means of treating and managing depression as self-help.

Mitigating Postnatal Depression: A Big Data Self-help Therapy

125

Fig. 1 Age of respondents

The rest of this paper is organised as follows: Sect. 2 evaluates the literature to enquire about therapies to cure postpartum depression. Section 3 provides the methodology of the research. The results of the research are provided in Sect. 4. Section 5 concludes the work.

2 Literature Analysis In the clinical setting, numerous therapies are employed to treat depressive disorders. Postpartum or parental depression can be treated with talking, audio/visual, and expressive art therapy treatments in mild to moderate cases. The treatment of severe mental and depressive problems involves clinical seizure treatments. In the United Kingdom, NHS provides talking treatments through the Improving Access to Psychological Therapy (IAPT) program. It includes guided self-help, counselling, and talking therapies like cognitive behavioural therapy that can help with common mental health concerns including mood disorders, anxiety, stress, or anger issues [8].

2.1 Talking Therapies The patient is urged by cognitive behavioural therapy to resist and eliminate negative thought patterns and to come up with constructive ways to handle difficulties. It is based on the patient’s current issues and ignores their history [9]. Anger management concerns, psychosis, anxiety or disturbed mood, bipolar disorder, borderline personality disorder, panic disorders, obsessive–compulsive disorders, and sleep disturbances are all addressed by cognitive behavioural therapy. The patient is urged to

126

A. Usman et al.

use and adhere to use the pattern provided during therapy sessions to solve the issues [10]. By analysing past events that upset the patient excessively, interpersonal psychotherapy helps patients to understand why their illness developed. The patient’s mood and the upsetting circumstances that led to the condition are analysed by interpersonal psychotherapy [11].

2.2 Audio/Visual Therapies The treatments for depression that employ audio and cinematic techniques are known as audio/visual therapies. Light therapy uses a lightbox that directs strong light towards daily tasks. Light therapy is a method of treatment. When used to treat seasonal and non-seasonal depression illnesses, it has produced effective and favourable results without any negative side effects, affecting mood through the retina [12]. A digital avatar that engages in peer-to-peer communication and has voice conversations with the patient in a therapeutic manner. The delivery of avatar therapy cannot be done completely or authentically. Robotics and machine learning are centred on establishing a personal relationship between a patient and a robotic voice. Patients are encouraged to enhance their daily life with this therapy [13]. Gamification therapy is a type of motivational treatment that uses video games to improve the mood of depressed patients. The therapist works with the client to assist them to understand the value of video games and how to go about living their daily lives. Clinicians should advise patients on the kind of video games to play and how much time should be spent on them [14].

2.3 Expressive Art Therapies Dance, movement, music, art, theatre, and archetypes can control postpartum depression illnesses. [15]. At the beginning of dance/movement therapy, the patients displayed feelings of embarrassment, contempt, and awkwardness. Women were under pressure to engage in social interactions, and before this, they had very little energy. These sentiments were purposefully reduced using dance and motion therapy [16]. Music therapy is a melody-singing activity that calms and soothes the patient and creates a strong relationship between them and their social group. This considerably aided them in reducing depression and postpartum stigma [17]. The Royal College of Music and Imperial College London investigated the “Breathe Melodies for Mums”, a form of therapy intended to cheer those experiencing stress, anxiety, anger, and social isolation as well as to heal the hearts of mothers. Through this therapy, 73% of the patients recovered from postpartum depression [18]. “My Time, My Space” therapy uses creative expression to treat and manage postpartum depression in group settings. The postnatal depression condition allows up to ten women

Mitigating Postnatal Depression: A Big Data Self-help Therapy

127

to engage in an activity, such as painting, sewing, or patchwork. Instead of emphasising the creation of products, it focuses on participant interest. Sessions last for two to three hours. Each session involves the use of several materials, including fabric, jewellery, and paint. This method of therapy raises patients’ spirits and aids them in managing their depression [19]. Participants in drama therapy use acting to express their stories, come up with answers to them, and establish a plan of action to reach their objectives. Therapists examine their patients’ issues, comprehend their requirements, and work to meet those needs. Depending on the needs, either a group session or an individual session may be held. Games, storytelling, case study resolution, or quizzes can be set up in clinical sites, counselling places, educational institutions, or mental health clinics. [20].

2.4 Clinical Procedures Several therapeutic techniques that involve surgical operations are coming under the canopy of clinical procedures. These days, the medical procedure known as electroconvulsive therapy is frequently used to treat serious depressive or bipolar illnesses. It sends a series of electrical signals that stimulate the brain. During the session, the patient is treated by skilled medical professionals. Electroconvulsive therapy is explored when other therapies, such as anti-depressant drugs or psychotherapy, do not work and the severity of the disease is escalating or causing suicidal thoughts. [21, 22]. Rapid application of magnetic fields to specific brain regions is known as transcranial magnetic stimulation. There are no seizure procedures, such as electroconvulsive therapy, in it. While receiving treatment, the patient is awake. As the patient receives treatment, it may have mild side effects like headaches or muscle soreness. A generator that produces electric pulses at predetermined intervals of time is surgically inserted under the patient’s chest skin during vagus nerve stimulation. It sends sporadic electric stimulation to the vagus nerve in the neck. [22].

2.5 Computerised Psychotherapy Legal and qualified experts should research and collaborate on the creation of computerised health systems. [23]. In rural places, computerised cognitive behavioural psychotherapy can be a useful and effective practise tool for providing psychotherapy remotely. More research is required to determine the best ways to give computerized psychotherapy in various parts of a country due to the differences in nature and potential local challenges that exist in each place [24, 25]. A robot as a counsellor can offer the patient independent advice on a variety of subjects. The effectiveness of a robot as a counsellor is equivalent to that of a human counsellor. Self-disclosure, especially when it comes to emotional ideas and opinions, has been found to increase patient’s

128

A. Usman et al.

openness to their negative emotions [25]. Online cognitive behavioural therapy is efficient and cost-effective. Programs for educating general practitioners about emerging electronic health models that could result in virtual clinics and primary care interventions are needed. Currently, most online counselling programs are focused on finding treatments for depressive disorders. Self-help is a practical and effective technique to offer therapies for sadness and anxiety. Cognitive behavioural therapy and automated psychotherapy are popular and effective treatments for anxiety and depression. More data is required in several areas to improve deliverable standards and the efficacy of various online treatments. [26]. Many smartphone apps on the digital market make claims to treat depressive illnesses; however, many of them have not undergone clinical testing. They are not associated with licenced psychiatrists. Because of this, there has been no proof of their therapeutic value. Due to limitations such as money and time inconvenience, patients may reject therapy treatment. For mobile application effectiveness, additional clinical trials are required [27].

3 Methodology We employed a mixed approach in this research. Thematic analysis of the literature is used as a qualitative research method to determine the significance of the problem and identify knowledge gaps [28]. Ten mothers who had more than one child and suffered from parenting or postpartum depression participated in a discussion. The goal of the debate was to confirm that a mobile application can be a therapy treatment. To analyse the statistical significance of the problem statement, this study used a Likert-scale survey as a quantitative research method. Concept comprehension, statement production, response outcome formulation, measurement gauging, and response gathering were the five key factors that were seen during the survey questionnaire’s creation [29]. This study employed the PRISMA technique to find mobile applications for analysis that purport to treat postpartum depression [30].

4 Implementation and Discussion A survey was done with moms who had children within the previous 10 years in July and August of 2022. All participants (n) received a briefing on the questions and the survey’s objectives via voice message. As indicated in Fig. 1, the participants (n = 40) in this study were 40 mothers between the ages of 18 and 59. The Office for National Statistics data bulletin, which is depicted in Fig. 2, was taken into consideration while choosing the respondents. According to statistics, the average age of women who became mothers in 2020 was 30.7 years, unchanged from 2019 [31]. Most postpartum depression cases, especially in the initial postpartum period, were shown to be unreported. Mothers felt uncomfortable revealing their postpartum

Mitigating Postnatal Depression: A Big Data Self-help Therapy

129

Fig. 2 Average age of parents in England and Wales, 1938–2020 [33]

sorrow because they were fearful of being criticized. Throughout their sadness, they wished to manage alone. [8]. Statistics were calculated by rating (r) the responses given to each statement or question that was posted on the platform, ranking them from 1 to 5 in descending order. For illustrating the relevance of the issue, the sentence “Every woman has postnatal depression” was put out as the hypothesis, as shown in Fig. 3. The survey presents this theory since several incidents are unreported. 16 out of 40 (or 40% of the respondents) were deemed to be highly likely, while 15 (37.50%) were deemed to be likely. However, just 4 out of 40 respondents (10%) believed it was unlikely, while 5 respondents (12.5%) were unsure. As demonstrated in Table 1, no one said very unlikely. Equation 1’s explanation of Function (x) states that it computes the sum of the product of responses (i) and gauge (r). x=

n 

i ∗r

i

x=

n  i=0

i ∗ r = 80 + 60 + 15 + 8 + 0 = 163

(1)

130

A. Usman et al.

Fig. 3 Women suffering from postnatal depression

Table 1 Partisan’s suffering from postnatal depression Responses

Very likely

Likely

Neither likely nor unlikely

Unlikely

Very unlikely

Responses (i)

16

15

5

4

0

Gauge (r)

5

4

3

2

1

Percentage

40

37.50

12.50

10

0

Total no. of responses n = 40

Equation 2, which displays the average result of the response gauge, demonstrates the application of another formula (Avg) for calculating the suffering of women with postnatal depression. n  x Avg = n i

Avg =

(2)

n  163 x = n 40 i=0

Avg = 4.1 It is evaluated that every woman experiences postpartum depression, either more or less, because 82% of respondents respond in favour of the argument, according to the average score of 4.1 out of 5.0, which indicates agreement with the statement. The second query examines the proportion of participants who experience postpartum depression, as depicted in Fig. 4. Eight responders (20%) indicated very likely. 13 respondents (32.50%) indicated that they likely experienced postpartum depression. 5 (12.50%) respondents indicated that they were unsure of whether they

Mitigating Postnatal Depression: A Big Data Self-help Therapy

131

had postpartum depression or not by choosing neither likely nor unlikely. 10 respondents (25%) said they may not experience postpartum depression. Four responders (10%) were certain that they had not experienced postpartum depression. In Table 2, this statistic is displayed.

x=

n 

i ∗ r = 40 + 52 + 15 + 20 + 4 = 131

(5)

i=0

Avg =

n  131 x = n 40 i=0

(6)

Avg = 3.3 According to Fig. 4., 66% of the participants reported experiencing postpartum depression, and the average score was calculated to be 3.3 out of 5.0. However, 82% of respondents to this poll feel that postpartum depression affects every woman. In

Fig. 4 Combination of participants who participated in the survey who had suffered from postnatal depression

Table 2 Statistics of participants suffering from postnatal depression Responses

Very likely

Likely

Neither likely nor unlikely

Unlikely

Very unlikely

Responses (i)

8

13

5

10

4

Gauge (r)

5

4

3

2

1

Percentage

20

32.50

12.50

25

10

Total no. of responses n = 40

132

A. Usman et al.

addition, 44% of those who responded to the study said they had not experienced postpartum depression. Figure 5, which claims that “Postpartum depression can lead to any other depression-like parenting depression or can disturb marital life if not monitored or treated properly” is another statement that was made. [7]. Seventeen respondents (42.50%) said that they strongly agree with the statement. Six people chose to be indifferent, whereas 16 (or 40%) agreed. The possibility that postpartum depression can trigger additional depression was denied by just one participant. As seen in Table 3, no one chose the option of strongly disagreeing.

x=

n 

i ∗ r = 85 + 64 + 9 + 2 + 0 = 160

(7)

i=0

Avg =

n  160 x = n 40 i=0

(8)

Avg = 4

Fig. 5 Consequence of postpartum depression

Table 3 Statics of the survey that demonstrate “Postpartum Depression can lead to other depression such as parenting or can disturb marital life if not monitored initially” Responses

Strongly agree

Agree

Neither agree nor disagree

Disagree

Strongly disagree

Responses (i)

17

16

6

1

0

Gauge (r)

5

4

3

2

1

Percentage

42.50

40

15

2.50

0

Total no. of responses n = 40

Mitigating Postnatal Depression: A Big Data Self-help Therapy

133

More than 80% of participants believed that postpartum depression could have negative impacts on their marital and physical lives, as indicated by the average score of 4 out of 5. Participants were questioned about informing medical experts about their experiences of depression as shown in Fig. 6. Out of 40 respondents, 4 (10%) indicated they always get in touch with their doctors. This is also determined by NHS UK records, one in ten moms experience postpartum depression. Seven (17.5%) respondents indicated they occasionally contact health experts, whereas 3 (7.5%) respondents stated they typically contact their therapist. Fourteen (35%) people reported that they rarely seek medical advice. According to Table 4, 12 respondents (30%) claimed they had never gotten in touch with a therapist or other medical services.

x=

n 

i ∗ r = 20 + 12 + 21 + 28 + 12 = 93

(9)

i=0

Avg =

n  93 x = n 40 i=0

(10)

Fig. 6 Seeking help from health professionals

Table 4 Result of how many participants contacted a health professional for their help in their postpartum depressive period Responses

Always

Usually

Sometimes

Rarely

Never

Responses (i)

4

3

7

14

12

Gauge (r)

5

4

3

2

1

Percentage

10

7.50

17.50

35

30

Total no. of responses n = 40

134

A. Usman et al.

Avg = 2.3 The average score is 2.3 out of 5.5, showing that only 47% of the patients contacted health professionals. About 66% of the participants suffered from postnatal depression as explained above so 47% out of 66% means 19% of patients did not contact health professionals when they suffered from postnatal depression, and cases have not been on record. “Do you employ mobile application therapy to improve your mood?” was the fourth query posed like Mindfulness or Wysa. Because automated therapy is simple to use and readily available, a questioning was done to assess how the patients used therapeutic mobile applications to cope in their postpartum period, as displayed in Fig. 7. 3 (7.5%) of the 40 respondents indicated they always utilise therapeutic mobile applications to improve their mood during postpartum time by selecting the first option, “Always.” 6 (15%) respondents selected “usually,” indicating that they typically use mobile applications as therapy for their low mood disorders. 7 (17.5%) responded occasionally, while another 7 (17.5%) said rarely. According to Table 5, 17 (35%) participants indicated they have never used a mobile application by selecting “Never.”

x=

n 

i ∗ r = 15 + 24 + 21 + 14 + 17 = 91

(11)

i=0

Fig. 7 Using mobile application practice for coping with depression

Table 5 Consideration of a mobile application for coping with postpartum depression Responses

Always

Usually

Sometimes

Rarely

Never

Responses (i)

3

6

7

7

17

Gauge (r)

5

4

3

2

1

Percentage

7.50

15

17.50

17.50

42

Total no. of responses n = 40

Mitigating Postnatal Depression: A Big Data Self-help Therapy

Avg =

135

n  91 x = n 40 i=0

(12)

Avg = 2.3 The participants used therapeutic mobile applications throughout their depressive postpartum time, as evidenced by the average score of 2.3 out of 5.5, which indicates that 45% of them used therapeutic mobile applications. While 19% of patients did not seek any medical assistance and made it through their difficult times and 21% of patients were unaware of mobile application therapy. As seen in Fig. 8, a claim was made to gauge the effectiveness of mobile applications that are advertised as being beneficial for treating depression and are accessible online. In the study, 4 (10%) participants picked very likely, while 11 (27%) selected likely, indicating that they think that Internet and mobile therapeutic applications offer therapeutic solutions. Participants were asked if they believed that therapeutic mobile applications that are available online are not offering therapeutic answers. Eight (20%) replied very likely, and four (10%) stated unlikely. As seen in Table 6, 13 (32.5%) people chose “neither likely nor unlikely”.

Fig. 8 Therapeutic mobile applications provide therapeutic solutions

Table 6 Available mobile applications provide therapeutic solutions Responses

Very unlikely

Unlikely

Neither likely nor unlikely

Likely

Very likely

Responses (i)

8

4

13

11

4

Gauge (r)

1

2

3

4

5

Percentage

20

10

32.50

27

10

Total no. of responses n = 40

136

A. Usman et al.

x=

n 

i ∗ r = 8 + 8 + 39 + 44 + 20 = 119

(13)

i=0

Avg =

n  x 119 = n 40 i=0

(14)

Avg = 2.9 The average score is 2.9 out of 5.5 verifying that 60% of the participants believe that the therapeutic mobile applications available online are not providing a significant solution in their depressive postpartum period. About 40% of participants believe that these applications are helpful in their depressive postpartum period. The final hypothesis presented was that a mobile app created with a therapist’s assistance could offer a potential treatment for self-help to regulate depression, as shown in Fig. 9. Nineteen individuals agreed with the statement, and five persons strongly agreed. About 60% of the audience agrees with the claim that a mobile application created with a therapist’s assistance can provide a therapeutic remedy for self-help in depressive illnesses. About 10% of the participants disagreed with the statement, as evidenced by the two people who disagreed and the 0 who disagreed severely. As indicated in Table 7, 12 (30%) individuals chose the option “neither agree nor disapprove.”

x=

n 

i ∗ r = 25 + 76 + 36 + 8 + 0 = 145

i=0

Fig. 9 Mobile application can be used as self-help therapy

(15)

Mitigating Postnatal Depression: A Big Data Self-help Therapy

137

Table 7 Self-help by mobile application therapy in depressive disorder Responses

Strongly agree

Agree

Neither agree nor disagree

Disagree

Strongly disagree

Responses (i)

5

19

12

4

0

Gauge (r)

5

4

3

2

1

Percentage

12.50

47.50

30

10

0

Total no. of responses n = 40

Avg =

n  145 x = n 4 i=0

(16)

Avg = 3.6 The participants’ average score was 3.6 out of 5.5, which confirms that 73% of them will accept and be prepared to use a therapeutic mobile application solution to help themselves in their difficult times of postpartum or parenting depression if it is created with the assistance of a therapist or clinical professional. A smartphone application cannot in any way be a therapeutic solution, according to 27% of respondents. While 73% believe it can be a therapeutic solution in their postpartum and parenting time.

4.1 Postpartum Mobile Applications Meta-Analysis This study presents a meta-analysis of postpartum-related mobile applications that can found on the Apple App Store and Android Play Store. The Apple App Store and Android Play Store yielded 286 mobile applications. The researchers narrowed the search by identifying criteria like mobile applications that are free on App stores, have at least 4.5 user review ratings, and are editor recommended. By using a keyword searching strategy, 109 mobile applications were found on the Apple App Store and 177 on the Android Play Store. By screening, 9 mobile applications were selected and nominated for meta-analysis as explained in Fig. 10. The definitions of postpartum depression symptoms, postpartum depression symptom solutions, user access to record one’s symptoms, and user access to preserve one’s symptom solutions are the main parameters for comparison. Balance The first app chosen is Balance: meditation and sleep, which has received 4.8 user ratings and more than 500 k downloads on the Android Play Store and 4.9 user ratings and more than three million downloads on the Apple App Store. It does not outline any indications of postpartum depression or suggest methods for dealing with

138

A. Usman et al.

Fig. 10 PRISMA flow diagram for identifying the mobile applications

it. Additionally, it does not permit users to add new signs of depression or their own treatment for already-defined symptoms. It may be useful for sleep disturbances. Minddoc On the Android Play Store, it has 4.3 user ratings and more than 1 million downloads. It has more than 3 million downloads and 4.6 user ratings on the Apple App Store. It does not specify any signs or treatment methods for postpartum depression. The user is not having the option to specify his or her own new depression symptoms or provide treatments for existing symptoms. It enables the user to keep track of negative thoughts that are making them anxious or stressed out. It cannot be an option as a substitute for therapy. Shine On the Android Play Store, it has 4.6 reviews and more than 100 k downloads. It currently has 4.7 user ratings on Apple App Store. Once more, neither the symptoms of postpartum depression nor the treatment options for it are specifically defined. The user is not given the option to specify new depression symptoms or provide treatments for existing symptoms. It offers a free subscription that enables the user to keep track of their new daily routine and create schedules for each new day of the week, which reduces their anxiety and stress. User also provided by the app an ability to communicate with a computer system to deliver appropriate reading material, but it cannot take the place of an automated solution for postpartum depression psychotherapy.

Mitigating Postnatal Depression: A Big Data Self-help Therapy

139

Insight Timer On the Android Play Store, it has 4.8 ratings and more than 5 million downloads. It has received 4.9 user reviews and more than 21 million downloads on the Apple App Store. Once more, neither the symptoms of postpartum depression nor the treatment options for it are specifically defined. The user is not given the option to specify new depression symptoms or provide treatments for existing symptoms. While this programme offers therapy classes that are not free and promoted by counselling experts, it also offers a free subscription that enables users to manage sleep difficulties and play meditation tracks. It does not take the place of a postpartum depression automated psychotherapy system. Mood Tracker On the Android Play Store, it has 4.8 stars and over 100 k downloads. With another name, Mood Balance—Daily Tracker on the App Store, it has 4.4 user ratings and more than 21 million downloads on the Apple App Store. It does not specify any signs of postpartum depression or treatment methods for postpartum depression. The user is not given the option to specify new depression symptoms or provide treatments for existing symptoms. It offers a free subscription that enables users to track their moods and provides tips for dealing with a sad mood, such as deep breathing or drinking water. Therapy courses that are not offered for free are available through this application. It does not serve as a substitute for an automated system of psychotherapy for postpartum depression. Simple Habit On the Android Play Store, it has 4.6 reviews and more than 1 million downloads. It has received 4.7 user ratings and more than 5 million downloads on the Apple App Store. Once more, neither the symptoms of postpartum depression nor the treatment options for it are specifically defined. The user is not given the option to specify new depression symptoms or provide treatments for existing symptoms. While the user can purchase paid counselling sessions, its free version offers meditation tracks that are available to play for lowering anxiety, and tension, or getting better sleep. Mindful Mamas On the Android Play Store, it has 4.7 stars and over 100 k downloads. It currently has 4.9 user ratings on Apple App Store. It contains built-in meditation tracks to play for lowering anxiety, but doesn’t describe any postpartum depression symptoms or provide clear tactics for dealing with it. The user can save their preferred mantras, or meditation tunes, on the mobile application. It is not a digital psychotherapy system and does not permit the user to add new depressive symptoms or treatments for existing depressive symptoms. Gratitude On the Android Play Store, it has 4.8 stars and more than one million downloads. It currently has 4.9 user ratings on Apple App Store. It does not specify any signs of

140

A. Usman et al.

postpartum depression or treatment methods for postpartum depression. The user is unable to add their own symptoms or treatment options. The user can keep track of his or her performance as well as a daily schedule. Additionally, the users can store their personalised images and movies that play back a calming meditation track. 29 k On the Android Play Store, it has 4.2 reviews and more than 100 k downloads. It has five user reviews on the Apple App Store. It is a free app, but it also asks the user, “What do you want to work on?” and has some built-in symptoms of general depression. Not all postpartum depression symptoms, such as sobbing, panic attacks, or an eating disorder, are present. Anxiety, self-compassion, war anxiety, stress, sleep, relationships, happiness, purpose, or leadership are among the options available to the user. The user is unable to add their symptoms. This program keeps track of the user’s thoughts using a journaling technique.

4.2 Implementation of Algorithm This research proposing an evidence-based mobile application following this algorithm as a combination of cognitive behavioural therapy and interpersonal psychotherapy. Step 1. Start mobile application. Step 2. Firstly, the mobile application will ask about user feelings. Step 3. If a user feels depressed then go to step 4, or step 7, otherwise go to step 10, or step 11 or step 13. Step 4. User can view symptoms that are already been defined by the mobile application. Step 5. User can view the already defined solution to handle a specific symptom. Step 6. User can find the solutions through internet surfing and can store them for future use. Step 7. User can store his own symptom, harmful thoughts, and troubles. Step 8. User can write his own strategy to tackle a specific symptom, harmful thought, and trouble. Step 9. User can save symptoms, harmful thoughts, and troubles written by him/her and data will be saved. Step 10. User can store symptoms of not feeling depressed. Step 11. User can view and edit the stored symptoms. Step 12. User can store the reason which makes the user happy.

Mitigating Postnatal Depression: A Big Data Self-help Therapy

141

Step 13. User can view and edit stored reasons, symptoms, thoughts, and stored solutions.

5 Results Postpartum depression affects many women and a small number of males each year. Many cases go unreported because women may not want to admit their weaknesses, may not be aware that they are experiencing postpartum depression, or may find relief from their symptoms of depression within three to six months. Background research also revealed that 30% of women may have depression even one year after the birth of their child. Various forms of therapy, including talking therapy, expressive art therapy, computer-aided therapy, and clinical procedures, were revealed in the literature review. The patient can engage in both spoken treatment and expressive art therapy independently. Meditation is offered as a treatment for anxiety or sleep disorders in several mobile applications. These mobile applications can offer predefined symptoms and associated fixes. Postpartum mobile applications have been found to not adequately explain postpartum depression symptoms and to not provide the user with personalised alternatives for recording their condition or symptoms. Additionally, they don’t allow the user the ability to store coping mechanisms of the patient’s choice. When a woman realised, she had postpartum depression, she could address it by outlining her remedies for the future time, if she feels depressed or uncomfortable in her daily routine. This was noted from the discussion among ten women who had postpartum depression and had more than one kid. She can therefore act as her own therapist by identifying her symptoms and possible coping mechanisms. She might find a self-help solution to her postpartum and parenting depression with a mobile app that lets her keep track of the signs and treatments for her unique symptoms. According to research, 66% of women had postpartum depression in the past ten years, and more than 80% of women think that postpartum depression affects all women. About 80% of the audience believes that postpartum depression can result in other depressions, such as parental depression. It poses a serious hazard, if improperly controlled, and has an impact on relationships too. About 19% of people with this illness did not seek help from a clinical practitioner. The NHS was not informed about 19% of the instances. In addition, the study discovered that 21% of the audience was unaware of or did not use medicinal mobile applications. About 15% of participants believe that mobile applications are not offering therapeutic answers, even though 45% of participants use mobile application treatment. About 73% of the audience said they would utilise the suggested mobile application, while 27% said it couldn’t offer a therapeutic solution.

142

A. Usman et al.

6 Conclusion and Future Work According to the results of this study, people are open to using self-help mobile application therapy methods to treat postpartum depression. Many incidents have gone unreported since most patients are reluctant to discuss their precarious circumstances. Solutions for meditation and several exercises, like journaling, are available through mobile applications on the Apple App Store and Google Play Store. Mobile applications, however, are currently unable to allow users to retain their condition’s symptoms and treatment plans. We described a fundamental algorithm that combines cognitive behavioural therapy and interpersonal psychotherapy. As a self-help method to manage and treat postpartum depression, it serves as a roadmap for the creation of a therapeutic mobile application. People who have had or are presently experiencing postpartum depression, as well as their relatives, can recover from the depressive disorder. The algorithm offers an evidence-based approach to developing mobile applications that can be tailored to a user’s particular needs. The user can store their symptoms, ideas, and remedies. The user can keep track of their symptoms over time, negative or positive thoughts, the causes of these ideas, and ways to deal with them. In the future, mobile application will be implemented based on the proposed algorithm as a combination of cognitive behavioural therapy and interpersonal psychotherapy.

References 1. Javed B, Sarwer A, Soto EB, Mashwani ZU (2020) The coronavirus (COVID-19) pandemic’s impact on mental health. Int J Health Plan Manage 35(5):993–996. https://doi.org/10.1002/ hpm.3008 2. Dadi AF, Miller ER, Mwanri L (2020) Postnatal depression and its association with adverse infant health outcomes in low- and middle-income countries: a systematic review and metaanalysis. BMC Pregnancy Childbirth 20(1):416. https://doi.org/10.1186/s12884-020-03092-7 3. Chen Q, Li W, Xiong J, Zheng X (2022) Prevalence and risk factors associated with postpartum depression during the COVID-19 pandemic: a literature review and meta-analysis. Int J Environ Res Public Health 19:4. https://doi.org/10.3390/ijerph19042219 4. Logsdon MC, Wisner K, Hanusa BH, Phillips A (2003) Role functioning and symptom remission in women with postpartum depression after antidepressant treatment. Arch Psychiatr Nurs 17(6):276–283. https://doi.org/10.1053/J.APNU.2003.10.004 5. Smith ML (2022) Depression and social isolation as risk factors for poor cardiometabolic health: considering cross-cultural generalizability and mechanisms. Accessed: 24 June 2022. https://www.proquest.com/openview/3abaa7e0a8e54fe01e1ed75117241595/1?pqorigsite=gscholar&cbl=18750&diss=y 6. Ownby RL, Crocco E, Acevedo A, John V, Loewenstein D (2022) Depression and risk for alzheimer disease systematic review, meta-analysis, and metaregression analysis. Accessed 06 July 2022. https://jamanetwork.com/journals/jamapsychiatry/article-abstract/209578 7. Eells TD, Barrett MS, Wright JH, Thase M (2014) Computer-assisted cognitive–behavior therapy for depression. Psychotherapy 51(2):191–197. https://doi.org/10.1037/a0032406 8. NHS (2019) Find an NHS psychological therapies service (IAPT) 15. https://www.nhs.uk/ser vice-search/mental-health/find-a-psychological-therapies-service/ (Accessed 06 July 2022).

Mitigating Postnatal Depression: A Big Data Self-help Therapy

143

9. Bhowmik D, Sampath Kumar KP, Srivastava S, Paswan S, Dutta AS (2012) Depression symptoms, causes, medications and therapies. The Pharma Innovation; New Delhi 1(3), Part A: 37–51. Accessed 28 June 2022. https://www.thepharmajournal.com/archives/2012/vol1is sue3/PartA/5.pdf 10. NHS (2019) Overview - cognitive behavioural therapy (CBT). 16 July 2019. https://www.nhs. uk/mental-health/talking-therapies-medicine-treatments/talking-therapies-and-counselling/ cognitive-behavioural-therapy-cbt/overview/ (accessed 07 July 2022) 11. Markowitz JC, Weissman MM (2004) Interpersonal psychotherapy: principles and applications. World Psychiatry 3(3):136–139. Accessed 13 July 2022. https://pubmed.ncbi.nlm.nih.gov/166 33477 12. Donmez M, Yorguner N, Kora K, Topcuoglu V (2022) Efficacy of bright light therapy in perinatal depression: a randomized, double-blind, placebo-controlled study. J Psychiatr Res 149:315–322. https://doi.org/10.1016/j.jpsychires.2022.02.027 13. Ward T et al (2020) AVATAR therapy for distressing voices: a comprehensive account of therapeutic targets. Schizophr Bull 46(5):1038–1044. https://doi.org/10.1093/schbul/sbaa061 14. Penn CA (2022) Using depth psychology to find meaning in video games: how can video games transform depression? Ann Arbour 2022. Accessed 05 July 2022. https://www.proquest.com/ docview/2654415772/fulltextPDF/76CE0DF2D360440EPQ/1?accountid=9653 15. Roulleau M (2022) Postpartum depression: healing through archetypes and the expressive arts therapies expressive arts therapies. Lesley University. Accessed 28 June 2022. https://dig italcommons.lesley.edu/expressive_theses/646?utm_source=digitalcommons.lesley.edu%2Fe xpressive_theses%2F646&utm_medium=PDF&utm_campaign=PDFCoverPages 16. Pylvänäinen P, Hyvönen K, Muotka J (2020) The profiles of body image associate with changes in depression among participants in dance movement therapy group. Front Psychol 11. https:// doi.org/10.3389/fpsyg.2020.564788 17. Williams R (2022) Singing the blues: how music can help ease postnatal depression. Accessed 04 July 2022. https://www.theguardian.com/society/2019/dec/06/singing-babies-postnatal-dep ression 18. Greenwood L (2021) Breathe melodies for mums. https://breatheahr.org/programmes/mel odies-for-mums/ (Accessed 05 July 2022) 19. Morton A, Forsey P (2013) My time, my space (an arts-based group for women with postnatal depression): a project report. Community Pract 86(5):31–34. Accessed 07 July 2022. https://www.creativityworks.org.uk/wordpress-content/uploads/2013/05/MTMSar ticleMay2013-1.pdf 20. North American Drama Therapy Association, what is drama therapy? https://www.nadta.org/ what-is-drama-therapy (Accessed 09 July 2022) 21. McDonald W, Fochtmann L (2019) What is electroconvulsive therapy (ECT). https://www.psy chiatry.org/patients-families/ect (Accessed 08 July 2022) 22. Baghai TC, Möller HJ (2008) Electroconvulsive therapy and its different indications. Dialog Clin Neurosci 10(1):105–117. https://doi.org/10.31887/dcns.2008.10.1/tcbaghai 23. Marcu G, Ondersma SJ, Spiller AN, Broderick BM, Kadri R, Buis LR (2022) The perceived benefits of digital interventions for behavioral health: qualitative interview study. J Med Internet Res 24(3):e34300. https://doi.org/10.2196/34300 24. Vallury KD, Jones M, Oosterbroek C (2015) Computerized cognitive behavior therapy for anxiety and depression in rural areas: a systematic review. J Med Internet Res 17(6):e139. https://doi.org/10.2196/jmir.4145 25. Uchida T, Takahashi H, Ban M, Shimaya J, Yoshikawa Y, Ishiguro H (2022) A robot counseling system–what kinds of topics do we prefer to disclose to robots?, 2017. Accessed 26 June 2022. https://ieeexplore.ieee.org/document/8172303 26. Christensen H, Batterham P, Calear A (2014) Online interventions for anxiety disorders. Curr Opin Psychiatry 27(1):7–13. https://doi.org/10.1097/YCO.0000000000000019 27. Serrano-Ripoll MJ, Zamanillo-Campos R, Fiol-DeRoque MA, Castro A, Ricci-Cabello I (2022) Impact of smartphone app-based psychological interventions for reducing depressive symptoms in people with depression: systematic literature review and meta-analysis of randomized controlled trials. JMIR Mhealth Uhealth 10(1):e29621. https://doi.org/10.2196/29621

144

A. Usman et al.

28. Braun V, Clarke V, Hayfield N, Terry G (2019) Thematic analysis. In: Handbook of research methods in health social sciences. Springer, Singapore, pp 843–860. https://doi.org/10.1007/ 978-981-10-5251-4_103 29. Wilson M (2004). Constructing measures Routledge. https://doi.org/10.4324/9781410611697 30. Page MJ et al (2021) The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ n71. https://doi.org/10.1136/bmj.n71 31. Bradford S (2022) Birth characteristics in England and Wales: 2020. https://www.ons.gov.uk/ peoplepopulationandcommunity/birthsdeathsandmarriages/livebirths/bulletins/birthcharacteri sticsinenglandandwales/2020#age-of-parents (Accessed 25 Aug 2022)

Improving Learning Effectiveness by Leveraging Spaced Repetition (SR) Aamir Mazhar Abbas, Thaier Hamid, Celestine Iwendi , Francis Morrissey, and Anchal Garg

Abstract The academic efficiency and knowledge retention of students can be improved by practising active recall testing and implementing spaced repetition techniques. The process of trying to recall information previously learned with the aim of increasing the chance of committing the information to long-term memory is called active recall. Spaced repetition is a technique which can help students to memorize and learn information by outspreading reviews of the topics over larger range of time revising the same topic multiple times in a single session. A qualitative method has been followed in this paper which takes a grounded theory approach while evaluating literature on different memory models, memory creation, and retrieval processes. Based on the literature review, an algorithm has been proposed with the aim of improving learning effectiveness by leveraging spaced repetition techniques. Keywords Experiential learning · Knowledge retention · Spaced learning

1 Introduction The world was bought to a standstill with the COVID-19 pandemic, and educational institutions had to implement tools and devise mechanisms to continue to impart A. M. Abbas (B) · T. Hamid · C. Iwendi · F. Morrissey · A. Garg School of Creative Technologies, University of Bolton, Bolton, UK e-mail: [email protected] T. Hamid e-mail: [email protected] C. Iwendi e-mail: [email protected] F. Morrissey e-mail: [email protected] A. Garg e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_10

145

146

A. M. Abbas et al.

education in the changing environment [1]. Students began using Learning Management Systems (LMS) which increase the adoption of e-learning during this period [2]. As students were learning remotely, there was a lack of student–instructor and peer-to-peer interaction [3]. Due to the reduced interaction between the participants, the opportunities to discuss topics being learned reduced. An implication of this was that the students had fewer instances in which they were compelled to reconnect the learned material, thereby reducing the chances of retention of knowledge and long-term memory. Knowledge retention can be improved by enabling students and learners to take part in quizzes that help students to actively recall pieces of information learnt in previous sessions. Studies have been carried out to show that active recall can help improve academic performance and has been proven to be more effective than repeated revisions by reading the material [4]. Spaced repetition (SR) help improve the quality of long-term learning techniques where a learner revises previously imparted knowledge multiple times over long period of time (days or weeks) at regular intervals. The diversity of learning can be improved by implementing SR techniques and has the additional benefits of improved problem-solving skills and memory retention. The major contribution of this research is the design of an algorithm that can improve the learning capabilities of a learner based on SR techniques. The rest of this paper is organized as such Sect. 2 reviews the related literature. In Sect. 3, the methodology of the research is discussed. Section 4 details the algorithm design. Section 5 concludes the research and provides the future scope.

2 Related Work Farmer and Matalin, who are the leading authorities in the field of cognitive psychology have describe memory as ‘the process of retaining information over time’ [5]. In the year 1969, Underwood described that by creating memories of words, other relevant information such as acoustics, pronunciation, frequency of usage, and the context of the usage are stored as well [6]. Individuals participating in verbal learning studies may not remember just the word, but might also remember external factors, and the emotional and mental circumstances under which the words were used during the period of the study. Memory processing is not dependent only on the representation of the original events but might also influence future behavioural patterns. Tests carried out a measure memory processing have shown a wide variability in terms of even three collection. Several factors such as hypnosis [7] and amnesia-inducing treatments [8] can alter memory processing. Psychological processes can remodel the information stored in the memory and can thereby determine if this information will have any influence on future behaviour. Manifestation of memory influencing an individual may be studied based on an index of single or multiple attributes of the memory. Humans are able to quantify

Improving Learning Effectiveness by Leveraging Spaced Repetition (SR)

147

their index by using vocal written illustration of previously learned events. On the other hand, the index in animals is less symbolic, and the memory attribute could be stored as a response to a stimulus. Retention is defined by the dissimilarity between the indexes of animals that had acquired the memory and animals that had not. Forgetting is most likely caused due to long intervals between testing and training. The concept of forgetting is relative where an assessment of performance is carried out after a specific period of time, with identical intervals similar to the ones in previous trials. Retention can be tested easily in humans by reconnecting previous critical events. In animals, the circumstances under which the memory was formed need to be recreated in order to test the animals reaction to stimulus. An ungenerous view of memory could consider memory storage as a single repository containing multiple memories that can be retrieved at a later stage. Other memory models could explain memory enough much more complex manner. They may designate multiple stores of memory such as short-term memory (STM) and long-term memory (LTM) storage [9]. Memory retrieval is requested after short intervals of time varying between a few seconds and minutes in STM. LTM, on the other hand, could request memory retrieval which is after many days, weeks, or even years. James, in 1890, in respect about memory and laid the foundation for modern cognitive psychology identifies two distinct memory groups, that is, the STM and the LTM [10]. Other researchers followed along and concluded that memory can contain multiple stores [11–13]. Many researchers have made arguments that the control for both STM and LTM are stored in a single unique memory system. There is also been put forth to state that the STM is an activated form of the LTM [14, 15]. Information is not easily retained after the first reading and can be forgotten with time as denoted in Fig. 1. Pieces of information will have to be regularly revised multiple times in order to remember [16–18]. Figure 2 represents a graph showing the retention of information with multiple repetitions. The formula to the present the rate of forgetting information is denoted by. R = e−t/s R—memory retention, t—time, s—relative strength of recollection, e—base of the natural logarithm. Hermann Ebbinghaus published the first report pertaining to the forgetting curve in the year 1885 in the German language. It was later translated into English by Henry Ruger and Clara Bussenius and published as Memory: A Contribution to Experimental Psychology in 1913 [19]. Ebbinghaus subjected himself to a number of tests, including memorization and recall of meaningless phrases made up of three alphabets. He would study the fictitious words and test his memory at various points. These findings were recorded and plotted in a graph, which is now known as the forgetting curve, as illustrated in Fig. 1, which is now referred to as the forgetting curve.

148

A. M. Abbas et al.

Fig. 1 Graph showing retained information without spaced repetition

Fig. 2 Graph showing retained information after spaced repetition

In his research, Ebbinghaus claims that various individuals may increase their memory strength through mnemonic method training. He proposed two possible approaches to enhancing memory strength: 1. better representational approaches, such as mnemonic techniques 2. using active recall approaches such as spaced repetition. Researchers in the field of neuroscience have conducted several studies on the spaced learning approach to assess the relationship between how frequently an idea is recalled and how effectively the information is kept. Because of the way memories are formed, SR is effective [20]. The Atkinson–Shiffrin (Fig. 3) memory model [11] demonstrates that for memory to be stored, it has to go through three steps. On the basis of the perception of the senses, such as sight or hearing, sensory memories are established in the initial stage. Based on the amount of attention given to a

Improving Learning Effectiveness by Leveraging Spaced Repetition (SR)

149

Fig. 3 Atkinson–Shiffrin memory model

particular piece of information, it may be stored in the short-term memory, or it may be entirely forgotten. If this memory is not encoded or committed to the long-term memory, it might be lost. When knowledge is repeatedly practised after it has been stored in the short-term memory, the neuronal circuits underlying memory storage are strengthened, which lowers the likelihood that the information will be forgotten. The model shows that memory must pass through three phases in order to be stored. Based on how the senses, such as sight and hearing, are perceived, sensory memories are formed in the initial stage. Depending on how much attention was given to the knowledge, it may be stored in the short-term memory, or it might just be forgotten. If information is not encoded or committed to the long-term memory, this memory risk being lost. Periodic repetition of a memory that has been stored in the short-term memory lowers the likelihood that it will be forgotten by preserving the neural pathways that are important for memory storage. Physiology and kinesiology students at the University of Pampa in Brazil participated in a study that encouraged the activation of previously learned material at the beginning of each lesson [21]. The so-called ‘retrieval activity’ was carried out in every class when a test was given based on material provided in prior sessions. Following the retrieval activity, the students were introduced to fresh material and then took a typical lesson. Following a survey at the end of the semester, the following findings were made: – Student engagement improved due to memory reactivation before introducing new topics in every. – Students achieved higher grades which could be attributed to better content recall capabilities. – There were changes in the behaviour patterns of students leading to an increased in the frequency of weekly studies. Feedback from research participants suggested that the active retrieval activities had an effect on their behaviour and increased the frequency of weekly study, thereby improving their performance at the end of the term. Various other kinds of technology have been used to enhance the learning process [24–27].

150

A. M. Abbas et al.

3 Methodology In order to use a qualitative method for this work, questions concerning memory generation, storage, and retrieval were left open-ended. The pursuit of a wide range of academic material resulted in the collection of data from interviews, graphs, tables, and observational sources as well as data from documentary sources. Although several methods for memory enhancement with SR were investigated, the main emphasis was on topics related to memory with a clear methodological focus on a qualitative approach [22]. A qualitative technique was chosen because it allows for flexibility in the data analysis process, enables the capture of insightful information, and may help spark the development of fresh ideas. To establish a well-informed perspective, research from a wide range of disciplines, including sociology, psychology, physiology, and education, was consulted. Data that was already accessible in the form of journal articles, text, and photographs was gathered using secondary research and a grounded theory methodology [23]. The available data that has been published in research papers, scholarly publications, and other reliable sources was compiled and summarized in the suggested method. The steps that were implemented to carry out the research work are as follows: 1. 2. 3. 4. 5.

Topic identification Source identification for research work Data collection from existing sources Collation and comparison Analysis.

4 Implementation and Results The SR algorithm proposed in this paper is designed to leverage the following aspects of memory creation for learning: – Memory creation and encoding – Memory storage and retrieval in STM – Memory transfer from STM to LTM. This algorithm assumes the following participatory roles: 1. Examiner: In the context of this algorithm, an examiner is a person who is engaged in teaching or who is motivated by the desire to gauge a candidate’s level of knowledge. The ability to create questions and an answer key may lie with the examiner. 2. Examinee: In the context of this algorithm, an examinee is an individual who is taking a test and is provided with questions for which he or she must select proper responses. The examinee may or may not be shown the performance score at the end of the examination.

Improving Learning Effectiveness by Leveraging Spaced Repetition (SR)

151

4.1 Assumptions The algorithm makes the following assumptions concerning its users and usage scenario: – – – –

Users will interact with the system using a personal computer or laptop Keyboard and mouse are used to provide inputs Screen or monitory will be used to display the output the examiner and examinee have basic computer literacy.

Data Representation Assuming that the algorithm would be implemented as a web application that will leverage Web services for data interchange, JavaScript Object Notation (JSON) is suggested as the appropriate information exchange format. JSON makes it simple for examiners to set questions because of the ease of reading and writing JSON objects and the ease with which computers can produce and transport this data. JSON is a native data type in JavaScript, which means that any web browser capable of rendering a webpage can produce, parse, and transport JSON data. Researchers from the University of Shanghai for Science and Technology examined the performance of web services based on JSON on a broad scale, focusing on JSON’s data binding and data mapping capabilities [23]. Question Types The different types of questions might be stored in a single array as denoted below: questionType = [‘sra’, ‘sba’, ‘eo’, ‘mcc’], where ‘mcc’, ‘eo’, ‘sba’, ‘sra’, are acronyms for ‘multiple correct choice’, ‘either option’, ‘single best answer’, and ‘single right answer’ questions, respectively. Question Metadata Similarly, each of the questions may be represented as JavaScript objects that may have characteristics related to the question type and its choices. An example of such a JavaScript object is denoted below: { id: 8,309,034, questionType: ‘sra’, stem: ‘Question string…’, noOfOptions: 4, optionList: [‘Opt A’, ‘Opt B’, ‘Opt C’, ‘Opt D’], multiKey: false, key: [‘Opt B’]. }. The above JavaScript object represents the question and contains the following properties:

152

A. M. Abbas et al.

– – – –

id: each question has its own unique ID. questionType: a string indicating the type of question stem: a string containing the question’s stem noOfOptions: an integer indicating the number of alternatives available in the query. A normal question will contain four alternatives, while questions such as true or false may simply have two. – optionList: an array containing the many choices that this lead may have – multiKey: a Boolean value indicating whether or not this question includes several right answer keys. – key: an array containing a list of valid response keys. Performance Metadata To assess how successfully the examinee has learned the information, the algorithm requires extra metadata. To monitor prior answer history, more attributes may be added to the JavaScript object. The performance history variable tracks the examinee’s performance history of the last ten tries for each question (pH). { …, perfHistory: [0, 0, 1, 1, 1, 0, 1, 0, 1, 1]. }. – perfHistory/pH: monitors the examinee’s performance for a specific question over the last ten times it has been answered. The pH array begins as an empty array. When a question is presented during the test, the user’s performance history is added to the array stack. If the user correctly answers the question, a value of 1 (one) is added to the array, as shown in Fig. 4. When the question is asked ten times, the array will be filled with ten values. When the question is asked again, the array’s old value is popped/removed, and the new value is pressed onto the array. During the re-calibration step, performance history is taken into account. When a question is repeatedly answered incorrectly, it indicates that the learner has difficulties retaining that piece of knowledge. Difficulty Quotient (dQ) Along with other bits of metadata, the examiner will be able to indicate a difficulty quotient (dQ) for each question. The dQ value ranges from 1 (one) to 10. (ten). The greater the dQ score, the more difficult the question should be to answer. A question with a higher DQ score should be more difficult to remember and recall, and it may take many revisions for the examinee to move money from their STM to their LTM. { …, difficultyQuotient: 5, }.

Improving Learning Effectiveness by Leveraging Spaced Repetition (SR)

153

Fig. 4 pH values updated based on examinee input

Recollection Metadata The algorithm must measure how well an examinee recalls material, which it accomplishes by tracking and computing a quantity known as the recall factor (rF). The value of rF ranges from 0 (zero) to 100 (hundred) and is updated each time the examinee answers a question. Every question’s rF value starts at the default value of 25 (twenty-five) and is increased if the user responds correctly or decremented if the user replies incorrectly. When a user answers a question incorrectly, the rF is simple to compute since it simply needs to be decremented. The quantity decremented is decided by the past answer history, which is saved in the pH. If the user’s performance has been poor in the most recent times when the question has been posed, the rF decrements significantly. However, if the user successfully answers the question, there is no way to determine how well the user has retained that piece of knowledge. In this scenario, the user is asked to rate how simple or difficult it was to recall that bit of knowledge. If the user indicates that memorizing this piece of knowledge was simple, the rF is increased by a significant number; if the user indicates that recalling this information was difficult, the rF is increased by a small value.

4.2 The Different Stages of Implementation This paper proposes an algorithm that is carried out in four (four) steps, as indicated in Fig. 5 and is detailed below: 1. The Seeding Phase consists of adding questions and answer keys to the question pool.

154

A. M. Abbas et al.

Fig. 5 Four phases of the algorithm

2. The Pre-Calibration Phase: The newly added questions and answer keys, as well as pre-existing questions and answers, are reviewed to establish the questioning sequence. 3. The Execution Phase: All tests are performed during this phase, during which questions are asked to the examinee and their replies are recorded. 4. The Re-Calibration Phase: After the test, the user’s replies are reviewed to assess how well the user understands individual questions. The Seeding Phase The examiner can enter a set of questions into the database. Every question will include a lead-in stem as well as a set of choices. The right answer, known as the key, should also be included in the question metadata. Different sorts of questions may be used in an exam, as covered in previous sections of this study. All questions that the algorithm can employ will be of the multiple-choice variety, in which the user is supplied with the stem and a selection of alternatives, one or more of which may imply the correct response. These sorts of questions are used because they are simple to examine and score using computer systems, giving them a viable method for assessing huge class sizes. The questions are sorted after they are added to the repository (Fig. 6) and then committed to memory. The Pre-calibration Phase Every time a question is added or edited in the question repository, the pre-calibration phase is launched. The question repository is empty the first time the system is run, and the examiner must replenish or seed the repository with questions (Fig. 7). The examiner can add questions at two points in time. The first instance of adding would be when there are no questions in the database. The input data is ingested and processed by sorting the question set from easy to hard questions based on the dQ value. Because the database is already empty, this would be a simple sorting operation.

Improving Learning Effectiveness by Leveraging Spaced Repetition (SR)

155

Fig. 6 Seeding Phase

Fig. 7 Pre-calibration phase

When a question repository already exists in the database and a new collection of questions is added, both sets are pooled together and the sorting process is repeated. The newest questions have the default rF value of 25 and are initiated with empty pH arrays. The Execution Phase When an examinee begins a test, the execution phase is initiated. The question repository had previously been sorted by rF values. Queries with low RF values are prioritized over questions with higher RF values and are placed to the top of the list. A subset of questions is formed and then sorted based on the pH array contents for questions with matching rF values. The sorting technique will be discussed in the next step (Figs. 8, 9 and 10), that is, the calibration phase of the algorithm. When the examinee begins a test, a series of cards with the question stem and alternatives is presented. The stem will give the information needed for the examinee to select a suitable answer for each question (Fig. 7). When presented with a question and its possibilities, the examinee will select the suitable option that they believe is the proper answer. If the answer is accurate, the examinee is asked to rate how simple or difficult it was to remember that material. If the user indicates that it was difficult for them to recollect the knowledge, the rF value is increased by a tiny amount (between 1 and 3). If, on the other hand, the user indicates that it was extremely simple for

156

A. M. Abbas et al.

Fig. 8 Execution phase

Fig. 9 Newly initialized repository

Fig. 10 Partially initialized repository

them to recall, it signals that the user is highly capable of remembering the piece of information. The rF value is then raised to a greater value (between 7 and 10). The Re-Calibration Phase There are three possible cases for the question repository: a. Newly initialized repository—all questions are unanswered: Because all of the questions in the repository would have been added recently, each question will

Improving Learning Effectiveness by Leveraging Spaced Repetition (SR)

157

have a rF value of 25 (twenty five). There will be no history entered into the pH array for any of the queries (Fig. 8). In this scenario, questions are sorted in truly random order and presented to the examinee in sequence. b. Partially initialized repository—if some questions have been answered and a few new questions are added: in this scenario, other questions in the repository would have previously been answered, and their rF values calculated and record of their history in the pH array will be available. The question set will be split in a 70:30 ratio, with 70% (seventy per cent) of the questions being older and 30% (thirty per cent) of the questions being new. The questions in the new set will be picked at random and will all have the identical pH arrays and rF scores. The already populated repository questions, on the other hand, will have pre-calculated rF values and pH arrays (Fig. 10). These questions are primarily grouped based on their rF values. Lower rF values indicate that the examinee is having difficulties recalling the knowledge. A few queries may have the same rF values, in which case the pH array is considered. Questions that have a greater number of 0’s (zeroes) in their pH arrays would have been answered incorrectly more frequently than others. For questions with the same rF values, a subset of questions with the same score is created and then sorted, sorting questions with the most 0’s (zeroes) at the top of the list. A few questions might have the same rF values, in which case, the pH array is taken into consideration. Questions with more number of 0’s (zeros) in their pH arrays have been answered wrong more often than others. For these questions, a subset of questions with the same score is made and then sorted by prioritizing questions with a higher number of 0’s (zeros) at the top of the list. c. Existing repository—all questions in the repository are answered and have relevant pH and rF values: in such a scenario, the questions are sorted based on their rF values (Fig. 11). Subsets of questions with the same rF values are constructed and then internally sorted depending on the pH array contents, with questions with a larger number of 0’s being prioritized (zeros). d. Summary The proposed algorithm is executed in four stages and attempts to activate initial memory storage in the STM before transferring the memories to the LTM at a later stage. The seeding step of the algorithm is responsible for populating the question repository with data. The pre-calibration step assesses the newly provided data and may build a combined pool of questions before sorting them in order, depending on the status of the repository. The third part of the algorithm, known as the execution phase, provides the user with the question stems and relevant replies in a sequential manner. The examinee completes the examination by selecting the correct response to each question. In the final stage, the recalibration phase, the responses are evaluated, and the question set is updated to appropriately represent the knowledge level of the examinee.

158

A. M. Abbas et al.

Fig. 11 Existing repository

5 Conclusion This work offers an algorithm that links sensory input with STM and LTM based on the Atkinson–Shiffrin memory model. When a person receives sensory information, it is stored in the STM for a brief period of time. Repetition of remembrance aids in the transmission of memories from the STM to the LTM. Examinees are offered questions on a screen, each of which triggers a sensory experience. As examinees attempt to recall information after receiving visual input, the neural pathways that hold the information are reinforced, ensuring that the information is meaningfully preserved in the LTM. The method presents a system for evaluating the examinee’s remembering abilities and prioritizing the remembrance of material that is more difficult to remember. The algorithm proposes to use SR techniques to improve learning effectiveness.

5.1 Future Work The algorithm given in this project’s scope can be expanded to include other features. This current algorithm solely takes into account question in which the examinee is provided options to choose. When the alternatives are offered to the examinee, they operate as visual signals, triggering memory retrieval. This may impede accurate assessment of the examinee’s ability. The algorithm may be extended to analyse textual responses and compare them to predicted answers to determine whether the examinee truly recalls the piece of knowledge. This algorithm is based on a single memory model and excludes research on the phonological loop and visuo-spatial cognition. The algorithm can be modified in future to incorporate audio snippets in conjunction with visual cues to help aid learning and improve memory recollection.

Improving Learning Effectiveness by Leveraging Spaced Repetition (SR)

159

References 1. Alam A (2020) Challenges and Possibilities of Online Education during Covid-19. https://doi. org/10.20944/preprints202006.0013.v1 2. Bashir A, Bashir S, Rana K, Lambert P, Vernallis A (2021) Post-COVID-19 adaptations; the shifts towards online learning, hybrid course delivery and the implications for biosciences courses in the higher education setting. Front Educ 6: 3. De Michele R (2020) Benefits, drawbacks and psychological impact of online lectures during quarantine due to Covid-19 pandemic. In: Proceedings of the 6th EAI international conference on smart objects and technologies for social good. Association for Computing Machinery, New York, NY, USA, pp 257–260 4. Roediger HL III, Agarwal PK, McDaniel MA, McDermott KB (2011) Test-enhanced learning in the classroom: Long-term improvements from quizzing. J Exp Psychol Appl 17:382–395. https://doi.org/10.1037/a0026252 5. Farmer TA, Matlin MW (2019) Cognition. Wiley 6. Underwood BJ (1969) Attributes of memory. Psychol Rev 76:559–573. https://doi.org/10.1037/ h0028143 7. Reiff R, Scheerer M (1960) Memory and hypnotic age regression. International University Press, Oxford, England 8. Bickford RG, Mulder DW, Dodge HW, Svien HJ, Rome HP (1958) Changes in memory function produced by electrical stimulation of the temporal lobe in man. Res Publ Assoc Res Nerv Ment Dis 36:227–240; discussion 241–243 9. Spear NE (2014) The Processing of Memories (PLE: Memory): forgetting and retention. Psychology Press 10. James W (1890) The principles of psychology, vol II. Henry Holt and Company, NY, US 11. Atkinson RC, Shiffrin RM (1968) Human memory: a proposed system and its control processes. In: Spence KW, Spence JT (eds) Psychology of learning and motivation. Academic Press, pp 89–195 12. Brown GDA, Preece T, Hulme C (2000) Oscillator-based memory for serial order. Psychol Rev Am Psychol Assoc (APA) 127–181. https://doi.org/10.1037/0033-295X.107.1.127 13. Lewandowsky S, Farrell S (2000) A redintegration account of the effects of speech rate, lexicality, and word frequency in immediate serial recall. Psychol Res 63:163–173. https://doi.org/ 10.1007/pl00008175 14. Cowan N (1988) Evolving conceptions of memory storage, selective attention, and their mutual constraints within the human information-processing system. Psychol Bull 104:163–191. https://doi.org/10.1037/0033-2909.104.2.163 15. Oberauer K (2009) Design for a working memory. The psychology of learning and motivation, vol 51. Elsevier Academic Press, San Diego, CA, US, pp 45–100 16. Baddeley AD (1997) Human memory: theory and practice. Psychology Press 17. Ellis NC (1995) The psychology of foreign language vocabulary acquisition: implications for call. Comput Assist Lang Learn 8:103–128. https://doi.org/10.1080/0958822940080202 18. Hulstijn JH (2001) Intentional and incidental second language vocabulary learning: a reappraisal of elaboration, rehearsal and automaticity. In: Robinson P (ed) Cognition and second language instruction. Cambridge University Press, Cambridge, pp 258–286 19. Ebbinghaus H (1913) Memory; a contribution to experimental psychology. Teachers College, Columbia University, New York city 20. Roffman JL, Tanner AS, Eryilmaz H, Rodriguez-Thompson A, Silverstein NJ, Ho NF, Nitenson AZ, Chonde DB, Greve DN, Abi-Dargham A, Buckner RL, Manoach DS, Rosen BR, Hooker JM, Catana C (2016) Dopamine D1 signaling organizes network dynamics underlying working memory. Sci Adv 2:e1501672. https://doi.org/10.1126/sciadv.1501672 21. Sosa PM, Gonçalves R, Carpes FP, Mello-Carpes PB (2018) Active memory reactivation previous to the introduction of a new related content improves students’ learning. Adv Physiol Educ 42:75–78. https://doi.org/10.1152/advan.00077.2017

160

A. M. Abbas et al.

22. Creswell JW, Creswell JD (2018) Research design: qualitative, quantitative, and mixed methods approaches, 5th edn. SAGE Publications Inc., Thousand Oaks, California 23. Peng D, Cao L, Xu W (2011) Using JSON for data exchanging in web service applications 7 24. Badshah A, Iwendi C et al (2022) Use of regional computing to minimize the social big data effects. J Comput Ind Eng (Q1 in Computer Science). ISSN 0360-8352. https://doi.org/10.1016/ j.cie.2022.108433. (SCI Impact factor 7.180 and journal indexed in 18 international databases. 2022, 108433) 25. Iwendi C et al A novel unsupervised ensemble framework using concept-based linguistic methods and machine learning for twitter sentiment analysis. Pattern Recognit Lett IF-3.756. ISSN 0167-8655. https://doi.org/10.1016/j.patrec.2022.04.004 26. Krishnan R, Nair S, Saamuel BS, Justin S, Iwendi C, Biamba C, Ibeke E (2022) Smart analysis of learners performance using learning analytics for improving academic progression: a case study model. Sustainability 14(6):3378. https://doi.org/10.3390/su14063378. ((Q2). SCI Impact factor 3.251) 27. Ngabo D, Dong W, Ibeke E, Iwendi C, Masabo E (2021) Tackling pandemics in smart cities using machine learning architecture. Math Biosci Eng 18(6):8444–8461. https://doi.org/10. 3934/mbe.2021418. ((Q2). SCI Impact factor 2.080)

Snooping for Fake News: A Cascaded Approach Using Stance Detection and Entailment Classification Ebenezer Ajay Williams, M. Karthik, A. Shahina, Harshithaa Murali, N. Safiyyah, and A. Nayeemulla Khan

Abstract In this digital era, the dissemination of news is increasingly done online among the exploding population of Internet users. Due to the low cost involved and vanishing journalistic integrity, the spreading of unverified news, also called “fake news,” has become commonplace, often with long-lasting consequences to society. In these times of “digital deceit,” identifying fake news has become very important. This paper proposes a two-step approach using linguistic techniques in a deep learning framework to tackle the problem of fake news. Fake news may comprise fabricated news stories as well as “click baits.” Our approach involves identifying both these fake categories through a cascaded approach employing stance detection and entailment classification by building a complex model involving multiple deep neural networks. We use Term Frequency (TF), Cosine Similarity, and Word2Vec as features at different stages in the model. The datasets from Fake News Challenge 1 (FNC 1) and the Sentences Involving Compositional Logic (SICK) are used for stance detection and entailment classification, respectively. The focus of this work is, given an article, not to establish its credibility by identifying whether the article is E. A. Williams · M. Karthik · A. Shahina · H. Murali (B) Department of Information Technology, Sri Sivasubramaniya Nadar College of Engineering, Kalavakkam, Chennai, India e-mail: [email protected] E. A. Williams e-mail: [email protected] M. Karthik e-mail: [email protected] A. Shahina e-mail: [email protected] N. Safiyyah Department of Computer Science and Engineering, Sri Sivasubramaniya Nadar College of Engineering, Kalavakkam, Chennai, India e-mail: [email protected] A. Nayeemulla Khan School of Computer Science and Engineering, Vellore Institute of Technology, Vandalur-Kelambakkam Road, Chennai, Tamil Nadu 600127, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_11

161

162

E. A. Williams et al.

counterfeit or not but to identify and eliminate the spread of fake news articles with respect to any particular spurious source. Keywords Entailment classification · LSTM · Natural language processing · Stance detection · Word2Vec

1 Introduction With the development of the Internet, news that was once transmitted through conventional media such as newspapers and television is now being passed instantly with the help of various platforms across the Internet. However, this advancement comes at a cost. Being a powerful influence of the views of society, these platforms unwittingly serve as a medium for the propagation of manipulated information for one’s secondary gain. Platforms such as Facebook and Twitter are subjected to harsh criticism for being the nodal center for spreading the news that lacks veracity. In other words, they become easy conduits of Fake News. There exist quite a number of websites producing fake news articles on a daily basis. Several other news websites cite these articles as the source of their publications and thus evolve the source of fake news. It is also believed that the propagation of fake news articles played a major role in the outcome of the 2016 US Presidential Elections [1]. Facebook is currently trying to curb the spread of these deceptive news articles and has implemented features that enable users to report fake news articles. However, this feature may only be biased to what the user determines as fake and would still suffer from bias on either end of the political spectrum. In [2], the authors state that the traditional methods of checking and vetting political deceptions are impossible with the plethora of articles that are fed by deceptive content creators. They broadly classify the techniques to identify fake news into two: Linguistic Approaches and Network Approaches. In addition to the fabricated news mentioned above, there is also a proliferation of headlines whose main purpose is to attract attention and encourage visitors to click on a link to a particular web page in order to generate advertising revenue. The authors in [3] refer to these deceptive contents as “Click baits.” It is evident that every fake news website cannot generate fake content that is uniquely novel. It is usually one website that creates the content and other websites build on that fake content, and thus, it becomes the origin of multiple sources of fake news. One such example is “Email Scandal: Clinton Foundation’s Ties to Financiers of Terrorism” by “investors.com” and “WikiLeaks: Hillary Clinton knew Saudi, Qatar were funding ISIS—but still took their money for Foundation” by “http://21stcenturywire.com.” We see that both articles aim at diverting the reader’s attention toward Hillary Clinton’s association with terrorism. We understand that fake news may be of different forms and is second to no other global problem. In this article, we explore an approach to detect and identify

Snooping for Fake News: A Cascaded Approach Using Stance Detection …

163

fake news that is circulated on the Web. Instead of approaching the problem from the perspective that a news article should be classified as fake or not solely based on its characteristics, we explore the possibility of tackling this problem from the perspective of an organization that has identified a particular fake news topic and wishes to eliminate the circulation of all other similar fake news articles that are in circulation on the Internet.

2 Related Work After the influence of fake news in the 2016 US Presidential elections, research in this field has been gaining momentum. There exists a considerable amount of work on deception detection employing machine learning and rule-based models for classifying them, particularly in social media and online reviews. The authors in [4] employed the Naive Bayes Classifier to detect fake news based on the occurrences of words in the article. However, if all the words in the article were not found in the training set, then classification was not possible. Bajaj [5] employed CNN-based and RNN-based models to classify fake news with datasets obtained from Kaggle [6] and the Signal Media News datasets [7] and found that gated recurrent units outperform other models in the F1 score and recall. The updates to the hidden states of the gated recurrent units for an input sequence of data x1, x2, x3, …, xt with weights W, Wu, Wr, U, Uu, and Ur and bias-vectors b, bu, br are given by the following equations: ht = ut h + (1 − ut )ht − 1 t

(1)

h = tanh(xt W + (rt ht − 1)U + b) t

(2)

ut = (xt Wu + ht−1 Uu + bu )

(3)

rt = (xt Wr + ht−1 Ur + br )

(4)

Δ

Δ

where rt , ut , and ht act as the reset and update gates, respectively. Feng and Hirst in [8] perform a semantic analysis looking at “object: descriptor” pairs for contradictions in online reviews and implement it on top of a deep model based on unigram, bigram, and syntax features for additional improvement. In 2015, there was an online competition on fake news detection, known as Fake News Challenge 1 (FNC 1). The challenge was introduced to explore the use of artificial intelligence, especially machine learning and natural language processing, to combat the problems faced due to fake news. The first iteration of the contest focused on stance detection. Stance detection is to detect the stance of two pieces of text. In FNC-1, the challenge was to find the stance of a news article headline with the article body. The categories of classification were agrees, disagrees, unrelated, and

164

E. A. Williams et al.

discuss. The top three submissions were from Talos Intelligence, Team Athene, and UCL Machine Reading. The submission by Talos Intelligence [9] used an ensemble of a tree model (XG Boost) and a deep learning model. It used the Google pre-trained word2vec [10] vectors for the deep learning model, and Bag of Words (BoW) as well as word2vec for the tree model as features. It produced an accuracy of 82.02%. The model built by Team Athene [11] followed a BoW approach with a few additional features extracted through non Nearest Neighbor Fields (NNF), Latent Semantic Indexing (LSI), Latent Semantic Analysis (LSA), and paraphrase detection on word embedding. The model was a simple dense neural network with a softmax output. The accuracy of the model was 81.97%. UCL Machine Reading [12] took a more simplistic approach. They deployed the features Term Frequency (TF) of the headline and the article and cosine similarity between them. The model used was a single hidden layer dense network with a softmax output. Even though the model and the features were much simpler compared to other top submissions, it was able to achieve results comparable to them with an accuracy of 81.72%.

3 Our Approach In our work, we take a cascaded approach to identify the “click baits” and fabricated news. By cascaded approach, we mean a sequence of modules to identify fake news of different types. We first check whether the article contradicts its heading to identify whether it is “click bait.” Once we have eliminated the “click baits,” we compare the headings of the non-click bait articles with the headings of articles established as Fake news to identify whether or not they are similar to the fake news topics. The articles that are similar are eliminated as fabricated news. Thus by addressing both “click baits” and fabricated news, our system would be more robust than the previous approaches.

4 System Architecture The architecture of the overall system is given in Fig. 1. We use selected articles from the Kaggle fake news dataset [6]. It contains 13,000 fake articles with headings and a corresponding body for each heading collected from various sources on the Internet. The main motive of the proposed system is to detect articles in the Internet or a single website that discuss the same topic and have the same stance as that of existing fake news. Consider the existing fake news as f, with fh denoting the title and fb denoting the body of the fake news. Articles similar to f are found using the Google custom search API denoted by s where si denotes the ith article in the list. Now the system is used to check whether the article si agrees with fake news f to classify it as fake news. The process denoted in the Fig. 1 is carried on for a fake news article f selected from the dataset and a single article si that is similar to f. The

Snooping for Fake News: A Cascaded Approach Using Stance Detection …

165

Fig. 1 Overall architecture of the system. The figure shows the cascaded approach taken to determine whether an article is fake or not

output could be fake, not fake or unknown. The overall system contains a cascade of two classifiers: (1) Stance Detector and the (2) Entailment Classifier.

4.1 Stance Detection Stance, in linguistics, is defined as the position the author or the writer takes with respect to the ongoing topic of interaction. Stance detection is the process of classifying two different pieces of text based on the relative claim or perspective or stance made by each on a particular topic. These pairs of texts could discuss the same topic or agree with each other or disagree with each other or are unrelated to each other. In the proposed system, stance detection is performed between the headline and the body of an article. It is used to detect whether the article body really discusses the same thing as the article headline. Stance detection could also be used to solve the clickbait problem on the Internet. With respect to articles, false headlines with increased sensationalism and emotion are used to attract Internet users to something that is completely unrelated to it. Model and Features The model used for stance detection is similar to that of UCL machine reading [12] used in FNC 1. This model was chosen due to its simplicity and accuracy comparable to the other top submissions in FNC 1. The input to the model is the heading and

166

E. A. Williams et al.

body of articles. The model classifies into the categories unrelated, agrees, disagrees, or discusses with respect to the corresponding headline. The features used for stance detection are BoW TF of headline and BoW TF of body. The cosine similarity of TF Inverse Document Frequency (IDF) of headline and body BoW considers a text, sentence, or a document as a set of words without regarding the grammar or order of it. TF is a BoW approach that represents the weight of the word in a sentence or document. Consider f (t, d) as the frequency of a word t in a document d then, term frequency of a word is denoted by TF(t, d) = Σ

f(t, d) f(t, , d)

(5)

t, ∈d

Cosine similarity between two vectors is the measure of similarity between them. It could also be represented as the cosine of the angle between the two vectors. The formula for cosine similarity is given as cos(θ) =

A.B ||A||||B||

(6)

/Σ n 2 where the norm ||A|| = i Ai . Inverse document frequency of a word is the inverse of the number of documents in which the word occurs. The formula is given as IDF(t) = log

N nt

(7)

where nt is the number of documents in which the word t occurs. TF-IDF is the dot product of the TF vector and the IDF vector of a sentence or document. TF − IDF(t) = TF(t).IDF(t)

(8)

The model used for stance detection is given in Fig. 2. The dataset used is the dataset provided in FNC 1. The model as shown in Fig. 2 has two hidden layers of 100 units each and dropout of 0.6. Output is a softmax layer with four categories: unrelated, agrees, disagrees, discusses. The input layer is of size 10,001. The TF of the headlines and the body are calculated based on the 5000 most frequently occurring words in the dataset. These are appended with the TF-IDF cosine similarity to form the 10,001 input array. The model is trained with a batch size of 500 and 90 epochs. An accuracy of around 91% is obtained. Stance detection is applied to the headline and body of each si . The primary purpose of this step is to check whether the body discusses the same issue as the title.

Snooping for Fake News: A Cascaded Approach Using Stance Detection …

167

Fig. 2 Stance Detection Model. The figure shows the architecture of the neural network used to determine the stance

4.2 Entailment Classification An entailment is a deduction or implication, that is, something that follows logically from or is implied by something else. Deducing whether two statements are contradictory or entailing is an integral part of differentiating between the information presented by the text in order to understand whether the text refers to the same incident or different incidents. For example “A soccer game with multiple males playing” entails “Some men are playing a sport” while “A black race car starts up in front of a crowd of people” contradicts “A man is driving down a lonely road.” An entailment classifier needs more than one sentence as input and will return a class ∈ {neutral, entailment, contradiction}. In our approach, we compare the headings sh of the articles obtained through the Google custom search API with the title fh of the fake news article. If sh entails fh , we assume that the news article s presents the same story as that of the fake article f. We do so because all click bait articles are ruled out in the stance detector phase and only the articles whose headings sh are related to their bodies sb are considered for entailment classification. Headlines sh that are

168

E. A. Williams et al.

neutral or contradictory to fh do not agree with the fake news and can be ruled out as non-fake articles. Features The complexity of entailment classification exists in handling input vectors of variable length and preserving the underlying semantic meaning of the sentence. This is because it is not possible to feed several classifiers with inputs of variable lengths. In a BoW model, complex features such as TF and TF-IDF have been engineered for the purpose of text classification. However, in the case of entailment classification, it is important to note that the context of the text must be preserved in order to ensure that the underlying meaning of the sentence stays the same. For example “The man stood in front of the boy” and “The boy stood in front of the man,” we see that the two sentences, though having the same words denote a different positioning of the individuals mentioned. Thus, the BoW model cannot be employed for this task. Since context needs to be preserved, a new way of representing text that preserves context as well as the meaning of the word is required. One solution to this problem is to use word2vec. Word2vec is a vector space that is used to represent words based on the context in which it appears in a corpus. It could be trained on a dataset to extract the vector representation of each word appearing in the dataset and also group similar words. One of the popular models used is the Google news pre-trained word2vec model that is trained on the Google news dataset. It has a vocabulary of around 300 million words and represents each word with a 300 dimensional vector. This vector representation of words preserves the meaning of words and also groups similar words together within the vector space. Words with similar meanings are closer to each other in the vector space than others. Another problem in representing text for machine learning is that the sentences to be given as input are of varied length. Popular machine learning models are trained on a fixed-length input. Since text is of variable length, it is not possible to use it in models with fixed input size. One way of representing a variable length text in a fixed size is using doc2vec [13]. Doc2vec takes as input the word2vec representation of each word in a corpus and outputs a vector of predefined size. It has a similar implementation to that of word2vec. A corpus is related to its doc2vec representation similar to the way a word is related to its word2vec representation. This solves the variable length problem, but recent use of doc2vec for text classification hasn’t provided much accurate results. Another simple way of solving this problem is to use padding. Since each corpus is of variable length, its word2vec representation of each word could be padded with zero vectors in the end or at the beginning to make it a fixed-length vector. This has provided good results in the past for text classification, and this is used as the feature for entailment classification. Model For the purpose of maintaining the context of the text, we propose a simple Recurrent Neural Network (RNN)-based model for the task. For a given sequence of data (x1 , x2 , x3 , ..., xT ), where at each t ∈ 1,…, T, updates to a hidden state vector ht are

Snooping for Fake News: A Cascaded Approach Using Stance Detection …

169

performed via the following equation: ht = sigmoid(Wxt + Uht − 1)

(9)

However, traditional RNNs suffer from the shortcoming that its back-propagated gradients become vanishingly small over long sequences. This is known as the vanishing gradient problem and is a major problem for networks to learn longterm dependencies. To eliminate this shortcoming, we turn to a special kind of RNN capable of learning long-term dependencies known as Long Short-Term Memory Networks or shortly LSTMs which was proposed in [7]. For a sequence of data (x1 , x2 , x3 , ..., xT ), where at each t ∈ 1,…,T, updates are performed in an LSTM with weights Wi , Wf , Wc , Wo Ui , Uf Uc Uo and bias-vectors bi , bf , bc , bo as follows: it = sigmoid (Wi xt + Ui ht−1 + bi )

(10)

ft = sigmoid (Wf xt + Uf ht−1 + bf )

(11)

ct = tanh (Wc xt + Uc ht−1 + bc )

(12)

ct = it ct + ft ct−1

(13)

ot = sigmoid (Wo xt + Uo ht−1 + bo )

(14)

ht = ot tanh (ct )

(15)

Graves in [14] provides a deeper explanation of the LSTM model. We see that authors in [15] employ a Siamese architecture-based network for the purpose of Natural Language Processing in identifying the similarity measures of the text. A Siamese network is composed of two identical sub-networks (similar weights and bias) whose outputs are merged by using a specific function for the task. The authors in [16] propose a Siamese Adaptation of the LSTM to classify entailment on the SICK [17] dataset. We build on the same approach for our task. In the Siamese adaptation of the LSTM, rich semantics can be learned from the highly structured space of word embedding extracted from pairs of sentences while being trained. Semantic features are extracted from the two input texts while the similarity metric between the two sentences is learned from the output of the two sub-networks by using a similarity function such as cosine similarity, Euclidean distance, or Manhattan distance. We employ Manhattan distance in this case. The SICK dataset has 9840 sentence pairs with 4439 for training, 495 sentence pairs for validation, and 4906 sentence pairs for testing with each pair containing an entailment label and relatedness score given by humans. The relatedness score is a value denoting the similarity measure between two sentences as labeled by humans

170

E. A. Williams et al.

as y ∈ [1, 5] The entailment label represents the class z ∈ {neutral, entailment, contradiction} to which the pair of sentences belong to. Considering our task of classifying news headings, we view the SICK dataset as apt for the purpose. We employ word2vec to the pair of input sequences (h1 (a), h1 (a), … hT (a))) and (h1 (b), h1 (b), … hT (b))) which are then transformed into fixed-size vectors of (Rdin ). Based on the maximum length n of the sentences being compared the input vector– matrix h(a,b) of shape din * n is created for each sentence, these matrices represent the word2vec embedding. The model produces a mapping from a general space of variable length sequences into a structured metric space of fixed dimensions that can be easily interpreted during prediction. Thus, examples not present in the training set would be classified correctly. For the first phase of our task, we use the relatedness measure y as the target output with (hi (a), hj (b)) as inputs. Here, hi (a) and hj (b) denote the word2vec representations of each word in the pair of input sentences. The process involved in entailment detection is shown in Figs. 3 and 4. Just as in [16], we propose a Manhattan LSTM model that uses the Manhattan distance as a similarity measure for this task. The Manhattan distance between two points M(x,y) and P(x,y) is given by d(M,P) = |Mx − Px|+|My − Py|. The hidden state of the LSTM at each sequence (x1 , x2 , x3 , ..., xT ) is updated by Eqs. (10)–(15) when passed through it. The LSTM uses 50-dimensional hidden representations ht and memory cells ct along with din-dimensional vectors = 300 and drep-loops in the network = 50 for this task. The following function is applied to the LSTM representations to obtain the representation space which is then generalized to infer the semantic similarity between the headlines: g : Rdrep × Rdrep → R

(16)

Fig. 3 Relatedness Score Prediction. The figure shows the architecture of the model used to predict the relatedness score of the sentences

Snooping for Fake News: A Cascaded Approach Using Stance Detection …

171

Fig. 4 Entailment Classification. The figure shows the architecture of the fully connected network used for entailment classification

Let g(h(a)) and g(h(b)) be the outputs of the two parallel LSTMs which act as submodels in the system. We use the following similarity function to force the LSTM to analyze the semantic differences during training: f(g(h(a)), g(h(b))) = (exp(−|| g(h(a)) − g (h(b)) ||1) ∗ 4) + 1 ∈ [1, 5]

(17)

We tune the gradient clipping ratio to avoid the exploding gradient problem and use the AdaDelta optimizer to compile the model. We train the model for 25 epochs with the Mean Squared Error Loss function to predict the relatedness score y. Given the target variable and the predicted variable ypred the mean square error between them for n instances is given by MSE =

1Σ (y − ypred )2 n

(18)

We obtain an MSE of approximately 0.336 on the testing set. In order to predict the entailment class of the sentences, we rely on the same representation space learned

172

E. A. Williams et al.

from predicting the relatedness score. As in [18], we extract each LSTM output of the pair of sentences g(h(a)) and g(h(b)) and engineer two main features |g(h(a)) − g(h(b))|, element-wise absolute differences and g(h(a))g(h(b)) element-wise products into a single feature matrix. These are then fed into a fully connected network as in Fig. 4. Finally, we train the model using the Adam optimizer [19] with categorical cross-entropy as the loss function for 10 epochs.

5 Results Our stance detector performs fairly well with an accuracy of approximately 91% on the test set. The Loss and Accuracy of the detector with a validation split of 0.3 are plotted in Figs. 5 and 6. Considering an example from the dataset: “NASA questions whether crater in Nicaragua caused by meteorite” as the heading with “A Bolivian nun gave birth in San Severino Marche after being taken to hospital where she complained of a bad stomach ache, Italian newspaper Corriere Adriatico said on Friday. The newspaper said the nun, whose age wasn’t given, gave birth last Sunday and intends to keep the Fig. 5 Stance detection—Accuracy versus Epochs

Fig. 6 Stance detection—Loss versus Epochs

Snooping for Fake News: A Cascaded Approach Using Stance Detection …

173

baby, whose sex wasn’t given. The nun had been staying at a cloistered convent in the province of Macerata since June. The hospital hasn’t confirmed the birth, and the bishop of nearby town Camerino, Francesco Brugnaro, hasn’t made any comment on the case. Another case of a nun giving birth took place in 2011 in Marche, when a 41-year-old Congolese nun gave birth to a baby girl in Pesaro. In that case, the nun had been raped abroad by a foreign priest and initially gave her daughter up for adoption. The case engendered strong protests from the adoptive family after it reached Italy’s highest court of appeals in February 2014 after the nun changed her mind and was granted custody, reversing a lower court decision.” we see that the article was about a Bolivian nun giving birth and had nothing to do with the heading which was concerned with NASA. The softmax output for prediction is shown in Fig. 7a. Our Entailment Classifier performs with an accuracy of 80.6% on the testing set. The graphs for the training accuracy and loss with a validation split of 0.3 are shown in Figs. 8 and 9. Considering an example in the training set “There is no man in a black jacket doing tricks motorbike” and “A person in a black jacket is doing tricks on a motorbike”, our classifier correctly predicts it as a contradiction. The output of the softmax layer is as shown in Fig. 7b and c. However, our classifier suffers from a drawback that any pair of sentences not recognized as contradiction or entailment would be ruled out as neutral. This is mainly due to the skew in the SICK dataset as well as different words accounting for the same meaning. Considering another example in the testing set, “Two people are kickboxing and spectators are watching.” and “Two people are fighting and spectators are watching.” We see that the class is wrongly predicted as neutral. This is primarily because our classifier fails to understand that the words “kickboxing” and “fighting” are more or less the same. We infer that classifying the sentences solely on their representations would be a difficult task. We ran the system on examples by extracting keywords and searching using the Google API. The following is a sample of the output produced by the entire system. At present, there exists no standard dataset to test this approach. Hence, we select news articles from several of the previously mentioned sources. In Table 1, it may be seen that the headings “WikiLeaks: Hillary Clinton knew Saudi, Qatar were funding ISIS—but still took their money for Foundation” and “Julian Assange: Isis funded by Saudi Arabia and Qatar—Clinton still takes money.” are classified as Neutral even though they are the same. This is primarily due to the skewness and the simplicity of the sentences in the dataset. Examples with complex similarities and contradictions are not present in the dataset. Further, enhancement of the classifier could tackle this issue. Additionally, to solve this issue, we propose the creation of a dataset solely based on news headings with a relatedness score and an entailment class.

174

E. A. Williams et al.

(a)

(b)

(c)

Fig. 7 Output of the Softmax Activation Function of an example. a, b, and c show bar graphs predicting the class as unrelated, predicting the class correctly as a contradiction, and predicting the class incorrectly as neutral, respectively. The red bar denotes the actual class

6 Conclusion The results shown in this paper are very promising. This method shows that a cascading approach has great potential in tackling the problem of Fake news and can be further improved by using other machine learning approaches. The stance detector shows promising potential in identifying “click baits.” Also, the entailment classifier shows great potential in identifying similar fake news articles. However,

Snooping for Fake News: A Cascaded Approach Using Stance Detection …

175

Fig. 8 Entailment model: Accuracy versus Epochs

Fig. 9 Entailment model: Loss versus Epochs

in spite of the high performance of both our classifiers, there are definite areas for improvement. Constant research following this approach is needed to achieve a robust system. Creation of a dataset to be used primarily for this approach would further improve the potential of this work. Future work is aimed at improving the accuracy of the model by using even more complex features such as Wordnet [20] embeddings to solve problems like recognition of words with similar meaning. A system such as a browser plugin that would search for all other similar articles and classify them as fake or not in real time could also be built with an improved system.

176

E. A. Williams et al.

Table 1 Output of the overall system on selected examples SH

FH

Stance

Entailment Classification

The $10bn question: what happened to the Marcos millions?

Muslims BUSTED They Stole Millions In Gov’t Benefits

Agrees

Neutral

True

Exiled former president Muslims BUSTED They Stole Millions In Gov’t Yahya Jammeh “stole $11.4 m” from the Gambia Benefits

Agrees

Neutral

True

Fifa inquiry: US attorney general announces 16 new indictments—video

Discusses Neutral

True

Hillary Clinton reacts to BREAKING Weiner Discusses Neutral new FBI investigation into Cooperating With FBI On her emails—video Hillary Email Investigation

True

Re Why Did Attorney Fifa officials pocketed $150 m from “World Cup General Loretta Lynch of fraud”—US prosecutors Plead The Fifth

Discusses Neutral

True

Judge accepts request from Judge accepts Stormy Stormy Daniels lawyer to Daniels request to depose depose Trump President Trump

Agrees

Entailment True

CBS 60 min Withheld Trump’s Appeal to “Stop Attacking Minorities,” and Ignored Reports of Attacks on Trump Supporters

CBS 60 min Withheld Agrees Trump’s Appeal to “Stop Attacking Minorities,” and Ignored Reports of Attacks on Trump Supporters

Entailment True

WikiLeaks: Hillary Clinton knew Saudi, Qatar were funding ISIS—but still took their money for Foundation

Julian Assange: Isis funded by Saudi Arabia and Qatar–Clinton still takes money

Neutral

Re Why Did Attorney General Loretta Lynch Plead The Fifth

Agrees

False

Author Contributions All authors discussed the contents of the manuscript and contributed to its preparation. While E. A. W and K. M designed the model and implemented the system, H. M and S. N helped with the manuscript. S. A and N. K. A provided critical feedback and helped shape the research, analysis, and final version of the manuscript.

Conflict of Interest The authors declare no conflict of interest.

References 1. Allcott H, Gentzkow M (2017) Social media and fake news in the 2016 election. J Econ Perspect 31(2):211–236 2. Conroy NK, Rubin VL, Chen Y (2015) Automatic deception detection: methods for finding fake news. Proc Assoc Inf Sci Technol 52(1):1–4

Snooping for Fake News: A Cascaded Approach Using Stance Detection …

177

3. Chen Y, Conroy NJ, Rubin VL (2015) Misleading online content: recognizing click bait as “false news”. In: Proceedings of the 2015 ACM on workshop on multimodal deception detection, pp 15–19. (Nov 2015) 4. Granik M, Mesyura V (2017). Fake news detection using naive Bayes classifier. In: 2017 IEEE first Ukraine conference on electrical and computer engineering (UKRCON), pp 900–903. IEEE. (May 2017) 5. Bajaj S (2017) The pope has a new baby! Fake news detection using deep learning 6. Kaggle, Getting real about fake news. https://www.kaggle.com/mrisdal/fake-news/. Accessed 2020 7. Corney D, Albakour D, Martinez-Alvarez M, Moussa S (2016) What do a million news articles look like? In NewsIR@ ECIR, pp 42–47. (March 2016) 8. Feng VW, Hirst G (2013) Detecting deceptive opinions with profile compatibility. In: Proceedings of the sixth international joint conference on natural language processing, pp 338–346. (Oct 2013) 9. Baird S, Sibley D, Pan Y(2017) Talos targets disinformation with fake news challenge victory. https://blog.talosintelligence.com/2017/06/talos-fake-news-challenge.html 10. Mikolov T, Sutskever I, Chen K, Corrado GS, Dean J (2013) Distributed representations of words and phrases and their compositionality. In: Advances in neural information processing systems, pp 3111–3119 11. Hanselowski A, Avinesh PVS, Schiller B, Caspelherr F (2017) Team Athene on the fake news challenge. https://medium.com/@andre134679/team-athene-on-the-fake-news-challenge-28a 5cf5e017b 12. Riedel B, Augenstein I, Spithourakis GP, Riedel S (2017) A simple but tough-to-beat baseline for the Fake News Challenge stance detection task. arXiv:1707.03264 13. Le Q, Mikolov T (2014) Distributed representations of sentences and documents. In: International conference on machine learning, pp 1188–1196. (Jan 2014) 14. Graves A (2012) Supervised sequence labelling. In: Supervised sequence labelling with recurrent neural networks. Springer, Berlin, Heidelberg, pp 5–13 15. Yih WT, Toutanova K, Platt JC, Meek C (2011) Learning discriminative projections for text similarity measures. In: Proceedings of the fifteenth conference on computational natural language learning, pp 247–256. (June 2011) 16. Mueller J, Thyagarajan A (2016) Siamese recurrent architectures for learning sentence similarity. In: Thirtieth AAAI conference on artificial intelligence. (Mar 2016) 17. Marelli M, Menini S, Baroni M, Bentivogli L, Bernardi R, Zamparelli R (2014) A SICK cure for the evaluation of compositional distributional semantic models. In: LREC, pp 216–223. (May 2014) 18. Tai KS, Socher R, Manning CD (2015) Improved semantic representations from tree-structured long short-term memory networks. arXiv:1503.00075 19. Kingma DP, Ba J (2014) Adam: a method for stochastic optimization. arXiv:1412.6980 20. Miller GA (1995) WordNet: a lexical database for English. Commun ACM 38(11):39–41

Early Planning of Virtual Machines to Servers in Cloud Server Farms is an Approach for Energy-Efficient Resource Allocation Kumar P. and VinodhKumar S.

Abstract Carbon dioxide emissions are a significant source of pollution in the atmosphere. The innovative hardware plays a fundamental part in the carbon release. This is huge in light of the fact that power utilization by mechanical apparatus has surpassed 40 k TWh in 2021 from 10 k TWh in 2000 and is expanding step by step. Server farms in the Cloud processing climate have involved a significant imperative situation in this class of mechanical hardware. In Cloud computing, computational assets are leased staying away from gigantic ventures on the business part. Because of this alluring contribution, reception and sending of Cloud computing have become exceptionally famous among ventures as well as in the exploration local area. Increased use of Cloud computing, on the other hand, has resulted in higher energy consumption and carbon emissions into the atmosphere. A server farm is a collection of servers with a big number of them, and there are a lot of them all over the world. In the modern era of Cloud computing, energy consumption is currently viewed as one of the most significant evaluation challenges. Service-Level Agreements are contracts between a customer and a vendor in which the vendor agrees to particular service qualities such as quality, availability, and accountability. One of the major challenges identified is reducing the amount of power consumed by server farms without affecting the Quality of Service (QoS). Consequently, it is proposed to optimize resource allocation and minimize energy consumption for the Cloud environment. This is completed by limiting dynamic servers in a server farm without compromising the exhibition of undertakings and client prerequisites. To verify the efficiency of the suggested calculations, CloudSim is used in conjunction with verifiable responsibility data obtained from over 1,000 virtual computers from Planet Lab. Keywords Virtualization · Utility computing · CloudSim · Service-Level agreement · Energy consumption

P. Kumar (B) · S. VinodhKumar Rajalakshmi Engineering College, Chennai, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_12

179

180

Kumar P. and VinodhKumar S.

1 Introduction As a result of the Internet’s massive success and the rapid advancement of computing and volume advancements, computing assets have become more affordable, viable, and all-around accessible. This innovative initiative has resulted in the acceptance of a new computer model known as Cloud computing. Cloud computing is a relatively new category of Internet-based services that has arisen in recent years. Grid computing, Utility computing, and Autonomic computing are just a few of the computing paradigms and innovations that make up Cloud computing. The cloud serves as a model for providing a common pool of adjustable processing assets with universal, advantageous, on-demand network induction that can be provisioned and communicated with little effort or specialist cooperative communication. It enables the client to pay-per-use lease or share computing equipment. The Cloud is a massive pool of instantly and effectively accessible virtualized computing resources that also serves as a platform for application development and various types of services. The pool can be powerfully reconfigured to change in accordance with competing responsibilities in terms of load balancing, scalability, and elasticity, allowing for optimal resource utilization. One of the most essential characteristics of Cloud computing is the ability to distribute and de-administer Cloud resources for Cloud customers on demand. A Service-Level Agreement (SLA) is a contract between a service provider and its clients that details the services the provider will deliver and the service standards to which the provider must adhere. Customers and service providers agree on specific characteristics of the service, such as quality, availability, and responsibility. The goal of the Cloud service is to meet client Service-Level Agreements (SLAs) while reducing functional expenses during the period spent allotting and de-assigning assets from server farms. Load balancing is accomplished utilizing virtual machine (virtual machine) movement, which chips away at the highest point of virtualization innovation. Virtual machine relocation in Cloud server farms also enables reliable and responsive asset provisioning. Consolidating virtual machines or workloads is an efficient way to maximize asset utilization while lowering energy consumption. Cloud computing also entails reducing energy use in order to improve overall energy efficiency. According to estimates, the cost of cooling and controlling server farms accounts for 53% of total activity utilization [1]. According to one study, server farms in the USA consumed more than 1.5% of total energy produced in 2006, with that percentage expected to rise to 20% in the future [2]. Subsequently, service providers are predominantly worried about the decrease in energy utilization. Server consolidation [3] and energy mindful undertaking planning [4] are two distinct ways to address the issue by merging the assignments into fewer machines and turning off or putting into rest mode unused machines. As a result, overall data center resource utilization is improved, and power consumption is reduced. Consider a scenario in which there are ten hosts. It can be seen that all hosts are in use, with usage ranging from 25% to 55%. We can move a few workloads from one host to another using workload consolidation so that the target host is not overcrowded. We were able to put

Early Planning of Virtual Machines to Servers in Cloud Server Farms …

181

five hosts into power-saving mode while leaving five others running, with utilization varying from 60% to 75%. Business Clouds make use of large-scale server farms and work in an integrated manner, achieving high sensitivity and lowering ongoing costs. It also necessitates a significant initial investment in the development of server farms and increases energy utilization [5]. Smaller-sized server farms are preferable to larger-sized server farms for a variety of reasons, including: (a) a little server farm consumes less power than a large server farm, necessitating the absence of a massive and costly cooling framework; and (b) more modest-sized server farms are less expensive to set up and are better geologically suited when compared to largersized server farms. The actual assets, like CPU centers, data transmission, and disk storage, should be cut and divided between virtual machines. In Cloud server farms, over provisioning of assets is a typical wonder, as Cloud ensures limitless asset provisioning through elasticity, reliability, and availability. Amazon, Google, and Microsoft, for example, are sending massive amounts of server farms all over the world to meet the rapidly expanding capacity interest of clients. They have roughly more than 1 million servers between them in their server farms. As an outcome, a tremendous amount of energy is devoured by these immense server farms to run the servers and keep the cooling framework working. Thus, enormous scope server farms are more costly to support, just as they have additional consequences for the climate because of high energy utilization. The issue of asset and energy failure is tended to with the utilization of virtualization advancing. Virtualization advancements permit the formation of different virtual machines on a solitary actual server, each virtual machine totally disconnected from the other, addressing a runtime environment. Virtualization advances additionally permit live movement of virtual machines [6], starting with one machine, then onto the next, and along these lines, further developing asset use. Power consumption can be diminished by transforming inactive actual machines into power-saving modes to save energy while fulfilling a client’s execution requirements.

2 Related Work Energy utilization is one of the significant concerns distinguished by numerous analysts, in the Cloud environment. It is recommended to change every one of the inactive servers to suspend or wind down mode. Yet, this might raise different issues like tradeoff in execution, infringement of SLA, the expense of reconfiguration, and figuring/correspondence cost during virtual machine relocation. So, to resolve these issues, it is prescribed to monitor server use and responsibility migration. Furthermore, by developing a competent method, there will be an improvement in asset utilization and energy utilization. Verma et al. [7] viewed the trial of power as a careful, one- of-a-kind situation of usage as a major issue. Canisters are viewed as a variable in terms of size and cost. Live migration is used for virtual machine development, starting with one host and progressing to the next at a typical arranging stretch. Regardless, the creators do not discuss the SLA. Prakash et al. [8] propose Optimized

182

Kumar P. and VinodhKumar S.

Energy Utilization in Deployment and Forecast (OEUDF) to reduce energy consumption in the cloud. Makers propose two stages of instruments: first, computing the best data gathering to energy path (ODEP) to pass on and orchestrate the virtual machines and second, computing the situation to design the work interaction. Makers have created power/energy models and proposed an estimation method that uses Directed Acyclic Graphs (DAGs). Lee et al. [9] recognized the problem of underutilized resources consuming energy. Makers propose that latent resources can be put into rest/power-saving mode to maximize resource use while reducing energy consumption. Two task association heuristics, ECTC and MaxUtil, for determining energy, have been presented. It is proposed that a computation can be used to join the obligations using cost work (separate for ECTC and MaxUtil). Manufacturers pledged to reduce energy use, as well as utility prices and their carbon footprint. Beloglazov et al. [10] presented a method for identifying overloaded and under loaded servers by carefully putting up upper and lower utilization borders. Makers recommend transferring some virtual machines from this host to avoid SLA infringement if the host client excels as far as possible (SLAV). If utilization falls below the lowest possible level, all virtual machines on this underutilized host must be transferred, and the host must be shut down to save energy. In any event, no formal technique or process for determining the upper and lower margins has been provided. Wu et al. [11] propose a booking calculation with a Dynamic Voltage Frequency Scaling (DVFS) procedure to improve asset utilization and thus reduce overall energy utilization. The key concept is determining loads for each virtual machine and allocating a virtual machine to the work based on the weight (in expanding requests). Creators guarantee the strategy’s effectiveness in reducing energy consumption based on experimentation results. Melody et al. [12] use virtualization for dynamic resource segmentation according to the obligation’s requirements and work on the number of dynamic hosts to achieve energy viability in the Cloud server ranch. The estimation of variable thing size holder squeezing (VISBP) has been proposed as a model for resource dispersal subject to the major difficulty of free online repository. As long as the gathering (of virtual machines and PMs) requirements are followed, VISBP can handle an acceptable sized assortment. When the movement and weight offset in troublesome places show differently than the current estimate, the makers ensure improved execution. In any case, all PMs are treated as homogeneous, and the application’s cardinal limitation may be the unit cutoff. Lee et al. identified the challenges of expanding cloud system use while reducing total cost [13]. Makers focused on resource allocation for the board and virtual machine and provided an execution assessment-based resource assignment method (using the best fit strategy).The creators claimed that by providing the computation and subsequent experiments, they were able to distribute virtual machines to the best central location, promoting resource use. Workload Consolidation to Achieve Higher Resource Utilization and Energy Efficiency in Cloud Data Center has been proposed by Patel and Patel [14]. In a server farm, the idea of a responsibility union restricts dynamic servers without jeopardising the presentation of tasks and client requirements. An effective technique for assigning users to a pool of computers in an energy-efficient manner was put forth by Habib Ben Abdallah et al. [15]. The

Early Planning of Virtual Machines to Servers in Cloud Server Farms …

183

allocation model highlights the crucial value of non-dominant resource types, such as memory, which are typically wasted by homogenous allocation methodologies. The algorithm’s performance qualifies it for usage in real-time situations where splitsecond judgments must be made. The process of combining responsibility typically entails (a) selecting a small number of virtual machines from overused hosts and attempting to move them to other hosts so that the source host returns to normal and the target hosts are not overused; and (b) selecting all virtual machines from underutilised hosts, attempting to move them to other hosts so that the target hosts are not overused, and turning off these underutilised hosts.

3 Proposed Work The objective of this paper is to address the executives’ asset use and energy use in Cloud computing. Saving energy might result in a compromise in execution. As a result, depending on the client’s requirement for quality of service (QoS) [14], an adaptable choice for energy utilization by the client is dependent on their execution requirement. The proposed system architecture is displayed in Fig. 1. It shows that Cloud clients are on top and they present their prerequisites to the Cloud Service Provider (CSP) through the online interface. In the base, the Cloud server farm involves various actual machines as registering components. These actual machines are given as virtual machines for the execution of errands. Each actual machine is composed of different virtual machines. Each host has a hypervisor (otherwise called a “Virtual Machine Monitor”), which is responsible for dealing with and managing all the virtual machines on the host. The planning part deals with virtual machine booking for the host, and the home specialist monitors the situation with all the virtual machines and the status of the execution of errands. The widespread specialist monitors the whole status of the server farm by totaling data from every one of the home specialists. Workload consolidation has been identified as a process that can be divided among four sub-processes.

3.1 Early Planning The most generally used method of installing virtual machines on servers can be divided into two parts: (i) initial virtual machine mapping on servers during the startup stage and (ii) virtual machine assurance, development, and placement during the cementing stage. Early Planning of virtual machines to servers assumes an imperative part in energy utilization in impending time cycles and ensuing tasks in the Cloud environment. On the off chance that the underlying planning is not productive, ensuing activities might prompt pointless virtual machine relocations, which

184

Kumar P. and VinodhKumar S.

Fig. 1 Cloud system architecture

may thus bring about an expansion in movement costs and SLA infringement. Proficient planning of something very similar during the beginning planning will keep (a) the quantity of the live host as negligible as could possibly be expected and (b) the live host as occupied as conceivable by effectively using them. The trial is completed with a changing number of virtual machines and hosts to grasp the present starting preparation of virtual machines on servers in CloudSim. There are minimal drawbacks to the present early-on virtual machine location method. In any event, the strategy reliably allocates all of the virtual machines to open hosts. During the basic stage, it does not consider the stack of a single host while putting a virtual machine on the host. This resulted in a number of changes in the blend time and, as a result, a degradation in the system’s execution. Second, each host can support up to two virtual machines. This restriction limits the number of virtual machines that can be installed on open hosts, and from time to time, distributions of virtual machines that are more than twofold in number seem to behave differently with respect to host size. Third, while setting up a virtual machine on it and commencing preparation, the default technique does not consider a cutoff. It merely determines whether the true host has sufficient capacity to meet the virtual machine’s requirements. It does not function when it comes to selecting a suitable host for a certain virtual machine, taking into account the host’s capacity, usage, existing weight, and soon. As a result, the goal of this research is to identify these concerns and conduct tests using established appropriation methodologies such as

Early Planning of Virtual Machines to Servers in Cloud Server Farms …

185

the Default technique, Least strategy, Top strategy, Earliest technique, and Backpack approach. (a) Default technique: For the most part, this method correctly assigns all virtual machines to accessible hosts. The following are a few examples to help you understand the practical scenario, using various virtual machines and the number of hosts as boundaries. It is typical that the host has sufficient capacity to meet the virtual machine’s requirements. Instance 1: Regardless of the host limit or utilization, coordinated planning would be done with a total number of virtual machines of 500 and a total number of hosts of 500. Instance 2: The total number of virtual machines is 500; the total number of hosts is 250: Each host will receive two virtual machines. Instance 3: The total number of virtual machines is 750; the total number of hosts is 250: Even if the hosts have the capacity to accommodate more virtual machines, after distributing two virtual machines per host, the remaining 250 virtual machines would remain unallocated. Instance 4: The total number of virtual machines is 400, and the number of hosts is 250: Initially, 250 virtual machines would be distributed among 250 hosts based on a balanced strategy. With one virtual machine already disbursed, 150 virtual machines would be allocated to 150 hosts. As a result, 150 hosts will have two virtual machines each, while 100 hosts will have one virtual machine each. As a result, the default methodology disregards the single host constraint when establishing a virtual machine on it in the first phase. It simply guarantees that the genuine host has adequate resources to provide the virtual machines with what they need. It does not function when it comes to selecting the best host for a specific virtual machine based on the host’s capacity, usage, weight, and other factors. This causes a lot of changes in the time it takes to blend and, as a result, contamination in the overall system’s execution. Finally, because each host has two focuses, there can be all things considered ridiculous two virtual machines on each host during initial preparation. This barrier limits the number of virtual machines that can be placed on accessible hosts, and it manifests itself to a large extent when virtual machines are assigned that are more than two times the size of the hosts. (b) Least strategy: The Least strategy method chooses the host with the maximum raised breaking point for any virtual machine to be propagated on available hosts. A virtual machine of 500 MIPS, for example, must be installed on a host. 500 MIPS, 1000 MIPS, 1500 MIPS, and 2000 MIPS are the open host possibilities. The Least strategy approach technique would place a 500 MIPS necessary virtual machine on a host with a limit of 2000 MIPS. This will free up the most space on the host when the virtual machine duty is completed, increasing the chances of obliging one more virtual machine in the host’s farthest reaches. In any case, this method slows down the production of small openings by as much as MIPS on the server; the disadvantage is that if a virtual machine with higher requirements emerges later, it would not be obligated because the host with the most resources is now involved.

186

Kumar P. and VinodhKumar S.

(c) Top strategy: When allocating virtual machines to existing hosts, the Top strategy method chooses the host with the least MIPS limit that can hold the required virtual machines. A virtual machine with 500 MIPS is necessary, for example, to be installed on a host. 400 MIPS, 600 MIPS, 1000 MIPS, and 1500 MIPS are the available host options. The Top system strategy would place a 500 MIPS essential virtual machine on a 600 MIPS capable host. Although this strategy makes use of accessible hosts, it is computationally slower when it comes to finding a true host for a virtual machine. It is possible that it will frequently create small, insufficient air pockets of unutilized MIPS. (d) Earliest technique: The Earliest strategy methodology selected first has enough experience to meet the virtual machine requirement for any virtual machine to be designated on available hosts. A virtual machine with 500 MIPS necessary, for example, are to be installed on a host. 400 MIPS, 1000 MIPS, 600 MIPS, and 2000 MIPS are the open host options. The Earliest methods’ methodology would place a virtual machine using 500 MIPS on a host with a limit of 1000 MIPS. However, while this method is more powerful in terms of finding a suitable host for a virtual machine, it results in inefficient assignment of existing host limit, as the excess idle constraint of host after designation becomes waste if it is unreasonably more modest, and as a result, virtual machine requests with greater essential cannot be satisfied. (e) Backpack technique: It has a problem with combinatorial upgrades. Choose the mix of virtual machines to remember for a host from a large number of virtual machines, each with its own essential, so that the hard and fast need for virtual machines is not actually or comparable to the host’s capacity, and as far as feasible is just as broad as could truly be expected. Rather of seeking for the necessary for each virtual machine as in the previous situations, this method looks for an open decision for all expected mixes of virtual machines for a particular host. The computation of combinatorial advancement using Backpack is shown as calculation Sect. 3.2 below.

3.2 Algorithm Input: PossibleHost (For virtual machine Assignment), virtual machines_to_Transfer Result: Designated virtual machines_to_Transfer, Host Selected Sort PossibleHost and virtual machines_to_Transfer list in descending order based on consumption for eachHost Selected in PossibleHost do Assign Backpack [virtual machines_to_Transfer.length] [virtual machines_to_Transfer.length] = 0 Assign ConsumptionDifference [virtual machines_to_Transfer.length] = 0 Assign Total virtual machines_to_TransferOnHost [virtual machines_to_Transfer.length] = 0 for J=0 to virtual machines_to_Transfer.length with increment 1 do OutstandingHostConsumption = Tu – U (Host Selected)

Early Planning of Virtual Machines to Servers in Cloud Server Farms …

187

for K = 0 to virtual machines_to_Transfer.length with increment 1 do If U (virtual machines_to_Transfer (K)) < OutstandingHostConsumptionBackpack [J] [K] = 1 Outstanding Host Consumption = OutstandingHostConsumption – (virtual machines_to_Transfer (K))Total virtual machines_to_TransferOnHost [J] ++ ConsumptionDifference [J] = OutstandingHostConsumption Add all virtual machines to Selected virtual machines_to_Transfer for migration with lowest_ConsumptionDifference Update the list of PossibleHost Update virtual machines_to_Transfer list return Selected virtual machines_to_Transfer, Host Selected The Backpack approach seeks out the finest optimum mix of virtual machines with the goal of relegating the greatest number of virtual machines possible, with the ensuing host utilization remaining high and low for a long time. Furthermore, this solution would reduce the size of the powerful host as little as possible, reducing future migrations, lowering energy consumption, and promoting SLA.

4 Results and Discussion Cloud computing environment should provide a perspective on limitless processing assets to clients. To evaluate the proposed calculations, a massive scope of virtualized server farm foundation is required. However, it is quite difficult to complete a substantial degree in order to investigate a recognized establishment that it is necessary to survey and examine the recommended calculations. Diversions have since been enjoyed as a means of evaluating the presentation of planned work and ensuring the repeatability of assessments. The CloudSim [16] tool compartment has been chosen as a recreation stage. To comprehend the impact of beginning planning (of virtual machines on servers) on the energy utilization of server farms, a progression of experimentation with existing default strategy and different methodologies, including dynamic writing computer programs, was conducted [17]. The results of the experiments imply that there has been some progress in the default planning strategy used by CloudSim. A movement of experiments with an arrangement of blends of different hosts and virtual machines with changed MIPS limits was coordinated [18]. Because the Cloud is inherently active, with various proportions of hosts and virtual machines sent to a server farm, the trials are divided into two social occasions: (a) for few hosts and virtual machines and (b) for incalculable hosts and virtual machines which are shown in Tables 1 and 2. Figure 2 shows the underutilized after introductory planning of various methods. Regardless of the number of virtual machines, the default procedure is used on all hosts open in server ranches. A basic number of hosts remain unused, which will be abandoned in the following stages, and the virtual machines on them will be relocated to other hosts. This cycle is long, and unending development could result in

188

Kumar P. and VinodhKumar S.

Table 1 Selection of virtual machine and host for experimentation (small in size) during Early Planning Trials

T1

Virtual machine

22

22

22

22

22

22

22

8

10

12

14

16

19

20

Host

T2

T3

T4

T5

T6

T7

Table 2 Percentage of underutilized host during Early Planning (small in size) using CloudSim Trials Default

T1

T2

T3

T4

T5

T6

T7

2

2

3

3

3

4

4

2

18

29

33

49

50

54

Earliest

18

32

42

47

50

55

56

Top

19

35

44

49

52

57

58

Backpack

22

37

47

51

55

59

61

Least

Percentage of underutilized host

SLA encroachment. When the Backpack is displayed differently in comparison to the default arrangement, the degree of unutilization increases [19]. This indicates that there is still more capacity that might be converted to a power-saving mode, resulting in increased energy adequacy. Therefore, one more virtual machine segment methodology is recommended to be used rather than the default procedure during the starting preparation of virtual machines to reduce planning time and SLA encroachment. Cloud server farms are lively environments, and the quantity of hosts and virtual machines is colossal in size. From now on, incalculable hosts and virtual machines are seen as experimentation. Table 3 shows the quantity of virtual machines and hosts in the server farm, which is huge in size.

Percentage of underutilized host during Early Planning (small in size) using Cloudsim 70 60 50 40 30 20 10 0 T1

T2

T3

T5

T4

T6

T7

Trials Default

Least

Earliest

Top

Backpack

Fig. 2 Underutilized hosts after Early Planning (small in size) of different techniques

Early Planning of Virtual Machines to Servers in Cloud Server Farms …

189

Table 3 Selection of virtual machine and host for experimentation (large in size) during Early Planning T1

Trials

T2

T3

T4

T5

T6

T7

Virtual machine

22

50

100

500

1000

1500

2000

Host

12

40

65

350

760

1200

1800

Percentage of underutilized host

Figure 3 shows values of exploratory outcomes for various portion strategies like default, Least, Top, Earliest, and Backpack technique. The names of the operations are highlighted in the X-pivot, and the Y-hub displays various underutilized rates. The results are shown in Table 4, which demonstrate that the degree of underutilization has improved from the default system to Backpack, implying that the strategies retain a greater number of hosts underutilized even after spreading the existing weight among the hosts. These underutilized hosts can be reprogrammed to regulate the energy-saving mode, resulting in increased energy viability. Along these lines, SLA insurance will be achieved indirectly.

Percentage of underutilized host during Early Planning (large in size) using Cloudsim 70 60 50 40 30 20 10 0 T1

T2

T3

T5

T4

T6

T7

Trials Default

Least

Earliest

Top

Backpack

Fig. 3 Underutilized hosts after Early Planning (large in size) of different techniques

Table 4 Percentage of underutilized host during Early Planning (large in size) using CloudSim Method

T1

T2

T3

T4

T5

T6

T7

Default

1

1

1

1

1

1

1

Least

30

14

31

47

14

38

47

Earliest

38

33

45

54

35

52

62

Top

38

35

46

55

36

53

64

Backpack

40

37

48

57

38

55

66

Kumar P. and VinodhKumar S.

Average resource underutilized(MIPS)

190

Average resource underutilization per active host during Early Planning 3500 3000 2500 2000 1500 1000 500 0

T1

T2

T3

T5

T4

T6

T7

Trials Default

Least

Earliest

Top

Backpack

Fig. 4 Average resource underutilization per active host during Early Planning

Figure 4 shows the average resource underutilization per active host during Early Planning. The X-pivot illustrates the method used, and the Y-hub delineates the normal, underutilized MIPS per dynamic host in a Cloud server farm. From Fig. 4, the graph observed that the measure of underutilization per dynamic host decreased from the default technique to the Backpack technique. Hence, the Backpack technique improves the usage of accessible dynamic hosts. Therefore, higher resource utilization is achieved with the existing default technique. The research outcomes, which are shown in Table 5, show that effective planning during beginning designations keeps the quantity of disconnected hosts as extreme as possible, which thusly brings about less energy utilization. Table 5 Average resource underutilization per active host during Early Planning Method

T1

T2

T3

T4

T5

T6

T7

Default

2250

1850

2300

2750

1900

2600

3000

Least

1750

1600

200

1850

1650

1800

2300

Earliest

750

650

180

740

660

660

660

Top

740

Backpack

720

625

160

725

650

650

650

600

150

710

640

640

640

Early Planning of Virtual Machines to Servers in Cloud Server Farms …

191

5 Conclusion and Future Work Effective designation of virtual machines on a host will support a decrease in total energy utilization for server farms, without compromising general activity and without disregarding SLAs in Cloud environments. One of the critical features for resolving the issue of energy utilization in a Cloud server farm is the underlying planning of virtual machines. Backpack is used as the initial virtual machine arrangement for combinatorial development. The proposed computations are evaluated by simulations of a large-scope analysis setup utilising responsibility results from over 1,000 Planet Lab virtual computers. Through the reenactment result, we could accomplish an upgrade in the energy effectiveness of the server farm without compromising the SLA. In the future, the proposed calculations can be reached from reproduction arrangement to ongoing server farm climate, for example, OpenStack with more intricate responsibility models like Markov Chains.

References 1. Hamilton J (2009) Cooperative expendable micro-slice servers (CEMS): low cost, low power servers for internet-scale services. In: Conference on innovative data systems research (CIDR) 2. Li B, Li J, Huai J, Wo T, Li Q, Zhong L (2009) EnaCloud: an energy-saving application live placement approach for cloud computing environments. In: IEEE international conference on cloud computing. IEEE, pp 17–24 3. Srikantaiah S, Kansal A, Zhao F (2008) Energy aware consolidation for cloud computing. Proceedings of Power aware computing and systems 10:1–5 4. Vasic´ N, Barisits M, Salzgeber V, Kostic´ D (2009) Making cluster applications energy-aware. In: ACDC, Proceedings of the 1st workshop on automated control for datacenters and clouds, pp 37–42 5. Kumar P, Anand S (2016) Multi criteria based task scheduling in cloud environment. Asian J Res Soc Sci Hum (2249–7315), 6(11):659–675 6. Clark C, Fraser K, Hand S, Hansen J, Jul E, Limpach C, Pratt I, Warfield A (2005) Live migration of virtual machines. In: ACM proceedings of the 2nd conference on symposium on networked systems design & implementation, vol 2, pp 273–286 7. Verma A, Ahuja P, Neogi A (2008) pMapper: power and migration cost aware application placement in virtualized systems. In: Proceedings of the 9thACM/IFIP/USENIX international conference on middleware, pp 243—264 8. Prakash P, Kousalya G, Vasudevan SK, Sangeetha KS (2015) Green algorithm for virtualized cloud systems to optimize the energy consumption. Artificial intelligence and evolutionary algorithms in engineering systems, Springer, pp 701–707 9. Lee YC, Zomaya AY (2012) Energy efficient utilization of resources in cloud computing systems. J Supercomput Springer 60(2):268–280 10. Beloglazov A, Buyya R (2010) Energy efficient resource management in virtualized cloud data centers. In: Proceedings of the 2010 10th IEEE/ACM international conference on cluster, cloud and grid computing. IEEE Comput Soc 826–831 11. Wu CM, Chang RS, Chan HY (2014) A green energy-efficient scheduling algorithm using the DVFS technique for cloud datacenters. J Fut Gen Comput Syst Sci Direct 37:141–147 12. Song W, Xiao Z, Chen Q, Luo H (2014) Adaptive resource provisioning for the cloud using online bin packing. IEEE Trans Comput 63(11):2647–2660

192

Kumar P. and VinodhKumar S.

13. Lee HM, Jeong YS, Jang HJ (2014) Performance analysis based resource allocation for green cloud computing. J Supercomput Springer 69(3):1013–1026 14. Patel N, Patel H (2020) Energy efficient resource allocation during initial mapping of virtual machines to servers in cloud datacenters. Int J Distrib Syst Technol 15. Abdallah HB, Sanni AA, Thummar K, Halabi T (2021) Online energy-efficient resource allocation in cloud computing data centers. In IEEE 24th conference on innovation in clouds, internet and networks and workshops (ICIN) 16. Calheiros RN, Ranjan R, Beloglazov A, De Rose CA, Buyya R (2011) CloudSim: a toolkit for modeling and simulation of cloud computing environments and evaluation of resource provisioning algorithms. J Softw: Pract Exper Willey 41(1):23–50 17. Huang J, Wu K, Moh M (2014) Dynamic virtual machine migration algorithms using enhanced energy consumption model for green cloud data centers. In IEEE international conference on high performance computing & simulation (HPCS). IEEE, pp 902–910 18. Kumar P, Anand S (2013) An approach to optimize workflow scheduling for cloud computing environment. J Theor Appl Inf Technol 57(3):617–623 19. Ghamkhari M, Mohsenian H (2013) Energy and performance management of green data centers: a profit maximization approach. IEEE Trans Smart Grid 4(2):1017–1025

Performance Analysis of Distributed Algorithms for Big Data Classification K. Haritha

and M. V. Judy

Abstract In present data age, massive amounts of data are being produced at breakneck speed, and such data, usually obtained from multiple sources, can be in various formats and in most cases must be analyzed in near real time. Data may be interpreted, classified, or forecasted using a variety of machine learning methods. As far as efficiency and outcomes go, these algorithms are unique. It is necessary to do a side-by-by-side comparison of the various approaches in order to determine which one is best for a certain situation. Also, an enormous dataset contains a tremendous amount and spectrum of information. Traditional tools are incapable to processing it. Several alternate methods can be considered to process it, including constructing a distributed environment or contracting cloud-based isolation. As a result, improved approaches and instruments are necessary to make information credible. The practical usefulness and relative performance of a variety of classification algorithms are evaluated with respect to different types of large datasets. Different criteria are used to evaluate algorithms, including accuracy, precision, speedup, scale-up, and data scalability. The study indicates that there are no universally best classification tools; a better tool will depend on the size and characteristics of the dataset. This work presents a comparative study of multiple distributed classification algorithms analyzed using the Apache Spark Data Analytics platform for various categories of datasets. Keywords Big data classification · Spark · Directed acyclic graphs · Fuzzy cognitive map

K. Haritha (B) · M. V. Judy Department of Computer Applications, CUSAT, Ernakulam 682023, KL, India e-mail: [email protected] M. V. Judy e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_13

193

194

K. Haritha and M. V. Judy

1 Introduction Increasing volumes of data are being created at an alarming rate. And, deriving valuable information and meaning from such bulk volume of data is not a trivial task. A number of factors play into this, such as the size of the data, the heterogeneity of the data, the speed of generation, the inconsistencies and biases, the noise and abnormalities in data, and so on. In response to this, traditional relational databases and enterprise data warehouses have been challenged as they cannot easily handle this growing magnitude of data, and intricacy of the information makes analysis a challenge. The massive volume of the dataset makes it challenging to process it using an individual system. Either the resources of the system should be expanded until they are sufficient enough to accommodate large datasets. This is known as vertical scaling. As the size of the data grows, vertical scaling is economically infeasible, due to the high cost of loading more resources. Another alternative is horizontal scaling, wherein instead of expanding the resources of a single computer, multiple computers are connected together and the job is performed by utilizing the combined resources in all the connected machines. The massive dataset is distributed to chunks, and then, those chunks of data are passed onto individual machines for processing. The components are spread across multiple computers, but they function as one system. This process is known as distributed computing. For big data analysis to be effective, sophisticated tools, strategies, and environments are required. In the literature, there are a number of distributed data mining methods created particularly for large datasets, such as Apache Hadoop, Apache Spark, based on the MapReduce paradigm. Following the horizontal scaling technique, MapReduce uses a distributed computing approach. Single-threaded algorithms should be converted to parallel algorithms to apply Machine learning in distributed environments. Implementing parallel algorithms is the second phase. An understanding of system semantics and runtime is required to ascertain that parallel execution is executed in an accurate and efficient manner. We used Spark’s MLlib in our work, which is a machine learning library that is capable of performing large-scale data classification. This study seeks to investigate different classification algorithms based on their performance on various types of large-scale data. The considerable contributions of the proposed work are: – Analyzed the dependency of speedup and scalability of the algorithms in regard to the number of cores and the size of the dataset. – The impact of the number of attributes and the count of instances on the performance of classification algorithms was examined. – Directed Acyclic Graphs for each classification algorithm depict the resilience, and lazy evaluation process adopted was studied. The remaining parts of the paper are structured in the following manner. Section 2 provides insight on the current literature. Section 3 is dedicated to discussion of the distributed algorithms used in the article. Section 4 demonstrates experimental setup

Performance Analysis of Distributed Algorithms for Big Data …

195

and Sect. 5 gives the performance evaluation. The paper’s conclusion is presented in Sect. 6.

2 Literature Review A high performance, in memory computational engine that enables users to perform analytics on large-scale data is Apache Spark. One of Apache Spark framework’s core component is Resilient-Distributed Dataset (RDD) which enables implicit data parallelism and fault tolerance across clusters in a distributed environment [1]. Spark provides a machine learning platform known as MLlib that enables parallel machine learning [2]. Richter et al. made a comparison of various open-source machine learning tools for big data [3]. The algorithms were compared in terms of scalability and speed. H20 and MLlib scored the best, according to the authors. A survey of open-source machine learning technologies in the Hadoop environment was also undertaken by researchers Landset et al. and arrived at the conclusion that H20 and MLlib perform best speed-wise, applicability, number of algorithms included, and ability to scale to different dataset sizes [4]. Several studies have been conducted evaluating the performance of different machine learning algorithms in Spark-based environment. Mo Hai et al. compared the performance of Naïve Bayes and Random Forest algorithms using four metrics, classification accuracy, speedup, scale-up, and size-up [5]. Jesus M. et al. present a new Spark-based approach for performing a k-nearest neighbor classification. In memory operations are utilized to classify the dataset [6]. Categorization of text documents based on Apache Spark-distributed computing architecture was accomplished by Semberecki and Maciejewski [7]. Vettriselvi et al. evaluated the performance of RDD-based algorithms of Decision Tree, Random Forest, Logistic Regression, gradient-boosted tree (GBT), Linear regression Ridge and Lasso Regression models [8]. The results were compared in both Spark and Hadoop frameworks, and the results depict the clear dominance of spark in terms of performance. In [9], the authors evaluated the performance of various common big data models using Spark and Weka. The results indicated that Spark is able obtain faster results as compared to Weka. Seema Purohit and Neelam Naik conducted a research analyzing and contrasting the various binary classification techniques such as Decision Tree, GB tree, and Random Forest [10]. The authors arrived at a conclusion that Random Forest algorithm performs the best for the dataset under consideration. Using large datasets, this study aims to evaluate the performance of several machine learning algorithms. All of these enormous datasets are distinct in size and composition. In the experiment, Apache Spark technology is utilized. Accuracy, precision, speedup, scale-up, and data scalability are considered as the evaluation metrics to compare the algorithms.

196

K. Haritha and M. V. Judy

3 Distributed Classification Algorithms for Big Data Classification methods are capable of analyzing vast data and predicting the data class. The probabilities can be utilized as input into Naïve Bayes classification, which relies on statistical techniques. Classification model such as a Decision Tree, that uses machine learning techniques, is focused on computational operations. A network of interconnected nodes is used in the case of a neural network-based classification task. Nonlinear function of inputs is executed at each node to produce the output. This research examines the Apache Spark-supported distributed classification methods Logical Regression, Naïve Bayes, Random Forest, Multilayer Perceptron, Decision Tree, and Fuzzy Cognitive Map. In Spark, when a job is submitted to the master node, a DAG is built from the RDD’s lineage graph. There are various stages in a DAG. Both the shuffle map and result stages are separated by this step. It is important to understand the difference between shuffle map stages and outcome stages, because the results of one stage might be used as input for another stage. The term “Naive-Bayes Classifier” [11] refers to a set of Bayes theorem-based basic probabilistic classifiers. Based on assumption-based prior probabilities, the Bayesian theorem calculates the posterior probability of different types of data objects based on given assumptions. Bayesian algorithms are simple and widely used classification methods grounded in mathematical theory. A MapReduce paradigm-based implementation of Naïve Bayes is employed in this paper. The training samples are scanned from the source file in HDFS and mapped as key–value pairs of class labels and feature vectors and are stored as RDDs. Another map operation maps each feature to its corresponding class label. This mapped value using a hash function is grouped based on feature values and is used to determine the count of attributes with respect to classifications and find the prior probability of each sample. The reduce function finds the count and determines the conditional probability of each feature and produces the probability table. This probability table is used to find the likelihood of each class to create the Naïve Bayes Classifier (Fig. 1). The DAG depicting the distributed processing of Naive Bayes Classifier is specified in Fig. 2.

Fig. 1 Parallel Naive Bayes classification using RDDs

Performance Analysis of Distributed Algorithms for Big Data …

197

Fig. 2 DAG depicting Naive Bayes Classifier

Decision Tree Classifiers are widely used tree-based classifiers that predict unknown class labels. To build a tree, input characteristics are grouped into categories based on the information gain measure and iterate until every data point has a class. In Decision Tree Classifier, binary partitioning is done recursively to classify the features using a greedy algorithm. It takes a map of feature vector and class label as input from the source file stored in HDFS. Additionally, the impurity measure, max depth of the tree, max number of bins, etc. can also be passed on as inputs. The whole training set is considered as RDDs. The map of features and their labels is passed to the reduce function to get the count of each feature which is needed to get the gain ratio of each feature. The attribute with the maximum information gain is set as the root of the tree and in the next iteration remove the root node from the set of attributes and find the node with maximum information gain value. The information thus obtained is used to identify the attribute to be split to form the node. Nodes are recursively spilt and the tree is built until either all the attributes have been exhausted or the maximum entropy is attained. DAG generated for distributed Decision Tree Classifier is depicted in Fig. 3. Random Forest is a classifier model using ensemble learning. Random Forest generates an ensemble of Decision Trees. The dataset is partitioned into different subsets and spread across different nodes. Each tree is trained using a different subset of the data. The Random Forest trees are actually trained on different parts of the same training set. A TreePoint structure is used to save the memory by storing

198

K. Haritha and M. V. Judy

Fig. 3 DAG depicting Decision Tree Classifier

the replica count of each instance in each subset. The number of mappers created is the same as count of Random Forest trees. Parallel training of a variable set of trees is optimized depending on memory constraints. Random Forest models reduce the risk of overfitting. The DAG depicting Random Forest Classifier is given in Fig. 4. The regression approach that uses several independent variables to predict dependent variables is known as Logistic Regression. Multinomial logistic regresion is the method of choice since many datasets contain more than two distinct categories. When using multinomial Logistic Regression, the method generates an X*Y matrix, or X sets of coefficients, for the number of outcome classes and Y sets of features. A length X vector of intercepts is accessible if the algorithm is fitted with an intercept term. The softmax function is used to model the outcome classes’ conditional probabilities. The model produces the intercept vector and the coefficient matrix as output.

Performance Analysis of Distributed Algorithms for Big Data … Fig. 4 DAG depicting Random Forest Classifier

199

200

K. Haritha and M. V. Judy

Multilayer Perceptron classifier (MLPC) is a feedforward neural network classifier. MLPC has layered nodes. Network layers are completely linked. Input layer nodes represent input data. All other nodes translate inputs to outputs by linearly combining inputs with weights, bias, and an activation function. The model takes the composition of layers as input. According to Fuzzy Cognitive Maps, a system appears akin to human perception of things. The nodes of the FCM represent concepts of the system, and connections between the nodes represent causal relationships. FCMs portray the system using causal weighted diagraphs [12]. The features of a Fuzzy Cognitive Map are utilized in the determination of an initial state vector. On the state vector, FCM learning is applied until it converges to deliver the predicted outcome. A DenseMatrix data structure is used to represent the weight matrix, and a one-dimensional vector is used to represent the state vector. The weight matrix is distributed across the cluster and the state vector is broadcasted to each node containing weight matrix. The FCM learning is thus performed in parallel at each node, and the reduce function cumulates the result and produces the classifier (Fig. 5). The DAG for FCM-distributed processing is given in Fig. 6.

Fig. 5 Parallel fuzzy cognitive map classifier

Performance Analysis of Distributed Algorithms for Big Data …

201

Fig. 6 DAG depicting FCM Classifier

4 Experimental Setup An open-source framework that facilitates massive data analytics is Apache Spark. Faster alternatives than Hadoop MapReduce are provided by the Apache Spark framework, which is notably useful for machine learning algorithms. With Spark, you can perform a wide variety of tasks, including data analysis, machine learning, streaming data, database management, parallel computation, graph operations, etc. Java, Scala, Python, and R are all supported by Apache Spark. A large amount of data can be processed by Spark, which can be analyzed using advanced machine learning algorithms. Spark is capable of analyzing and applying complex machine learning algorithms to vast volumes of data. Using the Apache Spark framework, we construct a machine learning pipeline to manage the benchmark datasets in this model. The experiment was performed on a high-performance Hadoop cluster with one Name node server and two Data node servers which have a combined capacity of 768 GB RAM and 144 core processor. The cluster supports Hortonworks Data Platform, HDP 3.0. The software used is Apache Spark 2.3.0. In our model, we process and analyze the benchmark datasets using six machine learning classification algorithms (i) Decision Trees (DT), (ii) Random Forests, (iii) Naïve Bayes (NB), (iv) Multilayer Perceptron, (v) Fuzzy Cognitive Map, and (v) Multinomial Logistic Regression.

5 Performance Evaluation and Discussion In this study, benchmark datasets were taken from University of California at Irvine (UCI) repository as given in Table 1. The performance of the selected classifiers Logistic Regression, Naive Bayes, Random Forest, Decision Tree, Multilayer Perceptron, and Fuzzy Cognitive Map is evaluated based on different performance metrics mainly accuracy and precision: – Accuracy: The accuracy metric is described as follows:

202 Table 1 Details of benchmark datasets used

K. Haritha and M. V. Judy Dataset

Features 5

Iris Covtype2

54

Instances

Size

150

4.4 KB

581,012

10.7 MB

Epsilon

2000

500,000

3.6 GB

Colon cancer

2000

62

595 KB

48,842

3.8 MB

14

Adult HIGGS

28

11,000,000

5.74 GB

HEPMAS

28

10,500,000

4.82 GB

SUSY

18

5,000,000

2.23 GB

A=

TP + TN . TP + TN + FP + FN

(1)

A true positive (TP), a true negative (TN), a false positive (FP) and a false negative (FN) are all represented by the letters TP, TN, FP, FN respectively. In order to evaluate classifier performance, accuracy is used as a metric, since accuracy is a good indicator of classifier performance when the dataset is evenly distributed. Accuracy measure depicts the closeness of a result obtained to the actual or expected value. Figure 7 depicts the comparison of accuracy values obtained for the different benchmark datasets considered for each classification algorithm taken into consideration. From the figure, it can be deduced that for datasets with fewer instances, Random Forest, Decision Tree, and Naïve Bayes algorithms tend to perform the best, but as the number of instances increases, Multilayer Perceptron algorithm and multinomial Logistic Regression algorithm produce the best results. When number of features are taken into consideration, as depicted in Fig. 8, Naïve Bayes, Random Forest, Decision Tree perform the best when number of features

Fig. 7 Accuracy measures of the benchmark datasets with respect to number of instances in dataset

Performance Analysis of Distributed Algorithms for Big Data …

203

Fig. 8 Accuracy measures of the benchmark datasets with respect to number of features in dataset

are less, but as the number of features increases, all the classification algorithms considered produce almost similar accuracy values. – Precision: The precision metric is described as follows: P=

TP TP + FP

(2)

where TP represents a true positive and FP a false positive. Precision represents the reproducibility of a given result. Figure 9 depicts the variation in accuracy values as a function of the dataset size. It can be observed that when the number of instances is less, Naïve Bayes, MLP, and Logistic Regression give the maximum precision values, whereas when the number of instances increases, Decision Tree and Logistic Regression give the best results. Similarly, when the number of features of the dataset is taken into consideration, Fig. 10, when number of features are less, MLP and Naïve Bayes algorithms give the best results, and when the count of features is large, Decision Tree and Logistic Regression give the best results. To analyze the performance of Spark, three aspects are taken into consideration: 1. Analysis of System Speedup An improvement in a parallel algorithm’s speed over its serial counterpart is called speedup. It is an important way to assess how efficiently parallel processing works and how much of an impact parallelization has. Assuming the serial algorithm’s (i.e., single node) duration is Ts, and the parallel algorithm’s (i.e., multiple nodes) duration is Tp, the speedup would be Sp = Ts/Tp. The higher the increase in speed, the better the parallel efficiency and performance. We take into consideration two of the largest datasets analyzed, HEPMSS dataset and HIGGS dataset. As depicted in Figs. 11 and 12, we can see a considerable gain in speedup as the number of cores is increased. This is because as the number of cores increases, so does the utilization of parallel processing in the algorithms. It is evident that the time required by Decision Tree in the two modes varies greatly. The gap between the periods that

204

K. Haritha and M. V. Judy

Fig. 9 Precision measures of the benchmark datasets with respect to number of instances in dataset

Fig. 10 Precision measures of the benchmark datasets in concern with the count of features in dataset

it takes different modes to complete Logistic Regression gets increasingly clear as the number of nodes grows. This trend is less distinct in the Random Forest method, as the time spent in different modes decreases as the number of nodes increases. In all modes, the Naive Bayes method takes roughly the same amount of time. 2. Analysis of system scalability Scalability refers to a method’s ability to boost performance as the number of slaves increases. When using a parallel approach, it shows the cluster’s consumption rate. Scalability equation: J = Sp/p; Speedup is represented by Sp, and slave number by p. J is a positive integer equal to one or less. As it approaches one, scalability

Performance Analysis of Distributed Algorithms for Big Data …

205

Fig. 11 Speedup of training time of Hepmass dataset

Fig. 12 Speedup of training time of HIGGS dataset

improves. The scalability curve of a parallel program demonstrates a decreasing tendency as the number of slaves increases. Figure 13 represents the scalability trend of different distributed algorithms taken into consideration for HIGGS dataset in Spark cluster. Scalability for the distributed algorithms taken into consideration appears to be stabilizing with the increase in cores and dataset size, Fig. 14. 3. Analysis of data scalability Training time varies widely depending on the size of the dataset. Data scalability refers to this variation. From Fig. 15, it can be observed that Random Forest algorithm produces best scalability as its training time increases linearly with the size of the dataset.

206

K. Haritha and M. V. Judy

Fig. 13 Scalability trend of distributed algorithms for HIGGS dataset

Fig. 14 Scalability trend of Random Forest algorithm with respect to size of dataset

Fig. 15 Analysis of data scalability

Performance Analysis of Distributed Algorithms for Big Data …

207

6 Conclusion In this study, we make an effort to evaluate and contrast the effectiveness of six different classification algorithms. These include the Naive Bayes algorithm, the Decision Tree algorithm, the Random Forest algorithm, the Logistic Regression algorithm, the Multilayer Perceptron algorithm, and the Fuzzy Cognitive Map algorithm. We used the Apache Spark machine learning framework for this comparison. The study concluded that as the number of instances in the dataset grows, the Multilayer Perceptron method, Fuzzy Cognitive Map, and multinomial Logistic Regression produce the best accuracy scores. When the number of instances increases, Decision Tree and Logistic Regression give the best precision values, and when the number of features is high, Decision Tree and Logistic Regression give the best precision. Random Forest algorithm produces the fastest results when datasets with large number of features are considered, and when datasets have large number of instances, Decision Tree and Naïve Bayes algorithms produce the fastest results. There is a considerable speedup in the execution times of the machine learning algorithms when the number of cores increases. Scalability of the algorithms tends to stabilize as the number of cores and the size of the dataset increase.

References 1. Zaharia M, Apache spark: a unified engine for big data processing. Commun ACM 59(11):56– 65 2. Meng X, Mllib: machine learning in apache spark. J Mach Learn Res 17:1–7 3. Richter AN, Khoshgoftaar T, Landset S, Hasanin T (2015) A multi-dimensional comparison of toolkits for machine learning with big data. In: 2015 IEEE international conference on information reuse and integration 4. Landset S, Khoshgoftaar T, Richter A, Hasanin T, A survey of open source tools for machine learning with big data in the hadoop ecosystem. J Big Data 2(1):1–36 5. Hai M, Zhang Y, Zhang Y, A performance evaluation of classification algorithms for big data. Procedia Comput Sci 122:1100–1107 6. Maillo J, Ramírez S, Triguero I, Herrera F, knn-is: an iterative spark-based design of the k-nearest neighbours classifier for big data. Knowl-Based Syst 117:3–15 7. Semberecki P, Maciejewski H (2016) Platform, distributed classification of text docu- ments on apache spark. In: ICAISC 2016: artificial intelligence and soft computing. Lecture notes in computer science book series 8. Vettriselvi A, Dinadayalan P, Sutha S, A comparative study of machine learning algorithms using rdd based regression and classification methods. Ann Rom Soc Cell Biol 25(4):4249– 4259 9. Assefi M, Behravesh E, Liu G, Tafti A (2017) Big data machine learning using apache spark mllib. In: 2017 IEEE international conference on big data (Big Data) 10. Naik N, Purohit S, Comparative study of binary classification methods to ana- lyze a massive dataset on virtual machine. Procedia Comput Sci 112:1863–1870 11. Hand D, Yu K, Idiot’s bayes—not so stupid after all? Int Stat Rev 69(3):385–398 12. Kosko B, Cognitive fuzzy maps

IoT

Achieving Sustainability by Rectifying Challenges in IoT-Based Smart Cities Neha Bhardwaj, Celestine Iwendi, Thaier Hamid, and Anchal Garg

Abstract It has become mandatory for traditional cities to be converted into smart cities for the benefit of giving citizens a better lifestyle. The large amounts of people migrating toward smart cities have increased the requirements of a sustainable smart city with the assistance of cloud-based Internet of Things (IoT) devices such as active sensors and smartphones. When implementing these cloud applications that are IoT-based, there are several challenges including security and privacy of data in the devices and satisfaction with the smart services provided. Previous research done revealed that the primary challenge was privacy and security, while current research emphasizes on the security of the smart devices and the awareness and satisfaction of customers with them. In this paper, we used a qualitative and quantitative approach to attain sustainability. Under the qualitative analysis, we proposed a five-layer smart city framework based on IoT, making use of various technologies such as fog computing, Artificial Intelligence, edge computing, Machine Learning, actuators, cloud computing, Deep Learning, sensors. In the quantitative analysis, we carried out a smart city survey to gather knowledge about the citizens’ level of satisfaction and awareness. The study was carried out on a minimal scale because of limited resources and time; however, it still provides a possible solution to the issues encountered. Keywords Smart cities · Cloud computing · Internet of Things (IoT) · Smart services · Security · Privacy · Cyberattacks · Sustainability

N. Bhardwaj (B) · C. Iwendi · T. Hamid · A. Garg School of Creative Technologies, University of Bolton, Bolton, UK e-mail: [email protected] C. Iwendi e-mail: [email protected] T. Hamid e-mail: [email protected] A. Garg e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_14

211

212

N. Bhardwaj et al.

1 Introduction A smart city is a technologically advanced urban region which makes use of physical and digital devices for the collection and storage of data. Internet of Things (IoT) devices collect data, store them over the cloud, and ensure their delivery whenever they are required [1]. This data collected is used to manage revenue, assets, and resources. It is also used to enhance performance across the city. It is estimated that by year 2030, over 60% of people would rather live in a smart city [2]. The smart city concept was proposed to increase the citizens’ quality of life through the provision of highly digitalized and interactive services using Information and Communication Technologies (ICTs) [3]. The “smartness” is the ability to collectively and effectively make use of the available resources to achieve the set goals. Smartness is applied in every sector including health care, homes and industries, transportation, agriculture, etc. It is an importance for citizens to be made aware of novel technologies and how to use them [4]. Cloud computing delivers computer resources on demand including servers, memory, database, networking, applications, and software with a pay-as-you-go system. The mental and financial burdens on clients are reduced because they do not need to pay for the setup or for the upkeep of the software and hardware devices [5]. Several cloud providers offer cloud services including Infrastructure as a Service (IaaS), Software as a Service (SaaS), and Platform as a Service (PaaS). They also offer several models for cloud deployment including Private, Public, Hybrid, and Community clouds [6] (Fig. 1). Fig. 1 Smart city components [4]

Achieving Sustainability by Rectifying Challenges in IoT-Based Smart …

213

IoT is an advanced and modern technology that connects persons-to-persons (P2P), machine-to-machine (M2M), and persons-to-machine (P2M) at any place or time. IoT has affected every part of everyday life through innovations of smart devices in the environment, health care, agriculture, automation, transportation, etc. It has been predicted that IoT has the ability to generate a revenue of over ₤714 billion, while others believe that it could go up to $6.2 trillion by 2025 [7]. IoT comprises all physical objects including software, sensors, and other wireless gadgets that are capable of connecting different devices to collect and exchange data when it is required. Cloud computing based on IoT helps the attainment goals of sustainability in a smart city [8]. Although cloud computing based on IoT is rapidly developing in every sector, some people are still not familiar with or aware of the smart services [9]. The network of a smart city generates large amounts of data that should be properly managed in order to prevent cyberattacks or risks. In the extraction, organization, storage, and retrieval of data processes collected by IoT devices, low latency and increased security are the main issues [10]. In this paper, we provide an in-depth study of IoT, cloud computing, and smart cities. We also examine the challenges mentioned above and propose solutions to them. The following are the main contributions of this research: • We examine the components of the cloud computing environment. • We explore the various IoT technologies based on their application. • We study how implementing IoT-based cloud computing can affect privacy and security. • We analyze the means by which the challenges can be resolved. We also propose an effective framework for the security of the smart city. • We evaluate the satisfaction and awareness of citizens in a smart city of government-offered smart city services. The rest of this paper follows this structure: Sect. 2 reviews the related work. Section 3 presents the structure of smart city based on IoT. Section 4 explores the challenges faced in the implementation of smart services through a survey. Section 5 provides the results and discusses solutions to the challenges. In Sect. 6, we conclude the research and briefly discuss the future work.

2 Literature Analysis In this section, we highlight previous research done on cloud computing, IoT, and smart cities. The table below presents a brief review of research done on smart cities with IoT-based cloud computing. It examines the perspectives of several researchers, starting with cloud computing and ending with smart city (Table 1). A report issued in [16] estimated the market growth of smart cities in the next few years to be 837.20 billion GBP in 2021. It is expected to go up to 5845.03 billion GBP by 2030. Other researchers have proposed different kinds of security

214

N. Bhardwaj et al.

Table 1 Analysis of previous research Citation

Research focus

Research contribution

“A survey paper on cloud The research focuses on the security issues and challenges” concept and services of cloud [11] computing. It introduced the cloud services such as SaaS, PaaS, and IaaS with its security architecture

The information extracted from this paper is used in the introduction of the current thesis as the fundamental source of knowledge regarding cloud structure

“Building smart cities applications using IoT and cloud-based architectures” [3]

The study discusses the need of a smart city. It defined IoT and cloud computing as the best solution for smart cities

The current thesis is all about smart cities, so this paper guides the requirements and advantages of making a city smarter

“Internet of Cloud: Security and Privacy Issues” [5]

This study explored the Internet of Cloud (IoC) technology and the challenges in security it faces. It also examined sensor technology

This paper provides an insight into IoC and IoT and discusses sensor technologies used in the implementation of IoT technologies

“Security and Privacy Issues in This paper studied cloud, edge, This paper provides the and fog computing and their concept of decentralized Cloud, Fog and Edge security challenges cloud functioning used in the Computing” [12] introduction section “An overview of security and privacy in smart cities’ IoT communications” [2]

This paper studies how cyberattacks affect information and how to prevent them using blockchain, Machine Learning (ML), Artificial Intelligence (AI), and Deep Learning (DL)

This paper’s explanation of cyberattacks and how to prevent them are used while discussing security challenges in the literature review

“Security Challenges and Strategies for the IoT in Cloud Computing” [13]

The study supports the Confidentiality, Integrity, and Availability (CIA) trio and examines internal data security threats using privacy policy, user access control list, and a datagram transport layer security framework

The CIA triad in this paper played an important role in achieving privacy and security in the current thesis, as well as the methodology

“IoT and Cloud Computing The paper explores the Issues, Challenges and migrating techniques from IoT Opportunities: A Review” [14] to IoT solutions that are cloud-based making use of load balancing, fog computing, and Mobile Edge Computing (MEC) offloading methods for migration of virtual machines

The techniques in this article help in the management of virtual machines, as well as the issues considered in the methodology of the current paper

(continued)

Achieving Sustainability by Rectifying Challenges in IoT-Based Smart …

215

Table 1 (continued) Citation

Research focus

“IoT in Smart Cities: A Survey This research focuses on the of Technologies, Practices and framework of a smart city with Challenges” [15] IoT-based cloud computing using sensors, application programming interface, network topologies, Artificial Intelligence, Deep Learning, and Machine Learning

Research contribution The five-layer framework in the current paper is theorized from this paper, with the addition of cloud, edge, and fog computing layers to tackle challenges such as security, latency, and privacy

and privacy mechanism for the Internet of Things using Artificial Intelligence, Deep Learning, and cloud computing [17–20]. After analyzing the research done above, we can conclude that the researchers mostly focused on the smart cities’ privacy and security issues, while the current research focuses on attaining the highest level of citizen satisfaction through the maintenance of sustainability in smart cities.

3 Methodology Research methodology is the phase of research where full focus is on the methods or way to carry out the research. These can be of any type depending on the area you are doing research. The most frequent categorization of methodology is on the basis of nature of information of the research. This can be classified as qualitative research or quantitative research depending on the qualitative or quantitative data used for research, respectively. Although the phenomena of smart city have been broadly explained and studied in recent years, smart city services need a better methodology to connect the world and provide the society a better lifestyle [21]. In this research, we used a combination of quantitative and qualitative approaches. The qualitative approach is based on the secondary data gathered from the previous research done, while the quantitative approach used is based on the collected primary data from a survey.

4 Implementation of IoT-Based Smart City In the qualitative approach, a novel five-layer framework of smart cities making use of IoT, ML, AI, cloud, fog, and edge computing and other technologies is proposed [15]. The proposed architecture introduced altered layers for transfer of data with extra layers of fog, edge, and cloud computing for attaining low latency and increased security missing in the previous studies (Fig. 2).

216

N. Bhardwaj et al.

Fig. 2 Proposed five-layer architecture for an IoT-based smart city

4.1 IoT Layer (Layer 1) Also referred to as the perception layer, it consists of sensors, mobile elements and actuators. The sensors collect data from the real world through physical objects like the RFID readers and tags. The actuators deliver the information from the system database to the real-world environment. Edge computing provides support to the sensing layer as it operates on the edges (sensor nodes) (Figs. 3 and 4). Different kinds of IoT technologies are used in IoT-based smart cities. They are the building blocks of IoT that help in the identification, storage, processing, and display of data. They include [17]:

Achieving Sustainability by Rectifying Challenges in IoT-Based Smart …

217

Fig. 3 How sensors operate

Fig. 4 How actuators operate

• Radio Frequency Identification (RFID): They are divided into two: active and passive. In active RFID chips, when there is a change in the environment like a change in the temperature, the response will be activated. In passive RFID, no such battery is placed, and information is provided when it is located closely to the chip. • Wireless Sensor Networks (WSNs): These are standalone devices which can store and process small piles of data. It comprises one or more sensors which stores limited data and can interact with a base station. • Middleware: It acts as a middleman between the applications and the developers who analyze them. It enables developers work with sensors through the provision of a level of abstraction absent of the need to know the specifics of their implementation. • Applications of IoT: This is the application software which provides end users with the facility to do any work smartly. For instance, a smartphine app helps drivers to find and reserve free parking space in a parking lot.

4.2 Transport Layer (Layer 2) The data collected in the perception layer is then sent to the transport layer and processing layer via wireless network technologies and topologies as discussed below: Network technologies exchange and manage the computer resources for the fulfillment of a user’s request over a network. They include Wireless Area Networks (WANs) like Wireless Fidelity (Wi-Fi), cellular internet, and Bluetooth.

218

N. Bhardwaj et al.

Network topologies interconnect all logical devices and physical nodes’ devices over a communication network. They include bus, ring, star, hybrid, and mesh topologies. There are several fog nodes that allow the connection and sharing of data, so fog computing supports the network layer.

4.3 Processing Layer (Layer 3) This layer provides a general interface for the IoT layer hardware to the application layer through an Application Programming Interface (API). It also possesses several database managements’ services for the management of data and to make them available when the user requests it.

4.4 Application Layer (Layer 4) This is referred to as an abstraction layer which enables the introduction of the communication protocols to provide the communication method of computer systems over the network.

4.5 Business Layer (Layer 5) This layer is linked to the application layer and is utilized for the development of new policies and strategies which offer users the best services. It comprises the following: Artificial Intelligence (AI): This technology is utilized to train a computer to smartly perform a task that usually needs human intelligence. It makes decisions on the data based on its intelligence. As presented below (see Fig. 5), Machine Learning and Deep Learning are the subsets of Artificial Learning. Machine Learning (ML): It allows the possibility for a computer to provide a response to a request for data without being explicitly and clearly programmed using an algorithm. It is mostly used to examine data based on their patterns matching. Deep Learning (DL): DL is where the main artificial neural network is built. It develops the algorithm which AI and ML use for selection of the data. IoT services required decentralized computing. Thus, our architecture supports cloud computing with edge and fog computing (see Fig. 6) to tackle issues including bandwidth and resource limitation and high latency.

Achieving Sustainability by Rectifying Challenges in IoT-Based Smart …

219

Fig. 5 Dependencies of AI, ML, and DL

Fig. 6 How the fog layer operates

5 Challenges of Smart Cities and Their Solutions The figures below illustrate the challenges faced by smart cities in different stages of implementation. They have different levels of complications which require that effective and efficient actions can be taken to tackle them [22–25].

220

N. Bhardwaj et al.

5.1 Making People Aware of Smart City Services The most important challenge is making the citizens aware of smart city services, their necessity, and usage. Smart cities need to make their citizens smart as well, by familiarizing them with the cutting-edge applications provided by the government to enhance the quality of life. The survey was conducted among the citizens of the UK to know the awareness of individuals in different smart sectors and how they feel about facilities provided by the government in different areas. This survey was conducted on a minimal scale as a result of limited time. The survey stakeholders should be the public using the smart services, the businesses that create them, the researchers and organizations conducting research on them, and the governments that develops smart city policies on a large scale (Table 2). Table 2 Feedback received from survey Survey questions

Scaling feedback (1 to 5) 1

2

3

4

5

1. Are you aware with the Smart City concept?

1

13

16

13

2

2. How often you use Smart City services?

3

9

18

10

5

3. Are you satisfied with the Smart Transport services offered by the government?

2

6

16

16

5

4. How is the Smart Health services in your city?

2

6

24

11

2

5. How committed is your city to the Smart Homes and Industries concept?

4

6

21

11

3

6. Is your city taking initiatives in utilizing energy efficiently under 1 Smart Energy

4

16

20

4

7. What is the quality of Smart Agriculture services provided by the government in your

1

3

20

16

5

8. In your opinion, is it important to evaluate the performance and progress of a Smart City

2

3

16

18

6

9. Do you think, it would be relevant for a Smart City to know its rank out of several

2

5

6

22

10

10. Do you think, it insecure to use smart services in all the sectors 0 as they fetch your

4

22

16

3

11. How likely is it that the information saved in smart devices can 2 be hacked by

6

19

15

3

12. “I always make sure that my personal information is secured while using Smart

1

5

10

18

11

13. “Introducing Smart City concept increases the cost of living and hence increases the

3

3

16

16

7

14. Do you find it safe to link up different sectors such as healthcare, industrial and

0

6

12

22

5

15. Will you approve the idea of taking feedback from the citizens of a Smart City

0

1

12

17

15

Achieving Sustainability by Rectifying Challenges in IoT-Based Smart …

221

The above survey is done over the citizens of the UK. There are questions based on the awareness, type of services offered, satisfaction level to measure the qualityof-service level of the smart city. The overall response will be noted to recommend the changes required. To calculate Likert Scale result presented in appendix B of the survey recorded, the following method was used with the uppercase Greek letter sigma (S) which means: N Σ

means sum up, or add up

i

Strongly disagree

Disagree

Neutral

Agree

Strongly agree

1

13

16

13

2

1

2

3

4

5

N Σ

X i = 1 + 26 + 48 + 52 + 10 = 137

i N Σ i

Xi =

N Σ

X

i

Total responses = 45. Average response = 137 ÷ 45 = 3.04. As average response is nearby 3, so the average answer is neutral (Figs. 7 and 8). As seen above, only 2% of citizens are not aware of the smart city services. Ninety-three % of people are aware of and use the smart services provided by the government. Only 7% have never used the smart city services (Figs. 9, 10, 11, 12 and 13). From the survey above, it can be observed that people are almost completely satisfied with the services the government provides to make their city smarter. It can thus be concluded that the government is making all possible and considerable efforts to attain the end users’ maximum satisfaction. The only challenge is the citizens’ awareness and proper utilization of the smart city applications and services. To familiarize people with the services in the smart city, awareness programs or trainings can be organized on “How to use smart services”. The services with below par performance as per the need of the citizens can be altered or improved through the necessary measures.

222

N. Bhardwaj et al.

Fig. 7 Awareness of citizens of the smart city

Fig. 8 Usage of smart city services by citizens

5.2 Cloud, Edge, and Fog Computing Attacks and Their Prevention Techniques Cloud computing: This is well-known for its storage of big data, utilization of resources, and elasticity. The customer is allocated resources whenever they require them. Big data storage, on the other hand, stores large amounts of data. Elasticity represents the scaling up and down as per the customer’s need. Despite these benefits, cloud computing still has some challenges including [12]:

Achieving Sustainability by Rectifying Challenges in IoT-Based Smart …

223

Fig. 9 Satisfaction of the citizens with smart transport services

Fig. 10 Opinions of citizens on standards of smart healthcare services

Phishing attack: This is a common cybersecurity attack used to steal sensitive and private information such as password or user ID. It is mostly done via mail, where the user might be inclined to click the link because he/she believes that it is sent by a trusted entity or person. Prevention techniques include anti-phishing software and educating people on spam mails. Denial-of-Service (DoS) attack: Here, the unauthorized user transmits large requests to the network to obtain access from an unauthorized return address which prevents

224

N. Bhardwaj et al.

Fig. 11 Commitment of citizens to the smart homes’ concept

Fig. 12 Opinion of citizens on initiatives in the smart energy concept

authorized users from accessing the cloud services. Prevention techniques include anomaly detection monitors the network for any irregularities. Cloud malware injection: Here, the system is infected through the injection of a malevolent virtual machine or cloud service, by insisting that it implements it. It can be prevented by protection of the end points for the detection and halting of the malware before its entrance into the cloud.

Achieving Sustainability by Rectifying Challenges in IoT-Based Smart …

225

Fig. 13 Opinion of citizens on the quality of smart agriculture services

Communication interception: This attack has the intent of eavesdropping into users’ conversation and exploiting it. To prevent it, a user access list should be provided to allow only authorized users to participate in the conversation. Fog computing: It gives a solution to latency through the provision of real-time communication, heterogeneity, and by setting up a similar pattern for every fog node. It also provides interoperability through the offering of all services combined. However, it still has a few challenges as stated below: Denial-of-Service (DoS) attacks: It is identical to its occurrence in cloud computing. Virtual Machine-based attacks: It is identical to the injection of malware in cloud computing. Session hijacking: Here, an unauthorized user hacks other users to obtain their user data, which he can use to collect information. To prevent this, use strong passwords and IDs with multilevel authentication. Edge Computing: Edge computing operates on the edges where smart gadgets like actuators and sensors generate the data by themselves and offers real-time processing of data absent of latency. The delicate data to be sent to the cloud for safe storage is filtered because they are vulnerable to cyberattacks. Some of its challenges include: Eavesdropping: Here, the malicious attack hides himself and observes the network to steal user data. To prevent this attack, it is important to follow authentication protocols before any communication. Denial-of-Service (DoS) attacks: It is identical to its occurrence in cloud computing.

226

N. Bhardwaj et al.

Data tampering attack: The hacker can tamper with the data sent over the cloud storage during the users’ conversation in order to use, alter, or delete it. It can be prevented by using a Copy-On-Write (COW) system to monitor every attempt in the database.

5.3 Providing Privacy and Security in the Smart City Network Attacks: IoT-based smart cities are also susceptible to network attacks such as those explained below [15]: Denial-of-Service (DoS) attack: Here, the unauthorized user transmits large requests to the network to obtain access from an unauthorized return address which prevents authorized users from accessing the cloud services. Prevention techniques include anomaly detection monitors the network for any irregularities. Eavesdropping: Here, the malicious attack hides himself and observes the network to steal user data. To prevent this attack, it is important to follow authentication protocols before any communication. Only users who can pass the authentication steps can participate in the conversation. Spoofing attack: This is a dangerous attack for smart city services because there are numerous IoT devices linked to the smart city network. Here, an attacker joins the network under the guise of a legitimate device. When it obtains access to receive and send data, it starts transmitting abnormal data to obstruct the regular operation of the smart city system. Prevention techniques include hybrid encryption, blockchain, and cryptography to validate the devices and legalize the exchange of data. Man in the Middle attack: Here, the network data is obstructed by a network node with a false identity. It operates as the middleman to collect the information sent from one end to the other end. To prevent this, all cryptographic protocols should be adhered to. Encrypted data is less vulnerable to attacks, and they can be decrypted through the use of a unique cryptography algorithm. Access Control: Different applications provide different smart city services that are used by different enterpises. For a smooth operation in smart cities, this data should be managed by an access control scheme, which would allow only legitimate users access. Leakage of Data: This refers to the accidental transfer of data which may have sensitive information such as contact details or passwords in smart cities. Application managers in smart cities use this information to improve their services and offer the users a better experience. The data could be shared by any attackers or third party.

Achieving Sustainability by Rectifying Challenges in IoT-Based Smart …

227

5.4 Sensor Performance IoT devices are vulnerable because of their inability to protect themselves from hackers. To tackle these issues, smart sensors that have high reliability, good data storage, and more robustness should be installed. Reliability here refers to the smart sensors’ consistency, i.e., if it can be fully depended on. Robustness ensures the full functioning of the devices despite issues in the system. Power consumption and memory are also challenges in IoT devices. Thus, during the installation of IoT devices, these factors should be taken note of to achieve security. There should also be proper maintenance and update of the devices.

5.5 Development of Novel Data Analytics As smart cities grow, so do their services. To monitor the data and ensure the provision of good services, it is necessary to create an algorithm which can be applied to varying types of data. Machine Learning and Deep Learning play a crucial role in this process, but it requires a high level of training, which presents a challenge in implementing today’s smart services.

6 Results In this study, we proposed a five-layer framework which resolves the challenges of smart cities through the use of modern technologies including fog, cloud, and edge computing for tackling security, latency, and privacy issues; AI, DL and ML for creating intelligent systems; and smart sensors for the improved operation of smart services. We carried out a Strength Weaknesses Opportunities Threats (SWOT) analysis to support the quantitative and qualitative analyses. Here, strength stands for the merits of the current research with the aid of technologies, while weaknesses are the limitations. Opportunities represent the research scope in the area of smart city services, while threats represent the challenges faced during the implementation of the IoT-based smart services. The table below illustrates the SWOT analysis (Table 3). Strengths and opportunities are the positive and helpful sides, while weaknesses and threats are the negative and harmful sides. Strengths and weaknesses relate internally to the project and determine the negative and positive aspects of the research itself. Opportunities and threats, on the other hand, are directly related to the external execution of the research.

228

N. Bhardwaj et al.

Table 3 SWOT analysis of the current research

7 Conclusion and Future Directions In this study, we presented an overview of the challenges and workings of IoT-based smart cities making use of cloud computing. To attain the goal of a sustainable smart city, all challenges faced need to be tackled and resolved. Thus, we proposed a five-layer framework to support the smart city services. This study also presents countermeasures to the challenges which include citizens’ lack of awareness, IoT device latency, and privacy and security deficiencies. We carried out a survey among the smart city citizens to monitor their level of awareness of and satisfaction with the smart services offered by the government.

Achieving Sustainability by Rectifying Challenges in IoT-Based Smart …

229

Future research can focus on the implementation of enhanced cloud-IoT technology in the development of a smart city. Future research can also work on training the system to deal with cyberattacks and altering the inappropriate services based on the citizens’ feedback. Also, a survey can be carried out on a large scale in different smart cities of a country, and their ranking would be evaluated based on their quality of service. The survey stakeholders should be the public, research organizations, businesses, and the government. This would increase the competition level and ultimately result in improved smart city services. Acknowledgements This project is supported by Dr. Celestine Iwendi, Dr. Thaier Hamid, and Dr. Anchal Garg, Department of School of Arts and Creative Technologies, University of Bolton, UK. A sincere thanks to all of you for your valuable feedback and continuous guidance throughout the study. Another thanks to the University of Bolton for providing resources to support the current research.

References 1. Alam T (2021) Cloud-Based IoT applications and their roles in smart cities. Smart Cities 4(3):1196–1219 2. Al-Turjman F, Zahmatkesh H, Shahroze R (2019) An overview of security and privacy in smart cities’ IoT communications. Trans Emerg Telecommun Technol 33(3) 3. Kaur M, Maheshwari P (2016) Building smart cities applications using IoT and cloud-based architectures. In: 2016 international conference on industrial informatics and computer systems (CIICS) 4. Maltezos E, Karagiannidis L, Dadoukis A, Petousakis K, Misichroni F, Ouzounoglou E, Gounaridis L, Gounaridis D, Kouloumentas C, Amditis A (2021) Public safety in smart cities under the edge computing concept. In: 2021 IEEE international mediterranean conference on communications and networking (MeditCom) 5. Cook A, Robinson M, Ferrag M, Maglaras L, He Y, Jones K, Janicke H (2018) Internet of cloud: security and privacy issues. Stud Big Data 271–301 6. Khoda Parast F, Sindhav C, Nikam S, Izadi Yekta H, Kent K, Hakak S (2022) Cloud computing security: a survey of service-based models. Comput Secur 114:102580 7. Gulzar M, Abbas G. Internet of Things Security: A Survey and Taxonomy. 2019 International Conference on Engineering and Emerging Technologies (ICEET). 2019. 8. Han H, Liu Z, Wang X, Li S (2021) Research of the relations among cloud computing, internet of things, big data, artificial intelligence, block chain and their application in maritime field. J Phys: Conf Ser 1927(1):012026 9. Tahirkheli A, Shiraz M, Hayat B, Idrees M, Sajid A, Ullah R, Ayub N, Kim K (2021) A survey on modern cloud computing security over smart city networks: threats, vulnerabilities, consequences, countermeasures, and challenges. Electronics 10(15):1811 10. Sahmim S, Gharsellaoui H (2017) Privacy and security in internet-based computing: cloud computing, internet of things, cloud of things: a review. Procedia Comput Sci 112:1516–1522 11. Saxena T, Chourey V (2014) A survey paper on cloud security issues and challenges. In: 2014 conference on IT in business, industry and government (CSIBIG) 12. Parikh S, Dave D, Patel R, Doshi N (2019) Security and privacy issues in cloud, fog and edge computing. Procedia Comput Sci 160:734–739 13. Mohiuddin I, Almogren A (2020) Security challenges and strategies for the IoT in cloud computing. In: 2020 11th international conference on information and communication systems (ICICS)

230

N. Bhardwaj et al.

14. Mohammed Sadeeq M, Abdulkareem N, Zeebaree S, Mikaeel Ahmed D, Saifullah Sami A, Zebari R (2021) IoT and cloud computing issues, challenges and opportunities: a review. Qubahan Acad J 1(2):1–7 15. Syed A, Sierra-Sosa D, Kumar A, Elmaghraby A (2021) IoT in smart cities: a survey of technologies practices and challenges. Smart Cities. 4(2):429–475 16. Smart Cities Market Size, Growth, Trends, Report (2022–2030) [Internet]. Precedenceresearch.com. 2022 [cited 12 September 2022]. https://www.precedenceresearch.com/smart-cit ies-market 17. Iwendi C, Wang G-G (2022) Combined power generation and electricity storage device using deep learning and internet of things technologies. Energy Rep 8:5016–5025. ISSN 2352–4847, Q1. SCI Impact factor 6.870 https://doi.org/10.1016/j.egyr.2022.02.304 18. Anajemba JH, Iwendi C, Razzak I, Ansere JA, Okpalaoguchi IM (2022) A countereavesdropping technique for optimized privacy of wireless industrial IoT Communications. IEEE Trans Ind Inf 18(9):6445–6454. https://doi.org/10.1109/TII.2021.3140109.(Q1).SCIImp actfactor10.215 19. Iwendi C et al (2021) Sustainable security for the internet of things using artificial intelligence architectures. ACM Trans Int Technol 21(Article No. 73):1–22. https://doi.org/10.1145/344 8614(Q1).SCIImpactfactor3.750 20. Shabbir M et al (2021) Enhancing security of health information using modular encryption standard in mobile cloud computing. IEEE Access 9:8820–8834. https://doi.org/10.1109/ACC ESS.2021.3049564.ImpactFactor3.745.Q1.ISIIndex.Supervisor 21. Valencia-Arias A, Urrego-Marín M, Bran-Piedrahita L (2021) A methodological model to evaluate smart city sustainability. Sustainability 13(20):11214 22. Gupta B, Madan G, Quadir AMd (2022) A smart agriculture framework for IoT based plant decay detection using smart croft algorithm. Mater Today: Proc 23. Sathish S, Ashwin S, Quadir MdA, Pavithra LK (2022) Analysis of convolutional neural networks on Indian food detection and estimation of calories. Mater Today: Proc 24. Quadir MdA, Prassanna J, Jackson JC, Sabireen H, Gupta G (2021) Efficient algorithm for CSP selection based on three-level architecture. Artif Intell Technol 515–531 25. Md AQ, Agrawal D, Mehta M, Sivaraman AK, Tee KF (2021) Time optimization of unmanned aerial vehicles using an augmented path. Fut Int 13(12):308

Framework for Implementation of Smart Driver Assistance System Using Augmented Reality K. Baskar, S. Muthuraj, S. Sangeetha, K. Vengatesan, D. Aishwarya, and P. S. Yuvaraj

Abstract This research is to investigate momentum innovation for constantly detecting street signs from a moving vehicle. The most encouraging innovation for perceptive vehicle frameworks is vision sensors and image preparation; therefore, this is examined the most thoroughly. Various handling calculations and study the world over concerned with sign acknowledgment are investigated. A functional framework has also been implemented using a regular web camera installed in a testing car. This framework is limited to speed signs and achieves outstanding displays due to rapid but hearty computations. The division is based on shading data, and the recognition is based on a model coordinating computation. The human–computer interface is a voice that announces which sign has been discovered. Driver carelessness is a key element that commonly results in distortion of surroundings, for example, erroneous recognition or disregard of street and traffic signs. Although significant progress has been made in developing a vehicle capable of autonomous guided driving, development has been slow because to concerns of speed, safety, and the ever-changing complexity of the on-road situation. Street plays a key role for the motorist in gathering data from the sign and then acting in a similar fashion. Keywords Augmented reality · Smart driver · Road accident · HCI · Street sign K. Baskar Department of Computer Science and Engineering, Kongunadu College of Engineering and Technology, Tholurpatti, India S. Muthuraj · P. S. Yuvaraj Department of Information Technology, Sanjivani College of Engineering, Kopargaon, India S. Sangeetha Kongunadu College of Engineering and Technology, Tholurpatti, India K. Vengatesan (B) Department of Computer Engineering, Sanjivani College of Engineering, Kopargaon, India e-mail: [email protected] D. Aishwarya Department of Electronics and Communication, College of Engineering and Technology, Kattankulathur, Chennai, India Department of AIDS, Panimalar Engineering College, Chennai 600123, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_15

231

232

K. Baskar et al.

1 Introduction With the improvement of the astute vehicle, safe driving help frameworks are turning out to be increasingly significant. In safe driving help framework, the traffic sign acknowledgment is a key innovation and has been broadly utilized. The precision and short preparing time are significant for traffic sign acknowledgment. Be that as it may, in genuine driving circumstances, the different circumstances of traffic signs including the pivot, perspective, scale, and enlightenment are unpredictable and bothersome. Accomplishing vigorous traffic sign acknowledgment with short handling occasions is a difficult undertaking. Traffic sign discovery incorporates traffic sign acknowledgment and traffic sign arrangement [1, 2]. So as to accomplish fast and hearty traffic sign identification, planning a processing effective and exceptionally discriminative element is fundamental. In the interim, so as to accomplish fast and hearty traffic sign arrangement, setting up a grouping procedure that can diminish the measure of highlights and continue order precision is likewise significant [3–5]. Current deals with the traffic sign location and acknowledgment can be separated into three unique sorts. To begin with, pre-handling strategies are inquired about to find and perceive the traffic signs. Second, pre-handling strategies consolidating with groupings are received to accomplish fast and strong traffic signs acknowledgment. Third, explicit structure highlights brushing with the classifiers are utilized to accomplish the vigorous and registering proficient acknowledgment [6–8]. This research is driven by the desire to devise a better system for driving where the smartness and efficacy of the driving may be improved to provide better outcomes [9–12]. This new system that uses state-of-the-art technology such as Augmented Reality (AR) and Embedded Systems will allow drivers to get a better understanding of the findings, making it easier for them to act on it by providing relevant alerts [13–15].

1.1 Background The Road Sign Recognition (RSR) is a field of applied PC vision look into worried about the programmed location and grouping of traffic signs in rush hour gridlock scene pictures procured from a moving vehicle. The aftereffect of the RSR look into exertion can be utilized as an emotionally supportive network for the driver. At the point when the neighbor condition is comprehended, PC backing can help the driver in pre-crash forecast and evasion [16–19]. Driving is an assignment put together for the most part with respect to visual data and picture handling. The street signs and traffic signals characterize a visual language translated by drivers. Street signs give numerous data essential to effective driving—they depict the present traffic circumstance, characterize option to proceed, disallow or license certain headings, caution about unsafe components, and so on. Street signs additionally help drivers with route (Fig. 1).

Framework for Implementation of Smart Driver Assistance System …

233

Fig. 1 Intelligent infrastructure and intelligent vehicles to warn about slippery road ahead

1.2 Motivation Out of the car, Smart Driver Assistance Systems are comprehensive frameworks that assist the driver in a variety of ways. These frameworks may be used to provide important data regarding agreements, conclusions, and clog clearance. • Frameworks may be used to determine whether the human driver has reached his or her limit, and if so, implement preliminary alerts or assess the driving. Larger frameworks may take away the human’s ability to evaluate danger (like surpassing and stopping). • The primary benefits of using the assistance framework are that they enable communication across different cars, vehicle frameworks, and transportation, allowing the executives to concentrate on their respective areas of responsibility. By making this available, valuable data for carmakers will be exchanged, allowing for improved vision, limitation, arrangement, and fundamental leadership of cars. • Advanced Driver Assistance Systems and self-driving vehicles depend on locating and recognizing traffic signs to function properly. To make this announcement, we offer a system that can see and recognize traffic signs as a motorist is looking at them. This method relies on the driver’s real-time 3D view gathered using a stereo imaging framework to see ahead and a non-contact 3D gaze tracker. • For identification, we applied a straightforward Support Vector Machine and made highlights via the Histogram of Situated Gradients. Scale-Invariant Feature Transforms and coloring data are used to conduct acknowledgment. The approach we have developed both recognizes and detects warning indicators the driver is facing (and identifies signs which may have been ignored by the driver).

234

K. Baskar et al.

2 Literature Review When applied to the issue statement, a variety of approaches may be used. We have referenced many scholarly publications. When we go over the papers for our final decision, we utilize different criteria to come up with the best solution approach. A summary of these review papers is presented below: [Mehdi Mekni, Andre Lemieux, 2017] has introduced the paper on “Increased Reality”, in which virtual setting is initially coordinated with the presentation of genuine environment visuals, a developing field of aggregate plan. The PC’s continuous blending video produces content with live exhibitions. AR depends on advancements created in Augmented Reality and collaborates with the virtual world as well as has a level of reliance with this present reality. The expression “expanded the truth” was first begat at Boeing in 1990 by specialist Tom Coddell, who was approached to improve the costly graphs and checking apparatuses used to direct laborers on the manufacturing plant floor. Coming up next are its favorable circumstances in that it has an assortment of utilizations—medicinal, instruction, military, amusement and sports, and so on. There are sure constraints with AR that must be survived. AR frameworks need to manage a lot of subtleties in all actualities. In this way, little equipment ought to be utilized, effectively versatile and light and quick enough to speak to designs. Additionally, the battery life utilized by these intricate AR gadgets is another constraint for AR utilization. [K. Govindaraju, 2017] proposed a paper on Embedded Systems. An embedded system computer system that perform certain pre-determined program that are typically used on a large scale by an mechanical or electrical systems. Typically, it is introduced from small MP3 players to large-scale hybrid vehicle systems. Some other examples of embedded systems often used in our daily life are—keyboards, mouse, ATMs, TV, PDA, cell phones, printers, elevators, smoke detectors, DVD player, refrigerator, cameras, GPS navigators, radios, TV remotes, telephones, game controllers, monitors, digital images processors, bar code readers, SD cards, washing machines, anti-lock braking systems, blender, and more. We especially use embedded systems due to its dependent, efficient and it meets real-time constraints. Examples of embedded systems illustrate that we now take embedded systems for granted. Due to the implementation of smart embedded technologies in our house, we are quite acquainted with the phrase “smart home”. Embedded systems are linked to the Internet nearly all the time now. Embedded system applications may be roughly divided into two groups: household applications, such as dishwashers, washing machines, and microwaves, and commercial applications, such as gaming and movie playback devices. Embedded systems have a major drawback since they are very cost-sensitive. Small adjustments in the price of a piece of construction equipment may have a significant impact on the case. [Ronald Azuma, 2016] an Augmented Reality-based research proposal an AR system’s ultimate aim is to increase the amount of information the user sees and interacts with by augmenting their perception and interaction with the actual world with virtual 3D objects that seem to be physically located alongside the real world.

Framework for Implementation of Smart Driver Assistance System …

235

Augmented Reality works in synergy with other technology such as Internet of Things (IoT) and many other to give people a rich view on their reality. Augmented Reality is exactly what the name suggests, and this is when a various reality dominates your already existing reality using certain device such as smartphone and smart glass. Computer-generated image of a different setting is superimposed on your device to change the view of reality. AR has limitations in that it may require a large amount of hardware. Despite many recent advances in AR, much remains to be done. [Hassan Fleyeh, Mark Doughtery, 2016] a study describing an advanced way of roadside sign detection and recognition was given. This phrase explains the features of road signs, the procedures, and obstacles that need to be overcome while trying to find road signs, and finally, road sign detection and recognition are essential in ITS. This is because traffic signs and road signs are vital in everyday life. These icons make it possible for motorist to comprehend the visual languages presented to them. They display current traffic conditions on the road, dangers, and potential obstacles to drivers and provide valuable information that helps drivers with their navigations. Color-based and shape-based road sign detections are among the features of Road Sign Recognition. Because it needs more gear and is costly, it has certain restrictions. [Lotfi Abdi, Meddeb, September 2016] The Application of Augmented Reality Inside a Vehicle” TSR, which incorporates In-Vehicle Augmented Reality (AR) For many years, the design of TSR has been a challenging issue. The road signage classifications and localizations provide as the groundwork for further methodology that is utilized for proper TSR. Weather conditions, angles of view, sign distortions, and different backdrops all contribute to an obstruction of signals. It is critical to develop computing-efficient and highly discriminating features in order to produce a quick and robust TSR. In order to enhance driving safety and cut down on the effort, information that is easier to comprehend and process must be made available. The AR-TSR completes the vehicle’s external view of traffic conditions with information about those conditions, which is shown on the driver’s monitor. [Felipe Jimenez, Aurelio Ponz, April 2016] presents a paper on “Advanced Driver Assistance Systems for Road Environments to Improve Safety and Efficiency”, which deals with two sides that were detected as key themes are: “Safety and Efficiency”. It facilitates the flow of details between vehicles and assists in the identification and data transmission process. Applications developed include: “Adaptive control with optimization, overtaking support systems and driver assistance”. To give an early reply to risk conditions, and optimize safety measure for occupants of the vehicles and the characteristic of collisions, pre-collisions in system arise. Nevertheless, it should be noted that these systems demand high amounts of reliable data about the vehicle and surrounding. It also requires support systems with motion control during maneuvers, collision avoidance system with the potential for developmental maneuvers which includes pedestrians, cyclists, and motorists, and cooperative optimization consideration of other vehicle and road verticals’ side and traffic signal detection characteristic. [Mauricio Hincapié, Andrea Caponio, Horacio Rios, Eduardo González Mendívil] made a presentation on AR as a branch of VR (Virtual Reality Technology), which is sometimes referred to as “the use of unusual equipment and

236

K. Baskar et al.

computer programming to create a replica of a replacement environment, which customers perceive as real or legitimate”. It is innovation in virtual reality (VR) that enables a scenario in which the customer experiences himself inside a PC-generated virtual world while immersed in the world. Customers who are in the world feel and appear to be inside the PC-generated virtual world, as people do while immersed in the real world. While AR makes the client able to perceive this current world, it expands it with virtual information overlaid on top of it. On the other hand, since virtual reality supplants reality, Augmented Reality complements it, creating a realm in which genuine and virtual content may coexist. The positive aspects of AR are that it is extraordinarily flexible, particularly in industrial settings, and also may very well be realized in many applications. Despite its many positives, there are a few negatives that may jeopardize its authenticity in real-world applications. An additional important point to bear in mind is the weight of the equipment. [Filip Malawski, July 2018] exhibited that as of late, propelled picture preparing calculations have been utilized to investigate nature while driving and furnish the driver with helpful data. Driver help frameworks can identify people on foot, perceive street signs, and give route directions. Then again, the methods for showing these data are somewhat unrefined and ordinarily incorporate a little show. Along these lines, the driver needs to isolate consideration between the presentation and this present reality. Tending to this problem may be done via Augmented Reality (AR). We propose a system in which a substantial amount of relevant data for the drivers is displayed on semi-straightforward glasses that are then inserted into the current world using AR technology. We conduct a condition inquiry on a cell phone, which is then connected to the headset. Thus, the headset is responsible for determining the location of each individual phone and changing the phone’s viewpoint. Evidence of an idea in the context of a person just dozing off includes a passer-by finding as well as someone nodding off, with the use of an implicit accelerometer. [Paul George, Indira Thouvenin, Vincent Fr’emont and V’eronique Cherfaoui] By looking at how the drivers feel, we can assess how robust our planned new propelled driver assistance frameworks will be. A driving assistance architecture presented in this article is inherently linked to the client. Daria is useful for autonomous vehicles since it uses Augmented Reality (AR) to enhance the functionality of the sensor. Snags and their risk quantification are in the middle of the identification. Driver behavior is the other option. The driver is able to see the risky areas while keeping his eyes open by means of a suitable perceptual representation At this point, the preliminary results indicate that our approach may be extended to automobiles, ships, or water routes, while the organizing assistance allows the client to get at its objective without earlier information out. [Lucas Morillo Méndez, June 2018] proposed a paper on Augmented Reality. AR is progressively being created in the car space as an ADAS framework. This inexorably mainstream innovation can possibly decrease the fatalities out and about which include HF, anyway the psychological segments of AR are as yet being examined. This survey gives a brisk outline of the investigations related with the intellectual systems engaged with AR while heading to date. Related research is changed, and a scientific categorization of the results is given. AR frameworks ought to pursue

Framework for Implementation of Smart Driver Assistance System …

237

certain criteria to stay away from bothersome results, for example, subjective catch. Just data related with the fundamental driving errand ought to be appeared to the driver so as to stay away from impediment of the genuine street by non-drivingrelated undertakings and high mental remaining task at hand. Notwithstanding, data ought not be appeared consistently so it does not influence the driving aptitudes of the clients and they do not create overreliance in the framework, which may prompt hazardous practices. Some famous employments of AR in the vehicle are route and as security framework. AR subjective results ought to be considered in these specific settings later on. This article is expected as a small-scale control for makers and architects so as to improve the quality and the effectiveness of the frameworks that are presently being created. One of the potential disadvantages of AR inside the vehicle is that it can prompt intellectual catch if the AR territory is overloaded with data. This may make the driver react to the offered data instead of to the driving errand and an issue of overreliance, and subsequently, that clients take part in dangerous driving. [Chin-Chen Chang, Shu-Chun Huag, June 2018] showed a presentation about where traffic signs are placed and how to use an assistance system. Driving along the street might provide several types of messages from street signs. To capture the attention of drivers, traffic signs have eye-catching colors and simple graphics. Drivers overlook signals from traffic signs when they drive in complicated situations or if their mental state is poor. In the case that there is a programming framework for reporting the discovery and acknowledgment of traffic signs, it can pass that information on to the driver and lower the burden of having to memorize the different types of traffic signs. When the vehicle misses a traffic sign, the platform is capable of showing a cautioning signal. This framework can help the driver software navigate the roadway, by figuring out the road state. So, the driving experience is improved in huge ways, with the dangers of errors being minimized. In this research, we focus on the identification and finding of traffic signs to support vehicle assistance applications. We tried several different types of frameworks, although also checking to see if the framework execution was satisfactory. Street signs may convey numerous signals to the motorist, even while they are simply operating a car. To make sure the traffic signs stand out, they are made with striking colors and simple graphics. Despite this, a motorist can have an extremely poor understanding of traffic signage, especially if the driver is perplexed about what they are supposed to do or if they are in a poor mental state. Should there be a traffic sign recognition and reporting system in place, the motorist will get relevant notifications quicker and also find it easier to navigate the road. The framework has the ability to see whether the motorist misses a traffic sign. An encouraging warning This framework can help the driver software navigate the roadway, by figuring out the road state. It follows that a driver’s comfort is much increased, and the risk of getting into an accident is reduced. This study focuses on how traffic indicators for driving assistance programs may help the driving population can be found and distinguished. Additionally, we investigate several design blends of the framework in order to confirm the framework’s performance.

238

K. Baskar et al.

2.1 Comparative Study See Table 1.

3 Design and Implementation The main focus is to develop a rotating inductive road sign detection system by Augmented Reality and advanced POSIT algorithm. Glyph-based Augmented Reality is used in our project in conjunction with the POSIT algorithm. The POSIT algorithm is extended by our novel approach. The objective is to use image processing with glyph recognition technology and better and efficient recognition and detection of road signs to aid the driver and reduce road accidents. Some glyph recognition markers have been created that show street signs in particular. (These markers should be placed on the road and made standard for real-life application). When these markers are placed in front of the camera, they are identified and information is stored in the system. Information will be displayed in front of the driver in an automobile system using any device. Notify using an LCD or mobile phone.

3.1 POSIT Algorithm The purpose of 3D pose estimation is to determine the change required to map an object from a 3D coordinate system to a 2D image. It plays a very important role since the beginning of computer vision. One specific application is to help mobile robots self-localize in unknown environments. Another application area that I am particularly interested in is Augmented Reality, and it can help visual perception of 3D in areas such as hospitals, navigation, and admissions. These techniques have already been used and will be widely used in the future. Therefore, it is very important to study pose estimation. Many algorithms can perform currency estimation. One of the most popular and widely used algorithms is POSIT (from orthography to scaling with pose and repetition). Unlike some other algorithms, POSIT does not require preliminary estimation. It just requires four or more corresponding attribute points between the 3D object and the 2D image. In the POSIT algorithm, the 3D model should not have the feature points coplanar (in the same plane), otherwise it does not work. However, there is a low algorithm called coplanar POSIT that can deal with this situation.

Andre Lemieux, Mehdi Mobile Technology, Mekni Virtual Environment, and Augmented Reality

FilipMalawski, 2018

Ms. YaminiGowardhan, Dr. RamchandHablani, 2016

Augmented Reality: Applications, Challenges and Future Trends

Driver Assistance System Using Augmented Reality Headset

Traffic Sign Detection and Analysis

Color Segmentation, Shape Analysis, Conversion of color models

Image processing algorithms, Augmented Reality

Prototype of driver assistance system in Augmented Reality has been presented based on the weathervane metaphor adapted for a use in real conditions metaphor adapted for a use in real conditions

Paul George, Indira Thouvenin, Vincent Fr’emont and V’eroniqueCherfaoui

DAARIA: Driver Assistance by Augmented Reality for Intelligent

Technique used

Author and year

Title

Table 1 Comparative analysis of exiting work Conclusion

Driver needs to divide attention between the display and the real world

AR systems have to deal with large amounts of details in reality. Therefore, small hardware should be used

(continued)

Using HIS color model to detect the traffic sign followed by circle detection

Detect pedestrians, recognize road signs, and provide navigation instructions

It describes work performed in different application domains

Axe for future research is Prototype of driver focused on the metaphor assistance system in improvement Augmented Reality has been presented

Limitation

SVM Classifier and Regions detected by Edge Detection provides color detection cannot be improved results determined to the exact sign region

Proof-of-concept scenario includes pedestrian detection as well as falling asleep detection, with use of a built-in accelerometer

Various Applications in military, medical, manufacturing, and entertainment

A lot of problem of integration (calibration, synchronization—not described here—and data exchange) has been resolved to conduct first experiment and prove the feasibility

Proposed work

Framework for Implementation of Smart Driver Assistance System … 239

Lucas Morillo Méndez, Cognitive and Perceptual Provides safety, Occlusions of the road Cognitive outcomes of 2018 Implications of AR navigation, and take over elements may disturb the interacting with AR inside the car request assistant availability of AR while driving information Image processing techniques

Haar cascade and the bag Good compromise of visual words approach between resource efficiency and overall performance

Mohammad S. Aminian, Christina olaverri-Monreal,2017

Lotfi Abdi, ArefMeddeb,2017

Augmented Reality as an Advanced Driver Assistance System: A cognitive Approach

Smartphone-based traffic sign detection using OpenCV

In-Vehicle Augmented Reality TSR to improve driving safety and enhance driver’s experiences

Conclusion

Color models, Neural Networks, Template Matching, Classical Classifiers

High object detection rate (84%) within few milliseconds of time

List of candidate objects that could be probable road signs

Does not make use of augmented markers, instead uses spatial information

A new AR-HUD approach to create real-time interactive traffic animations is introduced

A high-resolution device Displays detected needs more time to traffic signs and assists process the driver in not missing important traffic information

Input does not need to be Training overhead, transformed into another multilayer neural representation state networks cannot be adapted due to its architecture

Hierarchical grouping Selective search in the methods make it little bit color space with the complex accuracy and recall rates, and the red filter screening results

Hassan Fleyeh, Mark Doughtery

Limitation

Road And Traffic Sign Detection And Recognition

Proposed work

Convolutional Neural Utilized two types of Networks, color features deep learning structures to train our system

Chin-Chen Chang, Shu-Chun Huag, 2018

Traffic Sign Detection and Recognition for Driving Assistance System

Technique used

Author and year

Title

Table 1 (continued)

240 K. Baskar et al.

Framework for Implementation of Smart Driver Assistance System …

241

3.2 Details of Input/Data Used Knowledge is available that can be used to deal with the difficulty of recognizing and detecting road sign in an effective manner. Every road signs are design, built, and installed as per stringent set of rules set forth by federal committee. Color is set not only for the signal range (Red = stop, Yellow = hazard, etc.), but also according to the blot of color or paint that shadows the signal, with a tolerance that is visible, at a specific wavelength should correspond to spectrum. This is certainly an important detail, but caution will be exercised for getting use of it as the standards are set as per the controlled illuminations that prevails while performing experiments, whereas during practice, weather situations are subjected to external illumination, and of course, results will be affected by the colors captured by the camera. Color over road signs also fades over time. Text fonts as well as character heights are also regulated for shapes and dimensions as well as picture lines (Fig. 2). Signs are mainly located at the right side of the road, in usual practice two to four meters from the roadside, which are not strictly followed and their lies expectation of overhead and clearance signs which are located on the central lane. This phenomenon is beneficial for sign recognition because a big portion of image of road can be neglected and thus speed up process. • Signs could appear in a variety of situations, including damaged, partially cloudy, and exposed to sunlight. Signs may also occur, for example, three or four signs appear above/beside each other (Fig. 3).

4 Experimental Results and Analysis The Proteus Design Suite is a device suite, including one component in particular, the component that will robotize electrical structure. Schematic design and electronic print preparation are the common uses of the solution for electronic structure architects and experts. Only a few subnets of the roadway network can benefit from the concept of ad hoc architecture. A great illustration of this is a completely automated roadway in which only self-driving cars may be operated. The fact that the first steps of intelligent vehicle research have yielded positive results, such as vehicles being able to operate themselves, means that it is technologically possible to have a complete vehicle and road automation that includes things like traffic lights and road signs (at least on highways or very structured roads). As a designer, you should be aware of and attentive about several factors outside of technical aspects, including any difficulties. Factors include laws dealing with responsibilities, failure to operate correctly, and the influence of automated driving (both human and passenger). A good car should have an elegant, well-proportioned design and provide a comfortable environment. This is mostly down to how well-designed the passenger

242

K. Baskar et al.

Fig. 2 Localizing signs

compartment is. The framework’s interface will have a huge effect on how the framework’s frame is seen and understood. There are many lengthy evaluations and reconfigurations that are required before these frameworks are open to the public. Additionally, the public has to endure more years of manually driven cars and markets before the street network can be transformed into one that uses intelligent and adaptable drivers and other means of data trade (Fig. 4).

Framework for Implementation of Smart Driver Assistance System …

Fig. 3 Augmented marker

Fig. 4 Screenshot of Proteus Design Suite

243

244

K. Baskar et al.

5 Results and Future Scope The developed system quickly recognizes road signs and notifies to drive. The program also keeps track of the already passed public places such as petrol pumps and restaurants on the way and notifies the same to the driver (Figs. 5 and 6).

Fig. 5 Screenshot “Fifty”

Fig. 6 Recognition of augmented marker

Framework for Implementation of Smart Driver Assistance System …

245

5.1 Data Set Used Dataset In the back end (embedded system) whenever the road sign will be recognized, it will show a code on the display corresponding to that sign like, for example, for right turn, it will display R; for left turn, it will show L, etc. (Table 2). • AR is a colossal innovation at the same time; right now, it is as yet having significant issues by causing major issues that endanger its usage in mechanical world. In this report, we have displayed the fundamental advantages that AR can offer to modern methods, with incredible consideration regarding look after tasks. AR could improve human exhibitions without a doubt, and this can give us extraordinary advantages not just from a practical perspective. A superior support of a car does not only mean lower price, but also great reliability and, thus, leads few failures and corresponding accidents. In our project, Embedded Systems would be used at the back end and Microsoft Visual Studio would be used at the front end. • The developed system quickly recognizes road signs and notifies to drive. The program also keeps track of the already passed public places such as petrol pumps and restaurants on the way and notifies the same to the driver. • In the back end (embedded system) whenever the road sign will be recognized, it will show a code on the display corresponding to that sign like, for example, for right turn, it will display R, for left turn, it will show L, etc. • In-vehicle contextual Augmented Reality (AR) provides users with real-time visual and visceral input for a more efficient and impactful driving experience. To aid the driver in various driving situations, this technique is used to improve the traffic sign recognizer. It also makes drivers more comfortable and also reduces the risk of vehicle accidents. Through this analysis, we have shown that AR can enhance TSR greatly. To improve driving safety and decrease driving stress, AR may be utilized. A visualization of the information presents it in a manner that is simple for the driver to comprehend, requiring minimal cognitive effort. • To control the traffic for Road Signs’ Recognition using mobile phones or other smart devices, applications can be developed. Results based on calculations show quite high object detection rate, i.e., 84% within a time of few milliseconds. The time used in detecting road signs is different from time used with the help of resolution of the device which is being used. More time is required to process more number of pixels in high resolution, but the detection comes out to be much accurate. By improving the application in future analysis, one can reduce the amount of false and misleading detections of road signs. • A latest application using Augmented Reality approach is developed for creating real-time interactive traffic animations is introduced, in terms of rules for placement and visible to users, various traffic signs and movement of these is done to an in-built vehicle display of the device. • Important drawbacks that are majorly taking AR technology behind in the industrial environment were detected and solutions are proposed. Scientific solutions to solve these drawbacks are required to make AR more breakthrough technology.

246 Table 2 Data used for train the model

K. Baskar et al. Name No parking

Right turn

Left turn

U turn

No left turn

No right turn

Road signs

Augmented markers

Framework for Implementation of Smart Driver Assistance System …

247

Much better substances, faster algorithms for solutions, small hardware elements are taken. The research committee takes responsibility of these needs and provides solutions in order to solve the problems.

5.2 Conclusion Complete mechanization will be done distinctly to uncommon foundations, for example, in uses of businesses or open transportation for the present. At that point, mechanized innovation of vehicles will be gradually and bit by bit stretched out to numerous other significant transportation regions, for example, delivery of products. For instance for expensive trucks, where the expense of the vehicle itself and the administration it gives is a lot higher than the expense of an autopilot which is practically oblivious before the previous one. At long last once, innovation has been created and kept up and the most worthwhile arrangement and best calculations are utilized, huge. The developed model can recognize road signs, but traffic lights had to be done in more detailed way that is the future scope of this project. The Road Sign Recognition systems will probably become an important element of future support systems of drivers in a very short period of time. Many experts and researchers have concluded that the initial generation of individual driver information and warning systems will be visible in the market in upcoming six to nine years. One of the first on the market would probably be Daimler Chrysler. The improvement of upcoming cars could be gained depending on both infrastructure and vehicles as well. Based on the particular app, user can get either some of the advantages or disadvantages. Improving the infrastructure and condition of road may give advantages to the architectures of transportation which are dependent on looping and already scheduled routes for driving, like transportation used in public and robotics in industries. On the contrary, a difficult and extensive organization and maintenance would be required for extended road networks for private vehicles which may become hard to understand and ambiguous as well as extremely costly.

References 1. Abdi L, Meddeb A, Abdallah FB (2015) Augmented reality based traffic signs detection for better driving experience. In: Communication technologie for vehicles. Springer, Cham, pp 94–102 2. Bark K, Tran C, Fujimura K, Ng-Thow-Hing V (2014) Personal Navi: advantages of an augmented reality. navigational aid using a see-thru 3D volumetric HUD. In: Proceeding of the 6th international conference on automotive user interface and interactive vehicles application. ACM, New York, NY, USA, pp 1:1–1:8 3. Lin JH, Lin CM, Dow CR, Wang CQ (2011) Implementation and designing of Augmented Reality for better Driving Visual Guidance. In: 2011 second international conference on innovation in bio-inspired computing and application

248

K. Baskar et al.

4. Müller Schneider S (2009) Augmented reality in asmart driver assistance system. In: Workshop proceeding–32nd annual conference on artificial intelligence, KI 2008—workshop on planning, scheduling, design, and configuration, PuK 5. Pauzie A (2014) Head up display in automotive: A new reality for the user. In design, users usability and experience: Interactive experience designing. Springer International Publishing, pp 506–517 6. Kim S, Dey AK (2016) Augmenting human senses to enhance the user experience in car: augmented reality is applied and haptic approach to reduce cognitive distance. Multimedia Tool Appl 7. De la Escalera, Moreno LE, Salich MA, Armingol JM, Road traffic sign recognition and detection. IEEE Trans Ind Electron 8. Greenhalgh J, Mirmehdi M, Real-time recognition and detection of roads traffic sign. IEEE Trans Intell Transp Syst 9. Huang SC, Lin HY, Chang CC (2016) An in-car camera’s system for traffic signs. In: Proceeding of the joint 16th world congress of international fuzzy system association and 8th international conference on soft computing and intelligent systems, recognition and detection, Japan, June 2016 10. Kellmeyer DL, Zwahlen HT (1995) Determination of highway warning sign in usual video image using color image processing and neural network. In: Neural Network 11. Abdi, Meddeb L (2015) Augmented reality based traffic sign recognition for improved driving safety. In: Communication technologies for vehicles 12. Wang J, Söffker D (2016) Improving driving efficiency for hybrid electric vehicle with suitable interface 13. In: 2016 IEEE international conference on systems, man, and cybernetics (SMC), pp 928–933 14. Fang C, Chen S, Fuh C (2003) Road-sign detection and tracking. IEEE Trans Veh Technol 52:1329–1341 15. Escalera A, Armingol J, Mata M (2003) Traffic sign recognition and analysis for intelligent vehicles. Image Vis Comput 21:247–258 16. Miura J, Kanda T, Shirai Y (2000) An active vision system for real-time traffic sign recognition. In: Presented at 2000 IEEE intelligent transportation systems, Dearborn, MI, USA 17. .Houben S, Stallkamp J, Salmen J, Schlipsing M, Igel C (2013) Detection of traffic signs in real-world images: the German traffic sign detection benchmark. In: The 2013 international joint conference on neural networks (IJCNN). IEEE, pp 1–8 18. Kun Z, Wenpeng W, Guangmin S (2014) An effective recognition method for road information based on mobile terminal. Math Probl Eng 2014:1–8 19. Mogelmose A, Trivedi MM, Moeslund TB (2012) Vision-based traffic sign detection and analysis for intelligent driver assistance systems: Perspectives and survey. IEEE Trans Intell Transp Syst 13(4):1484–1497

IoT-Based Mental Health Monitoring System Using Machine Learning Stress Prediction Algorithm in Real-Time Application Md Abdul Quadir, Saumya Bhardwaj, Nitika Verma, Arun Kumar Sivaraman, and Kong Fah Tee Abstract With the primary focus of healthcare technologies being on the physical health of a person, mental health issues sometimes go unattended. Stress, anxiety, and depression are becoming increasingly common problems in our community leading to serious heart-related problems such as high blood pressure, episodes of heart attack and can even lead to chronic illness. Prediction of stress or depression at an earlier stage can prevent serious consequences as sometimes patients suffering from mental illness are not aware of the severity of their condition or do not keep up with counseling for a longer period of time. In this context this paper proposes a stress prediction method using machine learning to detect the development of stress or anxiety problems at an early stage. Our proposed method observes any changes in the human body under stress or depression by monitoring the ECG values and other physiological factors to predict any kind of possible stress or depression. The proposed model provided high accuracy of 98% in predicting stress. On detecting stress, appropriate actions such as informing the patient’s guardian and doctor are taken. As compared with other models, our model outperforms the other state of the art models, making it a real-world predication model.

M. A. Quadir (B) · S. Bhardwaj · N. Verma School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India e-mail: [email protected] S. Bhardwaj e-mail: [email protected] N. Verma e-mail: [email protected] A. K. Sivaraman Digital Engineering Services, Photon Inc., Chennai, India e-mail: [email protected] K. F. Tee Department of Civil and Environmental Engineering, King Fahd University of Petroleum and Minerals, Dhahran, 31261, Kingdom of Saudi Arabia e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_16

249

250

M. A. Quadir et al.

Keywords IoT · Stress prediction · Machine learning · Anxiety · Depression · Temperature sensor

1 Introduction In recent times, we can see the rapid changes happening around us. Everyone is blending in the new era of technology and advancing alongside, as a result, both physical and mental health of a person keep getting neglected. Lifestyle is another factor that disturbs a person’s health, including disturbed sleep schedule, poor eating habits, social media usage, exercise, and various others. In today’s scenario, technology has become indispensable, and we are all bound to it now. To combat the harmful effects of technology on a person’s health, Internet of Things (IoT) comes into play. IoT can be defined as a network of numerous sensors or devices that collects and transfers data over the network. In the real world, smart watches and Alexa are some of the commonly used IoT devices for collection of a person’s data. Technologies such as smart watches offer a way to detect a person’s heart rate, oxygen level, BP, and others to monitor their physical health. Stress detection is becoming an important research domain as it is relevant in studies based on monitoring of various health issues, mental illnesses like depression and anxiety disorders, as well as physical illnesses like migraines and high blood pressure, as well as heart attacks and strokes. Some of the major challenges in the diagnosis of stress are: • The reluctance of patients to receive continuous health monitoring and the limited timeframe of specialists to monitor every patient with quality time could result in a late diagnosis or even untreated conditions. • Providing immediate assistance to patients during emergencies such as anxiety or panic attacks becomes difficult sometimes and can lead to unfortunate consequences. • Accurate diagnosis becomes difficult while dealing with patients that are reluctant to share their conditions properly. To overcome the above issues, the proposed system comprises the following: • The proposed system would obtain data through sensors and wearable devices, so the patient needs not visit the doctor regularly. The data collected is stored in the cloud facilitating regular collection of updates regarding fluctuation in the stress level of a patient monitoring through the algorithm hence not requiring the doctor’s intervention at each stage of data collection which would result in a more efficient and thorough diagnosis. • The proposed system is provided with the facility to send an alert message to the patient’s guardian and the assigned doctor in case the heart rate of the patient shoots above a safe limit.

IoT-Based Mental Health Monitoring System Using Machine Learning …

251

• Our proposed system employs a machine learning algorithm trained with a preexisting dataset to provide stress detection with high accuracy. In this paper, we use the existing technologies to propose a method using data collected from GSR and other physiological sensors for stress detection with the help of machine learning algorithms and publicly available Wearable Stress and Affect Detection (WESAD) dataset. The rest of the paper is organized into the given following sections: Sect. 2 presents the related work in the area of mental health and the development of various systems to combat the same. In Sect. 3, the architectural framework and modules for the Stress Prediction Model are proposed with a Random Forest machine learning algorithm. Section 4 discusses the various experimental results, accuracy, and performance analysis followed by Sect. 5 that lastly concludes the entire paper with future directions of the proposed method.

2 Literature Review Priyadharsan et al. [1] propose the various medical parameters in predicting the physical health of a patient using IoT and machine learning techniques. In this, they have performed the initial training and validation of machine learning algorithms. Their testing phase estimated the prediction of possible health problems using the data collected by sensors in their proposed IoT framework. Here, Gunti et al. [2] proposed a multi-model system that combines ECG and GSR sensors for intimating and monitoring the stress, anxiety, and depression levels of a person. They used the ECG for monitoring the heart rate of the person by displaying the value on the LCD and as well sending this data via Bluetooth module to an app to plot the ECG graph on the mobile phone. They placed the GSR sensor on the finger, and it works on the skin conductance, i.e., based on the sweat secretions and with the help of certain threshold values. Deepika et al. [3] proposed an architecture that would help individuals suffering from depressive disorders and will safeguard them from a seriously devastating scenario. Their system would change the mood and mental state by diverting the individual by automated playback of songs or display of favorite pictures or videos when triggered by the symptoms via the mobile app as well as send the notification to the psychiatrist or nearby hospital and to the caretaker of the individual in emergencies. Also, Patil et al. [4] have recommended a system that works on the security of healthcare data and faster communication. They retrieved the data stage from the cloud. The cloud server was defined such that it was shared with only an authenticated user as per request. During an emergency, an emergency mode is activated and the device is updated every 1 min. The wearied device sends the results to the user’s cellphone using a built-in Bluetooth connection or NFC technology. Md Haffer et al. [5] have worked on performance analysis of monitored health data through low-power WAN networks. They have also shown that power consumed by the LoRaWAN network was less than that of GPRS/3G/4G by a factor of ten. They concluded that LoRaWAN gave the better overall efficiency of IoT

252

M. A. Quadir et al.

for health monitoring systems. Masud et al. [6] use ECG signals for measuring at different intervals and situations. They have analyzed the limited resources available in the computing domain and challenges related to loss of network continuity. To solve them, they developed a mathematical model to execute all the tasks sequentially using three different approaches to work out the process which was a mobile-based monitoring approach, the second was using data mining, and the third was being a machine learning approach. Seigmund et al. [7] have used speech recognition for determining a person’s positive and negative phase with the help of an emotional user interface to make a wearable device to help a patient suffering from Major Depressive Disorder (MDD). McGinnis et al. [8] have used a real-time demonstration with wearable sensors on children for 20 s using machine learning to analyze the output. This resulted in a high fraction of accuracy which shows that a high number of children are suffering from anxiety and depression. Zhou et al. [9] have proposed a deep regression network to understand depression through better representation, through a visual explanation that will provide a clinical prediction of depression with the help of facial images. Also, Jazaery et al. [10] proposed a framework of deep spatiotemporal depression analysis that predicts the level of depression in an individual using the visual expression data obtained by Beck Depression Inventory-II. Thangaraj et al. [11] talked about “Digital hospitals”. They worked on enabling automatic electronic medical records in a standard form. They also talked about its implementation in real-world scenarios in smart autonomous hospitals with IoT. Islam et al. [12] worked on developing an IoT-based healthcare system to monitor the basic physiological signs of patients such as heart rate, body temperature, along with physical conditions such as the humidity, CO, and CO2 gases’ level of the patient’s room. Researchers have also paid attention to psychological. In [13–29], the authors propose different models based on QoS parameters.

2.1 Problem Statement The Internet of Things (IoT) has provided the world with new opportunities in a variety of different types of fields, most importantly, the healthcare sector improving the quality of life, providing an effective patient care system (using data collected from the IoT devices), and patient health safety. Technology has proven to be useful in almost every domain of health care. Previously, the primary focus of technological development was in the field of physical health, whereas mental health was considered irrelevant. With the increasingly stressful and hectic lifestyle of people, there is a need to develop a system that could monitor the physical as well as the mental health of a person for their well-being. Cases related to mental illness are increasing dramatically around the world, emerging as an urgent global health threat. Over 500 million people around the world suffer from mental health disorders, among which depression and stress are the most common. In such a scenario, the Internet of Things (IoT) can provide us with the best healthcare service for early patient care.

IoT-Based Mental Health Monitoring System Using Machine Learning …

253

In cases of patients suffering from mental illness, sometimes they might be unaware of the severity of their health condition which could lead to serious and unfortunate consequences. With the existing technology consisting of smartwatches and smart wristbands that monitor the physical health of a person, there is a need to develop a system that would monitor mental health through the data collected from their daily activities with the help of IoT devices. One of the major challenges in mental health care is the reluctance of patients to receive continuous health monitoring and the limited timeframe of specialists to monitor every patient with quality time which could result in a late diagnosis or it could even go untreated if the patient is not willing to get monitoring at regular intervals. One solution to this problem that we would like to explore is to develop a system that would collect regular updates regarding fluctuation in the stress and anxiety level of a patient through various “physiological factors and sensors” that could be monitored through an algorithm, not requiring the intervention of a specialist at every stage of data collection, hence providing a much efficient and thorough diagnosis to the patient so that mental illness can be treated at a very early stage, improving the quality of life.

3 Proposed Stress Prediction Model Architecture The data collected by the sensors along with the data from the patient records is fed into a machine learning algorithm that has been trained with existing datasets on mental health issues such as stress. The model predicts whether the person is likely to suffer from stress or not, hence providing detection of these mental issues at an early stage so that the patient can seek help and the doctor would have the necessary data and analytics to provide better health care. The system also has the facility of an alert system in which an alert in the form of notification is sent to the patient’s caretaker’s mobile phone or to the doctor assigned to the patient if the system detects any symptoms that could lead to depression or anxiety.

3.1 Arduino-ESP8266 This is a self-sustaining SOC (System on Chip) with an inbuilt TCP/IP protocol stack that allows the microcontroller to communicate to the Wi-Fi network. This acts as an interpreter between sensors and the smartphone device.

3.2 GSR Sensor GSR sensor stands for galvanic skin reaction, and it is used to estimate the electrical conductance of the skin. This sensor uses the sweat produced by our body and

254

M. A. Quadir et al.

changes it into skin conductivity. This is taken as a sign of mental and physiological excitement produced by a body. It also helps to detect stress as during stress there is an expanded emission in perspiring organs.

3.3 ECG Sensor An electrocardiogram (ECG) sensor is used to detect a patient’s heartbeat. During the depression, it has been noted that a person’s heartbeat drops to 10–15 heartbeats per minute, while during stress that causes anxiety, it increases up to 38–45 beats per minute. With an ECG sensor that records electrical moments of the heart, it is possible to detect depression and anxiety in a person.

3.4 Temperature Sensor—LM35 The LM35 temperature sensor series consists of precision integrated circuits, and the output obtained from then is in the form of voltage vs. centigrade temperature. The advantage of using the LM35 device over the kelvin-calibrated linear temperature sensors is that the difficult calculations and conversion include the subtraction of a large constant voltage from the original output to get the final output that is a centigrade scale. This sensor is used here as the temperature in the room/environment plays a great role in the fluctuation of the heartbeat as well as the production of sweat from the body.

3.5 GSM Module A GSM module is a GSM modem that connects our mobile phone device to PCB/Arduino. It is used to send and receive messages. This will help us to send the alert messages to the mobile phone of the patient’s guardian in case of any severity detected and keep a record of the data obtained.

3.6 Web Interface A Web Interface gives a user-friendly platform to the patient, their guardian, and the doctors. Here, the patient’s record is shown which has been obtained through the WiFi–Cloud module in the Arduino. Every patient has a unique ID and password. This is done for data confidentiality. The patient’s report will be visible to his/her doctor. Through this, the patient’s medical history is available directly to the doctor with all

IoT-Based Mental Health Monitoring System Using Machine Learning …

255

the data to make a clearer diagnosis, and all the data displayed on the interface are stored in the cloud for easier retrieval when required. In Fig. 1, the architectural setup for our proposed method is shown, and we observe that the user’s data has been obtained through sensors using Arduino and is being sent to our machine learning predictive algorithm. After obtaining the predictive results, the user’s health report is generated based on the factors of ECG, skin response, and user’s body temperature. This data is made available on the user web interface for easy accessibility. Every person will have their account, which can be accessed by the user, his/her guardian, and doctor for data confidentiality purposes. Also, in case of an emergency, an alert notification will be sent to the user and the doctor. All the data will also be saved in the cloud. The mental health monitoring phase. This phase deals with receiving the patient’s data. The sensors that are used to obtain the data are GSR, ECG, and temperature sensors. Through this, we obtain the patient’s heartbeat information, body temperature as well as skin sweat production. If the heartbeat rate of a person shoots up high during any time of the day, it could be an indication of depression, panic attack, anxiety attack, or any other serious heart complication due to depression or any sudden stress causing the heart rate to rise beyond a safe level, in such a case the system would send an alert to the patient’s guardian or the doctor in-charge so that immediate assistance can be provided. Mental health prediction phase. In this phase, the data that has been received from the patient is processed using machine learning algorithms and with various conditions to obtain the final value of whether the patient is under stress or anxiety.

Fig. 1 Proposed architecture for stress prediction

256

M. A. Quadir et al.

The temperature sensor also plays an important role over here to tell us about the environment as the temperature is also one of the key factors behind the fluctuation of GSR sensors. The analysis will also be done based on the age of the patient as the heartbeat (ECG sensor) range differs with age. Analysis of the data collected will be done using the Random Forest Classifier algorithm. Random Forest is considered as “a highly accurate and robust method because it involves several decision trees participating in the process”, interconnected to give the final prediction. In addition to that, it does not suffer from overfitting due to high variance in the data. Using Python as the language with Jupyter Notebook as the platform, given below is the algorithm for the proposed system. The basic working of the Random Forest consists of branching the root node to internal nodes and further to the leaf node. If we use a Random Forest Classifier to solve the regression problem, then the mean squared method (Eq. 1) will be used. MSE =

N 1 Σ ( f i − yi )2 , N i=1

(1)

where N consists of the total number of data points, f i is the value returned by the model, and yi is the actual value for the respective data point. For solving classification problems, the Gini index [30] and entropy [31] are used. Stress Prediction Algorithm. Input: Read Ei , Gi , Ti ∈ sensor_data, that is, the values collected by the respective sensors (ECG, GSR and Temperature) for a person. Output: Li ∈ label, that is, the prediction of whether the person has stress or not. Algorithm Begin: Step 1: Import and Initialization Import the required libraries, then read the dataset and initialize it to variable data Step 2: Pre-processing to bring the data in the format in which is available in the dataset ECG (mV): ((raw_data/chan_bit-0.5)*vcc) GSR(μS): (((raw_data/chan_bit)*vcc)/0.12) Step 3: Import different function and libraries to split the dataset import train_test_split and RandomForestClassifier from sklearn Step 4: Split dataset into features and labels X = data[[‘ECG’, ‘GSR’, ‘Temp’]] //features under study y = data[‘Label’] // Value to be predicted

IoT-Based Mental Health Monitoring System Using Machine Learning …

257

Step 5: Split dataset into training set and test set in the ratio of 70:30 as training data: testing data. X_train, X_test, y_train, y_test = train_test_split(X, y, test_size = 0.70, random_state = 10) Step 6: Create a Gaussian Classifier and train the model using the training sets. clf = RandomForestClassifier(n_estimators = 100) clf.fit(X_train,y_train). Step 7: Prediction to be made on test set and accuracy check y_pred = clf.predict(X_test) from sklearn import metrics print(“Accuracy:”, metrics.accuracy_score(y_test, y_pred)) Step 8: Stress Prediction based on sensor data sensor_data = [[E,G,T]] L = clf.predict(sensor_data) End

4 Experimental Results In this phase, the data that has been received from the patient is processed using machine learning algorithms and with various conditions to obtain the final value of whether the patient is under stress or anxiety. The temperature sensor also plays an important role over here to tell us about the environment as the temperature is also one of the key factors behind the fluctuation.

4.1 Dataset In our paper, we have used WESAD data as our dataset. It is publicly available. This is a multimodal dataset [32]. This dataset has been recorded from professional wearable devices like Raspberry and Empatica E4 (a wearable device that is used to collect physiological data of the person wearing it primarily for research work). Preprocessing was performed on this dataset to remove any null values and the columns with values of sensors, not under our study and our purpose of detecting mental health in a person. Basic characteristics of our modified dataset as shown in Table 1 depict the details about the columns understudy from the dataset, namely ECG, GSR, temperature (sampled at 4 Hz. Data is provided in °C), and lastly, label. Inside the column label,

258 Table 1 Information about the columns of the dataset

M. A. Quadir et al. Column

Example

ECG

Float64

GSR

Float64

Temperature

Float64

Label

Float64

Total entries: 4,165,000 (0–4,164,999)

there are binary values (0 and 1), where 0 denotes that the person under study was not in stress at those values of other sensors, while 1 indicates that person was stressed. In our dataset, we observe that we have “Experiencing stress” data points (4,30,500) denoting that a person is suffering from stress or anxiety at that moment, while another set is for “other” symptoms data points (37,34,500) for when the person is not under stress. We ran a Random Forest Classifier for the given dataset with 41,65,000 data values to train the model in predicting whether the person is stressed or not and measured the model’s efficiency.

4.2 Our Approach and Results The general approach of the proposed method includes the collection of physiological data from the patient which gets stored in the cloud to provide the doctor with regular updates about the patient’s condition and progress which helps in coming up with a better diagnosis and treatment for the patient. The guardian and doctors would be immediately notified in case of an emergency. From Fig. 2, we observe that the patient’s data is being collected through Arduino. Heart rate is being calculated through ECG value. If the heart rate is greater than 120, it denotes that a person is short of breath which causes severe chest pain or is suffering from a panic or anxiety attack. In this case, an emergency alert is sent to the patient’s guardian and his/her doctor for immediate help assistance. Else if the heart rate is below 120, then all the sensors’ data are sent to a backend to pass it through our Random Forest machine model which is pre-trained on stress and depression values. The patient’s data is then used for predicting whether he/she is suffering from stress/depression. If the prediction yields 1 (indicating that the patient is suffering from mental illness), then an alert notification is sent. Else, no alert notification is sent. This data gets stored in the cloud which can be accessed by the doctor for diagnosis or monitoring, based on which the patient’s health report is generated. Mental illness is an area of interest, and research in recent times has led to many innovations. Using the dataset, we have proposed a method to analyze and then predict the occurrence of stress. We have split the dataset in the ratio of 7:3 (as training data and testing data) and have used Random Forest Classifier with n_estimator = 100 to train our model. On testing the trained model with the test data, we achieved an accuracy of 98% with favorable values in the confusion matrix (Table 2).

IoT-Based Mental Health Monitoring System Using Machine Learning …

259

Fig. 2 Experimental architecture of the proposed system

Table 2 Confusion matrix

Actually positive

Actually negative

Predicted positive

1,106,864

13,465

Predicted negative

11,413

117,758

On analysis of this result based on the importance of the sensor data, we conclude that the data obtained from the GSR sensor had the largest contribution to our prediction with an importance score of 0.506. The bar graph given below shows the contribution of each feature toward the prediction. The features represented in Fig. 3 and their respective importance scores are given, Feature: 0 (ECG sensor data), Score: 0.08422; Feature: 1 (GSR sensor data), Score: 0.50644; Feature: 2 (Temperature sensor data), Score: 0.40934. It is shown in Fig. 4 how GSR sensor values play a major role in the prediction of stress followed by temperature and ECG values. To calculate the efficiency of our model, we have calculated precision, recall, and F1-score using Eqs. (2), (3), and (4). Pr ecision =

T r ue Positive T r ue Positive + False N egative

(2)

260

M. A. Quadir et al.

Fig. 3 Importance score of different sensor values

Recall =

T r ue Positive T r ue Positive + False N egative

F1 scor e = 2 ×

(3)

Pr ecision × Recall Pr ecision + Recall

(4)

Precision tells us total positive prediction in the model that belongs to true-positive class. Recall calculates the correct positive prediction in the class. F1-score balances the precision and recall metrics. The precision, recall, and F1-score percentages of our model are 94, 95, and 95%, respectively. Many researchers around us have worked with WESAD dataset to predict stress with other machine learning predictive algorithms and deep learning approaches. So, we did a comparative analysis of all the models in Table 3. We can observe from Table 3 that Stewart et al. [33] used multiple machine learning algorithms such as K-nearest neighbors, Lasso, SVC and Neural process that achieved the average accuracy of 74, 88.1, 88.2 and 95.7% for stress prediction. Table 3 Comparative analysis of our model

Model

Accuracy (%)

K-nearest neighbors [33]

74

Residual-temporal convolution network (Res-TCN) [34]

86

Deep neural network + IoT-based biomarkers [35]

87.7

Lasso [33]

88.1

Support vector classifier (SVC) [33]

88.2

Neural process (Random choice) [33]

95.7

Our proposed model

98

IoT-Based Mental Health Monitoring System Using Machine Learning …

261

Kenneth et al. [34] used Residual-Temporal Convolution Network and achieved 86% accuracy, whereas on using Deep Neural Network along with IoT-based biomarkers gave the result of 87.7% to Kumar et al. [35]. Our model gave the highest accuracy of 98% for prediction of stress on WESAD dataset.

5 Conclusion and Future Work In this paper, an ML-assisted IoT framework has been proposed to serve the purpose of early detection of stress and depression with the help of three physiological signals (ECG, GSR, and temperature) to prevent any serious future complications or health issues. Random Forest was used to predict based on the sensor values if the person has stress. The proposed method aims at predicting if a patient is showing symptoms of being at risk of developing or is suffering from stress, anxiety, or depression; furthermore, it also strives at improving the communication between the patient, their guardian, and the doctor, so that all the necessary attentions and assistances, if required, are provided to the patient, which in turn leads to a better and more thorough diagnosis of mental illness. The proposed method was performed using an already available dataset in which the values were obtained using wearable devices that might have their limitations; in future, to increase the accuracy of our analysis and prediction in the proposed system, we plan to validate our approach by using actual data collected under proper observations using the system architecture proposed in this paper and use that in our prediction. As IoT and machine learning continue to grow, we anticipate that better sensors will become available soon to serve our purpose even better, and we intend to keep modifying our proposed solution as new and better technologies emerge in the ever-growing field of scientific research.

5.1 Declarations Fundings The authors did not receive support from any organization for the submitted work. Conflict of Interests No conflict of interest for this study. Acknowledgements None.

262

M. A. Quadir et al.

References 1. Priyadharsan DJ, Sanjay KK, Kathiresan S, Karthik KK, Prasath KS (2019) Patient health monitoring using IoT with machine learning 2. IoT Application: Human Emotions Management System (2020). Int J Recent Technol Eng 9(1), 1261–65. https://doi.org/10.35940/ijrte.F9971.059120 3. Deepika Mathuvanthi P, Suresh V, Pradeep C (2019) IoT powered wearable to assist individuals facing depression symptoms 4. Patil S, Pardeshi S (2018) Health monitoring system using IoT. Int Res J Eng Technol (IRJET) 5(04) 5. Mdhaffar A, Chaari T, Larbi K, Jmaiel M, Freisleben B (2017) IoT-based health monitoring via LoRaWAN. In: IEEE EUROCON 2017–17th international conference on smart technologies. IEEE, pp 519–524 6. Masud MM, Serhani MA, Navaz AN (2016) Resource-aware mobile-based health monitoring. IEEE J Biomed Health Inform 21(2):349–360 7. Siegmund D, Chiesa L, Hörr O, Gabler F, Braun A, Kuijper A (2017) Talis—a design study for a wearable device to assist people with depression. In: 2017 IEEE 41st annual computer software and applications conference (COMPSAC), vol 2, pp 543–548. IEEE 8. McGinnis RS, McGinnis EW, Hruschak J, Lopez-Duran NL, Fitzgerald K, Rosenblum KL, Muzik M (2018) Wearable sensors and machine learning diagnose anxiety and depression in young children. In 2018 IEEE EMBS international conference on biomedical & health informatics (BHI). IEEE, pp 410–413 9. Zhou X, Jin K, Shang Y, Guo G (2018) Visually interpretable representation learning for depression recognition from facial images. IEEE Trans Affect Comput 11(3):542–552 10. Al Jazaery M, Guo G (2018) Video-based depression level analysis by encoding deep spatiotemporal features. IEEE Trans Affect Comput 11. Thangaraj M, Ponmalar PP, Anuradha S (2015) Internet of Things (IOT) enabled smart autonomous hospital management system-a real world health care use case with the technology drivers. In 2015 IEEE international conference on computational intelligence and computing research (ICCIC). IEEE, pp 1–8 12. Islam MM, Rahaman A, Islam MR (2020) Development of smart healthcare monitoring system in IoT environment. SN Comput Sci 1:1–11 13. Priyadarshini R, Quadir MdA, Rajendran N, Neelanarayanan V, Sabireen H (2022) An enhanced encryption-based security framework in the CPS Cloud. J Cloud Comput 11(1):1–13 14. Md AQ, Varadarajan V, Mandal K (2019) Correction to: efficient algorithm for identification and cache based discovery of Cloud Services. Mob Netw Appl 24(4):1198–1198 15. Md AQ, Vijayakumar V (2020) Combined preference ranking algorithm for comparing and initial ranking of cloud services. Recent Adv Elect Electron Eng (Formerly Recent Patents on Electrical & Electronic Engineering) 13(2):260–275 16. Md AQ, Vijayakumar V (2019) Dynamic ranking of cloud services for web-based cloud communities: efficient algorithm for rating-based discovery and multi-level ranking of cloud services. Int J Web Based Commun 15(3):248–270 17. Mathew SA, Md AQ (2018) Evaluation of blockchain in capital market use-cases. Int J Web Portals 10(1):54–76 18. Christy Jackson J, Prassanna J, Quadir MdA, Sivakumar V (2022) Stock market analysis and prediction using time series analysis. Mater Today: Proc 19. Quadir MdA, Prassanna J, Jackson C, Anusooya G, Premalatha M, Sathyarajasekaran, Sabireen H (2020) A review on DNS amplification attacks towards building an efficient prevention strategy. J Crit Rev 13:34–43 20. Quadir MdA, Jaiswal D, Daftari J, Haneef S, Iwendi C, Jain SK (2022) Efficient dynamic phishing safeguard system using neural boost phishing protection. Electronics 11(19):3133 21. Kiran M, Matthew, Quadir A (2017) Computer network and information security. Comput Netw Inf Secur 2:36–42

IoT-Based Mental Health Monitoring System Using Machine Learning …

263

22. Panja A, Jackson C, Quadir MdA (2021) An approach to skin cancer detection using Keras and Tensorflow. J Phys: Conf Ser 1911(1):012032 23. Matthew KM, Muhammed AQ, Varadarajan V (2019) An improved key management scheme in cloud storage. Int J Adv Intell Paradigms 14(3/4):197 24. Prassanna J, Quadir AMd, Christy Jackson J, Prabakaran R (2020) SecrecyProtector: a novel data analytics based credit score management system. Int J Sci Technol Res 9(03) 25. Matthew KM, Quadir AMd (2018) Analysis framework for logs in communication devices. Int J Web Portals 10(1):15–26 26. Srinivasan A, Quadir MA, Vijayakumar V (2015) Hybrid cloud for educational sector. Procedia Comput Sci 50:37–41 27. Quadir MA et al (2020) An efficient algorithm to detect DDoS amplification attacks. J Intell Fuzzy Syst 39(6):8565–8572 28. Basit Z, Tabassum M, Sharma T, Furqan M, Quadir AMd (2022) Performance analysis of OSPF and EIGRP convergence through IPsec tunnel using Multi-homing BGP connection. Mater Today: Proc 29. Gupta B, Madan G, Quadir AMd (2022) A smart agriculture framework for IoT based plant decay detection using smart croft algorithm. Mater Today: Proc 30. Gini coefficient—Wikipedia. En.wikipedia.org (2021) https://en.wikipedia.org/wiki/Gini_coef ficient. Accessed 14 July 2021 31. Entropy (information theory)—Wikipedia. En.wikipedia.org (2021). https://en.wikipedia.org/ wiki/Entropy_(information_theory). Accessed 14 July 2021 32. Schmidt P, Reiss A, Duerichen R, Marberger C, Van Laerhoven K (2018) Introducing wesad, a multimodal dataset for wearable stress and affect detection. In Proceedings of the 20th ACM international conference on multimodal interaction, pp 400–408 33. Stewart CL, Folarin A, Dobson R (2020) Personalized acute stress classification from physiological signals with neural processes. arXiv:2002.04176 34. Lai K, Yanushkevich SN, Shmerko VP (2021) Intelligent stress monitoring assistant for first responders. IEEE Access 9:25314–25329. https://doi.org/10.1109/ACCESS.2021.3057578 35. Kumar A, Sharma K, Sharma A (2021) Hierarchical deep neural network for mental stress state detection using IoT based biomarkers. Pattern Recogn Lett 145:81–87

Fish Feeder System Using Internet of Things Jyotsna Malla, Preetham Lekkala, Rishi Raghu, J. Jayashree, J. Vijayashree, and Vicente Garcia Diaz

Abstract Pets have become quite common in almost every household. In the present busy world, it has become quite a hassle to take good care of the pets due to the lack of time and knowledge about their eating patterns and schedules. People having fish as pet need to devote more time and attention to them owing to their complex and timely food requirement. The goal of designing an IoT-based fish feeder is to eliminate manual labour involved in feeding the fish regularly at certain intervals of time. This proposed system can offer regular feeding without the interference of the owner. This system also helps to reduce costs by preventing the overspending on extra fish feed than required. This study is based on the concept that the fish will be fed the owner is not present at home. Wireless connectivity is used in the fish feeder system. The system can be built by programming a fish feeder to feed fish at a specific time, and then commanding it to deliver the food through stepper motors. It removes the existent need of a manual aquarium maintenance. The fish feeder system can be atomised and can be easily managed using a Graphical User Interface (GUI). Keywords Fish feeding system · Internet of things · Graphical user interface · Wireless · Monitoring system automation · Fish monitoring system · Aquaculture · Arduino · Sensors · Micro controller

J. Malla · P. Lekkala · R. Raghu · J. Jayashree (B) · J. Vijayashree School of Computer Science and Engineering, Vellore Institute of Technology, Katpadi, TN, India e-mail: [email protected] J. Vijayashree e-mail: [email protected] V. G. Diaz Department of Computer Science, University of Oviedo, Oviedo, Spain e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_17

265

266

J. Malla et al.

1 Introduction The modern world has a lot of distractions and having pets is a great way to ensure that people remain mindful and calm. Fish keeping is a popular way of having a pet that is low maintenance and easy to manage. However, there is one issue with fish. This is the process of fish feeding. It may seem to be easy but is quite complicated. Fish require a strict schedule to feed and any deviation can cause problems with the health of the fish. Numerous automated fish feed systems have been developed to help the fish owner in feeding tasks by automating them. These systems have limited functionalities and there is always a possibility of overfeeding or underfeeding the fish which can lead to severe consequences. There is also a lack of interaction of these systems with the owner [1, 5, 6, 8]. In the event, the owner leaves the home, feeding the fish is close to impossible as they cannot be transported. Neither can they be left in the care of someone else as delivering an entire aquarium is tough. This makes feeding them very difficult. Dropping extra food in the aquarium in hopes that it suffices for the entire duration of the leave is not possible either as this could result in the issue of overfeeding. This could also cause fatalities in the fish. The excess food also spoils the quality of the water through the release of ammonia as well as nitrites in the water due to decomposition [2, 3, 11, 15]. In order to combat this problem, the Internet of Things-based fish feeder for a home-based aquarium has been developed. This system automatically dispatches the fish feed to the aquarium once the container gets empty via a stepper motor. The owner can set the amount of fish feed required for the particular type of fish in the mobile application. The system continues to send fish feed till the daily requirement is met. The fish feeder system alerts the owner a day before the fish feed storage gets depleted. The owner can remotely operate the oxygen pump attached to the aquarium using the mobile application. The proposed system is also equipped with a water pump to filter the aquarium regularly in order to maintain proper hygiene for the fish. Monthly reports are generated in the mobile application regarding the quantity of fish feed consumed to reduce overspending. The proposed fish feeder system for a homebased aquarium has been found to eliminate the limitations of the existing systems and can be used to help automate most of the manual tasks to be performed by the owner such as feeding, changing water regularly and keeping the tank oxygenated. The remaining part of the paper is organised as follows. Section 2 contains the Literature Survey. The Proposed Model is discussed in Sect. 3. Section 4 contains the Experiments and Results. Lastly, the Conclusion and Future Directions is presented in Sect. 11.

Fish Feeder System Using Internet of Things

267

2 Literature Survey Limited research has been done in the field of automation of aquaculture tasks such as feeding, maintaining dissolved oxygen levels, regular water filtering and pumping and maintaining the optimum pH levels required for the fish. These help in understanding the existing systems to develop a better IoT-based feeder system for fish. The objective of the authors in [7] was to provide a hygienic environment to which would improve the breeding activity and help in producing good quality fish. This is done by maintaining the optimum water levels in the aquarium. Fish dies due to unhygienic and stagnant water. This prevalent problem was solved by ensuring regular pumping in the aquarium to remove the waste water and pump in clean water. The flow of water needs to be closely monitored. Several parameters need to constantly checked to ensure proper maintenance of the fish such as water flow speed, the dissolved oxygen concentration, the pH levels, optimum food requirement and ammonia level in the aquarium tank. The main aim of the system was to ensure the regular supply of fresh and flowing water into the aquarium without the interference of the fish owner. The authors suggested that ensuring proper water quality standard is the most crucial step in the proper development of aquaculture. Research work done in [2] was to develop and automated small scale aquaculture environment using IoT. The authors developed a prototype of a small aquaculture environment. The system was connected with IoT in order to automate the fish feeding task. The prototype was mainly concerned with automating the task of fish feeding in the aquaculture. The authors suggested the existing model can be equipped with an automatic water pumping system to remove the excreta from the water and pump in fresh water in the tank. This would help in ensuring proper hygiene of the fish. An automatic koi fish feeder system using IoT was proposed in [9]. The system automated the task of feeding the fish. The owner could remotely monitor the feeding activity and the number of food pellets being delivered to the fish was constantly displayed to the fish owner using a mobile application. The proposed system could also weigh the fish food and alert the owner if there is a deficit in the food storage tank. There are various sizes of fish food which can be given to the koi fish as per its requirement. The size of the pellets to be given to the fish can be remotely varied by the user in the mobile application. The motor would take up then dispatch the instructed size pellets to the tank. The authors in [1] designed an automated fish feeding system. The system is able to automate the feeding task owing to an underlying algorithm. The tank water quality parameters such dissolved oxygen concentration, ammonia levels and pH are taken from the installed sensors and the optimum quantity of fish feed required and the schedules for feeding is computed based on the type of the fish. Aquaponics is a combination of fish breeding and plant cultivation in the same environment. Maintaining proper water quality standard, proper pH, acidity and dissolved oxygen levels is necessary to stimulate fish breeding. In this study [13], an automated fish feeder system equipped with pH, and TDS monitoring module

268

J. Malla et al.

was developed. The user can remotely monitor the pH and TDS levels in the mobile application. The pH values were recorded using a pH IoT sensor, and the TDS values were recorded using a TDS IoT sensor. The user schedules the specific food weight and the food timings in the mobile application. The fish feeder system feeds the fish automatically according to the given specifications. Ornamental fish in the office is currently popular among fishers and fish farmers since it can help to produce a healthy environment in a given location. As a result, anglers must pay close attention to and control a variety of elements such as feeding, air pumps, light and water pH, all of which can impact water quality and cause fish to die. This research conducted in [12] produced an autonomously fish feeding system that operates through a smartphone application, thanks to the growth of IoT technology that is used to manage wireless equipment. By programming device control commands onto the microcontroller device, users can automatically manage the timing and amount of fish food based on the number of fish, as well as receive notifications when the pH value of the waterfalls below certain threshold water is not proper. The Internet of Things (IoT) is a technology that enables control, communication, collaboration and data transmission across the Internet network. Technology is now more important than ever for the development and use of natural resources, including fisheries. The goal of this study [6] is to see how IoT technology is progressing in aquaculture management. According to a literature review, the most common IoT for aquaculture is a water quality instrument, such as temperature, DO and salinity monitoring. Other IoT applications include automatic feeders, feed control devices, fish laureate design, a marketplace that connects fish farms and consumers, and aquaculture model design that uses IoT technologies. When people who are rare owners of pet fish are away from the fisheries field setting, they are frequently upset because they will not be able to feed them on time and check the field’s diverse circumstances. These factors may cause starving and overfeeding, endangering the fish’s health and resulting in poor water quality in the fishery. As a result, monitoring the fisheries area is quite important and useful to the owners. The authors in [14] propose a design that utilises the Raspberry Pi to execute a variety of tasks in the field. The fish feeder serves as the primary controller in the development of a fully automated feeder system that allows owners to customise their feeding process and monitoring various aspects of the field. Fish keeping has become popular among people in recent days. But the fish needs to take proper care and the food requirements and schedules must be precisely met which has become a problem for busy fish pet keepers. The authors in [10] proposed the FishTalk system which uses aquarium sensors to operate actuators in real time, based on an IoT solution dubbed IoT Talk. The link between aquarium sensors and actuators is described, with real examples of various water conditions provided. As an example, we implement an intelligent fish feeding mechanism that ensures that the fish are not over or underfed while also allowing the fish owner to enjoy remotely watching the fish feed. The authors in [4] proposed a IoT-based smart monitored aquarium system for ensuring fresh water flow in the aquarium tank. The system is used to monitor the

Fish Feeder System Using Internet of Things

269

water quality standard in the tanks to ensure proper hygiene for the fish. The owner can remotely monitor the water quality standard via the mobile application. The system is developed in Arduino and a mini micro controller. Wi-Fi module is used to bride the communication between the IoT system and the mobile application. Water pH levels are monitored using pH sensor. This system can be used to monitor fish efficiently. This system can be developed on a larger scale which can be economically beneficial for the country.

3 Proposed Model 3.1 Fish Feeder System Module Overview There are three components in our system on which the core project idea is based. Figure 1 shows the architecture diagram of the proposed fish feeder system. Figure 6 shows the circuit diagram of the proposed feeder system. The modules are listed in the table below. Module 1: Storage and Dispatch Module 2. Module for Wi-Fi Module 3. Graphical User Interface (GUI). Module 4. Notifications and Alerts Module 5. Generating Reports. These modules have the potential to make the system more intelligent and efficient. Module for Storage and Dispatch: The storage module informs the user whether there is an enough amount of food in the container. Then send container, keeping in mind that the food will be dispatched as soon as the container is empty. Module for Wi-Fi manages the system’s Internet access and allows it to stay connected at the user’s demand. The Arduino Uno server accepts the command, Fig. 1 Architecture diagram of the proposed fish feeder system

270

J. Malla et al.

Fig. 2 Circuit diagram of the proposed fish feeder system

making the Wi-Fi module the brains of our system, allowing us to connect with both the user and the system. Graphical User Interface: The GUI module is intended to allow the user to remain connected to the system and provide commands to it. It is important to have when running the software on multiple devices in a cross-platform environment. Alerts and Notifications: The notification and alert module send a notification to the user if there is a shortage of food in storage. The message arrives a day before the food storage runs out, saving the owner time and ensuring that he never runs out of supply. Generating Reports: The report will be created, as well as an acknowledgement for the user, so that the owner may acquire an overall pricing of fish food. How much fish food do you think you’ll need? The owner will also receive an overall cost estimate. The programme intelligently saves all generated values for cost and total fish food eaten throughout the feeding process, as well as the number of fish fed during the procedure over a monthly period. A database is used to store all of the essential values (Fig. 2).

3.2 Design and Development of Feeder System The proposed prototype involves a small fish pond, installed food dispatching machine, required controller kits, oxygen motor and pump, water purifying pump and a mobile application installed on owner’s mobile. These were the components used in the development of an automatic fish feeder system using IoT. Micro controller NodeMCU NodeMCU is a microcontroller with analogue and digital capabilities similar to Arduino. UART, SPI and I2C are also supported. I2C may be used by NodeMCU to communicate with a powered LCD display, a magnetometer, an accelerometer, GPS module, MPU-6050 Gyro metre, RTC chips, SD

Fish Feeder System Using Internet of Things

271

cards, screen displays, etc. Wi-Fi interfaces and pins were used to bridge the communication between the NodeMCU board and the other devices which can be used by feeders for its tasks. Relay: The relay module of the microcontroller is a digital switch that allows the microcontroller to function with high-voltage equipment. This module has a 10A electrical capacity. It may be used in both direct and alternating circuits. It can withstand the NodeMCU’s 5 V pressure. The relay module’s status is shown via an LED. The relay module is utilised in turning ON and OFF the oxygen and water pump of the tank. Water Pump: The water in the aquarium tank is purified using a water pump. It has a 55-W motor and a 4500-L-per-hour pump. Food is consumed by fish, who then excrete it. As a result, the feeders must filter the water. Oxygen Pump: It is of utmost importance to ensure that the fish tank is oxygenated for the fish to breathe properly. Through the pipe, this equipment will provide dissolved oxygen to the tank. Because the water passes through a filter, air will bubble into it. While the bubbles rise to the surface, oxygen will dissolve in the water. As a result, the number of bubbles produced by the oxygen pump is determined by the type of filter utilised. The dissolved oxygen concentration increases with the depth of the fish pond. The air pump boosts oxygen levels in the pond, allowing fish to develop at a healthy pace. Wireless Internet at Home: The Internet service providers can provide home WiFi which needs to be installed where in houses. Nowadays, there are many Internet service providers available in the market. fibre Internet is more reliable and faster than the non-fibre Internet sources. In fibre Internet, the supplier installs a dual-band Wi-Fi router with two 2.4 or 5 GHz signal ports. This allows the appliances to be connected to the Internet. A wide range of appliances can be connected to the Internet such as mobile phones, TVs, computers, ACs, laptops and refrigerators.

3.3 Feeder System Flow The proposed home-based fish feeder system consists of three components which are, namely the fish feed dispatch module, the oxygen pump module and the water filter pump module. Figure 3 shows the flow chart of the proposed fish feeder system. Figure 4 shows the communication between the feeder system and the user via the mobile application. The Wi-Fi router is used to send the information gathered from the ultrasonic sensors to the Internet which is then displayed in the mobile application installed in the user mobile. The proposed feeder system main advantage is its automatic fish feed dispatch to the tank. Figure 5 shows the feeder system fish feed dispatch system. The ultrasonic sensor calculates the amount of food in the container. Once the container gets empty, the stepper motor starts to dispatch fish feed from the storage to the tank automatically without the interference of the owner. The motor status is displayed as ON in the mobile application. Additional alert is given to the user regarding the food dispatch. The owner can remotely now monitor

272

J. Malla et al.

Fig. 3 Flow chart of the fish feeder system

the food dispatch without manually doing it. Figure 6 shows the user interaction with the mobile application. The user can remotely monitor many automated tasks using the mobile application. Reports are generated for the daily and monthly fish feed quantity consumed. This helps the user from overspending on fish feed than required and helps increase savings. The ultrasonic sensor gives alert to the user a day before the fish feed storage runs out. This helps the user to ensure that the fish are not deprived of food anytime. The owner can also switch ON and OFF the oxygen tank using the mobile application. This helps to ensure that the fish get the required amount of dissolved oxygen in the water for their proper functioning. The water filter pump can also be switched ON and OFF using the mobile application. This ensures that the excreta from the fish is regularly pumped out and fresh water is pumped into the tank. This helps in maintaining proper hygienic conditions for the fish in the tank.

4 Experiments and Results 4.1 Dispatching Fish Feed Experimentation The fish food dispensing system has an ultrasonic sensor that measures the quantity of food consumed. The sensor calculates the distance between the fish feed in the container and the sensor to record the quantity of fish feed consumed. The amount of fish feed consumed is updated dynamically in the mobile application for the user. The

Fish Feeder System Using Internet of Things

273

Fig. 4 Fish feeder system connected to the mobile application

Fig. 5 Fish feed dispatching system

stepper motor is mainly responsible for dispatching the fish feed from the storage to the aquarium tank. The automatic fish feeder was used to feed some catfishes in the fish tank for 14 days. The number of fishes and the weights for each category are depicted in Table 1. The fish was placed across 21 different aquariums for with one for each weight category.

274

J. Malla et al.

Fig. 6 User interaction with the fish feeder system

Table 2 gives the number of fish dead in the 14 days period of automation of the fish feeder system. Survival rate of the fishes in the tank was 93% that is 8 dead fishes in the 14 day period.

4.2 Oxygen Pump and Water Pump Experimentation The user can remotely switch on and off the oxygen and water filtering pump via the mobile application. The user can check the dissolved oxygen concentration and switch ON the oxygen tank in the mobile application. The oxygen pump starts dispersing dissolved oxygen into the tank automatically. Similarly at periodic intervals, the user can switch ON the water filtering pump from the mobile application. The water pump removes the waste water from the tank and pumps in fresh water.

4.3 User Interaction with the Mobile Application The mobile application allows the user to remotely monitor and automate the manual tasks involved in fish keeping. Figure 7 shows the login page for FishCare mobile application. A new user can sign up for a new account and an existing user can login using his credentials. Figure 8 shows the homepage of the mobile application. The user can choose which automated function he wants to monitor at a particular instant

Fish Feeder System Using Internet of Things Table 1 Total number of fishes and their weights

Table 2 Total dead fish count

Sl. no

275 Weight

Number of fish

1

125

8

2

135

7

3

145

8

4

155

6

5

165

6

6

175

5

7

185

4

8

205

7

9

215

3

10

225

6

11

235

5

12

245

8

13

255

7

14

285

6

15

305

4

16

345

5

17

395

5

18

405

4

19

415

6

20

425

4

21

445

5

Total

5490

114

Sl. no

Date

Number of fish

1

4 March 2022

2

2

10 March 2022

1

3

11 March 2022

1

4

12 March 2022

0

5

14 March 2022

1

6

15 March 2022

2

7

17 March 2022

1

Total

8

of time. Figure 9 shows the feeder monitoring page. The page shows the quantity of fish feed consumed dynamically and shows the motor status in the page. It also shows the fish feed storage left. When the fish feed storage reached, refill status becomes 100% an alert is sent to the user a day before to ensure the storage is refilled before

276

J. Malla et al.

Fig. 7 Login page of the mobile application

Fig. 8 Home page of the mobile application

the supply completely runs out. Figure 10 shows the oxygen pump automation page. The user can switch ON and OFF the oxygen tank using the mobile application.

5 Conclusion and Future Directions Fish keeping or fish farming requires a lot of care and attention by the owners. This has become quite difficult in the present busy world. The proposed fish feeder system is able to perform all the basic functionalities and ease the owner from doing all the

Fish Feeder System Using Internet of Things

277

Fig. 9 Fish feeder dispatch monitoring page

Fig. 10 Oxygen pump automation page

manual tasks himself. The proposed system dispatches the food automatically to the fish and informs the user in case of shortage of supply. Additionally, the manual tasks of oxygenating the tank and filtering the water in the tank are eliminated with the fish feeder system. The system is able to perform more functionalities and eliminate most of the manual tasks of the fish keeping owners than the existing state-of-art systems. The proposed model can incorporate certain new functionalities such as allowing any form of fish food to be used in the system. Medium-sized pellets are just one of the numerous types of fish food available. Our design, in a future iteration, should be able to fit any type of food, allowing customers to use what they already have rather of buying more.

278

J. Malla et al.

References 1. Adegboye MA, Aibinu AM, Kolo JG, Aliyu I, Folorunso TA, Lee SH (2020) Incorporating intelligence in fish feeding system for dispensing feed based on fish feeding intensity. IEEE Access 8:91948–91960. https://doi.org/10.1109/ACCESS.2020.2994442 2. Binti Hasim HN, Ramalingam M, Ernawan FRP (2017) Developing fish feeder system using Raspberry Pi. In: 2017 third international conference on advances in electrical, electronics, information, communication and bio-informatics (AEEICB), pp 246–250. https://doi.org/10. 1109/AEEICB.2017.7972422 3. Dada E (2018) Arduino UNO microcontroller based automatic fish feeder 4. Deepa TP, Khadka B, Chatterjee N, Rahul SN, Sridhar SK (2022) Fish tank monitoring system using IoT. In: García Márquez FP (ed) International conference on intelligent emerging methods of artificial intelligence & cloud computing. Springer International Publishing, Cham, pp 306– 314 5. El Shal AM, El Sheikh FM, Elsbaay AM (2021) Design and fabrication of an automatic fish feeder prototype suits tilapia tanks. Fishes 6(4). https://doi.org/10.3390/fishes6040074. https:// www.mdpi.com/2410-3888/6/4/74 6. Idachaba F, Olowoleni O, Ibhaze A, Oni O (2017) IoT enabled real-time fishpond management system 7. John JRMP (2021) Automated fish feed detection in IoT based aquaponics system. In: 2021 8th international conference on smart computing and communications (ICSCC), pp 286–290. https://doi.org/10.1109/ICSCC51209.2021.9528186 8. Lee C, Wang YJ (2020) Development of a cloud-based IoT monitoring system for fish metabolism and activity in aquaponics. Aquacult Eng 90:102067. https://doi.org/10.1016/j. aquaeng.2020.102067 9. Lekswina F, Widjaja D (2021) IoT based two levels feeding system for koi fish pond. IOP conference series: materials science and engineering 1115:12051. https://doi.org/10.1088/175 7899X/1115/1/012051 10. Lin YB, Tseng HC (2019) FishTalk: an IoT-Based mini aquarium system. IEEE Access 7:35457–35469. https://doi.org/10.1109/ACCESS.2019.2905017 11. Noor MZH, Hussian AK, Saaid MF, Ali MSAM, Zolkapli M (2012) The design and development of automatic fish feeder system using PIC microcontroller. In: 2012 IEEE control and system graduate research colloquium, pp 343–347. https://doi.org/10.1109/ICSGRC.2012.628 7189 12. Pasha Mohd Daud AK, Sulaiman NA, Mohamad Yusof YW, Kassim M (2020) An IoT-Based smart aquarium monitoring system. In: 2020 IEEE 10th symposium on computer applications industrial electronics (ISCAIE), pp. 277–282. https://doi.org/10.1109/ISCAIE47305.2020.910 8823 13. Riansyah A, Mardiati R, Effendi MR, Ismail N (2020) Fish feeding automation and aquaponics monitoring system base on IoT. In: 2020 6th international conference on wireless and telematics (ICWT), pp 1–4. https://doi.org/10.1109/ICWT50448.2020.9243620 14. Rizal A, Aditya G, Nurdiansyah H (2021) Fish feeder for aquaculture with fish feed remaining and feed out monitoring system based on IoT. Procedia Eng Life Sci 1. https://doi.org/10. 21070/pels.v1i2.983 15. Uddin M, Rashid M, Mostafa M, Salam S, Nithe N, Rahman M, Aziz A (2016) Development of automatic fish feeder. Glob J Res Eng 16:15–23

An Efficient and Recoverable Symmetric Data Aggregation Approach for Ensuring the Content Privacy of Internet of Things L. Mary Shamala, V. R. Balasaraswathi, M. Shobana, G. Zayaraz, R. Radhika, and Thankaraja Raja Sree

Abstract Internet of Things devices collect many personnel information such as user identity, behavior, location, which may disclose a lot about end-user’s day-today activity. Though IoT offers tremendous economic values, the data gathered by IoT devices produce significant privacy and security issues. The intermediate honest but inquisitive nodes help IoT devices in the data collection as well as aggregation processes due to their resource constraints. This further aggravates the privacy and security issues resulting in life-threatening activities by attackers. Aiming to secure the privacy of end-user data gathered by IoT devices, this article provides a simple, privacy-preserving solution for data aggregation. The sensitive data gathered by IoT devices are enciphered using lightweight symmetric functions in particular modular addition, bitwise rotation, bitwise exclusive OR, and hash. The intermediate aggregator node performs aggregation operations on the concealed data without decrypting it. As a result, nothing is revealed to the curious aggregator nor it is compromised. It employs a recoverable end-to-end aggregation to protect data integrity. Simulation results show that the proposed method decreases the storage, communication, and computation overhead, and hence, an efficient scheme is to be used in IoT-enabled networks. L. Mary Shamala (B) · T. Raja Sree Vellore Institute of Technology, Chennai, India e-mail: [email protected] T. Raja Sree e-mail: [email protected] V. R. Balasaraswathi · M. Shobana · R. Radhika SRM Institute of Science and Technology, Chennai, India e-mail: [email protected] M. Shobana e-mail: [email protected] R. Radhika e-mail: [email protected] G. Zayaraz Puducherry Technological University, Puducherry, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_18

279

280

L. Mary Shamala et al.

Keywords Content-oriented privacy · Symmetric data aggregation · End-to-end encryption · IoT security · IoT devices · Data collection · Lightweight cryptography

1 Introduction The current era of Internet of Things enables any object or thing to get connected to the Internet, thereby allowing anyone to enjoy network-based services. The foremost benefit of IoT is the extensive influence it will have on quite a few facets of users’ day-to-day interactions and immediate surroundings. This will ultimately enhance our standard of living in various fields. However, the highly varied sphere of applications, intrinsic resource-limited devices, and divergence platforms directly affect the choice of technology pertinent to sustain user privacy. IoT devices collect many personnel information such as user identity, location, which may divulge lots of user’s day-to-day activity [1]. Due to the security threat imposed by IoT devices, privacy is becoming a greater concern in IoT setup [2]. An intruder can not only collect personnel information about users but can also control their environment leading to huge security outbreaks. IoT privacy aims at ensuring that sensitive data of individuals remain safe from getting revealed and misused in IoT environment [3]. The constantly growing IoT devices generate a huge volume of data. As IoT devices are resource-limited, cloud servers and fog nodes usually assist them in ensuring efficient data collection and aggregation process. Since the data are handled by untrusted environments, it brings additional privacy and security threats, ever since the merits of these computing systems. The traditional way to secure against these threats is through data encryption. Also, concealed data aggregation is applied to the data received from different IoT devices to save a considerable amount of power or energy, thereby extending the lifespan of IoT devices. However, the process of data aggregation hinders the security benefits offered by security mechanisms as it is carried out by intermediate curious aggregator nodes [4]. To address the aforementioned issues, several interesting solutions have been identified by researchers to perform secure data aggregation on the data collected by IoT devices [2] [5]. Most of the existing data aggregation solutions for content privacy are found to be insecure against disclosure of either the private information at IoT nodes or confidential data under decryption at aggregator nodes. Besides, most of them incur additional overhead in respect of computation along with communication costs. Since IoT-enabled systems are resource-limited owing to a limited battery charge, communication bandwidth as well as storage space, they are not suitable for IoT environment. Hence, end-to-end privacy conserving symmetric data aggregation to protect the content of individual IoT devices is presented in this paper. Our scheme focuses on protecting the content privacy of collected data by IoT devices. This is achieved by using symmetric encryption of IoT-reported data which is then aggregated by the intermediate aggregator nodes without decrypting it. Thus,

An Efficient and Recoverable Symmetric Data Aggregation Approach …

281

it performs efficient data aggregation by saving data transmission energy. Also, the freshness of data is ensured by adding timestamps to every transmitted data over the public channel. The reminder of the paper is organized in the following manner. Section 2 examines the underlying work. Section 3 defines the core terminologies and system model used that will be used in the paper. Section 4 describes the suggested work. The proposed system’s performance is analyzed and compared to related solutions in Sect. 5. Finally, in Sect. 6, we bring our work to close.

2 Related Work Alghamdi et al. [5] presented a secure data aggregation scheme in wireless sensor networks for Internet of Things. Their solution adopts an elliptic curve-based algorithm and Hilbert curve-based data transformation concealing the sensor data, thus making it impossible for an adversary to eavesdrop the communication. Through performance analysis, the authors claimed that the scheme improves the lifespan of network and preserves the privacy of participating nodes. For remote health monitoring systems, the authors of [6] presented the secure privacy-preserving data aggregation (SPPDA) scheme. It uses bilinear pairing to achieve privacy as well as improved data aggregation efficiency. Underneath the decision-making bilinear Diffie Hellman assumption, this technique is confirmed to be reliable. However, the scheme suffers from communication and computational overhead due to homomorphic operation. In [7], the authors designed a lightweight data aggregation scheme (LPDA) considering the fog computing-enhanced devices. Their method aggregates the collected data from various entities by using the Chinese Remainder Theorem and Paillier cryptosystem. The scheme relies on one-way hash chain for ensuring source authentication. Though proved to be lightweight, the solution is not evaluated under real IoT applications and adapts a weak threat model. Chunqiang Hu et al. [8] introduced an efficient and reliable data aggregation method for Internet of Things. Their solution preserves sensitive data of users by slicing and confusing the data before they are reported to the aggregator. The analysis shows that it reduces the computational cost and improves communication efficiency. The problem is with the additional overhead of frequent data exchange between IoT devices making it unsuitable for IoT devices with limited resources. For cloudenabled IoT devices, the authors of [9] framed a simple data aggregation strategy. ElGamal encryption and identity-based signing are used for secure aggregation by cloud. The security study demonstrates that this method achieves privacy with partial homomorphic encryption and is safe from false data injection. The efficiency of the scheme against existing schemes is also verified, but it fails to provide authentication features. For use in smart grid contexts, Lyu et al. [10] developed a privacy-preserving fog-enabled aggregation (PPFA). This approach relies on fog computing for doing complicated calculations, thereby consumes more resources, and hence, is unsuitable for use in lightweight IoT devices.

282

L. Mary Shamala et al.

Sunday Oyinlola Ogundoyin [11] presents a simple aggregation signature method for IoT environments. This scheme adopts a set homomorphic signature approach for aggregating data received from IoT devices. By doing do, the aggregator node does not require the knowledge of private keys of participating entities. The problem with this method is data exchange between the different participating IoT devices takes place in plaintext form, thereby leaks the private data of customers. In [12], Li et al. implemented the n-out-of 1 OT algorithm that can be used for secure data aggregation in smart grids and many other practical applications. The authors constructed hidden permutation circuits for implementing an anonymous communication system. A rigorous theoretical analysis is also given for identifying the performance and security strength of the protocol, but experimental results were not given. Pu et al. [13] presented two effective and secure data aggregation strategies for user privacy protection. Both the schemes support plug and play as well as mixing the device data while reporting ensures the privacy of data. The first technique used the Paillier cryptosystem and AES encryption to ensure the secrecy and integrity of the data gathered. In the second technique, noise data are combined with the actual user data rather than being encrypted. Deployment of the schemes in the practical IoT scenario is not explored. In 2018, a novel fog-assisted healthcare data aggregation (EHDA) is presented in [14]. Specifically, privacy is implemented by two different algorithms, namely the message receiving algorithm to generate aggregated data and the message extraction algorithm to extract the aggregated message at the fog server. Further compression is performed to lower communication costs. But, the dynamic topology of the WSNs has not been considered by this approach. A technique for authentication and Anonymous Privacy-Preserving Aggregation (APPA) in fog-enhanced IoT networks was recently published in [2]. It uses Paillier algorithm to protect data privacy during aggregation. Security and execution analysis of APPA demonstrates the suitability of the approach in fog-enhanced IoT systems with privacy preservation properties. The comparison of the scheme with related works shows a decrease in computation and communication overhead. In summary, researchers have identified many techniques for doing secure data aggregation on the data collected by IoT devices. Few of them either apply privacy at IoT devices only or leak aggregation results to intermediate nodes. Further, most of the available solutions employ asymmetric primitives such as elliptic curve cryptography, bilinear pairing, Paillier cryptosystem, and homomorphic property of ElGamal cryptosystem. These solutions incur storage costs, communication, and computation overhead. Since IoT-enabled gadgets are resource-limited in respect of energy, communication bandwidth as well as storage space, they are not suitable for IoT environment. This study presents a data aggregation strategy that is both efficient and privacy-preserving in which collected IoT device data are concealed using lightweight symmetric encryption preceding aggregation.

An Efficient and Recoverable Symmetric Data Aggregation Approach …

283

3 Preliminaries This section briefly reviews few important terminologies used in the proposed work along with the system model.

3.1 Content-Oriented Privacy and Data Aggregation The protection of content privacy is given a lot of consideration in the Internet of Things context. Secret details on individuals, companies, and high-value assets are contained in the data that the network collects and transmits. By its very nature, content-oriented privacy protects these data from intruders and attackers. There are two instances in which basic information manipulation and encryption mechanisms are not sufficient to ensure the content-oriented privacy, data aggregation and data query [15]. The activity of merging data collected from multiple IoT devices when data move toward the control server by carrying out operations such as maximum, minimum, sum, and average is called data aggregation [4]. The process of data aggregation reduces the computation overhead of the control server from processing a large number of messages and also lowers the communication overhead and power consumption of IoT devices [16]. Therefore, it is an important technique for increasing the lifespan and efficiency of IoT-enabled networks as nodes are typically resource-constrained [17]. The structure of the typical data aggregation activity is depicted in Fig. 1. Here, the local gateway usually acts as the aggregator node. The two main categories of data aggregation methods are end-to-end aggregation and hop-by-hop aggregation [6]. The aggregator node in hop-by-hop aggregation must first decipher the ciphertext before aggregating it in plaintext form. It is a computationally expensive approach as well as each node must exchange secret keys with every other node which in turn requires complex key management. Further, the

Fig. 1 Architecture of data aggregation process

284

L. Mary Shamala et al.

decryption of data by intermediate nodes may reveal sensitive data to the curious aggregator node. The aggregator nodes in end-to-end aggregation execute the aggregation function on ciphertexts without decrypting it [18]. End-to-end data aggregation is also called as Concealed Data Aggregation (CDA), which saves data transmission time and therefore is a desirable choice to be adopted in IoT-enabled networks [19] [20]. By CDA, privacy is maintained among the connected systems and the control server as the personal details of the device are not stored at the intermediate aggregator nodes.

3.2 System Model Our system model adopts a three-level, tree-based structure and is represented in Fig. 2. Control server (CS), aggregator node (AN), and IoT device (D) are the three types of nodes in the system model. IoT device (D). IoT devices denoted as {D1 , D2 , D3 ,…, Dn } are things or objects such as sensors, smart meters, health monitors, smart TVs, RFID readers, vehicle trackers, etc. These devices communicate and exchange data over the network. It is the responsibility of IoT devices to report sensed data to the aggregator node. Devices are typically resource-limited in respect of storage, computation capabilities, and battery power [21].

Fig. 2 System model of proposed data aggregation

An Efficient and Recoverable Symmetric Data Aggregation Approach …

285

Aggregator node (AN). Aggregators are denoted as {A1 , A2 , A3 ,…, Am }. The local network gateway usually acts as the aggregator node. IoT devices are connected to the aggregator through wireless communication. It is responsible for collecting the sensed data from IoT devices into an integrated unit; performs aggregation on them; and reports the aggregated result to the control server. They are semi-trusted but honest entities that are curious about the individual IoT device data. Control server (CS). Control server is an individual powerful server in a remote location that is maintained by a trusted authority. It provides storage for collected data by IoT devices which can then be retrieved by the application users.

4 Privacy-Protecting Symmetric Data Aggregation Method The issues identified with the prevalent solutions are addressed by the proposed privacy-protecting symmetric data aggregation (PPSDA) method. The requirement of content-oriented privacy between IoT devices, aggregator node, and control server is achieved by aggregating the sensed data in encrypted form, thus protecting the contents of data from curious intermediate nodes or any other adversary. The suggested method is predicated on the assumption that connected devices are setup with the relevant private keys since smart objects have limited capacity for storage, computing, and communications. Important symbols employed in the proposed approach are listed in Table 1. Our scheme performs secure data aggregation through three different stages: the initialization and data generation phase, the data aggregation phase, and the data extraction phase. The data generation phase is performed by the IoT nodes, whereby Table 1 List of symbols

Symbol

Description

IDi

Unique identity of the ith IoT device

ANj

jth aggregator node

CS

Control server node

Ki

The pre-shared secret key of IoT device

Kj

The pre-shared secret key of AN

T

Timestamp generated by the IoT device

σi

Encrypted message of ith IoT device

σ Agg

Aggregated data of AN

Mi

Sensed data by IDi

Ek (M)

Lightweight encryption function

h()

Secure hash function

⊕

Bitwise exclusive OR operation

(A, B)

Concatenation of data A and data B

286

L. Mary Shamala et al.

it performs its role of sensing the data as well as reporting it to the intermediate nodes (AN in this case) at a regular interval of time. Before data transmission, initialization of system parameters such as the establishment of permanent identity, secret keys for data exchange with CS is done. Further, IoT devices encrypt the generated data by using a lightweight symmetric encryption algorithm and transmit the generated data to the AN. During the data aggregation phase, AN performs an aggregation function on the gathered ciphertext data from various IoT devices and reports the aggregated result to the CS. Finally, the data extraction phase is carried out by the CS on the aggregated data received from the AN. In this process, the CS extracts individually encrypted IoT device data and then either decrypts using a pre-shared secret key with the corresponding IoT device for further processing and/or stores it in its local repository for access by application users. The details of the three phases of the scheme are discussed below.

4.1 Initialization and Data Generation The system administrator generates the master key for the control server and stores it in its memory. The master key in turn is used to establish secret parameters for further communication with the variety of IoT devices registered with CS and preconfigured in their memory. Under the assumption that sensitive data from the IoT devices are reported to the AN at periodic time slots, each IoT device IDi collects the data, forms them into message Mi, encrypts using lightweight symmetric algorithm with a preshared key Ki to form the ciphertext Ci = E K i (Mi ). It then generates the timestamp T and computes the hash value Ai = h(Ci , T ). Finally, the message σi containing device ID (IDi ), encrypted sensed data Ci , hash value Ai are all concatenated together with the timestamp T which is transmitted to the nearby AN. The detail of the message flow in the proposed privacy-protecting data aggregation method is illustrated in Fig. 3.

4.2 Data Aggregation The data aggregation scheme is applied to ciphertext Ci instead of message Mi to ensure end-to-end security and not to disclose the message Mi at AN. Even when AN is aware of the ciphertext, it is challenging for an opponent to decipher the message’s original content. Thus, data privacy is maintained by the aggregation process. The pseudocode of the data aggregation activity is elaborated in Algorithm 1. The data aggregation algorithm takes the set of encrypted data chunks {σi }, where i ∈ [1, n] as input from different participating IoT devices. It generates an aggregated encrypted message σ Agg for a global message as output.

An Efficient and Recoverable Symmetric Data Aggregation Approach …

287

Fig. 3 Message flow of PPSDA scheme

The data aggregation algorithm is executed by the AN which collects sensed data from IoT devices {D1 , D2 , D3 ,…, Dn } connected to it. Normally, the aggregation operation is carried out by taking a single message at a time. Firstly, it receives the encrypted message σi from the device IDi. as {σ i = (I D i , Ci , Ai , T )}. Then, checks for the condition that Tnew −T < Δt holds or not. Here, T represents message generation time, Tnew is the time at which message was received, while Δt is the permissible communication delay. This step helps to maintain the freshness of the message and thus prevents replay attacks by an adversary. AN next recomputes the hash value Ai ’ = h(Ci ,T) and checks with the received one Ai . This step ensures the integrity of the received message so that the tampering of data by an adversary can be identified. It terminates the session if the check fails and discards the received message. On successful integrity check, AN performs aggregation by concatenating the encrypted data Ci with the previously aggregated result as {σ Agg = (σ Agg , σi )}. In this manner, it aggregates all collected data by various IoT devices, encrypts σ Agg by Advanced Encryption Standard algorithm [22] with the secret key Kj as σ Sign Agg = E K j (σ Agg ), and finally, transmits the aggregated ciphertext message σ Sign Agg along with timestamp T1 to the CS as {σ Sign Agg , T1 }.

288

L. Mary Shamala et al.

Algorithm 1 Data aggregation algorithm

4.3 Data Extraction The data extraction phase is implemented at the control server (CS). The CS receives aggregated messages from all aggregator nodes. The data extraction algorithm is applied to each of the reported messages and is detailed in Algorithm 2. It takes an aggregated message σ Sign Agg as input and outputs individual IoT device data {M1 , M2 ,…,Mn }. Firstly, checks for the freshness of the message with the condition that Tnew − T1 < Δt holds or not. Here, T1 represents message generation time, Tnew is the time at which message was received, while Δt is the permissible communication delay. It then decrypts the aggregated ciphertext σ Sign Agg to retrieve the aggregated data σ Agg = E K j (σ Sign Agg ). Individual data {σ i = I D i , Ci , Ai , T } is identified with the delimiter and the extraction process are applied to every chunk of data as follows. CS checks that if I D i is a valid IoT device. If so, retrieves its secret key from the CS database. Next, it determines the hash value Ai ’ = h(Ci , T) and checks with the received hash Ai . Upon a successful integrity check, CS proceeds with the extraction step as Mi = D K i (Ci ). Finally, CS stores the data in its local storage for further processing and/or access by a trusted authority.

An Efficient and Recoverable Symmetric Data Aggregation Approach …

289

Algorithm 2 Data extraction algorithm

5 Performance Analysis In terms of storage expense, communication burden, and computational complexity, the efficacy of the suggested secure data aggregation strategy is assessed in this section. Additionally, a comparison with pertinent techniques like APPA [2] and ASAS [22] is discussed to demonstrate its effectiveness. For evaluation, we conducted all experiments with MIRACL C/C + + libraries and Java on an Intel Core i5 PC with a 2.5 GHz processor and 4 GB RAM executing on a 64-bit Windows operating system.

5.1 The Storage Cost In the proposed work, each IoT device needs to store IDi, ANj , the identity of CS, and the symmetric key Ki . In total, the memory usage of each device is 32 + 32

290 Table 2 Storage cost of PPSDA scheme

L. Mary Shamala et al. Node

Storage cost (in bits)

IoT device

448

AN

32n + 160

CS

160(n + m)

Fig. 4 Comparison of storage cost

+ 32 + 128 = 448 bits. Each AN needs to store the ID of IoT devices, secret key Kj, and the ID of CS, which occupies (32n + 160) bits. Here, n denotes the number of devices managed by them. The CS needs to store the ID of IoT devices, ANs as well as the secret parameters Ki and Kj of all devices and ANs. Assuming there are m aggregator nodes and n devices per AN are connected to the network, the entire storage cost of CS is about [160(n + m)]. Table 2 provides the outline of storage cost of our scheme. A comparison of proposed work for storage overhead by varying the message sizes is done with ASAS and APPA schemes as shown in Fig. 4. Hence, PPSDA achieves better storage as compared to ASAS and APPA methods.

5.2 The Communication Cost In the proposed lightweight aggregation method, communication cost can be determined in two steps: (i) communication from ith IoT device IDi to ANj , the jth aggregator node, and (ii) communication from ANj to the control server CS that can be represented as ID-to-AN and AN-to-CS, respectively. At first, the IoT device creates the data collection report as {σi = (I D i , Ci , Ai , T )} and sends it to AN. The report size is denoted as Sσi = |I D i | + |Ci | + |Ai | + |T |, , which acquires the

An Efficient and Recoverable Symmetric Data Aggregation Approach …

291

Fig. 5 Comparison of communication cost

communication cost of (32 + 128 + 256 + 32) = 448 bits. Assuming n number of IoT devices connected |to AN, | the communication cost of an aggregator node is calculated as n × |Ci | + | A j | + |T | which takes n × 416 bits. Secondly, { }the AN sends the aggregated result {σ Agg , Ai , T } in the encrypted form σ Sign Agg to the CS. This step takes 128n + 160 bits. Hence, the total communication cost of AN is T otCC AN = 544n + 160bit. Figure 5 elucidates the comparison of communication overhead of the APPA, ASAS, and PPSDA schemes against the total number of connected devices. From the plot, it is quite clear that PPSDA has relatively lower communication overhead when compared with APPA and ASAS schemes.

5.3 The Computation Complexity The asymptotic time required for performing one XOR operation, AND operation (modular addition), rotation operation, and the hash function is indicated by the symbols Txor , Tadd , Trot , and Th , respectively. The computational complexity involves encryption at IoT devices, data aggregation at AN, and extraction by CS. The computation cost of the major components of the proposed scheme is mentioned in Table 3. By contrasting the aggregated findings with the non-aggregated results, our method’s effectiveness is shown in Fig. 6 for 1 to 20 users. Contrary to the nonaggregated approach, where the number of users grows rapidly, the computing overhead of the proposed aggregated scheme advances gradually. Even though the complexity of the suggested scheme advances proportionally to the number of users,

292

L. Mary Shamala et al.

Table 3 Computation complexity Node type

Phase involved

Computation complexity

IoT device

Initialization and data generation

4Tadd + 2Txor + 6Tr ot + Th

AN

Data aggregation

Txor + nTh

CS

Data extraction

4T add + 7T xor + 6Tr ot + nTh

Total cost

8T add + 14T xor + 12Tr ot + (2n + 1)Th

Fig. 6 Total computation cost of PPSDA scheme vs. non-aggregated scheme

the total computing cost is reduced than those of non-aggregated schemes, confirming its efficiency. By contrasting the suggested method for computation complexity with the most recent ASAS and APPA, the effectiveness of the proposed method is also assessed. Figure 7 depicts the comparison of computation overhead with other schemes. The ASAS and APPA systems have execution times of 198 and 187 ms, respectively, for 50 users. Although the proposed scheme rises linearly in length, the least computing effort for 50 users is 178 ms, which is less than comparable designs. Furthermore, such techniques have greater initialization and data generation parameters than our scheme. From Fig. 6, it is evident that the proposed approach is more effective in IoT systems as it is built with a wide range of resource-limited IoT devices.

An Efficient and Recoverable Symmetric Data Aggregation Approach …

293

Fig. 7 Comparison of computational complexity

6 Conclusion The ever-growing IoT devices collect many personnel information such as user identity, location, which can divulge a lot about the user’s day-to-day activity. Despite the enormous economic advantages of IoT, privacy and security concerns are its primary concern. A lightweight privacy-protecting symmetric data aggregation (PPSDA) scheme for protecting the content privacy of data gathered by IoT devices is presented in this paper. End-to-end encryption is carried out at the data collection stage, and aggregation of reported data without decryption reveals nothing to the malicious adversary. This increases the efficiency as well as ensures the privacy of the message transmission between the end node and the control server. The performance analysis shows that the suggested scheme considerably lowers storage, communication, and computation costs as well as security risks.

References 1. Seliem M, Elgazzar K, Khalil K (2018) Towards privacy preserving IoT environments: a survey. Wirel Commun Mobile Comput 2018, (2018). 2. Guan Z, Zhang Y, Wu L, Wu J, Li J, Ma Y, Hu J (2019) APPA: an anonymous and privacy preserving data aggregation scheme for fog-enhanced IoT. J Netw Comput Appl 125:82– 92(2019) 3. Chanal PM, Kakkasageri MS (2020) Security and privacy in IoT: a survey. Wirel Person Commun 115(2):1667–1693

294

L. Mary Shamala et al.

4. Tonyali S (2018) Privacy-preserving protocols for secure and reliable data aggregation in IoT-enabled smart metering systems. Futur Gener Comput Syst 78:547–557 5. Alghamdi A, Alshamrani M, Alqahtani A, Al Ghamdi SSA, Harrathi R (2016) Secure data aggregation scheme in wireless sensor networks for IoT. In: 2016 international symposium on networks, computers and communications (ISNCC) 6. Ara A, Al-Rodhaan M, Tian Y, Al-Dhelaan A (2017) A secure privacy-preserving data aggregation scheme based on bilinear ElGamal cryptosystem for remote health monitoring systems. In: IEEE access, vol 5, pp 12601–12617 7. Lu R, Heung K, Lashkari AH, Ghorbani AA (2017) A lightweight privacy-preserving data aggregation scheme for fog computing-enhanced IoT. In: IEEE access vol 5, pp 3302–3312 8. Hu C, Luo J, Pu Y, Yu J, Zhao R, Huang H, Xiang T (2018) An efficient privacy-preserving data aggregation scheme for IoT. In: Chellappan S, Cheng W, Li W (eds) Wireless algorithms, systems, and applications - WASA 2018, LNCS, vol 10874. Springer, Heidelberg, pp 910–928 9. Tripathi A, Pasupulet SK (2018) A secure lightweight data aggregation scheme for cloud assisted IoT. In: Proceedings of fifth international conference on parallel, distributed and grid computing (PDGC), Solan Himachal Pradesh 10. Lyu L, Nandakumar K, Rubinstein B, Jin J, Bedo J, Palaniswami M (2018) PPFA: privacy preserving fog-enabled aggregation in smart grid. IEEE Trans Industr Inf 14(8):3733–3744 11. Ogundoyin SO (2020) A lightweight privacy-preserving data aggregation scheme with provable security for internet-of-things. Int J Comput Appl 42(3):315–327 12. Li R, Sturtivant C, Yu J, Cheng X (2019) A novel secure and efficient data aggregation scheme for IoT. IEEE Internet Things J 6(2):1551–1560 13. Pu Y, Luo J, Hu C, Yu J, Zhao R, Huang H, Xian T (2019) Two secure privacy-preserving data aggregation schemes for IoT. Wirel Commun Mob Comput 2019:1–11 14. Ullah A, Said G, Sher M, Ning H (2020) Fog-assisted secure healthcare data aggregation scheme in IoT-enabled WSN. Peer-to-Peer Netw Appl 13(1):163–174 15. Sarwar K, Yongchareon S, Yu J (2019) A brief survey on IoT privacy: taxonomy, issues and future trends. In: ICSOC 2018 workshops, LNCS, vol 11434. Switzerland 16. Lopez J, Rios R, Bao F, Wan G (2017) Evolving privacy: from sensors to the internet of things. Futur Gener Comput Syst 75:46–57 17. Pourghebleh B, Navimipour NJ (2017) Data aggregation mechanisms in the Internet of things: a systematic review of the literature and recommendations for future research. J Netw Comput Appl 97:23–34 18. Sang Y, Shen H, Inoguchi Y, Tan Y, Xiong N (2006) Secure data aggregation in wireless sensor networks: a survey. In: Proceedings of 7th international conference on parallel distributed computing applications and technologies (PDCAT). IEEE, Taipei, Taiwan, pp 315–320 19. Girao J, Westhoff D, Schneider M (2005) CDA: concealed data aggregation for reverse multicast traffic in wireless sensor networks. In: IEEE international conference on communications, ICC 2005, vol 5. Seoul, Republic of Korea, pp 3044–3049 20. Castelluccia C, Chan ACF, Mykletun E, Tsudik G (2009) Efficient and provably secure aggregation of encrypted data in wireless sensor networks. ACM Trans Sens Netw 5:20–36 21. Sethi P, Sarangi SR (2017) Internet of things: architectures, protocols, and applications. J Electr Comput Eng 2017, 25 22. Bos JW, Osvik DA, Stefan D (2009) Fast implementations of AES on various platforms. In: IACR, cryptology. ePrint Archieve 23. Wang H, Wang Z, Domingo-Ferrer J (2018) Anonymous and secure aggregation scheme in fog-based public cloud computing. Future Gener Comput Syst 78(2):712–719 24. Mary Shamala L, Zayaraz G, Vivekanadan K, Vijayalakshmi V (2020) A tweakable key alternating lightweight cipher for internet of things. Int J Inf Secur Priv 14(4):113–133

Wireless Networks

Implementation and Comparative Analysis of Various Energy-Efficient Clustering Schemes in AODV S. Dhanabal, P. William, K. Vengatesan, R. Harshini, V. D. Ambeth Kumar, and S. Yuvaraj

Abstract The wireless sensor network allows for the surveillance of physical environments as well as the implementation of timely actions based on sensor data. The major objective of this sort of network is to extend the life of the network by utilizing energy-efficient connection technologies. Cooperation between sensor nodes and agents, on the other hand, must occur as quickly as feasible. This study offers a clustering methodology for building an automated framework for a WSN network’s organization. We explore MANET security problems and suggest a few potential research areas. Many key and trust management systems have been devised to avoid external assaults from outsiders, and various secure MANET routing protocols have been proposed to prevent internal attacks from within the MANET system. A novel intrusion detection framework built particularly for MANET has been studied in terms of intrusion detection. Both preventive and detection approaches will be utilized to address the security problems in MANET. A hierarchical, salable energy optimization-based network model is used in the proposed clustering protocol. Because sensor node battery resources are limited and the physics detection cycle is brief, one of the most significant wireless sensor network techniques is to

S. Dhanabal Department of Computer Science and Engineering, Kongunadu College of Engineering and Technology, Tholurpatti, Trichy, India P. William Department of Information Technology, Sanjivani College of Engineering, SPPU, Pune, India K. Vengatesan (B) Department of Computer Engineering, Sanjivani College of Engineering, Kopargaon, India e-mail: [email protected] R. Harshini · V. D. A. Kumar Department of AI&DS, Panimalar Engineering College, Chennai 600123, India S. Yuvaraj Department of Electronics & Communication, SRM Institute of Science and Technology, Kattankulathur - Chennai, India V. D. A. Kumar Department of Computer Engineering, Mizoram University, Aizawl 796004, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_19

297

298

S. Dhanabal et al.

create energy-efficient network life routing protocols (WSN). This study presents a technique to heterogeneous clustering that is energy efficient. Keywords WLAN · WSN · MANET · Energy efficient · Throughput · End-to-end delay

1 Introduction Today, wireless sensor networks (WSNs) are of great interest not only to the scientific community but also to the business world. The development of this type of network poses new problems in software and hardware engineering, but at the same time brings scenarios that were previously thought of as science fiction closer to reality. Sensor networks have become a reality as microelectronic-mechanical systems technology, wireless communications, and digital electronics converge [1]. This has led to an increase in the number, scale, and complexity of existing WSNs [2]. Wireless sensor networks are composed of sensor nodes and a sink node. In this network architecture, the sensor nodes are passive devices that only collect data locally and the sink in a global way, that is, it collects information from the whole network and does not interact in the environment to modify it according to the information obtained. In this case, human personnel are required to review the data collected in the sink and execute actions in response to such information. The need of WSN to guarantee a fast response to sensor nodes in an automatic way has allowed the development of wireless sensor and actor networks (WASNs), which represent an important extension of WSNs [3]. WSNs present a new paradigm based on the collaborative effort of a large number of sensors deployed near or within the phenomenon to be observed, which allows the provision of diverse services for numerous applications. This new network architecture can be applied to a variety of commercial, scientific, and military applications such as environmental monitoring, industrial plant sensing and maintenance, military surveillance, medical sensing, among many other new applications. In addition, the technologies developed can also be applied to aerospace industries [4]. A wireless sensor network is typically made up of a collection of low-cost, energy efficient, and multifunctional sensor nodes deployed in an area of interest. These nodes are small in size, but they contain sensors, integrated microprocessors, and transceivers, which allow them to detect, process data, and communicate. They communicate over short distances via a wireless medium and collaborate to accomplish a specific task, such as environmental monitoring or industrial process control, among other things.

Implementation and Comparative Analysis of Various Energy-Efficient …

299

1.1 WSN Applications The sensors can be used to monitor a variety of variables such as temperature, humidity, pressure, light, sound, and water quality, among others. Wireless sensor networks are less expensive and take less time to set up than wired networks, and they can be used in inhospitable environments such as battlefields, outer space, and oceans. WSNs were initially developed for military applications, but due to the low cost of sensors and the wireless communication system, a wide range of areas were explored, with significant advances made. Environmental monitoring, military applications, health, industrial process control, security and supervision, and, finally, home automation are among the most important [5].

1.2 Parameters of a WSN Network It is important to understand the features and needs of the application for which you want to create the network when developing a WSN, because certain elements will be considered that will directly impact your influence [6]. It is crucial to remember that each application has unique requirements, therefore, it is not required to take into account all of the elements throughout the design process.

1.3 Topologies in WSN In a WSN network, there can be from tens to thousands of nodes deployed in the area of interest, with a high density, being many inaccessible and unattended, prone to frequent failures, which entails a complicated maintenance of the network’s topology area. To study the maintenance of this topology, three phases can be established. 1. Pre-Deployment and Deployment: Nodes can be launched or placed individually on site. They can be thrown from a plane, catapulted, placed one by one by a Person or a Robot, etc. 2. Post-deployment: After the deployment of the nodes, changes may occur in the topology tambie´n due to the position, accessibility, available energy posicio´n and/or failures of them, as well as dina´mica and according to the dynamics of the tasks. Redeployment: New nodes can be added to the network to replace failed nodes, or due to changes in task dynamics.

300

S. Dhanabal et al.

1.4 Energy Consumption in WSN The wireless sensor nodes have limited, self-contained power sources, and it is often impossible to recharge the batteries. As a result, the lifespan of a node is inextricably linked to that of its battery. When a node fails to function, it causes significant changes in the network’s topology topolog a. As a result, energy conservation and management in the network are critical, and many research projects have been developed focusing on protocols and sensor network algorithms that optimize the use of energy administration. Among the various approaches, work has been done with sensor groups or clusters, studying sensor mobility and routing [7], and even changing the position of base stations. The communication protocols, which have been developed to ensure reliable communication and efficient use of energy energ a within the network, are among the important aspects that have been developed in WSN networks. Among these protocols are those for access to the medium, links, and even routing protocols [8].

1.5 Routing Algorithms In the case of the TCP/IP (Transmission Control Protocol/Internet Protocol) architecture, this task is entrusted to the network level, in which the so-called routing algorithms are implemented, responsible for determining the path followed by each packet until reaching to the target. There are several routing algorithms, those most used are Link State, Distance Vector, and Source Routing [8]: 1. Link State Algorithm: A cost is assigned to each link or connection and each node manages a complete map of the network topology. Periodically, each node disseminates the cost of the links to which it is connected, and the remaining nodes update the network map and the routing table, applying an algorithm that takes into account the path at the lowest cost. 2. Vector-Distance Algorithm: The node already knows the cost of the links to which it is connected. Each node communicates to its neighbor which other nodes it can reach and at what cost. Thus, each node recalculates its own routing table following the information it has received, and using an algorithm that takes into account the path at the lowest cost. 3. Source Routing: Routing decisions are made from the source, and packets follow an established path.

Implementation and Comparative Analysis of Various Energy-Efficient …

301

2 Literature Review A mobile ad hoc network is a wireless networking device that may be customized and arranged dynamically to form a mobile, multi-hop network. The reliance on fixednetwork infrastructure is removed from mobile ad hoc networks by treating a mobile node as an intermediate transfer, which extends mobile node ranges well beyond the base transceivers. Mobile ad hoc networks extend the range of moveable nodes beyond their current level. However, if the nodes want to connect with other nodes that are not within their range, the packets must be routed from source to destination using routing algorithms. We selected the AODV routing protocol because it works well in a complicated network environment like a MANET. What happens in this scenario will result in specific security concerns. In the event of such an occurrence, wormhole attacks are one type of successful denial of service attack. Because routing is the foundation of all mobile ad hoc communication, routing failures will bring the entire system to a halt. LEACH [9] is an acronym for Low Energy Adaptive Clustering Hierarchy. This protocol chooses some nodes at random, such as CH, and rotates this function to distribute the energy load evenly among the network nodes. Each CH compresses the data from its group’s nodes and sends the resulting packet to the base station at env a. Data collection is done on a regular and central basis, and it is useful when the sensor network needs to be constantly monitored. After a set amount of time, the role of CH is randomly rotated, resulting in uniform energy dissipation in the network (disipacio n). There is a variant of this protocol (LEACH-C [9]) that uses a centralized algorithm for cluster formation and can produce better results. Power-Efficient Gathering in Sensor Information Systems (PEGASIS) [10]: In this protocol, nodes need only communicate with their nearest node in the chain, until the last node communicates directly to the base station. The collected information is transferred node by node in the chain, merged, and transmitted to the sink by the final node. The protocol works on a round basis, alternating the node that has direct communication with the base station, so that power consumption is distributed evenly among all the nodes. On the other hand, the bandwidth required for communication is reduced due to the local coordination between nearby nodes. Threshold-Sensitive Energy-Efficient Protocols (TEEN) and Adaptive Periodic Threshold-Sensitive Energy-Efficient Protocols (APTEEN) [11, 12]: These protocols were created for applications that require a quick response to sudden changes in the parameters to be measured. TEEN nodes constantly monitor their surroundings, but data transmission is sporadic. Each CH sends a hard threshold (HT) to its group, which is the threshold value of the measured attribute; however, env an also sends a soft threshold (H S), which is a minor variation in the value of the attribute that causes the node to activate its transmitter and transmit. Only when the measured attribute is in the range of interest (above HT) and the difference from the previous transmitted measurement is greater than HS will the nodes transmit. The threshold

302

S. Dhanabal et al.

values in APTEEN are modified on a regular basis based on the needs of the user and the type of application. Minimum Energy Communication Network (MECN) [13]: This protocol adapts dynamically to node failures or the deployment of new sensors, although it assumes that all the nodes can transmit to any other node, which is not always possible. This is why an extension of this protocol was developed, called Small Minimum Energy Communication Network (SMECN) [13], which considers the possible existence of obstacles between nodes. Self-organizing Protocol (SOP) [14]: Used in networks with heterogeneous sensors of varying capacities and functionalities, as well as mobile and stationary nodes. The router nodes are fixed; they self-organize in a network to implement the algorithm, whereas the special nodes only keep track of the active routers and those nearby. The algo-rhythmo is divided into four stages: discovery, organization, maintenance, and self-organization. This algorithm has low table maintenance costs and a balanced routing hierarch. However, it can cause an overload, particularly in the organization phase of the algorithm, and if there are many outages in the network, the likelihood of using the self-organization phase increases, increasing operational costs. Sensor array routing [15]: The goal of these protocols is to collectively monitor a target’s behavior in a particular environment. A sensor suite is made up of network nodes that meet specific grouping criteria for a collaborative processing activity. [29] suggested three methods based on this concept: Distributed Aggregate Management (DAM), Energy-Based Activity Monitoring (EBAM), and Expectation–Maximization Like Activity Monitoring (EMLAM). Virtual Grid Architecture (VGA) [16]: Within the network, data aggregation and processing are utilized to extend network life. Topologism is based on the assumption that the nodes have a fixed topology. To construct a linear virtual topology with a CH for each zone, we generally utilize square clusters. Aggregation takes place on two levels: local and global. The set of CH is referred as the local aggregator (LA), and it is used to conduct local aggregation, while a subset of these is referred to as the master aggregator (MA), and it is used to do global aggregation. Hierarchical Power-Aware Routing (HPAR) [17] splits the network into clusters of geographically adjacent sensors, each of which is considered as a separate entity. Each zone is permitted to select how to route a message hierarchically via the other zones in order to maximize the battery life of the requirement system nodes. The messages are routed through the max min route, which is the maximum energy reserve route maxima of all remaining energy routes. • Two-Tier Data Dissemination (TTDD) [18]: This technique offers data delivery to multiple mobile base stations. The logic behind this notion is because utilizing high residual energy nodes can be more expensive than using the quickest way with the least amount of remaining energy consumption. Each information source builds a mesh ahead of time that is used to distribute data to mobile sinks, assuming

Implementation and Comparative Analysis of Various Energy-Efficient …

303

that the sensors are fixed and their position is known. For data retransmission, multi-hop techniques are employed. • Energy-Efficient Clustering Routing (EECR) [19]: The method suggests that the sink and member nodes work together to build clusters and choose the CH. The sensors are organized into clusters and can function in one of two modes: sensor mode or CH mode. This lets the algorithm to be more sophisticated, allowing for a more balanced connection between modes and energy savings. Location-based routing: The nodes are identified from their locations [20]. The distance between neighboring nodes can be estimated on the basis of the levels of the entrance and their relative coordinates through the exchange of information between neighbors [21]. The location of the nodes can also be obtained from a low-powered global positioning system (GPS) contained in the nodes. In many of these protocols, when there is no activity, the nodes are kept asleep, in order to save energy.

2.1 Traditional Routing Protocol in WAN Network This section will review the protocols traditional developed for solve the problem of routing in networks wireless ad hoc mobile. These protocols are divided into two groups (Royer and Toh, 1999): proactive or table-driven and reactive or on demand. A third group is also added which is a combination of the previous ones, known as hybrids [22] The proactive protocols are those that store information corresponding to the entire network topology and the routes to each of the possible destinations. Therefore, they periodically emit packets to discover new devices on the network and routes to them, so that they can be used in the future. To save this information and keep it updated, they use routing tables. The main advantage of these protocols is that if a device wants to send a packet to another device, it can obtain the routing information quickly and easily. Among its disadvantages is that to store this information, which is proportional to the size of the network, it requires a lot of memory in each of the devices. Reactive protocols are distinguished by the fact that they perform route discovery only when a device wishes to communicate with a destination device, utilizing the route request message flooding process to reduce the number of control messages in the network. When a route is discovered, it is kept until the destination is inaccessible or the route is no longer needed, which reduces the size of the routing tables. A disadvantage of reactive protocols is that if the devices do not have a baby an available path to communicate with a target device, this must be discovered, adding a significant time delay in the shipping of the first package, as well as increasing the possibility of net saturation due to the flood process. The reactive protocols Ad hoc On Demand Distance Vector (AODV) and Dynamic Source Routing (DSR) will be examined again. The last type of protocol to examine is hybrid protocols, which are preferred in networks of a certain size and with a certain number of devices. The majority of

304

S. Dhanabal et al.

these protocols use the zone concept, in which routing within the zone is performed proactively.

2.2 Optimized Link State Routing (OLSR) The OLSR protocol proposed by [26] is a proactive routing protocol based on the state of the links, designed to work in a completely distributed way and that does not require an or—none of your packages. Being a proactive protocol, OLSR maintains routes for all destinations on the network, so it periodically sends control messages that allow it to learn the network topology. To reduce the amount of control messages, sent by broadcast or flood, he introduces the concept of Multipoint Relays (MPR). These are a set of devices selected from the same neighborhood within a hop of each device, capable of covering devices that are at a distance of a hop as shown in Fig. 5. Also, during the transmission of control messages, the protocol generates traffic and extra before the thereof a link or a new addition. The OLSR protocol manages three kinds of tables: neighbor tables, topology tables, and routing tables. The neighbor table is built from HELLO messages that are broadcast and contain information about neighbors at one hop, the state of the links, and information about their neighborhood at two hops. While the topology table is fed by Topology Control messages, which contain the addresses of all MPR-marked devices. Finally, the routing table is built using the information from the previous two tables, with each entry containing the destination’s address, the address of the next hop, and the estimated distance to the destination (Fig. 1). Fig. 1 Multipoint relay in OLSR

Implementation and Comparative Analysis of Various Energy-Efficient …

305

2.3 Destination Sequenced Distance Vector (DSDV) The DSDV protocol is a hop-to-hop distance vector protocol proposed by [27], which allows multi-hop routing between network devices. Each device in the network maintains a routing table listing all possible destinations and the jump required to reach them. Each of these entries is linked to a sequence number originated by the destination device which indicates how old the route is, the higher the sequence number the more up to date the route information use of this sequence number makes DSDV a cycle-free protocol. DSDV must be periodically transmitting information corresponding to the network topology periodically and in case of a change in the network topology, immediately notify each device to keep its tables updated. When a device discovers through these control packages that a route on its table is no longer available, the device creates a new package with a higher sequence number and the jump number marks, immediately after it notifies its neighbors of the change in topology. DSDV requires many control packages to keep route information up to date. To reduce the amount of information contained in the packages, two types of packages are defined below as: Full Dump: This package carries all the information available about the routing in the network. Incremental Dump: This package only contains information that has changed since the last full dump received. Due to the constantly changing topology of an ad hoc mobile wireless network, the DSDV protocol requires all devices to periodically send all routing information they have learned periodically or to do so immediately when a change in information occurs. The main disadvantage of this protocol is that the information must be received by every device in the network, so it will take a time for the information to converge, producing a low in high density networks.

2.4 Ad hoc Demand Distance Vector The Ad hoc Demand Distance Vector (AODV) protocol proposed by (Perkins and Royer, 1999) is based on the DSDV protocol; however, unlike the latter, AODV is purely reactive, i.e., a device does not need to discover or maintain a route with another device until the two need to communicate, in other words, route acquisition is purely on demand. Only send discovery packages as required. Distinguish between the upkeep of local connectivity (neighborhood detection) and the overall upkeep of the topology. Distribute information about changes in local connection to neighboring devices that may require it. A fundamental feature of the AODV protocol is that, before providing routing information, the destination devices generate a destination sequence number (a concept borrowed from the DSDV protocol), which serves as a tool for determining when a specific route has been updated while avoiding the formation of loops. As a

306

S. Dhanabal et al.

result, if a device must pick between many pathways to a destination, it will select the one with the higher sequence number, which corresponds to the most current routing information. The major advantage of AODV is that it does not generate additional traffic for communication; nevertheless, it has the restriction that it takes time to establish a connection and that the first conversation to build a route is more difficult than with other protocols.

2.4.1

Route Discovery in AODV

When a device wishes to send a packet to another device, it first searches its routing database for a route to the destination, if one exists in the data packet; otherwise, it begins the route discovery process. In this scenario, the gadget will broadcast an RREQ message. The various parameters including different source and destination address are all included in the RREQ message. The gadget will wait for a Reset Route Message (RREP) to arrive. If an RREP message is not received within a certain amount of time, the device will wait for it. The site will retry a certain number of times. If it fails again, the gadget concludes that there is no way to reach that location. When a device gets an RREQ message for a destination, it may do one of the following actions: Because the device that got the RREQ message is unaware of the mistake, it broadcasts the message to its neighbors. It also saves the route back to the destination device [28]. The destination device is the device that receives the message or has route information to destination t. In either scenario, it sends an RREP message back to the source device through the same way that the initial RREQ message was received. When the source receives the RREP message again, a route to the t-destination is formed and available for usage. All devices keep a routing table containing information about each destination of interest, which is generated from received RREQ messages and stays active if utilized by any active adjacent device. The following information is contained in each entry in this table: Address of the destination device. a. b. c. d. e.

Direction of the next jump. Number of jumps to reach the destination sequence number of the destination. Active neighbors for that route. Expiry time for the route. Route maintenance in AODV

The devices’ mobility has no effect on the active routing path to a good goal. If the moving device is the source device of an active route, the route discovery process can be resumed to find a new path to the destination. When the destination device or one of the intermediate devices is relocated, destination receives a special RREP message. Network devices that are part of an active route transmit HELLO messages (a specific RREP message) to their neighbors on a regular basis. Failure to re-enter these messages will be construed as a loss of device connection. It broadcasts an RERR message (error message), and any device that receives it will cancel the

Implementation and Comparative Analysis of Various Energy-Efficient …

307

routes that go via the device that has become unreachable. If a source device receives an RERR message, it may resume the route discovery process if the route is still needed. To ascertain this, the device must check to see if the route has recently been utilized. If it decides to reconstruct the route to the target device, it sends an RREQ message with a sequence number greater than the previously known one to guarantee that a new viable route is built (Tables 1 and 2). Table 1 Comparative analysis of existing protocol S. no

Pause time

No. of nodes

40 (Energy of node)

NOC

OLSR

AODV

DSDV

1

1

5

16.28

57.54

24.98

2

10

138.09

52.69

35.09

3

15

48.50

114.89

77.69

5

14.71

45.13

141.18

5

10

79.45

95.04

501.30

6

15

70.19

39.63

213.91

5

9.07

78.10

27.49

8

10

20.62

20.37

370.94

9

15

89.29

19.23

253.89

4

7

2

3

Table 2 Vulnerabilities exploitable by Sybil, sinkhole, and wormhole attacks Vulnerability

Enumeration of attacks Sybil

Sinkhole

Wormhole

Reason

A

X

X

X

Theft of encryption keys

B

X

X

X

Data from setting

C

X

–

–

Difficulty in order to detection

D

–

X

–

Injection from packages

E

X

–

–

Difficulty in order to detection

F

–

–

X

Facil decoded

G

X

X

X

Espionage/interception

H

X

–

–

False identity/impersonation

I

–

X

–

Execution two

J

–

–

X

Loss of integrity

Where: X = Applies, - = Does not apply

308

S. Dhanabal et al.

3 Proposed Work The suggested technique comprises two packet forwarding mechanisms. When possible, the voracious technique, which consists of forwarding a data packet to a destination by using the geographical position of nearby devices to determine the next hop, employs it. Face-based routing entails sending a data packet to a destination device via a group of neighboring faces. It is employed in areas where the voracious technique does not work. To discover its neighbors, each device sends a mediating broadcast packet including the device’s unique identification, its geographic coordinates (x,y), and the identifier of the cell in which it is presently situated. A jitter (i.e., delay) of 1 to 100 ms is introduced to avoid synchronization with other broadcasts from nearby devices. When one of these packets is received, the device adds or updates a record in its connection database. The following illustrates the operation of the voracious technique and face-based routing using the plane partition. Attack or Vulnerabilities in AODV The vulnerabilities of the nodes within a WSN that allow the successful execution of The aforementioned attacks are closely related to the nature of WSNs, since as indicated in [44], nodes must have low power consumption characteristics energy and self-management, which leads to reduced computing capacities and therefore traditional security solutions cannot be applied effectively in WSN devices. Taking into account the above, in [Four. Five], the features what make to the WSN vulnerable to various attacks, the more important they are enunciated to continuation. 1. Device cost: The cost of the equipment is relatively low. Therefore, applying specialized security mechanisms can make this technology more expensive, thus affecting its popularity. 2. Susceptibility to theft: It is common to deploy nodes outdoors, without adequate protection the devices are exposed to being alienated from the network. In this way, an attacker can steal one of the nodes to extract information related to the network configuration, such as encryption keys, authentication credentials, or modify the application software. This makes it easier to execute attacks since access to the data transported in the WSN is compromised, increasing the possibility of injecting malicious packets into the nodes. 3. Communication point-to-point or ad hoc: In the absence of central nodes to channel network traffic, attacks can be executed from any direction and affect any of the nodes. In this way, an attacker can quickly go from hijacking data to interfering with it, causing loss of confidentiality and integrity of the information. 4. Balance between functionality and security: It is due to the conflict of interest of maintaining the characteristics of the nodes in terms of low resource consumption and increasing security levels. This translates into the need to develop solutions that maintain a balance between security and functionality, which is complex and difficult to apply.

Implementation and Comparative Analysis of Various Energy-Efficient …

309

5. Half from transmission shared: Dice what the WSN they use channels wireless shared on combination with the limiting on size, under consumption energetic and under cost from the nodes. 6. Limited encryption mechanisms: For many of the applications in WSN, the use of asymmetric encryption is costly due to resource limitations at the nodes. Therefore, the use of symmetric encryption protocols becomes one of the most widely used options. However, the management and storage of encryption keys is limited by the reduced storage capacity of the nodes, which results in keys of shorter length and easier to decrypt. 7. Given the preceding features, the WSN I am familiar with is vulnerable to a number of flaws that, if exploited, can seriously jeopardize the availability, integrity, and confidentiality of end-user and application information and data. An approach to these vulnerabilities that I am aware of focuses on the potential weaknesses and vulnerabilities to which smart homes are vulnerable. Those that are more common in the WSN are presented in this case. 8. Vulnerabilities what they can to be exploited to obtain successful Sybil, sinkhole, and wormhole attacks on real WSNs. Additionally, those vulnerabilities that are involved in the development of the attack and detection phases of this work are selected. It can be seen, it is possible to map the relationship between Sybil, sinkhole, and wormhole attacks and vulnerabilities in WSN. In this way, the most common vulnerabilities that can be exploited by potential attackers are exposed. In these cases, each vulnerability allows the execution of one or more malicious activities, e.g., vulnerability A. facilitates the extraction of encryption keys from the physically intervened nodes, with this information. It is possible to decrypt the captured traffic or encrypt malicious packets to be sent within the network. On the other hand, vulnerabilities such as G. are typical of the WSN and although adequate controls are applied, it must be assumed as a latent risk that requires permanent monitoring since it is impossible to determine who or who are capturing WSN radio waves for malicious use. An exploration of the possible control measures to reduce the risk of exploitation of these vulnerabilities by Sybil, sinkhole, and wormhole attacks is presented.

310

S. Dhanabal et al.

Algorithm 1: Choice of the Next Jump using the Voracious Method 1: procedure RE-SENDING (package) 2: minim Local ← true 3: Current distance ← Distance (Current node, cell Destination) 4: for all neighbor in V do: D Where V is the set of neighbors 5: distance Actual ← Distance(neighbor, cell Destination) 6:if distance Actual < distance Actual then: 7:distance Actual ← distance V 8:nextJump ← neighbor 9:minimoLocal ← false 10:if minimum Local == true then: 11:Face Route(package) 12:else 13:Resend Package(next Shoot, package)

3.1 Environment of Attack Analysis with Wormhole and Without Wormhole The simulation was carried out on the NS 2 simulator, which is an event-driven simulation tool for wireless packet mode communication. It is a discrete event simulator that is object oriented and is used to investigate the dynamic dynamics of communication networks. It provides a full framework for building network protocols, generating and visualizing scenarios, and assessing their performance under specified conditions. We worked with 50 network nodes; the simulation lasted 90 s: I.

Firstly, open the command user interface (CUI) or terminal on screen and type command and press ENTER. Then we will enter the desktop where the coding (.tcl) file is saved named richaworm.tcl. II. Further when richaworm.tcl file runs, following figure occurs. After running this file, richaworm.nam file is generated which is basically an animator file and helps in showing the animation. III. After that we have shown the file which protects worm hole and the file is named as richaisolate.tcl and to run this file. Lastly, after running the richaisolate.tcl file, same happens as in step (iii). richaisolate.nam file is generated which is basically an animator file (Figs. 2 and 3). Now, we will discuss about the coding work done at frontend, i.e., in.tcl file. In richaworm.tcl file, following terms are used (Fig. 4): 1.

Channel—wireless channel is used

Implementation and Comparative Analysis of Various Energy-Efficient …

Fig. 2 After running step (ii) command

Fig. 3 Command to run the.tcl file for introducing security algorithm to worm hole

2. 3. 4. 5. 6. 7. 8. 9. 10.

Propagation—two way ground Internet information—physical, wireless MAC address—MAC 802.11 Queue used—Priority queue Link layer type—Data link layer Antenna model—Omni antenna Maximum packets used—50 Number of mobile nodes—50 Routing protocol—AOMDV

311

312

S. Dhanabal et al.

Fig. 4 Worm.tcl file coded which introduces worm hole in data communication

Fig. 5 Comparison of throughput with wormhole and without wormhole

11. 12. 13. 14.

X axis dimension of topology—2500 Y axis dimension of topology—2500 Time of simulation start—0 s Time of simulation end—90 s

Implementation and Comparative Analysis of Various Energy-Efficient …

313

15. Initial Energy value—50.

3.2 Parameters Used for Comparison Below given Table 3 shows the comparison made between throughput with wormhole attack and throughput after isolating wormhole time slot from 0 to 91 s. Further, a graph shows graphical representation of the same (Figs. 5 and 6 and Table 4). Table 3 Tabular representation of throughput with wormhole and after isolating wormhole attack

Time

Worm-throughput

Proposed work-throughput

1

3072

3072

3

6141

6144

5

9207

9216

7

12,270

12,288

9

15,330

15,360

10

18,387

18,432

12

21,441

21,504

14

24,492

24,576

16

27,540

27,648

18

30,585

30,720

19

33,627

33,792

21

36,666

36,864

23

39,702

39,936

25

42,735

43,008

27

45,765

46,080

28

48,792

49,152

30

51,816

52,224

32

54,837

55,296

34

57,855

58,368

36

60,870

61,440

37

63,882

64,512

39

66,891

67,584

41

69,897

70,656

43

72,900

73,728

44

75,900

76,800

46

78,897

79,872

48

81,891

82,944

50

84,882

86,016

314

S. Dhanabal et al.

Fig. 6 Comparison of throughput with Sybil and without Sybil Table 4 Tabular representation of throughput with Sybil attack and after isolating Sybil attack

Time

Sybil-throughput

Proposed-throughput

1

87,870

89,088

3

90,855

92,160

5

93,837

95,232

7

96,816

98,304

9

99,792

101,376

10

102,765

104,448

12

105,735

107,520

14

108,702

110,592

16

111,666

113,664

18

114,627

116,736

19

117,585

119,808

21

120,540

122,880

23

123,492

125,952

25

126,441

129,024

27

129,387

132,096

28

132,330

135,168

30

135,270

138,240

32

138,207

141,312

34

141,141

144,384

36

144,072

147,456

37

147,000

150,528

Implementation and Comparative Analysis of Various Energy-Efficient …

315

Fig. 7 Comparison of throughput with gray hole and without gray hole

From above comparison, it is clear that average rate of successful message delivery is more when worm hole is isolated rather than when worm hole is present over communication channel (Fig. 7). Utilization of energy can be in the form of heat or electricity. Above given Table 5 shows the comparison made between energy consumption during communication with worm hole attack and energy consumption after isolating Sybil attack during time slot from 0 to 91 s. Further, a graph shows (Fig. 9) the graphical representation of the same.

4 Conclusion Technological evolution and especially in the field of communications networks has allowed more and more devices to connect to the network either by guided or wireless means, allowing the deployment of networks to places where a few years ago, it was unthinkable that network services could be offered; in this context, ad hoc mobile networks have become essential for the implementation of services that require characteristics of this type of network such as their autonomy, their ability to establish dynamic topologies as more devices access the network, the lack of need for a previously established physical infrastructure; despite the benefits of these networks, it should be borne in mind that they have limitations, mainly in terms of energy and processing capacity.

316 Table 5 Tabular representation of throughput with gray hole attack and after isolating gray hole attack

S. Dhanabal et al. Time

Gray hole-throughput

Proposed work-throughput

1

12,270

12,288

2

15,330

15,360

3

18,387

18,432

4

21,441

21,504

5

24,492

24,576

6

27,540

27,648

7

30,585

30,720

8

33,627

33,792

9

36,666

36,864

10

39,702

39,936

11

42,735

43,008

12

45,765

46,080

13

48,792

49,152

14

51,816

52,224

15

54,837

55,296

16

57,855

58,368

17

60,870

61,440

18

63,882

64,512

References 1. Rani P, Kavita S, Verma S, Nguyen GN (2020) Mitigation of black hole and gray hole attack using swarm inspired algorithm with artificial neural network. In IEEE access, vol 8, pp 121755–121764. https://doi.org/10.1109/ACCESS.2020.3004692 2. Ochola EO, Mejaele LF, Eloff MM, van der Poll JA (2017) Manet reactive routing protocols node mobility variation effect in analysing the impact of black hole attack. SAIEE Africa Res J 108(2):80–92. https://doi.org/10.23919/SAIEE.2017.8531629 3. Cai RJ, Li XJ, Chong PH (2019) An evolutionary self-cooperative trust scheme against routing disruptions in MANETs. IEEE Trans Mob Comput 18(1):42–55. https://doi.org/10.1109/TMC. 2018.2828814 4. Kumar A, Sinha M (2019) Design and development of new framework for detection and mitigation of wormhole and black hole attacks in VANET. J Stat Manag Syst 22(4):753–761. https://doi.org/10.1080/09720510.2019.1609555 5. Li T, Ma J, Pei Q, Song H, Shen Y, Sun C (2019) DAPV: diagnosing anomalies in MANETs routing with provenance and verification. IEEE Access 7:35302–35316. https://doi.org/10. 1109/ACCESS.2019.2903150 6. Djahel S, Nait-abdesselam F, Zhang Z (2011) Mitigating packet dropping problem in mobile ad hoc networks: proposals and challenges. IEEE Commun Surv Tutor 13(4):658–672, Fourth Quarter 2011. https://doi.org/10.1109/SURV.2011.072210.00026 7. Li X, Jia Z, Zhang P, Zhang R, Wang H (2010) Trust-based on-demand multipath routing in mobile ad hoc networks. IET Inf Secur 4(4):212–232. https://doi.org/10.1049/iet-ifs.2009. 0140

Implementation and Comparative Analysis of Various Energy-Efficient …

317

8. Dadheech P, Kumar A, Choudhary C, Beniwal MK, Dogiwal SR, Agarwal B (2019) An enhanced 4-way technique using cookies for robust authentication process in wireless network. J Stat Manage Syst 22:4, 773–782. https://doi.org/10.1080/09720510.2019.1609557 9. Akande DO, MohdSalleh MF (2019) A network lifetime extension-aware cooperative MAC protocol for MANETs with optimized power control. IEEE Access 7:18546–18557. https:// doi.org/10.1109/ACCESS.2019.2895342 10. Song Y, Luo H, Pi S, Gui C, Sun B (2020) Graph kernel based clustering algorithm in MANETs. IEEE Access 8:107650–107660. https://doi.org/10.1109/ACCESS.2020.3001137 11. El-Semary M, Diab H (2019) BP-AODV: blackhole protected AODV routing protocol for MANETs based on chaotic map. IEEE Access 7:95197–95211. https://doi.org/10.1109/ACC ESS.2019.2928804 12. Chen Y, Hu C, Wu EH, Chuang S, Chen G (2018) A delay-sensitive multicast protocol for network capacity enhancement in multirate MANETs. IEEE Syst J 12(1):926–937. https://doi. org/10.1109/JSYST.2017.2677952 13. Chen Z, Zhou W, Wu S, Cheng L (2020) An adaptive on-demand multipath routing protocol with QoS support for high-speed MANET. IEEE Access 8:44760–44773. https://doi.org/10. 1109/ACCESS.2020.2978582 14. Kumar A, Dadheech P, Goyal D, Patidar PK, Dogiwal SR, Janu N (2020) A novel scheme for prevention and detection of Black Hole & Gray Hole attack in VANET network. In: Recent patents on engineering vol 14, p 1. https://doi.org/10.2174/1872212114999200512120211 15. Hao S, Zhang H, Song M (2018) A stable and energy-efficient routing algorithm based on learning automata theory for MANET. J Commun Inf Netw 3(2):43–57. https://doi.org/10. 1007/s41650-018-0012-7 16. Ojetunde B, Shibata N, Gao J (2019) Secure payment system utilizing MANET for disaster areas. IEEE Trans Syst, Man, Cybern: Syst 49(12):2651–2663. https://doi.org/10.1109/TSMC. 2017.2752203 17. Jiang N, Xu P, Yao Y, Bui T, Chen Q (2018) Exploiting radio irregularity for location verification in sparse MANETs. IEEE Commun Lett 22(6):1284–1287. https://doi.org/10.1109/LCOMM. 2018.2828406 18. Ahmad M, Hameed A, Ikram AA, Wahid I (2019) State-of-the-art clustering schemes in mobile ad hoc networks: objectives, challenges, and future directions. IEEE Access 7:17067–17081. https://doi.org/10.1109/ACCESS.2018.2885120 19. Rahman T, Ullah I, Rehman AU, Naqvi RA (2020) Notice of violation of IEEE publication principles: clustering schemes in MANETs: performance evaluation, open challenges, and proposed solutions. IEEE Access 8:25135–25158. https://doi.org/10.1109/ACCESS.2020.297 0481 20. Kumar A, Dadheech P, Beniwal MK, Agarwal B, Patidar PK (2020) A fuzzy logic-based control system for detection and mitigation of blackhole attack in vehicular ad hoc network. In: Chaudhary A, Choudhary C, Gupta M, Lal C, Badal T (eds) Microservices in big data analytics. Springer, Singapore. https://doi.org/10.1007/978-981-15-0128-9_15 21. Ponguwala M, Rao S (2019) E2-SR: a novel energy-efficient secure routing scheme to protect MANET-IoT. IET Commun 13(19):3207–3216. https://doi.org/10.1049/iet-com.2019.0039 22. Doss S et al (2018) APD-JFAD: accurate prevention and detection of jelly fish attack in MANET. IEEE Access 6:56954–56965. https://doi.org/10.1109/ACCESS.2018.2868544

Adaptive End-To-End Network Slicing on 5G Networks P. Sakthi Saravanakumar, E. Mahendran, and V. Suresh

Abstract In today’s world, 5G promises a variety of services such as enhanced Mobile Broadband (eMBB), ultra-Reliable Low Latency Communication (uRLLC), and enormous Machine-Type Communication (mMTC). These services provide different levels of QoS to users based on their business requirements. QoS is a network control technique used to prioritize different types of traffic by regulating bandwidth, latency, flapping, and packet loss. The two main technologies used to enable network slicing in 4G/5G networks are Network Function Virtualization (NFV) and Software-Defined Networking (SDN). Slicing is possible across all tiers, from 5GUE to the datacentre network. Each network slice can be controlled and managed by the different independent network controllers. The controller is the network’s brain, which can be centralized or federated. The routing path is determined by it, whereas the data planes forward packets into network paths. In this research, we present a middleware for adaptive network slicing for end-to-end QoS provisioning across datacentre networks to 5G-UE devices. Keywords 5G · SDN · Network slicing · QoS · Network virtualization

1 Introduction 5G wireless technology is designed to provide huge network capacity, ultra-low latency services, increased reliability, and a more consistent user experience to a

P. Sakthi Saravanakumar (B) · E. Mahendran (B) Centre for Development of Advanced Computing (C-DAC), Chennai, India e-mail: [email protected] E. Mahendran e-mail: [email protected] V. Suresh Centre for Development of Advanced Computing (C-DAC), Pune, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_20

319

320

P. Sakthi Saravanakumar et al.

larger number of users. Given the widespread adoption of industry-leading technology such as big data analytics and artificial intelligence, the number of devices connected to the cloud environment is growing at an exponential rate. The amount of network bandwidth required to send all of the data to the cloud will be enormous. In a big data-enabled cloud computing environment, data transfer is a bottleneck. Legacy mechanisms do not address the reconstruction of data framework to standard level, and SDN appears to be extremely suitable for solving those problems in terms of network efficiency, scalability, flexibility, agility, as well as operation and maintenance complexity. In SDN controller, the network is split up into two planes such as data and control plane. The control plane is responsible for ensuring the application quality of service. It ensures QoS traffic should have no packet loss, and QoS should optimize a cost function other than path length. It should have pre-emptive rights based on traffic priority and importance, and QoS routing may select optimal routes based on traffic patterns and packet loss estimation of best traffic route. QoS policing can be divided into two categories. Ingress traffic is rate limited, while egress traffic is guaranteed bandwidth. Virtualization is a cutting-edge technology that creates software environments in the form of virtual machines on the fly. It divided computing resources into multiple execution environments while hiding the physical characteristics of processing resources to improve the path taken by different frameworks, applications, or endusers when interacting with those resources. It is a collection of technologies that act as a barrier between computer hardware and the software that runs on it. Virtualization solutions allow multiple operating systems to run in parallel on a single critical Processing Unit by presenting a logical rather than a physical view of computing resources (CPU). Parallelism differs from multitasking in that it saves money on overhead. Monitoring is an essential notion in network management because it allows network operators to assess how a network behaves and how its components are performing. It is also used for decision-making in traffic engineering, quality of service, and anomaly identification. For network provisioning and administration, software-defined networking (SDN) is becoming more prevalent. SDN is becoming increasingly popular for network provision and management tasks. The main contribution of our research paper is A proposed QoS middleware architecture that works with a 5G core environment for network slicing and realizes QoS in software-defined networking (SDN)-based virtual infrastructure. The remainder of this study is arranged as follows: Sect. 2 summarizes the present state of the art and places our results in context. The suggested SDN controller reference architecture is discussed in Sect. 3. Section 4 shows real-time experiments with bandwidth allocation for healthcare applications. Finally, Sect. 5 wraps up the proposed research project and considers future possibilities.

Adaptive End-To-End Network Slicing on 5G Networks

321

2 Related Work In the current literature, several architectures for network slicing are presented. Some studies investigate the effects of network programmability and network slicing on specific 5G services. The goal of network slicing is to make 5G a flexible, scalable, and demand-driven network. [1] A novel testbed called 5GIIK provides implementation, management, and orchestration of network slices across all network domains and different access technologies. The design criteria are a superset of the features present in other state-of-the-art testbeds and determine appropriate open-source tools for implementing them. 5G Testbed for Network Slicing Evaluation [4]: The testbed utilizes OAI for both RAN and CN domains. Two CNs share radio resources of a single eNB in the RAN. It has been appraised for connection establishment for both normal LTE UEs and UEs with an implemented Network Slice Selection Assistance Information. In the context of RAN sharing, authors in [5] propose the concept of the on-demand capacity broker to enhance the RAN sharing flexibility. The ondemand capacity broker aims at enabling more flexibility in the resource allocation by allowing a host RAN provider to allocate a portion of network capacity for a specific time to MVNO or OTT service provider via signaling means. Mosaic5G [6]: Brings flexibility and scalability to service provision. The testbed architecture consists of five software modules along with hardware components. The Mosaic5G platform has been used for a few use cases such as critical eHealth, V2X communication for intelligent transportation systems and multi-service management/orchestration for smart cities. Slice-Aware Service Assurance Framework [7]: Measures Quality of Experience (QoE) of a specific service according to the several service dependability Key Quality Indicators (KQIs). It provides web content browsing and adaptive video streaming services to assess infrastructure performance and the KQIs alteration for each service. In this paper, we mainly focus on the QoS issue and we propose an adaptive end-to-end network slicing approach to validate the QoE property. Network slicing [8] is considered a key technology for the upcoming 5G system, enabling operators to efficiently support multiple services with heterogeneous requirements, over a common shared infrastructure. In the emerging development in communication technology and the emerging usage of Internet of Things (IoT) devices that produce a huge amount of data, the fifth generation (5G) mobile network [9] is introduced to support this development. SDN architecture provides higher flexibility, scalability, cost-effectiveness, and energy efficiency in 5G mobile networks. There are usually different architectures for the SDN control plane.

322

P. Sakthi Saravanakumar et al.

3 QoS Architecture In our 5G testbed architecture, [2] (based on open5gs) includes a 5G core network, gNB, and UE devices. 5G core components have separated into two main planes called the control plane and user plane. The control plane and the user plane are the two major planes that make up the 5G fundamental components. Session management, mobility management, paging, and bearers were the main concerns of the control plane. The Home Subscriber Server (HSS) authenticates and creates a subscriber profile after the mobility management entity connects to it. The QoS settings are also sent as an input into the 5G core during UE registration. SMF, which serves as a control plane gateway, is another important component. Data packets are carried be- tween gNB and the external WAN network by the user plane. In the user plane, there are two types of core components that are connected to the Fronthaul and Backhaul networks. gNB connects to the SGWC which is connected to front haul network UE devices. SGWC is connected to UPF on to the WAN backhaul network. In open5Gs, SA core components contain the following functionalities such as (1) Access and Mobility management function, (2) Session management function, (3) User Plane Function, (4) Authentication Server function, (5) NF Repository, (6) Unified Data management, (7) Unified Data repository, (8) Policy and charging function, (9) Network Slice Selection Function (NSSF), and (10) Binding support function. In the backhaul network, application hosted on SDN Environment at Datacenter Network. It has configured to listens Openflow protocol at the data plane and the control plane it uses the Openvswitch to control and shape the VMs traffic. Our Middleware was built on top of the infrastructure depicted in Fig. 1. Bharat Operating System Solutions (BOSS GNU/Linux) is an Indian Linux distribution adapted from Debian. KVM is a virtualization module in the Linux kernel that allows the kernel to serve as a hypervisor. KVM turns Linux into a hypervisor, allowing a host system to run numerous distant virtual environments known as guests or virtual machines. The virtual machine image will operate in a “sandbox” environment in which another operating system can be executed. OpenvSwitch is a multilayer virtual switch released under the Apache 2 open-source license [3]. It is developed

Fig. 1 5G-SDN middleware

Adaptive End-To-End Network Slicing on 5G Networks

323

in platform-independent C and can be readily converted to various contexts, such as hardware switch chipsets. Open vSwitch uses 802.1q VLAN, which can effectively segregate various VM users while maintaining inter-VM security. Floodlight is a community-driven open-source SDN controller [10]. It is a Java-based enterprise class OpenFlow controller. Floodlight offers a module loading system to make it simple to customize and set up with few dependencies. Floodlight is defined as a set of apps that run on top of the controller. A QoS middleware lies between virtualized infrastructures in the DCN network and is coupled to the actual data centre architecture. The figure depicts the QoS middleware, which is composed of three distinct planes, namely (1) Application Plane, where application services are hosted as a virtual appliance. (2) Data Plane, via which all virtual appliances’ packets are routed. In general, data plane layer protocols are vendor specific; however, open flow protocols can be specified in SDN controllers. Openvswitch is a well-known virtual soft switch that handles OpenFlow protocols at the southbound interface. (3) The control plane is the network’s brain; it provides a centralized view of the whole network, including network topology learning, host and link auto-detection, firewall, access control lists, and all network functions, including traffic control and QoS services. QoS is a technique for controlling bandwidth, delay, flapping, and packet loss in a network. It is used to assign various priorities to different traffic in order to regulate delay and flapping and reduce packet loss. When the network is overcrowded or congested, QoS helps ensure the regular transmission of vital business traffic. A QoS middleware in the SDN controller architecture is given below in Fig. 2, which comprises the following components: Flow pusher, QoS flow injector, and QoS Manager. Based on traffic monitoring, the QoS flow pusher dynamically alters the rate limit of the bandwidth based on the needs. The QoS manager collects all real-time traffic data and updates statistics tables. It can also control the number of queues that can be established. A QoS flow injector can assist in reserving the necessary bandwidth for an application. This enables the creation of QoS depending on IP or apps. Based on network traffic factors, QoS can be divided into two types: Guaranteed Bit Flow rate (GBR) and Adaptive QoS. GBR collects and maintains information about all VM ports and associated network slice information. It ensures that the flow rate does not fall below the specified rate limit. The queue was formed with static parameter values based on the statistics information. For example, consider the JITSI application, which is hosted on virtual infrastructure and has the following services running inside the virtual appliance: ngnix server, STUN/TURN services, video bridge, and other low-priority services. These services share traffic bandwidth in accordance with the fair share mechanism. Services that demand a lot of bandwidth, like video bridging and STUN/TURN, are suffocating due to the equitable sharing of bandwidth with other services. The traffic priority varies from service to service. To avoid these limitations, QoS is critical in addressing the bandwidth sharing issue. The traffic can be classified depending on the priority of the services, and the QoS has been set for each service based on the pre-specified parameters. This will help to reduce the latency of application services and improve the QoE to the end-users.

324

P. Sakthi Saravanakumar et al.

Fig. 2 QoS middleware in the SDN controller architecture

Figure 3 represents the network slicing in JITSI. An adaptive QoS middleware calculates latency and jitter for each application service. The sample flow was taken in order to detect the high latency services as well as the average jitter for an application. These details are constantly updated in the QoS manager’s real-time bandwidth statistics table. The QoS manager already has slice information as well as application traffic priority. When traffic falls below a certain threshold, the flow pusher quickly adjusts the bandwidth restriction to the maximum of predefined settings. At the same time, the least priority services restriction rate is changed to guarantee that packets go smoothly. All of these changes are confined to the VM bandwidth limit chosen during provisioning.

Adaptive End-To-End Network Slicing on 5G Networks

325

Fig. 3 Network slicing in JITSI

4 Implementation Details In the 5G environment testbed, the 5G core and gNB (RAN Simulator) were deployed as distinct virtual appliances, whereas UERANSIM (UE) was operated within the application virtual machine. UERANSIM must first be registered with the 5GC simulation vm. During registration, the WebUI will start on port 3000 on the 5GC vm, and you will be able to register a new subscriber with the IMSI: 901,700,000,000,001 and other default information such as QoS parameters. 5G QoS Identifier (5GQI), Allocation and Retention Priority (ARP), Reflective QoS Attribute (RQA), Notification Control, Flow Bit Rates, and Aggregate Bit Rates are among the 5G QoS parameters. This ensures the bit flow rate at the UE level. This technology establishes a private network (tun device) over virtual networks ( tap device). It establishes the PDU session with the tun device in 5G. Now all the 5gs appliances connect with tun devices exclusively. Figure 5 represents the datacentre view of virtual appliance (Fig. 4). In our 5G environment, we trail run with two distinct health information system. One employs audio video conferencing system for consultation and exchanging massive amount of data exchanged between servers and clients. Another application employs voice-based clinical prescription generating (ASR assistance) as well as remote consultation via the webRTC module. 1. 5 Concurrent users, the total of 163.92 MB of data has been draw down; the working average on lower resolution with observing freezes and frames drops, in higher resolution will face more issue. Average user experience has been depicted in Fig. 5 and the data set out in Table 1, Sl.No1. 2. Record exchange: Upload and download of records, the total of 18.72 GB of data has been use up; the average user experience for sharing the medical imaging been depicted in Fig. 6 and the data set out in Table 1, Sl.No 2. The trails have been conducted for limited user loads and extrapolated, since emulating workflows for actual users would require emulation of many UE nodes. Video calling simulations have been performed through web browser-based calls; practical use would involve doctors initiating video calls through mobile devices or web and patients joining through mobile. ASR and notifications are Internet-based services directly initiated from mobile, so traffic captured at SDN VMs does not consider them directly; the Fig. 7 depicted the workflow of user and the data set out in Table 2.

326

P. Sakthi Saravanakumar et al.

Fig. 4 Datacentre view of virtual appliance

Fig. 5 Concurrently 5 users involved in AV

Fig. 6 Records exchange: upload and download of records

The sequence of events record in healthcare application with data size (MB) vs workflow for different size of users (Fig. 8). The sequence of events record in Health Care Application with data size (MB) vs workflow for different size of users.

Adaptive End-To-End Network Slicing on 5G Networks

327

Table 1 Traffic statistics for healthcare application Sl. no

Activity details

Avg. active duration (min)

Network IN

Network OUT

1

Concurrently 5 users involved in AV

39

Avg: 150 KB Max: 217.19 KB

Avg: 396.3 KB Max: 790 KB

2

Records Exchange: 30 Upload and download of records

Avg: 61.89 MB Max:159.66 MB

Avg: 9.44 MB Max:159.6 MB

Fig. 7 Traffic pattern of healthcare application

Table 2 Traffic statistics for healthcare application 2 Sl. No

Workflow

1

1 User

10 Users

100 Users

In (MB)

Out (MB)

In (MB)

Out (MB)

In (MB)

Out (MB)

Patient document upload

7.59

0.92

75.9

9.23

759

92.29

2

Doctor approva/document view

4.89

1.69

48.9

16.9

489

169

3

Voice-based prescription

6.87

1.73

68.7

17.3

687

173

4

Prescription upload/view

7.57

1.65

75.7

16.5

757

167

5 Conclusion and Future Work This research proposes an SDN-based end-to-end network slicing in a 5G environment for efficient sharing of RAN and cloud resources. The proposed method has been tested on Open5GS as well as virtualized infrastructure. The current approach has the limitation of using emulators connected to VMs through SDN; true QoS values would be evaluated using physical testbeds incorporating UEs / physical devices in the workflow. User traffic is extrapolated based on a sequential process workflow;

328

P. Sakthi Saravanakumar et al.

Fig. 8 Workflow for single user, 10 and 100 users

however, in actuality, users generate traffic in simultaneously (with delays). Furthermore, the arrangement necessitates several physical devices (UE) connecting with servers. Video calling trials were carried out utilizing browser-based call initiations and do not take into account the effect of traffic or latency caused by mobile device use. Only data rate monitoring simulations are currently available in the current configuration. Future work is required to improve the noise or background traffic addition to account for numerous RAN parameters.

References 1. Esmaeily A, Kralevska K, Gligoroski D (2020) A cloud-based SDN/NFV testbed for endto-end network slicing in 4G/5G. In: 2020 6th IEEE conference on network softwarization (NetSoft) (2020), pp 29–35.https://doi.org/10.1109/NetSoft48620.2020.9165419 2. Open5GS https://open5gs.org 3. OpenvSwitch, http://www.openvswitch.org/ 4. Shorov A (2019) 5G testbed development for network slicing evaluation. In: IEEE conference of Russian Young researchers in electrical and electronic engineering (2019), pp 39–44 5. Samdanis K, Costa-Perez X, Sciancalepore V (2017) From network sharing to multi- tenancy: the 5G network slice broker. IEEE Commun Mag (2017) 6. Nikaein N, Chang C-Y, Alexandris K (2019) Mosaic5G: agile and flexible service plat- forms for 5G research. SIGCOMM Comput Commun Rev 48:29–34 7. Kim J, Xie M (2019) A study of slice-aware service assurance for network function virtualization. In: 2019 IEEE conference on network softwarization (NetSoft) (2019), pp 489–497 8. Costanzo S, Fajjari I, Aitsaadi N, Langar R (2018) DEMO: SDN-based network slicing in C-RAN. In: 2018 15th IEEE annual consumer communications and networking conference (CCNC) (2018), pp 1–2. https://doi.org/10.1109/CCNC.2018.8319321

Adaptive End-To-End Network Slicing on 5G Networks

329

9. Kim S, Lee T, Kim K (2020) Research on the traffic type recognition technique for advanced network control using Floodlight. In: 2020 14th international conference on ubiquitous information management and communication (IMCOM) (2020), pp 1–6. https://doi.org/10.1109/ IMCOM48794.2020.9001773 10. Akcay H, Yiltas-Kaplan D (2017) Web-based user interface for the Floodlight SDN controller. Int J Adv Netw Appl 8(5):3175–3180

Zero Trust Framework in Integrated Cloud Edge IoT Environment S. Kailash, Yuvaraj, and Saswati Mukherjee

Abstract The increased demand for computation and storage needs in recent days cannot be catered fully by single environment, be it public clouds or private clouds or edge nodes, as the request specification are heterogeneous in nature. This led to the need for integrated environment comprising heterogeneous service providers and execution environments with varied components that constitute the integrated environment to cater service requirements. This leads to security requirements such trust management, reliability, authentication and authorization issues. A Zero trust framework within the integrated environment is proposed to address the security challenges by adaptive access control, policy enforcement. Along with the Zero trust framework, a smart gateway between IoT devices and edge nodes is proposed to handle the heterogeneity in protocols, applications and platforms upon which the applications are built. Keywords Zero trust framework · Cloud security · Zero trust model · Trust management system

1 Introduction Zero Trust The utilization of information technology platform by various domains is increasing rapidly. The demand from these domains are heterogeneous in nature which cannot be catered by standalone environments like cloud or edge or IoT devices as these standalone environments lack certain functionalities or services to cater these heterogeneous requirements. This led to a stream termed as cloud interoperability where the services are interoperable across multiple cloud services, service providers and S. Kailash (B) Centre for Development of Advanced Computing, Chennai, India e-mail: [email protected] Yuvaraj · S. Mukherjee College of Engineering, Anna University, Guindy, Chennai, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_21

331

332

S. Kailash et al.

federation of clouds. On the other hand, with the penetration of Internet of Things (IoT) and Industrial Internet of Things (IIoT), in multiple verticals including smart cities and Industry 4.0, the need for edge computing rapidly increased. To satisfy the performance and Service Level Agreement (SLA) requirements, the edge computing environment with IoT gateways is integrated with clouds, leading to formation of the integrated environment. It is evident that the integrated environment can cater the computational storage demand from various verticals. The functional components of an integrated environment comprise the combination of all or any of the following—public and private clouds, edge nodes, IoT gateways along with IoT devices or sensors, application and platform services. As per the supply chain management strategies, to balance the service requirement from domains and the service offering from the integrated environment, the variety of services, platform and resources are induced into the integrated environment. On the other side, this variety of services lead to security issues such as lack of trust, lack of validation of the new elements or process into the integrated environment. To fill the security requirements, Zero trust security mechanism is introduced into the computation environments like cloud computing. Zero trust at the first instance does not trust any of the components in the underlying environment. Zero trust framework assumes that all the software, hardware and networking components in clouds, edge nodes, IoT gateways are not trusted by default. Zero trust frame work in the integrated environment performs the functions of authentication, authorization and continuous validation of each and every components and process such as network equipments, traffic, computational units, storage units, platform, operating system process, application, data, data movement, APIs, monitoring tools, security tools, management process such as backup recovery, user credentials, controller functionalities and so on. Although there are various security mechanism to protect the perimeter of the clouds, edge nodes and IoT gateways, there are possibilities of denial of service attacks, insider attack, threats due to misconfigurations and so on, which can be addressed by inclusion of Zero trust framework. Various breaches are due to the conventional security mechanisms were unable to track suspicious activities caused by either internal or external entities [8]. Zero trust is based on the principle that no entity whether internal or external is to be trusted by default. In a Zero trust model, every access request in the underlying environment is strongly authenticated, authorized based on policy constraints. Most of the devices that we play with in the day to day life started to sense us for the betterment and upliftment of the human race. IoT is one fine field which paves the way to lighten our future through smartness, remoteness and cost-effectiveness. A smart gateway will supplant the customary system gateway to associate the home arrange and the Internet. The smart gateways will tackle the heterogeneity in protocols, applications and platforms upon which the applications are built.

Zero Trust Framework in Integrated Cloud Edge IoT Environment

333

In this paper, to tackle the security issues in the integrated environment comprising of hybrid clouds, public clouds, private clouds, edge nodes, IoT gateways, a Zero trust framework is proposed. To handle the functionalities at IoT gateway, smart IoT gateways are proposed which upon positioned between IoT devices or sensor and edge nodes will certainly lead to improved performance as the gateway possess certain level of automation like decision making and so on. The architecture of the Zero trust framework in an integrated environment is depicted in the paper. The paper introduces the concept of Zero trust in the integrated environment along with various research works on Zero trust, integrated environment and IoT gateways in literature. The architecture of the proposed Zero trust framework is explained the Sect. 3 followed by the working methodology of the IoT gateway. The implementation methodology is explained in Sect. 4 followed by conclusion, future work and references for design of the architecture.

2 Literature Survey To ensure safety and security, it is mandatory that organizations ensure all requests are continuously vetted and validated prior to permitting access to any of the assets including user identity and credentials, endpoint protection to devices, protocols, software platform with patch levels, applications with identity verification and behavioural analysis, micro-segmentation, endpoint security and least privilege controls [1]. In the existing scenario, cloud applications and the mobile workforce have redefined the security perimeter. With the inclusion of bring your own devices, the services and data are accessed outside the secured network and shared with external collaborators [2]. National Institute of Standards Technology (NIST) emphasis the need for a new security model that more effectively adapts to the complexity of the modern environment, embraces the mobile workforce and protects people, devices, applications and data wherever they are located. This is the core of Zero trust [2]. The consensus algorithm enables the users to trust the network as malicious nodes fail to get approval from all nodes, thereby causing the transaction to be aborted [3]. Truong et al. explain that Internet of Things devices are always at risk of network security. It is therefore crucial for an IoT network design to identify attackers accurately, quickly and promptly and proposed an edge-cloud architecture that fulfils the detection task right at the edge layer, near the source of the attacks for quick response with a multi-attack detection mechanism called Low-Complexity Cyberattack Detection in IoT Edge Computing (LocKedge), which has low complexity for deployment at the edge zone while still maintaining high accuracy. In spite of this, a reliable IoT system must meet many security requirements such as access control and authentication at the edge layer and attack detection at the network layer [4]. In [5], authors combine edge and cloud computing for IoT data analytics by taking advantage of edge nodes to reduce data transfer. In order to process data close to the source, sensors are grouped according to location. Combining edge with cloud computing has the potential to reduce IoT network traffic and associated latencies

334

S. Kailash et al.

while still supporting complex data analytics tasks. A blockchain-based gateway is proposed in [6], to handle the following scenario—before a user access an IoT device, the administrator can gain access to store device information and the privacy policies of the device in the blockchain network. The devices should also be trusted to avoid insider attack. Wazid et al. in [7] explain that majority of existing user authentication schemes proposed for WSNs and IoT have several drawbacks and also they are insecure against various known attacks and most existing schemes lack to preserve user and sensing node anonymity property, and to support various functionality features, such as efficient login and authentication phases, dynamic sensing node addition phase and biometric and password update phase, which are essential for some critical applications including military and battlefield scenarios and tactical surveillance and suggest that an user authentication scheme is extremely needed to provide high security and additional functionality features to authentication protocols. The summary of various IoT gateways from literature is listed as follows, Author name Paper title and year

IoT Hardware environment used for and gateway application

Transport layer protocol

Mahmud Al-Osta, Bali Ahmed, Gherbi Abdelouahed (2017)

A Smart lightweight home, semantic Smart city web-based approach for data annotation on IoT gateways

Charbel El Kaed, Imran Khan, Andre Van Den Berg, Hicham Hossayni, and Christophe Saint-Marcel (2018)

SRE: Semantic rules engine for the industrial Internet of Things gateways

Byungseok Kanga, Hyunseung Choob (2017)

An Smart home Raspberry Wi-Fi, experimental Pi LTE study of a reliable IoT gateway

Application Gateway layer features protocols

Raspberry Not MQTT, Pi 3 mentioned CoAP

Smart Not ZigBee, Not home, smart mentioned Ethernet, mentioned industry Bluetooth, and Wi-Fi

IoT framework protocol similar to HTTP model

Optimized data annotation of IoT data, device monitoring and control

Device monitoring, rule and semantic engine for effective device control

Dynamic discovery and auto-registration of devices, device monitoring and control (continued)

Zero Trust Framework in Integrated Cloud Edge IoT Environment

335

(continued) Author name Paper title and year

Fatemeh Jalali, Olivia J Smith, Timothy Lynar, Frank Suits (2017)

Cognitive IoT gateways: Automatic task sharing and switching between cloud and edge/fog computing

IoT Hardware environment used for and gateway application Not clearly mentioned

Transport layer protocol

Raspberry Wi-Fi Pi

Application Gateway layer features protocols Not clearly mentioned (generic)

Device monitoring and control, automatic task sharing and switching between cloud and fog

3 Zero Trust Framework in Integrated Environment The proposed zero trust framework in integrated environment is depicted in the following Fig. 1.

3.1 Integrated Environment The integrated environment comprises, IoT devices such as sensors and other devices connected to IoT gateways. The IoT devices such as sensors transfer the data to the IoT gateways. The proposed gateway is smart enough to make a decision to process the data in edge nodes or in clouds based on the factors such as urgency of data to be processed, data retention period, volume and velocity of data. IoT gateways connected to edge nodes, acts as an extended mini cloud environment capable of processing the data, storing the data for a shorter life span and on other end connected to clouds - one or more of public, private or hybrid clouds.

336

S. Kailash et al.

Fig. 1 Zero trust framework

The cloud environments considered in the integrated environment are public, private and hybrid clouds. All the constituents of the clouds—network devices and networking controllers, storage devices, compute units such as servers, hypervisors, container orchestration engineer, cloud orchestration software, virtual machines, containers, software stack, operating system, application, data, monitoring and management tools such as backup, replication, security solutions, and any of the microservices or applications that access these components in the integrated environment such as IoT nodes, edge or clouds are included in the preview of Zero trust framework. The proposed Zero trust security framework includes the functional modules such as identity and access control, micro-segmentation, policy formulation and enforcement with configuration management, trust algorithm with continuous diagnosis and mitigation, threat protection, behavioural analytics engine, adaptive access control mechanism, credentials management and log management. Micro-segmentation breaks up the security perimeters into small zones to maintain separate access for individual parts of the network. The proposed framework includes network segmentation preventing lateral movement, providing Layer 7 threat prevention, and simplifying granular user access control. The access to individual resources are validated and granted on per connection/per session basis determined dynamically by the factors like policy, system state and so on. The credentials management service protects secrets needed to access applications, services and IT resources, enabling easy rotate, manage and retrieve data, API

Zero Trust Framework in Integrated Cloud Edge IoT Environment

337

keys and other secrets throughout the service lifecycle. The service has extensible API keys and OAuth tokens control access to secrets using fine-grained permissions and audit secret rotation centrally for resources. The proposed Zero trust framework ensures that when one resource or service compromised, will not affect other systems. The framework works on the principle that authenticate and authorize based on all available data points, use least privileged access, predict breach based on behavioural analytics and every request is authorized authenticated based on policy constraints. The intellectual architecture in which IoT devices is proposed works as, application forwards the data form devices; the gateway registers, authenticates and stores the privacy details of the user’s IoT device data in gateways for IoT devices. Initially, all the IoT devices are made to register with their default and most suitable gateway and there by setting the access control over the device, prevention over data leakage, etc. Whenever the IoT devices gathers some data, it will be directed to its corresponding IoT gateway that is registered previously so that gateway will verify the device information with the help of device description available in it. Then the gateway will look forward to find out the nature of the application that it is handling and the rate of time sensitivity in it. Usually, all the device data are pre-assigned to reach the nearest Fog network and only if there are any specific conditions applied over the data or the request is too rare from that particular device; it will be moved to cloud. In this situation, there could be two problems that any network faces. One is, when the registered fog is busy and if so to which of the other fog it should redirect the device request? Secondly, how to solve data computation issues, when the nature of application and the no. of data records in fog and cloud to satisfy any of that application’s request contradicts? For these we deploy intelligence added gateway for prediction of choosing cloud/fog depends upon application data.

3.2 IOT Gateway Management The new approach of designing the gateway assumes that all the rest of the nonfunctional requirements are already carried off and focus mainly on data storage and deciding the computational end for an application request. So this work divides the functionality of a gateway into two cycles, (i) Mapping Cycle an initial storage mechanism that reads device description elaborately and with the help of a SVM, binds them either in fog/cloud (ii) Routing Cycle—upon having the knowledge of the nature of the application and the backup of data available, gateway will send it to either fog or cloud.

338

S. Kailash et al.

Mapping Cycle The problem of under-utilizing a LAN of gateways and as well as the problem of gateway deficiency has been successfully taken care of through this approach. After studying many withstanding architectures like Nginx, this automatic tuning algorithm is what we suggest for the underlying architecture. Any device that has to send request will access any one of the registered gateway IP to authenticate and then transfer the request to any of the computing higher end. In this phase, we consider two scenarios. One is what if the incoming application is with heavy weight application and required to have more computational resources? Secondly, what if the request that is sent matters a lot about time factor and turns to be an alarmed issue in the gateway? So in this phase, after knowing that the incoming element is data packet directly coming from the end user (and not device request), the gateway will check, for the given application id is there any mapped computational unit for data storage? If so it will route the incoming data to the corresponding computational id. If it is going to be a new device or a device that needs alternative computational resource, then the application is sent to SVM module. It is applied over wide range of applications that has various contributing factors for their computation such as CPU utilization, network bandwidth, time responsiveness nature, frequency of request, payload size and device description. As a result, SVM will classify each application that passes the gateway as time sensitive or non-time sensitive. Depending upon this key value, the native location for data storage from each application is decided. This will determine the time sensitivity nature of the application and returns a binary value to the gateway. Upon receiving this value, the gateway either sends the data to the nearest cloud or to the fog server. In the fog server, the device will pick a fog through RR manner and updates its arrival in the mapping table. If no fog is available with the minimum threshold of holding the incoming, the fog server initiates new instance and allocated the device to the new fog for storage. Thus, all the incoming data elements are reaching their destination as per their features. Routing Cycle In this phase, the mapped devices and their computational units are re-analysed according to the request and brilliantly decided to get run on either Fog or Cloud depending upon the following factors (i) Nature of Request—Time consuming or not, (ii) time sensitive or not, (iii) availability of data, (iv) locating result of similar query in the past and (v) computational requirements of the request. Depending upon the above factors and the SVM algorithms suggestion, the request’s computation will be decided. There are two situations where the decision making has be done, one is fog and cloud with varying data records. Secondly fog and cloud with number of data records. In these situations, how to find out a perfect computing destination to match the end user needs. The idea is as follows.

Zero Trust Framework in Integrated Cloud Edge IoT Environment

339

Once the incoming element is found to be a request, immediately gateway will look for the similar request in that application log table. If found, the result is fetched by looking into the log of that application (where the computation of that previous request done) and will be immediately sent. If there are no similar requests, then find out the no. of data records collectively available in a fog or cloud for that particular request. If the application is too time sensitive, and fog records are very large in number than that of cloud records, move the application to fog. On the other hand, if the application is not time sensitive and there are more no. of records in cloud than that of fog, then choose computational destination as cloud and move the fog data to cloud. If there are equal no. of records in fog and cloud and fog, then do parallel processing and merge the result store it in the fog. The combined algorithm for both the mapping cycle and the routing cycle is mentioned below. Description 1. 2. 3.

Applications IDs—Let Aid be the number of applications ranges from A1 to An. Request IDs—Lets device request be ranging from R1 to Rg and denoted as Rs. Device Description—A tuple that contains detailed description about the device. Let the descriptions be DevDid and it ranges from DevD1 to DevDn , as one for each application. 4. Time—It is a categorical input that helps to make the decision making for storage and computing. Let Tsid be the time sensitivity, as one for each application. 5. Destination—Let Dst be the destination edge notation where storage and computation work will take place. 6. Gateways—Let GWj be the gateways ranges from GW1 to GWm. 7. Fog Nodes—Let Fid denote the fog Id which can range from F1 to Fk . 8. Cloud Nodes—Let Cid denote the fog Id which can range from C1 to Cb. 9. Utilization Rate—Let URfid be the utilization rate for all the fogs ranges from UR1 to URk . 10. DATA—Lets DAid be the device data from the user, upon which any number of Rs can be applied.

340

S. Kailash et al.

Zero Trust Framework in Integrated Cloud Edge IoT Environment

341

4 Implementation The integrated environment is implemented as follows – for public cloud, AWS environment [11] is considered. Private cloud is considered with 3 tier architecture where 10 servers and workstations with OpenStack [10] are considered of which 4 servers are positioned as primary data centre and 3 servers as nearline data centre and 2 servers as disaster recovery centre. Within primary and nearline data centre, totally three OpenStack controller instances are deployed as virtual machines on high availability mode. The disaster recovery centre cloud acts as standalone cloud. The primary data centre hosts the workload in form of virtual machines and containers. The VMs and containers are replicated between primary and nearline data centre on synchronous mode of replication and between nearline and disaster recovery centre, the replication is asynchronous. A shared storage is employed within primary data centre to achieve live migration of VMs. 3 Edge nodes with less computing capacity are considered with StarlingX [9] is employed. The edge gateways are micro computational units connected with temperature, humidity, fire and motion sensors for implementation. The Zero trust framework is implemented as microservices with Python. All the tools employed are free and open source tools.

5 Conclusion and Future Work The demand for integrated environment due to the adoption of cutting edge technologies such as Cloud, Edge, IoT and so on increases rapidly. The requirements are heterogeneous in nature which can be catered only by integrated environment comprising of various clouds, edge nodes. The increased utilization by various consumer category and various services lead to security threats. To address the security and trust concerns in the integrated Cloud Edge IoT environment, a Zero trust framework is proposed which address major security issues such as trust, authentication and authorization, privacy concerns and so on. A smart gateway is proposed to address the issues in existing conventional IoT gateways such as lack of decision making in routing the workloads. Few of the future work to be carried out on the proposed architecture are inclusion of micro-segmentation in the Zero trust framework, orchestration with software-defined security, auto-scaling of edge nodes.

References 1. https://www.crowdstrike.com/cybersecurity-101/zero-trust-security/ 2. https://www.nccoe.nist.gov/projects/building-blocks/zero-trust-architecture 3. Patil AP, Karkal G, Wadhwa J, Sawood M, Reddy KD (2020) Design and implementation of a consensus algorithm to build zero trust model. In: 2020 IEEE 17th India Council international conference (INDICON). https://doi.org/10.1109/INDICON49873.2020.9342207

342

S. Kailash et al.

4. Huong TT, Bac TP, Long DM, Thang BD, Binh NT, Luong TD, Phuc TK (2021) LocKedge: low-complexity cyber attack detection in IoT edge computing. IEEE Access 9 5. Ghosh AM, Grolinger K (2021) Edge-cloud computing for internet of things data analytics: embedding intelligence in the edge with deep learning. IEEE Trans Indust Inf 17(3) 6. Cha S-C, Chen J-F, Su C, Yeh K-H (2018) A blockchain connected gateway for BLE-based devices in the internet of things. IEEE Access 6:24639–24649. https://doi.org/10.1109/ACC ESS.2018.2799942 7. Wazid M, Das AK, Odelu V, Kumar N, Conti M, Jo M (2018) Design of secure user authenticated key management protocol for generic IoT networks. IEEE Internet Things J 5(1):269–282. https://doi.org/10.1109/JIOT.2017.2780232 8. Sen J, Innovation Labs, Tata Consultancy Services Ltd., Kolkata, India, Security and Privacy Issues in Cloud Computing. 9. https://www.starlingx.io/ 10. https://www.openstack.org/ 11. https://aws.amazon.com/

Ph.D. Track Paper

O2 Q: Deteriorating Inventory Model with Stock-Dependent Demand Under Trade Credit Policy P. Jayashri

and S. Umamaheswari

Abstract Trade credit is one of the marketing strategies, which is imperative in inventory management used by suppliers to minimize the deterioration rate. The credit period for retailers is interest-free. Beyond this credit period, the retailers have to bear the interest for the residual goods. In this research article, the influence of this trade credit on deteriorating inventory models that follow Weibull distributions under stock-dependent demand is analyzed. A retailer’s profits are maximized through permissible delay in payment, constant replenishment, and backlog. For the portrayed model, numerical examples are provided in order to comprehend the mathematical formulation. Analyzing sensitivity gives insight into the effects of modifying key parameters. Keywords Backlogging · Shortage · Stock dependent · Trade Credit · Weibull distribution

1 Introduction Inventory is an ideal resource that has an economic value to an organization. If a resource loses economic value, then it ceases to be an inventory. Every product in inventory undergoes deterioration; hence, it is necessary to utilize the product before it gets expired. To overcome deterioration, suppliers use a wide range of marketing strategies such as discounts, seasonal sales, delay in payment, advance payments, and so on to increase demand. Among these, delay in payment attracts more customers which reduces the risk of revenue loss for suppliers caused by outdated products. In business scenario, trade credit is a credit period extended by the manufacturers to purchase goods without immediate payment. The parameters considered by the P. Jayashri · S. Umamaheswari (B) Division of Mathematics, SAS, Vellore Institute of Technology, Chennai 600127, India e-mail: [email protected] P. Jayashri e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_22

345

346

P. Jayashri and S. Umamaheswari

researchers in designing the EOQ inventory model to achieve maximum profit are displayed below: (a) (b) (c) (d) (e) (f)

the new order is placed when an inventory falls to zero, constant replenishment, stock-dependent relationship between demand and supply, on-hand inventory deterioration with Weibull probability distribution, negative inventory level which indicates a shortage level, and Suppliers give retailers a limited period of time to settle the balance without charging interest.

The proposed inventory model is systematically segmented as: Sects. 2 and 3 examine the detailed list of assumptions and notations. EOQ’s mathematical formulation under delays in payment is provided in Sect. 4. In Sect. 5, numerical calculations are briefly discussed. A sensitivity analysis is performed in Sect. 6 by modifying the values of significant parameters. Section 7 provides a conclusion remark and the future scope of this study.

2 Literature Review Goyal and Giri [1] carried out a detailed study to determine the deterioration rate for perishable inventory items. To obtain a total cost that optimizes net profit, Kavitha Priya and Senbagam [2] suggested an ordering policy inventory model with Weibull deterioration and quadratic time-dependent demand. Patriarca et al. [3] analyzed the inventory model for perishable items under uncertainty to figure out the optimal order quantity. The author employed Monte Carlo simulation to interact with variables that are influenced by ambiguity in the perspective of a probability distribution function with time-dependent demand. Haley and Higgins [4] proposed an inventory policy with trade credit financing to obtain an optimal ordering strategy for the lot size model. In his research article, the author proved that the ordering quantity and payment time are interdependent with two independent decisions variable transition cost and banking arrangement. Goyal et al. [5] used trade credit to extend a traditional EOQ model, in which order quantity and replenishment period rise dramatically as total cost decreases. Mishra et al. [6] analyzed a deteriorating inventory model using delay in payment to obtain an optimal market value and cycle length. The author considered different demand functions to figure out the maximum revenue on remanufactured items. Daellenbach [7] and Chung [8] analyzed the principle of finance (inventory control) with trade credit to minimize the optimal replenishment period. Tiwari and Barron [9] proposed an EOQ with partial trade credit to minimize the total cost by stimulating demand on ordering quantity. The optimal replenishment time and the time to draw down the stock completely are determined theoretically, and then, based on these values, the optimal ordering and backlog policies are determined.

O2 Q: Deteriorating Inventory Model with Stock-Dependent Demand …

347

In recent years, researchers have combined trade credit with preservation techniques to reduce the rate of deterioration. To reduce total cost, YP Lee and CY Dye [10] developed a stock-dependent inventory with preservation technology and a replacement strategy. Rahman et al. [11] designed a perishable inventory model with partial backlogging under a certain ratio. For the non-linear maximizing problem, the author used both price and stock-dependent factors. The comprehensive studies of the various deteriorating items with the proposed inventory models are described below. Source

Demand

Payment replenishment shortage

Roy [12]

Stock

Delay

Kavitha [2]

Time

✓

✓ ✓

Mishra [6]

Hybrid price Delay and stock

Delay

Suganthi [13]

Constan

Delay

Rahman [11]

Price

Advance

This paper

Stock

Delay

✓ ✓ ✓

✓

3 Assumptions The predicted optimal inventory model is designed using the notations and assumptions listed below. • The demand D = a + b I(t) is a function of stock-dependent demand. • θ = ct d is the two-parameter Weibull deterioration rate where c and d denote the scale and shape parameters. • Lead time is zero. • Replenishment rate is uniform. • Delay in payment is allowed. • Shortage with complete backlogging.

4 Notations Notation

Description

B

Backorder level

BC

Backlogging cost

cb

Backorder cost

D

Demand (continued)

348

P. Jayashri and S. Umamaheswari

(continued) Notation

Description

DC

Deterioration cost

h

Inventory holding cost

HC

Holding cost

i

Unit cost

Ip

Interest charged

Ie

Interest earned

I(t)

Inventory level over a period of time t

IP

Interest paid

IE

Interest earned

M

Credit period

OC

Ordering cost

R

Replenishment

τ1

Duration of time while the item is in positive stock

τ

Inventory cycle length

θ

Deterioration rate

5 Mathematical Formulations The differential equation that governs inventory status at any given time is dI (t) + ct d I (t) = −(a + bI (t)), 0 ≤ t ≤ τ1 dt

(1)

dI (t) = −(a + bI (t)), τ1 ≤ t ≤ τ dt

(2)

Initial boundary condition are I (0) = R − S and I (τ ) = −S. Equations (1) and (2) can be modified as follows  dI (t)  d + ct + b I (t) = −a dt

(3)

dI (t) + bI (t) = −a dt

(4)

On solving the above differential equation, we get

O2 Q: Deteriorating Inventory Model with Stock-Dependent Demand …

349

     b τ12 − t 2 ct d +1 τ1d +2 − t d +2 − bt (−a) τ1 − t + C + I (t) = 1 − ; d +1 2 (d + 1)(d + 2)

0 ≤ t ≤ τ1 (5)

a I (t) = 2 e−b(2τ −τ1 −τ ) ; b

τ1 ≤ τ ≤ τ

(6)

Total demand during τ 1 . Dτ1 = (a + bI (t))τ1 .

(7)

The amount of material which deteriorates during one cycle of time (Dτ ) is Dτ = R − (a + bI (t))τ1

(8)

The maximum allowable backorder is given by s = D(τ − τ1 ).

(9)

At t = 0, I (0) = R − S and t = τ, I (τ ) = −S from Eq. (3), the following result is obtained   bτ12 τ1d +2 + R − S = (−a) τ1 + C (10) 2 (d + 1)(d + 2) S=−

a −b(τ −τ1 ) e b2



1 bτ 2 τ1d +2 + 1 + 2 e−b(τ −τ1 ) R = (−a) τ1 + c 2 b (d + 1)(d + 2)

(11)  (12)

Case 1: Payment before the total depletion (τ 1 ≥ m). The customer takes advantage of trade credit and earns profit throughout the inventory cycle. Retailer’s total cost would be the sum of the setup cost, deterioration cost, carrying cost, interest paid, and backlogging cost minus interest earned. TC(τ1 , τ ) = τ1 where, DC = θ i I (t)dt 0

1 [OC + DC + HC + IP − IE + BC]. τ

(13)

350

P. Jayashri and S. Umamaheswari τ −τ1

BC = Cb

Dt dt 0

τ HC = ih I (t)dt 0

τ1 IP = iIp

τ I (t)dt + iIp

m

τ I (t)dt = 0

I (t)dt, iIp τ1

τ1

τ1 IE = iIe

Dtdt. 0

Substituting these in the above equation, we get

 1 {OC + ic τ1d A1 (τ1 ) + ih[A1 (τ1 ) + A2 (τ − τ1 )] τ b(τ − τ1 )2 a(τ − τ1 )2 + I (τ − τ1 ) + b(τ − τ1 )A2 (τ − τ1 )] + cb [ 2 2

2 aτ1 (14) + bτ1 A1 (τ1 ) } + IP A(τ1 − m) − iIe 2

TC =

τ1 where A1 (τ1 ) = I (t)dt 0

m A1 [τ1 − m] =

I (t)dt τ1

T A2 [τ − τ1 ] =

I (t)dt τ1 τ −τ1

A2 [τ − τ1 ] =

I (t)dt 0

The ideal solution of τ 1 and τ is determined by solving the equations =0 and ∂TC ∂τ

∂TC ∂τ1



∂A1 (τ ) ∂A2 (τ − τ1 ) ∂A1 (τ ) ∂TC 1 + ih = {ic d τ1d −1 A1 (τ1 ) + τ1d + ∂τ1 τ ∂τ1 ∂τ1 ∂τ1

=0

O2 Q: Deteriorating Inventory Model with Stock-Dependent Demand …

351

b(τ − τ1 )2 ∂I (t) + cb −a(τ − τ1 ) − b(τ − τ1 )I (τ − τ1 ) + 2 ∂τ1



∂A1 [τ1 − m] ∂A1 (τ1 ) ∂A2 (τ − τ1 ) − iIe aτ1 + bτ1 + b(τ − τ1 ) ] + iIp ∂τ1 ∂τ1 ∂r1 (15)

 ∂TC = [−C0 + ic τ1d A1 (τ1 ) + ih[A1 (τ1 ) + A2 (τ − τ1 )] ∂τ

a(τ − τ1 )2 b(τ − τ1 )2 + I (τ − τ1 ) + b(τ − τ1 )A2 (τ − τ1 ) + cb 2 2

2 aτ1 + iIp [A1 (τ1 − m)] − iIe + bτ1 A1 (τ1 ) ] 2 1 ∂A2 (τ − τ1 ) + [ih + Cb [a(τ − τ1 ) + b(τ − τ1 )I (τ − τ1 )] τ ∂τ ∂A2 (τ − τ1 ) b(τ − τ1 )2 ∂I (t) + b(τ − τ1 ) ]. (16) + 2 ∂τ ∂τ Provided these values of ti , have obtain above equation satisfy the conditions of 2 2 1 ,τ ) 1 ,τ ) Hessian determinant Di ≤ 0 (i = 1, 2) is given by ∂ TC(τ < 0 and ∂ TC(τ < 0. ∂τ1 2 ∂τ 2

1 ,τ ) 1 ,τ ) 1 ,τ ) The Hessian matrix for TC ( ∂ TC(τ )( ∂ TC(τ ) − ∂ TC(τ > 0 is always nega∂τ1 2 ∂τ 2 ∂τ 1 ∂τ tive defined, and the function is strictly concave and differentiable. As a result, (τ 1 ∗ , τ ∗ ) meets the global maximum, and (τ 1 ∗ , τ ∗ ) is unique# . #Based on Roy and Chaudhuri [12] analyzed inventory model with Weibull deterioration and demand that is stock dependent. The results are discussed in the proposed model. 2

2

2

Case 2: Payment on the total depletion (t = m). The deterioration cost, carrying cost, and Penalty cost are same. As the supplier paid in full during the permitted in delay, the interest payable for the period τ 1 ≤ t ≤ τ is zero.

 1 {OC + ic τ1d A(τ1 ) + ih[A1 (τ1 ) + A2 (τ − τ1 )] τ

a(τ − τ1 )2 b(τ − τ1 )2 + I (τ − τ1 ) + b(τ − τ1 )A2 (τ − τ1 ) + cb 2 2

2 aτ1 (17) − iIe + bτ1 A1 (τ1 ) } 2

TC =

The ideal solution of τ 1 and τ is determined by solving the equations =0 and ∂TC ∂τ



∂TC ∂A1 (τ ) ∂A1 (t) 1 ∂A2 (τ − τ1 ) + ih = {ic d τ1d −1 A1 (τ1 ) + τ1d + ∂τ1 τ ∂τ1 ∂τ1 ∂τ1

∂TC ∂τ1

=0

352

P. Jayashri and S. Umamaheswari  + cb −a(τ − τ1 ) − b(τ − τ1 )I (τ − τ1 ) +

∂A1 (τ1 ) } − iIe aτ1 + bτ1 ∂r1

b(τ − τ1 )2 ∂I (t) ∂A2 (τ − τ1 ) + b(τ − τ1 ) ] 2 ∂τ1 ∂τ1

(18)

 ∂TC = [−C0 + ic τ1d A(τ1 ) + ih[A1 (τ1 ) + A2 (τ − τ1 )] ∂τ

a(τ − τ1 )2 b(τ − τ1 )2 ∂I (t) + I (τ − τ1 ) + b(τ − τ1 ) + cb (τ − τ1 ) 2 2 ∂τ

2 aτ1 − iIe + bτ1 A1 (τ1 ) ] 2 1 ∂A2 (τ − τ1 ) [ih + Cb [a(τ − τ1 ) + b(τ − τ1 )I (τ − τ1 )] τ ∂τ ∂A2 (τ − τ1 ) b(τ − τ1 )2 ∂I (t) + b(τ − τ1 ) ] (19) + 2 ∂τ ∂τ Provided these values of ti , have obtain above equation satisfy the conditions of 2 2 1 ,τ ) 1 ,τ ) Hessian determinant Di ≤ 0 (i = 1, 2) is given by ∂ TC(τ < 0 and ∂ TC(τ < 0. 2 ∂τ ∂τ 2 1   ∂ 2 TC(τ1 ,τ ) ∂ 2 TC(τ1 ,τ ) ∂ 2 TC(τ1 ,τ ) − ∂τ 1 ∂τ > 0 is always The Hessian matrix for TC ( ∂τ1 2 ) ∂τ 2 negative defined, and the function is strictly concave and differentiable. As a result, (τ1 *, τ *) meets the global maximum, and (τ 1 *, τ *) is unique# . Case 3: Payment after the total depletion (t ≤ m). The Penalty cost, carrying cost, and deterioration cost are same. The interest payable is zero for the period τ 1 ≤ t ≤ τ because supplier paid in full during permissible in delay. τ1 IE = iIe

 τ Dtdt + i(a + bI (t))tIe m − 2

(20)

0

 1 {OC + ic τ1d A(τ1 ) + ih[A1 (τ1 ) + A2 (τ − τ1 )] τ

a(τ − τ1 )2 b(τ − τ1 )2 + I (τ − τ1 ) + b(τ − τ1 )A2 (τ − τ1 ) + cb 2 2

2  aτ1 τ − iIe + bτ1 A1 (τ1 ) + +i(a + bI (t))τ1 Ie m − } (21) 2 2

TC =

The ideal solution of τ 1 and τ is determined by solving the equations = 0. and ∂TC ∂τ



∂A1 (t) ∂TC ∂A1 (τ ) 1 ∂A2 (τ − τ1 ) + ih = {ic d τ1d −1 A1 (τ1 ) + τ1d + ∂τ1 τ ∂τ1 ∂τ1 ∂τ1

∂TC ∂τ1

=0

O2 Q: Deteriorating Inventory Model with Stock-Dependent Demand …  + cb −a(τ − τ1 ) − b(τ − τ1 )I (τ − τ1 ) +

353

b(τ − τ1 )2 ∂I (t) ∂A2 (τ − τ1 ) + b(τ − τ1 ) 2 ∂τ1 ∂τ1





∂A2 (τ − τ1 ) ∂I (t) } − iIe aτ1 + bτ1 + (a + bI (t))(m − r1 ) − (a + bI (t))τ1 + τ1 (m − τ1 )b ∂r1 ∂τ1

(22)

 ∂TC = [−C0 + ic τ1d A(τ1 ) + ih[A1 (τ1 ) + A2 (τ − τ1 )] ∂τ

a(τ − τ1 )2 b(τ − τ1 )2 + I (τ − τ1 ) + b(τ − τ1 )A2 (τ − τ1 ) + cb 2 2

2 aτ1 − iIe + bτ1 A1 (τ1 ) ] 2 1 ∂A2 (τ − τ1 ) + [ih + Cb [a(τ − τ1 ) + b(τ − τ1 )I (τ − τ1 )] τ ∂τ ∂A2 (τ − τ1 ) b(τ − τ1 )2 ∂I (t) + + b(τ − τ1 ) ] (23) 2 ∂τ ∂τ Provided these values of ti , have obtain above equation satisfy the conditions of 2 2 1 ,τ ) 1 ,τ ) Hessian determinant Di ≤ 0 (i = 1, 2) is given by ∂ TC(τ < 0 and ∂ TC(τ < 0. . ∂τ1 2 ∂τ 2 1 ,τ ) 1 ,τ ) 1 ,τ ) The Hessian matrix for TC ( ∂ TC(τ )( ∂ TC(τ ) − ∂ TC(τ > 0 is always negative ∂τ1 2 ∂τ 2 ∂τ 1 ∂τ defined, and the function is strictly concave and differentiable. As a result, (τ 1 ∗ , τ ∗ ) meets the global maximum, and (τ 1 *, τ *) is unique# . 2

2

2

6 Numerical Example The numerical solutions to the developed model are shown in the table below: Cases

Cost OC

cb

a

b

c

d

h

m

Ip

Ie

i

T*

τ 1*

TC 576.78

1

220

10

120

0.04

0.2

6

6

1

0.12

0.13

20

3

2

2

220

10

120

0.04

0.2

6

6

–

0

0.13

20

6

20

52.8829

3

220

10

120

0.04

0.2

6

6

4

0

0.13

20

8

21

12.39

T* —Optimal cycle time τ1 * —Optimal positive inventory level TC—Optimal total cost Others—Refer notations

Inference: Fig. 2 shows that the concavity of net profit is strictly concave in relation to the dependent variables τ and τ 1 . The cases 1, 2, and 3 are illustrated in Fig. 2 as (a), (b), and (c), respectively.

354

Fig. 1 Graphical representation of inventory level

Fig. 2 Graphical representation of numerical analysis

P. Jayashri and S. Umamaheswari

O2 Q: Deteriorating Inventory Model with Stock-Dependent Demand …

355

7 Sensitive Analysis The variation due to the parameters is as follows: Parameter OC

a

b

c

d

h

i

Values

τ1

τ

TC (τ1 , τ )

% Change in total cost

0

6

20

42.35

−19

10

6

20

47.6158

−9

20

6

20

52.879

0

30

6

20

58.14

9

40

6

20

63.40

19

80

6

20

50.54

−4

100

6

20

51.72

−2

120

6

20

52.879

0

140

6

20

54. 04

2

160

6

20

55.25

4

0.02

3

8

72.11

36

0.03

5

17

62.74

18

0.04

6

20

52.879

o

0.05

6

18

22.04

−58

0.06

7

21

163.01

208

0

5

15

91.96

73

0.1

4

10

90.39

70

0.2

6

20

52.879

0

0.3

4

10

91.823

73

0.4

4

10

92.5

74

2

5

14

13.48

−74

4

4

10

101.89

92

6

6

20

52.879

0

8

4

10

89.81

69

10

4

10

89.66

69

2

4

10

7.21

−86

4

4

10

49.16

−7

6

6

20

52.879

0

8

5

15

104.83

98

10

5

15

105.31

99

45

4

10

122

130

50

4

10

106

100

55

6

20

52.879

0

60

5

15

73

38 (continued)

356

P. Jayashri and S. Umamaheswari

(continued) Parameter

Values

τ1

τ

TC (τ1 , τ )

% Change in total cost

65

5

15

42.79

−19

8 Conclusion In this research article, the inventory deterioration rate is influenced by Weibull distribution and demand is modeled as stock dependent. An optimal ordering quantity is accomplished by influencing a complete backlog and constant replenishment. It is evident from the research model, delays in payment not only decrease deterioration as well as increase retailers’ net profit when established in inventory. The research model is enhanced with numerical calculation and sensitive analysis. The future scope of this model can be extended for perishable items with upstream and downstream trade credit.

References 1. Goyal S, Giri B (2001) Recent trends in modeling of deteriorating inventory. Eur J Oper Res 134(1):1–16 2. Kavitha Priya R, Senbagam K (2018) An EOQ inventory model for two parameter Weibull deterioration with quadratic time dependent demand and shortages. Int J Pure Appl Math 119(7):467–478 3. Patriarca R, Di Gravio G, Costantino F, Tronci M (2020) EOQ inventory model for perishable products under uncertainty. Prod Eng Res Devel 15(5–6):601–612 4. Haley CW, Higgins RC (1973) Inventory policy and trade credit financing, management science. Manag Sci 20(4) 5. Goyal SK (1985) Economic order quantity under conditions of permissible delay in payments. J Operat Res Soc 36:35–38 6. Mishra U, Wu JZ, Tseng ML (2019) Effects of a hybrid-price-stock de- pendent demand on the optimal solutions of deteriorating inventory system and trade credit policy on re-manufactured product. J Cleaner Product 241:118282 7. Daellenbach HG (1986) Inventory control and trade credit. J. Operat. Res. Soc. 37(5):525–528 8. Kee HC (1989) Inventory control and trade credit revisited. J Oper Res Soc 40(5):495–498 9. Tiwari S, Barron LEC, Shaikh AA, Mark G (2020) Retailers optimal ordering policy for deteriorating item under order size dependent trade credit and complete backlogging. Comput Indust Eng 139:105559 10. Lee YP, Dye CY (2012) An Inventory model for deteriorating items under stock dependent demand and controlled deterioration rate 63:474–482 11. Rahman S, Khan A, Halim MA, Nofal TA, Shaikh AA, Mahmoud EE (2021) Hybrid price and stock dependent inventory model for perishable goods with advance payment related discount facilities under preservation technology. Alex Eng J 6:3455–3465 12. Roy T, Chaudhuri K (2009) A production-inventory model under stock-dependent demand, Weibull distribution deterioration and shortage. Int Trans Oper Res 16(3):325–346

O2 Q: Deteriorating Inventory Model with Stock-Dependent Demand …

357

13. Suganthi K, Jayalalitha G (2019) An inventory model with constant replenishment rate, triangular distribution deteriorating under constant demand and trade credit policy. Int J Anal Exper Modal Anal 9(10):131–137 14. Goyal A (2018) An inventory model with multivariate demand Weibull distribution Deterioration under the effect of inflation and trade credit. Int J Pure Appl Math 118(22):1309–1323 15. Jamal MM, Sarker BR, Wang S (1997) An ordering policy for deteriorating items with allowable shortage and permissible delay in payment. J Oper Res Soc 48(8):826–833

Embedding (K 9 − C 9 )n into Certain Necklace Graphs Syeda Afiya and M. Rajesh

Abstract In parallel and distributed computing, there are practically two networks: linear networks (also called paths) and rings (also called loops). Many efficient algorithms, such as signal and image processing, were first discovered by solving algebraic problems, graphical problems and parallel implementations involving linear networks and rings. As a result, a network having both good path and cycle embedding is crucial. In this paper, using embedding method, we simulate the cartesian product of (K 9 − C9 )n graph into certain necklace graphs. Keywords Embedding · Wirelength · (K 9 − C9 )n · Necklace graphs

1 Introduction A well-known technique for developing parallel algorithms and modelling various interconnection networks is graph embedding. In the parallel architecture literature, the concept of bandwidth is regarded as dilatation. Dilation, then, refers to the earliest practical delay. Another technique in embedding is Congestion, which restricts the number of edges that can be immersed on a single edge. The summation of each individual dilatation or congestion is represented by the wirelength of G onto H [13]. Many requirements are mutually incompatible when it comes to designing the topologies of an interconnection network. To build a network that is flawless in every way is almost impossible. Based on the needs and properties, a suitable network must be created. Finding the best way to incorporate different networks into a network is one of the most crucial steps in developing and evaluating it. This complexity can be modelled by graph embedding challenge: Finding a mapping system from V (G’) to V (H’) that can transform each edge of a guest graph G’ into a path in S. Afiya School of Advanced Sciences, Vellore Institute of Technology, Chennai, Tamilnadu, India M. Rajesh (B) School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, Tamilnadu, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_23

359

360

S. Afiya and M. Rajesh

a host graph H’. Two typical indices of embedding efficacy are dilatation, and this evaluates the new architecture’s slowness, and congestion, which measures processor usage. Several interconnection topology of path and cycle embedding features have been studied in the literature [5]. Sergei introduced (K 9 − C9 )n graph obtained from K 9 deleting the cycle on 9 edges [2]. Using the δ-sequence introduced by Sergei, (K 9 − C9 )n will be used in future construction of supercomputers. We construct the graph (K 9 − C9 )n using optimal labelling defined by Ashlswede and Cai [1]. In this paper, we have embedded (K 9 − C9 )n in various necklace graphs, namely star necklace, path necklace, cycle necklace and complete necklace.

2 Preliminaries Definition 1 ([3]) A combinatorial analogue of a classical problem is ) ( isoperimetric called the Edge Isoperimetric Problem (EIP): Let G , = V , , E , and S ⊆ V , , ⊝(S) = {{v, w} ∈ E : v ∈ S&w/ ∈ S ( ) |θ | G , ; l = min is called the edge boundary of S. Then, the EIP is to calculate | | } { , |θ (S)| : S ⊆ VG &|S| = l for every integerl, 0 ≤ l ≤ |V , |, and identify sets that achieve the minimum. Definition 2 ([4]) An embedding (of guest )network(G , (VG , , )E G ) into host network , , , , , , , , , , H (V H ,E H ) is a mapping f , : G , VG , E G → H , VH , E H ν f V : VG → E G and , , , , f E : E G → PH where PH is the set of paths in the host network. Lemma 1 ([13]) For an embedding f’o f G’ into H’, the wirelength of f” is. Σ

( ) W L ,f G , , H , =

C f (ei )

ei ∈E(H , )

C f is the congestion sum. Remark 1 For any set S , of H , edges, we have, ( ) Σ ( , ) C f S, = C f Sm e∈S ,

Lemma 2 ([11, 11]) Let G’ represent a r’ regular graph and f’ represent the embed, , ding G’ into H’ Let S’ be the edge cut of H’ that separates H’ into H1 and 2 when ( ,) ( H , , ,) , , , −1 −1 H1 and G 2 = f H2 be the the edges o. S are removed, and let G 1 = f two components. S , also fulfils the following requirements. ( ( ) ( ( ) ( )) ) 1. P ,f f , i , , f , j , has zero edges in S , ∀ i , , j , ∈ E G ,k , k = a, b.

Embedding (K 9 − C9 )n into Certain Necklace Graphs

361

( , , ( )) ) ( ) ( ( ) 2. P ,f f (i ) , f , j , has a single edge in S , ∀ i , , j , ∈ E G , νi , ∈ V G a, and j , ∈ ( ) V G ,b . , 3. G k must be a maximum subgraph, k = a, b. | | ( ) ( ) , | , | Then, C f S , is minimum and C f S , = r |V G k | − 2|V G k |, k = a, b. , Lemma 3 ([11, H , . Set}of edges is repre) Define an embedding as f {: G, → ( 11]) , , sented as k E H , in H , repeated k , times. Let S1 , S2 , . . . , Sm be a partition of [ , ( , )] , K , E H , Sm being an edge cut of H , . Then m ) 1 Σ , ( , C f Sm W L f G,, H , = , K a=1

Lemma 2 and 3 have made it possible to calculate the precise wirelength in different range of topologies [6–10, 15]. Definition 3 (Lexicographic ordering) Lexicographic ordering is defined on a set of n-tuples with integer entries as (x1, , x2, , ..., xn, ), is greater than (y1, , y2, , ..., yn, ) if , there exist an index i , ∍, x, j, = y, j, for 1 ≤ j , < i , andx ,j , > y ,j , i , ≤ n. Theorem 1 ([1, 11]) If lexicographic order is optimal forG × G, then it is optimal for G n for any. Theorem 2 ([14]) Let G be (K 9 − C9 )n[, n ≥ 1.] For m, 0 ≤ m ≤ 9n , let k = [dm/9n − 1], r = mmod9n−1 and s = m/9n−1 Then, the recursive formula for IG (m) is given [ by, ] IG (m) = 9n−1 × 3(n − 1) k +9n−1 × IG (k) + [δG (s) × r ] +IG (r ). Definition 4 (K 9 − C9 )n graph is defined as the cartesian product of (K 9 − C9 ) with (K 9 − C9 )n−1 , n ≥ 2, where (K 9 − C9 ) is a graph obtained by deleting a cycle on 9 vertices fromK 9 . The graph obtained is not optimal. By applying local global principle [1], we obtain the lexicographic labelling and arranged the graph as shown in Fig. 1.

3 Main Results 3.1 Wirelength of (K 9 − C9 )n in Star Necklace Graph Definition 5 ([11]) Let K 1,m be a star with m + 1 vertices v1 , v2 , ..., vm+1 . Let K ni is a complete graph for 1 ≤ i ≤ m − 1 on vertices n and K nm be a complete graph on n m K ni ) has only vi as a cut vertex. The resulting architecture − 1 vertices ∍ K l,m ⊎ (Ui=1 m i is a K l,m ⊎ (Ui=1 K n ) star necklace with the notation S N (K l,m , K ni ).

362

S. Afiya and M. Rajesh

Fig. 1 a K 9 , b (K 9 − C9 ), c Lexicographic ordering of (K 9 − C9 )

Embedding(Algorithm )A Input: S N K (1,2) ; K 29n and (K 9 − C9 )n . 2 Algorithm: 1. (K 9(− C9 )n vertices ) are labelled in lexicographic ordering. 2 2. S N K (1,2) : K 9n vertices are labelled as 2

1, 2, . . . ,

9n 9n 9n 9n 9n 9n − 2, − 1, , + 1, + 2, . . . 2 2 2 2 2 2

.

) ( 3. Define f , : (K 9 − C9 )n → S N K (1,2) ; K 29n by f (y) = y. 2 ) ( n 2 Output: (K 9 − C9 ) into S N K (1,2) ; K 9n induces minimum wirelength. 2 ) ( Proof of correctness: Let (K 9 − C9 )n and star necklace, S N K (1,2) ; K 29n be G 2

n

and H , respectively. Consider S l , Ri j and Tuv ; 1 ≤ l, i, j, v ≤ 2 and 1 ≤ u ≤ 92 − 1 ) ( are edge cuts of S N K (1,2) ; K 29n . See Fig. 2. 2 ) ( For 1 ≤ l ≤ 2, each edge cut S l partitioned the edge set of S N K (1,2) : K 29n , such ( ) ( ) 2 that H splits into Hl1 and Hl2 where G l1 = f −1 Hl1 and G l2 = f −1 Hl2 are optimal sets by Theorem 2 and each Sl satisfies the Lemma 2. Thus, C f((Sl ) is minimum. ) For 1 ≤ i, j ≤ 2, each edge cut R ij partitioned the edge set of S N K (1,2) : K 29n , ∍ , H 2 ( ) ( ) j j j j splits into two components Hi1 andHi2 where G i j = f −1 Hi1 and G i j = f −1 Hi2 1

j

2

j

are optimal sets by Theorem 2 and each Ri satisfies the Lemma 2. Thus, C f (Ri ) is minimumn n For 1 ≤ v ≤ 2 and 1 ≤ u ≤ 92 − 1, each edge cut Tuv partitioned the edge set ) ( v v of S N K (1,2) ; K 29n , ν, H splits into two components Hu1 and Hu2 where G u v1 = 2

Embedding (K 9 − C9 )n into Certain Necklace Graphs

363

) ( Fig. 2 Labelling in S N K (1,2) ; K 29n 2

( ) ( ) f −1 Huv1 and G u v2 = f −1 Huv2 are optimal sets by Theorem 2 and each Tuv satisfies the Lemma 2. Thus, C f (Tuv ) is minimum. Then by Lemma 3, optimal embedding is obtained. Theorem 3 The) minimum embedding wirelength (K 9 − C9 )n into star necklace, ( S N K (1,2) ; K 29n is given by, 2

=

2 ( Σ l=1

( n ) 2 Σ 2 ( ( ( ) ( )) Σ ) ( ) ) 9 6n 9n−1 l − 2I 9n−1 l + 6n 9n−1 i j − 2I 9n−1 i j + 12n −1 2 i=1 i=1

Proof By Lemmas 2 and 3 ⎞ ⎛ 9n )) ( ( 2 2 Σ 2 Σ 2 2 −1 Σ Σ Σ 1⎜ ⎟ j = ⎝ W L (K 9 − C9 )n , S N K (1,2) ; K 29n C f (Sl ) + C f Ti + C f Ruv ⎠ 2 2 l=1

=

2 ( Σ

2 Σ 2 ( Σ

( ) ( )) 6n 9n−1 l − 2I 9n−1 l +

l=1

i=1 i=1

i=1 j=1

v=1 u=1

( n ) ( ) ( ) ) 9 6n 9n−1 i j − 2I 9n−1 i j + 12n −1 2

3.2 Wirelength of (K 9 − C 9 )n in Path Necklace Graph Definition 6 ([12]) Consider a path, P = v1 , v2 , . . . , vm . Let K ni , be a complete graph m K ni ) contains just vi as a cut vertex for 1 ≤ i ≤ m. on n vertices, such that P ⊎ (Ui=1 m i The resulting graph P ⊎ (Ui=1 K n ) is a path necklace with the notation P N (Pm ; K ni ).

364

S. Afiya and M. Rajesh

) ( n Fig. 3 Edge cuts in P N P3n ; K 33n

Embedding Algorithm B

( n) Input: The path necklace, P N P3n ; K 33n and (K 9 − C9 )n . Algorithm: 1. (K 9 − C9 )n vertices are labelled in lexicographic ordering. 2. Label the vertices of P3n as {1, 2, . . . , 3n} beginning from leftmost vertex and n K 33n as 1, 2, . . . , 3n from the leftmost vertex. ( n) 3. Define f , : (K 9 − C9 )n → P N P3n ; K 33n ) by f (y) = y ( n) Output: (K 9 − C9 )n into the path necklace,P N P3n ; K 33n induces minimum wirelength. ( n) Proof of Correctness: Let (K 9 − C9 )n and path necklace, P N P3n ; K 33n be G and H , respectively. Consider Si , T jk and Rlm ; 1 ≤ i, l ≤ 3n, 1 ≤ k, m ≤ 2 and. ( n) 1 ≤ j ≤ 3n − 1 are the edge cuts of P N P3n ; K 33n . See Fig. 3. For 1 ≤ i ≤ 3n , the edge cut Si , for 1 ≤ j ≤ 3n and 1 ≤ k ≤ 2, the edge cut k T j , for 1 ≤ l ≤ 3n and 1 ≤ m ≤ 2 the edge cut Rlm individually partitioned the ( n) edge set of P N P3n ; K 33n , such that H splits into Hi1 andHi2 where G 1 = f −1 (Hi1 ) and G 2 = f −1 (Hi2 ) are optimal set by Theorem 2 and each edge cuts satisfies the Lemma 2. Thus by Lemma 3, optimal embedding is obtained. n Theorem 4 The ( ) minimum embedding wirelength (K 9 − C9 ) into Path necklace, 3n n P N P3 ; K 3n is given by, 2 Σ 2 Σ |( ) |) ( ( ( n) ) 1 6n 3 j − 2I | 3n jk | + 12n3n n3n+1 3n − 1 + 3n + 1 + (−1)n+1 + 2 i=1 i=1

Proof Lemmas 2 and 3 show that,

Embedding (K 9 − C9 )n into Certain Necklace Graphs (

W L (K 9 − C9 ) , P N n

(

n P3n ; K 33n

))

365

⎞ ⎛ n n −1 2 n −1 3n 3 3Σ 3Σ Σ Σ 1 ⎝Σ j m⎠ = C f (Sl ) + C f Ri + C f Tl 2 l=1

j=1 k=1

l=1 m=1

2 Σ 2 Σ |( ) |) ( ( ( n) ) 1 6n 3 j − 2I | 3n jk | + 12n3n = n3n+1 3n − 1 + 3n + 1 + (−1)n+1 + 2 i=1 i=1

3.3 Wirelength of (K 9 − C 9 )n in Cycle Necklace Graph Definition 7 ([12]) Consider Cm = v1 ,(v2 , . . . , )vm be a cycle. Let K ni be a complete m i graph with n vertices, ) that Cm ⊎ Ui=1 K n has just vi as a cut vertex ( m such ( for 1) ≤ i i ≤ m. The Cm ⊎ Ui=1 K n graph is a cycle necklace indicated by C N Cm ; K ni . Embedding Algorithm C

) ( Input: The cycle necklace C N C3 ; K 332n−1 and (K 9 − C9 )n . Algorithm: 1. (K 9 − C9 )n vertices are labelled in lexicographic ( ordering.) 2. Label the vertices of K 332n−1 in cycle necklace C N C3 ; K 332n−1 as (i − 1)32n−1 +1, (i − 1)32n−1 + 2, . . . , (i − 1)32n−1 + 32n−1 , such that (i − 1)32n−1 + 1 is the label cut vertex. ( ) 3. Define f , : (K 9 − C9 )n → C N C3 ; K 332n−1 by f (y) = y. ) ( Output: (K 9 − C9 )n into the cycle necklace,C N C3 ; K 332n−1 induces minimum wirelength. Proof of Correctness:

( ) Let (K 9 − C9 )n and cycle necklace, C N C3 ; K 332n−1 be G and H , respectively. ( ) Consider Si , Tlm and R j ; 1 ≤ i, j, m ≤ 3 and 1 ≤ l ≤ 13 32n − 3 are the edge ) ( cuts of C N C3 ; K 332n−1 . See Fig. 4. ( ) For 1 ≤ i ≤ 3, the edge cut Si , for 1 ≤ m ≤ 3 and 1 ≤ l ≤ 13 32n − 3 , the the edge(set of edge cut Tlm , for 1 ≤ j ≤ 3, the edge cut R j individually(partitioned ) ) ( ) 3 1 2 1 1 2 −1 H j and G j = f −1 H j2 C N C3 ; K 32n−1 splits into H j and H j where G j = f are optimal sets by Theorem 2 and each Si , Tlm and R j satisfies the Lemma 2. Then by Lemma 3, optimal embedding is obtained. n Theorem 5 The ) minimum embedding wirelength (K 9 − C9 ) into cycle necklace, ( 3 C N C3 ; K 32n−1 is given by,

) ( 3 ) ( ) ( ) 1 Σ ( 2n 1 2n (2n 3 − 3 i − 2i|I ( (3 − 3))| + 12 32n − 3 + 4 × 32n−1 2 i=1 3

366

S. Afiya and M. Rajesh

) ( Fig. 4 Edge cuts in C N C3 ; K 332n−1

Proof Lemmas 2 and 3 show that, ⎛ 3 )) 1 Σ ( ⎜ C f (Si ) + W L (K 9 − C9 )n , C N C3 ; K 332n−1 = ⎝ 2 (

i=1

⎞

1 2n 3 (3 −3)

Σ

3 Σ

l=1

m=1

C f Tlm +

3 Σ

⎟ C f Rj⎠

j=1

( 3 ) ) ) ( ) ( 1 Σ ( 2n 1 2n (2n 3 − 3 i − 2i|I ( (3 − 3))| + 12 32n − 3 + 4 × 32n−1 2 3 i=1

3.4 Wirelength of (K 9 − C 9 )n in Circular Necklace Graph Definition 8 ([12]) Let us say K m is a set of complete graphs ( m withi )m vertices. Let K ni be a complete graph with n vertices, ∍ (K m ⊎ U)i=1 K n has just vi as a m i resulting graph,K ν U K , is a circular necklace, cut vertex for 1 ≤ i (≤ m. The m i=1 n ) represented by K N K m ; K ni .

Embedding (K 9 − C9 )n into Certain Necklace Graphs

367

Embedding Algorithm D

( n) Input: The circular necklace, K N K 3n ; K 33n , and (K 9 − C9 )n . Algorithm: 1. (K 9 − C9 )n vertices are labelled in lexicographic ordering. ( n) 2. Label the vertices of K 3 n as 1, 2, .., 3n in circular necklace, K N K 3n ; K 33n , such every cut vertex gives a maximum ( subgraph. n) 3. Define f , : (K 9 − C9 )n → K N K 3n ; K 33n by f (y) = y. ( n) Output: (K 9 − C9 )n into the circular necklace, K N K 3n ; K 33n , induces minimum wirelength. Proof of Correctness:

( n) Let (K 9 − C9 )n and circular necklace, K N K 3n ; K 33n be G and H , respectively. ( n) j Consider Sm , Ti and R j ; 1 ≤ m, i, j, l ≤ 3n are the edge cuts of K N K 3n ; K 33n . See Fig. 5. j n For 1 ≤ m ≤ 3, the edge cut Sm , for 1 ≤ i, j ≤ 3n , the edge(cut Ti . For ) 1≤l ≤3 , 3n n the edge cut R j individually partitioned the edge set of K N K 3 ; K 3n , such that H splits into Hk 1 and Hk 2 where G k 1 = f −1 (G k 1 ) and G k 2 = f −1 (G k 2 ) are optimal j sets by Theorem 2 and each Sm , Ti and R j edge cuts satisfies the Lemma 2. Then by Lemma 3, optimal embedding is obtained. Theorem 6 The minimum embedding wirelength (K 9 − C9 )n into circular necklace, ( n) 3 K N K 3n ; K 3n is given by,

( ) n Fig. 5 Edge cuts in K N K 3n ; K 33n

368

S. Afiya and M. Rajesh

| ( | ( ( ( ) )|)) ) )|) ) 1 (( ( ( n 3 6n 3 − 1 − 2| I 3n − 1 | + (32n (6n) + 3n 6n 3n − 1 − 2| I 3n − 1 | ) 2 Proof Lemmas 2 and 3 show that, ⎛ n ⎞ 3 3n Σ 3n 3n )) 1 Σ ( ( Σ Σ n j = ⎝ C f (Sm ) + C f (Ti ) + C f (Rl )⎠ W L (K 9 − C9 )n , K N K 3n ; K 33n 2 m=1

i=1 k= j

l=1

) | ( | ( ) )|)) ) )|) ( ( 1 (( ( ( n = 3 6n 3 − 1 − 2| I 3n − 1 | + (32n (6n) + 3n 6n 3n − 1 − 2| I 3n − 1 | ) 2

4 Conclusion In this study, we calculated the exact wirelength of (K 9 − C9 )n in star necklace, path necklace, cycle necklace and circular necklace. These results are proven to be optimal and can be used in network embedding to reduce the network’s dimensional space.

References 1. Ahlswede R, Cai N (1997) General edge-isoperimetric inequalities, part ii: a local–global principle for lexicographical solutions. Eur J Comb 18(5):479–489 2. Bezrukov S, Bulatovic P, Kuzmanovski N (2018) New infinite family of regular edge isoperimetric graphs. Theoret Comput Sci 721:42–53 3. Bezrukov SL (1999) Edge isoperimetric problems on graphs. Graph Theory Combin Biol 7:157–197 4. Bezrukov SL, Chavez JD, Harper LH, Röttger M, Schroeder UP (1998) Embedding of hypercubes into grids. In: International symposium on mathematical foundations of computer science. Springer, pp 693–701 5. Bondy JA, Murty USR et al (1976) Graph theory with applications, vol 290. Macmillan, London 6. Dong Q, Yang X, Zhao J (2008) Embedding a family of disjoint multi-dimensional meshes into a crossed cube. Inf Process Lett 108(6):394–397 7. Fan J, Lin X, Jia X (2005) Optimal path embedding in crossed cubes. IEEE Trans Parallel Distrib Syst 16(12):1190–1200 8. Han YJ, Fan JX, You LT, Wang Y (2012) Embedding complete binary trees into locally twisted cubes. In: Advanced engineering forum, vol 6. Trans Tech Publication, pp 70–75 9. Heun V, Mayr EW (1996) Optimal dynamic edge-disjoint embeddings of complete binary trees into hypercubes. In: Proceedings of the 4th workshop on parallel systems and algorithms. World Scientific, pp 195–209 10. Kulasinghe P, Bettayeb S (1995) Embedding binary trees into crossed cubes. IEEE Trans Comput 44(7):923–929 11. Liu JB, Arockiaraj M, Delaila JN (2019) Wirelength of enhanced hypercube into windmill and necklace graphs. Mathematics 7(5):383 12. Manuel P, Rajasingh I, Rajan B, Mercy H (2009) Exact wirelength of hypercubes on a grid. Discret Appl Math 157(7):1486–1495

Embedding (K 9 − C9 )n into Certain Necklace Graphs

369

13. Rajan RS, Parthiban N, Rajalaxmi T (2015) Embedding of recursive circulants into certain necklace graphs. Math Comput Sci 9(2):253–263 14. Rajasingh I, Arockiaraj M, Rajan B, Manuel P (2011) Circular wirelength of generalized petersen graphs. J Interconnection Netw 12(04):319–335 15. Syeda A, Rajesh M (2023) MinLA of (K 9 −C 9 )n and its optimal layout into certain trees. J Supercomputing Mar 6:1–13 16. Wu AY (1985) Embedding of tree networks into hypercubes. J Parallel Distrib Comput 2(3):238–249

Lyrics Generation Using LSTM and RNN Aarthi Dhandapani, N. Ilakiyaselvan, Satyaki Mandal, Sandipta Bhadra, and V. Viswanathan

Abstract For years, both the music and AI industries have been interested in automatic lyric creation. Early rule-based techniques have mostly been supplanted by deep-learning-based systems as computing power and data-driven models have evolved. This paper explores the capability of deep learning models to generate lyrics for a designated musical genre and artist. Previous lyric generation research in the field of computational linguistics has been restricted to Recurrent Neural Networks (RNN) or Gated Recurrent Units (GRU). On the contrary, this paper focuses on the use of LSTM networks to generate lyrics for a certain genre/artist from a sample lyric (seed) as input. The study also looks at how modern deep learning techniques may be used to improve the songwriting process by learning from artists and their significant works. The LSTM model proposed by this paper was able to generate both raps as well as pop lyrics, capturing average line length, in-song, and across-genre word variation very closely to the text it was trained upon. By adjusting the model’s parameters, the neural network was pre-trained on the lyrics of different composers and musicians. A multilayer LSTM-based training model with bidirectional neurons and BERT integration is used in this study. To generate a comprehensive set of lyrics, the lyrics supplied as input are divided down into a word and rhyme index. This model is generative; therefore, each trial yielded a unique set of lyrics. The model produces a loss of less than 0.06 when the parameters are set correctly. The differences in the results of permutations of different dropout positions were analyzed. Some of the A. Dhandapani · N. Ilakiyaselvan (B) · S. Mandal · S. Bhadra · V. Viswanathan Vellore Institute of Technology, Chennai, Tamilnadu 600127, India e-mail: [email protected] A. Dhandapani e-mail: [email protected] S. Mandal e-mail: [email protected] S. Bhadra e-mail: [email protected] V. Viswanathan e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 N. Venkataraman et al. (eds.), Big Data and Cloud Computing, Lecture Notes in Electrical Engineering 1021, https://doi.org/10.1007/978-981-99-1051-9_24

371

372

A. Dhandapani et al.

model’s lyrics were determined to be of suitable quality. To summarize, the findings suggest that deep learning techniques can be utilized to support the creative process of songwriting. Keywords Lyric generation · Long Short-Term Memory (LSTM) · Recurrent Neural Network (RNN) · Deep Neural Network (DNN) · Dropout

1 Introduction One of the oldest, if not the oldest art that ever existed—Music, has been there since the very dawn of human civilization, starting with the caveman’s music which consisted mainly of grunts and the sound of banged objects; to the Renaissance, Baroque, classical, and finally to the pop culture in the twenty-first century. Music has existed in almost every known culture around the world [1]. It has grown with us over time by becoming a universal language that everyone understands irrespective of one’s race, religion, nationality, age, gender, language, or skin color. Lyrics writing is a crucial part of the process of songwriting, and good lyrics contribute to expressiveness and influence the emotional valence of the music [2]. Writing lyrics from scratch, however, does not come easily to everyone. The task is comparable in its complexity to poetry writing with similar demands on expressiveness, conciseness, and form. Additional constraints due to melodic properties require basic music understanding which complicates the task even more. Thus, automatic lyrics generation is a useful and important task that aims at inspiring lyricists and musicians toward songwriting. The study focuses on creating a lyrics generator using LSTM at both the character and word levels. A generative model assisting the songwriting process included a recurrent neural network (RNN) and transfer learning. Artists spend their entire careers perfecting their ability to write emotive, relevant, and innovative songs. This is a job that artificial intelligence might help with. Instead of having struggling artists sift through hours of viable material, what if one could automate the process by having a neural network create line after line of decent content that the artist may then choose from as inspiration? The issue was significant since it is difficult to write appropriate lyrics that follow a strong rhyme scheme, but artificial intelligence may not only speed up but also make the process easier for artists. The outcomes were promising and at times were indistinguishable from popular songwriting approaches [29].

2 Literature Review A. Natural Language Processing and Lyrics Generation. The lyrics were analyzed by Mahedero, Martinez, and Cano using simple natural language processing technologies. Experiments were carried out on the text to recognize the language,

Lyrics Generation Using LSTM and RNN

B.

C.

D.

E.

F.

G.

373

classify it according to different topics, extract structures, and hunt for commonalities. The 500 lyrics were chosen from a variety of websites. The total number of curates obtained was 92. The Naive Bayes classifier was employed. To determine similarity, the Inverse Document Frequency (IDF) and Cosine distance were utilized. Language identification proved to be a simpler assignment than the others [3]. Titular and LyriCloud. Titular and LyriCloud are two software programs created to produce intelligent and interactive lyrics composition tools for artists. Titular is a semi-automatic song title-generating method, whereas LyriCloud recommends comparable word clouds depending on the supplied seed. The authors adopted a model-based approach. Profanity and disparaging terms were filtered away, and the words that occurred most frequently in the database were the ones most likely to feature in the created song’s final title. They obtain good results, but occasionally the title has no semantic value and is confusing to the reader [4]. Automated Poem Generation WASP (Automated Spanish Poet) is the first poem production tool that blends natural language-generating techniques with artificial intelligence. The system used human input as a seed. The system is based on the forward inference rule. The outcomes were regarded as poor and unsuccessful [5]. Semantic similarity in lyrics. Logan, Kositsky, and Moreno investigated the use of lyrics to automatically identify and categorize music, as well as to detect artist similarities. The song lyrics were gathered from several websites. To analyze the content and semantics of the lyrics, techniques such as PLSA (Probabilistic Latent Semantic Analysis) and the k-means clustering approach were applied. The system’s similarity is determined by mixing it with another sound system. There were pros and downsides to both approaches utilized. As a result, a combination of the two strategies would almost certainly be far superior [6]. Classification of lyrics based on rhyme and style. Rudolf Mayer and colleagues are versed in rhyming and stylistic aspects for grading and processing lyrics. To analyze the lyrics, they employed a word group, lyrics tags, and other statistical aspects. A rhyme is two words that sound the same when spelled together. Words near the conclusion of a stanza are frequently utilized for this characteristic. The proposed method’s effectiveness needs to be assessed further [7]. Rap Lyrics Generation. Nguyen and Sa created a rap lyric generator. This application provides a database of over 40,000 existing rap lyrics. The words and verses from the current lyrics are then combined to form a new lyric. The lyrics were created using a linear interpolation technique. However, the outcomes were judged to be insufficiently fluent. As a result, they adopt the four-gram form. They also offered a database of rhyming terms from two separate phrases. They were able to compose phrases that rhymed with each other in this manner. Finally, all of the phrases were recreated to fit the song’s structure and arrangement. The generator works all right; however, the lyrics do not make sense and are not related to a certain issue [8]. Automated Sega lyrics generation. Bhaukauraly, Didorally, and Pudaruth collaborated to create a tool allowing lyricists to produce Sega lyrics in Mauritian

374

A. Dhandapani et al.

Creole. Sixty-six respondents were asked to judge ten songs as either man-made or machine-generated. Five of these ten tracks were written by myself. The created lyrics appear to be of high quality since around half of the respondents were unsure if the lyrics previously existed or were generated by a computer program. The fundamental shortcoming of this study is a lack of information on the implementation of how sentences are generated or produced [9].

3 Background 3.1 Recurrent Neural Network RNNs are a form of neural network that allows for the use of previous outputs as inputs. They employ the abstract idea of Sequential Memory, which states that a sequence of items provides greater information and efficiency in obtaining a model’s output (think the Alphabet). Sequences have an intrinsic property of structure and information. RNNs can recognize patterns and sequences and utilize them to create predictions. They do this by establishing a feedback loop in the network that uses prior data to guide the next iteration of inputs. The hidden state is a vector representation of prior inputs in this feedback loop. This permits information to remain across the model, which is impossible with standard feed-forward networks [10] (Fig. 1). RNNs have several advantages, including the ability to analyze any length of input sequence since the model’s size does not scale with the amount of the input, and the ability to take into account previous historical data while operating. RNNs identify patterns in data sequences such as text, audio, or numerical time series data because of their benefits. This study employs a method that has a narrower domain

Fig. 1 Conventional Feed-Forward Neural Network [11]: the first and most basic sort of artificial neural network invented. In this network, information flows only in one direction—forward—from the input nodes to the output nodes, via the hidden nodes (if any). The network has no cycles or loops

Lyrics Generation Using LSTM and RNN

375

Fig. 2 Recurrent Neural Network [13]: CNNs are faster than RNNs because they are developed for image processing, whereas RNNs are meant for text processing. While RNNs can be trained to handle pictures, separating contrasting elements that are closer together remains problematic

using RNNs; however, the Vanishing Gradient Problem might impair text (lyric) production [12] (Fig. 2).

3.2 Long Short-Term Memory (LSTM) As previously stated, RNNs have difficulties learning long-term dependencies across time steps. RNNs have a short-term memory and cannot access information from the past. The Vanishing Gradient Problem occurs when the gradient of the loss function (values used to update weights) diminishes rapidly during back-propagation and eventually disappears. A gradient that gets too thin (and finally zero) does not help you learn much. Because of the microscopic adjustments of the weights by exceedingly small gradients, the neural network’s earlier layers are unable to learn [14] (Figs. 3 and 4). One can utilize LSTMs to fix this issue. An LSTM (Long Short-Term Memory) is a form of RNN that can learn long-term dependencies. LSTMs are capable of preserving errors that can be transmitted back through the layers. They improve RNNs by allowing them to learn across multiple time steps by keeping the error value constant. LSTMs can do this by employing gates, which are tensor operations that can figure out what information to add or remove from the hidden state. The LSTM network has an input gate, a ‘forget’ gate, and an output gate in each unit. The input gate can evaluate the worth of provided data. The ‘forget gate’ can select whether information should be discarded or saved for later use. The output gate is responsible for determining whether or not the information is relevant at a given

376

A. Dhandapani et al.

Fig. 3 (Left). The various Sigmoid functions available to us. (Right). Vanishing Gradient Problem [15, 16]—The Vanishing Gradient Problem arises from the fact that, as the back-propagation process continues, the gradient of the early layers (the layers closest to the input layer are obtained by multiplying the gradients of the later layers) disappears (the layers that are near the output layer). As a result, if the gradients of subsequent layers are smaller than one, their multiplication disappears at an especially quick rate

Fig. 4 LSTM Network Unit Diagram [17]: Time series forecasting models may estimate future values using LSTM based on prior, sequential data. This improves the accuracy of demand forecasts, resulting in better business decisions

phase. The gates take inputs as sigmoid function parameters throughout each phase. The sigmoid returns a number between 0 (allow nothing through the gate) and 1 (permit everything through the gate) (let everything through the gate). This idea is used in back-propagation (updating layer values) (Fig. 4).

Lyrics Generation Using LSTM and RNN

377

LSTMs may pick which information to take forward and which information to drop by employing these gates. These gates assist and govern information flow inside the network and allow the error value to remain relatively constant throughout the network, giving them a significant benefit over RNNs. It is so concluded that adopting LSTMs would be the best line of action for the lyrics generation model [18].

3.3 CuDNN LSTMs There are several types of recurrent layers in Keras, the high-level deep learning library, including LSTM (Long short-term memory) and CuDNNLSTM. A CuDNNLSTM, according to Keras documentation, is a: Fast LSTM implementation backed by CuDNN. Can only be run on GPU, with the TensorFlow backend. Most linear algebra operations may be parallelized to enhance speed, and vector operations like matrix multiplication and gradient descent can be applied to huge matrices that are done in parallel with GPU support. The Compute Unified Device Architecture (CUDA) interface allows vector operations to make use of GPU parallelism. CuDNN uses CUDA to construct kernels for massive matrix computations on GPU. CuDNNLSTM is built for CUDA parallel processing and will not run if a GPU is not present. However, LSTM is intended for standard CPUs. Parallelism is responsible for the quicker execution time. Unfortunately, these layers have certain limitations. For example, customized activation function, no recurrent dropout option, and masking, but if you want conventional layers, they are virtually the same. However, some of these traits may be brought into the model using additional layers on other levels. Consequently, CuDNN LSTMs were used in the model instead of ordinary LSTMs. The new total training duration was almost five times faster than conventional LSTM execution with this simple adjustment.

3.4 Model Building Input Lyrics Data: Finding a dataset to feed into the LSTM network was the first step in creating the lyric generator. A collection of song lyrics by several artists from different genres was sourced from the public domain [19]. Lyrics from prominent musicians such as Taylor Swift, Coldplay, Linkin Park, Eminem, The Chain Smokers, and others were included in the dataset. The goal was to create a lyrics generator that closely matched a certain artist’s songwriting style. Due to system’s memory limitations, the difficulty with the dataset was extracting a portion of it that was

378

A. Dhandapani et al.

sufficient but not too large for the system to handle. To train the models using this information, CSV files for each unique artist were built. Pre-processing: The lyrics of a particular artist were joined to form a single file, and standardized, by making all characters’ lowercase. Certain non-alphanumeric characters like quotations, full-stop, etc., were removed. Followed by, making two dictionaries, one that maps all unique characters to integers (starting from zero) and another that maps integers (starting from zero) to all unique characters. Both dictionaries are arranged in ascending order. The ‘characters to integer’ dictionary was used to convert the lyrics into a stream of integers, to be used as a 2D array (made up of several sentences, each of which is represented as an individual array of integers) input to the model. On the other hand, the ‘integer to character’ dictionary was used for the exact opposite process. The model produces a two-dimensional matrix, where each integer subarray represents a newly generated lyric. A sequence of 100 characters, in the form of meaningful sentences, is used as the input lyrics (seed) for the model. The corresponding integers to these characters from the dictionary are added to the sample array. The compiled textual data of the corresponding genre/artist is used as labels. This allows the model to take into consideration even the wordplay and the rhyme sequences the genre/artist might possess. This has been discussed in a later section. The Samples array is then reshaped into (number of characters, 100, 1) format. Where ‘number of characters’ is the number of characters in the compiled data, 100 is the number of characters in the input sequence. The reshaped array is then normalized with respect to the Labels array, followed by one-hot encoding of the output targets. Architecture: The neural network has four 256-node bidirectional LSTM layers, one input layer with the 100-character sequence formed during data pre-processing, a flattened, dense, dropout, and lastly an activation layer with a ‘soft-max’ activation function. A call-back function is used to determine which character is most likely to appear given the preceding character. In this example, the ‘checkpoint call-back’, a call back, is a function that is invoked after each epoch. A checkpoint call back saves the model’s weights each time the model improves. Loss = −

1 out putsi ze



out putsi ze

yi . log yˆ I + (1 − yi ) log(1 − yi )

i=1

(1)

Categorical Cr oss Entr opy f unction e zi σ (z )i =  K j=1

ez j

(2)

So f tmax activation f unction. Dropout is a regularization strategy that excludes input and recurrent connections to LSTM units from activation and weight updates when training a network. This

Lyrics Generation Using LSTM and RNN

379

reduces overfitting while also boosting model performance [20]. The training parameters include 30 epochs, a batch size of 128, and a validation split of 20% (4:1). Once the models were trained and made fit to the data, an input of 100 or more character sequences was given to the model, and a number of new lines (as specified by the user) of lyrics were generated, as output. The model predicts and generates one character at a time, since it is a characterbased system. The new character along with the original sequence forms a new sequence that is then inputted back into the model to generate the next character. The process is continued for the next 100 characters. To facilitate this operation, a tensor was generated by mapping unique characters to their frequency in the data. This was performed using Keras’ Tokenizer function, which allowed us to compress and vectorize characters effectively based on their input frequency. Every character that appears in an artist’s total song lyrics is assigned to two different dictionaries. The data is divided into samples and labels. After that, the data is reshaped, the size is normalized, and one-hot encoding is applied. The generated tensor is then converted into the corresponding character and concatenated to the ‘key’ string. This new ‘key’ string is then inputted back into the model to generate the tensor of the subsequent character. The loss function used is ‘Categorical Cross Entropy’ because several articles before this study had also used the same loss function. The ‘Adam’ optimizer was because it is computationally efficient and does not have too many memory requirements. BERT integration: BERT (Bidirectional Encoder Representations from Transformers) is a new NLP approach developed by Google researchers. BERT works on a high level by employing transformers (an NLP approach that uses Attention) to analyze and learn the relationship between words. It is a bidirectional approach that can comprehend the context or relationship of a word to its surrounding words in a corpus. The BERT model’s major benefit is that it allows for transfer learning, which substantially speeds up the training process while preserving accuracy. Consequently, ‘Bidirectional CuDNNLSTM layers’ were introduced into the neural network [21]. This integration allowed the network to generate more nuanced, human-like lyrics. Also, as mentioned in the previous sections, a shift was done from LSTM to CuDNNLSTM, because the former uses the CPU and the latter uses the GPU, that is why CuDNNLSTM is much faster than LSTM. Results were generated almost 15 times faster. Batch Sizes: Initially, a default batch size of four was used; however, this resulted in an extremely noisy stochastic gradient descent of the loss function. It was found that raising the batch size to 64 or 128 units resulted in a steady gradient descent and greatly reduced training time.

380

A. Dhandapani et al.

Unidirectional versus Bidirectional Neurons: An LSTM’s hidden state preserves information from earlier inputs. The information stored by the unidirectional LSTM is restricted to the past since the only inputs it has seen are from the past. Bidirectional processing sends your inputs in two directions: one from the past to the future and one from the future to the past. This differs from unidirectional processing in that the LSTM running backward maintains information from the future, whereas the two hidden states together preserve information from both the past and the present, it could be easier for the network to understand the next word location. Bidirectional LSTMs offer outstanding results since they are better at comprehending context. For example, trying to predict the next word in a sentence, on a high level what a unidirectional LSTM will see is, [23] The boys went to…

And will try to predict the next word only by this context, with bidirectional LSTM you will be able to see information further down the road for example, Forward LSTM: The boys went to … Backward LSTM: … and then they got out of the pool.

You can see that using the information from the future, it could be easier for the network to understand the next word.

4 Results and Inferences 4.1 Generated Lyrics With Dropout layers: The use of a dropout layer in the model was able to affect the overall overfitting of the testing results only very slightly. However, its implementation brought out an interesting feature in the result. As evident from the attached results, oftentimes it was seen to generate a set of words/phrases several times in succession. Although this effect in no way affected the overall appeal of the generated lyrics, it did add a slight poetic edge to the lyrics. Such use of phrases in literature is a poetic device called ‘Repetition’, which is employed to make a phrase/text more noticeable (Table 1). Seed: “ takeaway, yeah, yeah, yeah-yeah. your heart for takeaway. your heart for takeaway, yeah-yeah, yeah-yea”. Generated lyrics: h your heart for takeaway, yeah, yeah, yeah-yeah. before i love you, na, na, na.

Lyrics Generation Using LSTM and RNN

381

Table 1 Table captions should be placed above the tables Designation

Dropout layer position

Training loss

Validation loss

Control

Nil

0.0420

6.0612

Experiment 1

After dense layer

3.5126

4.1771

Experiment 2

After every hidden layer

0.0655

5.7330

Experiment 3

After the hidden layers, before dense layer

0.0427

6.6893

i’m gonna leave you, na, na, na. before i’m someone you leave behind. i’ll break your heart so you don’t break mine. before i love you, na, na, na. i’m gonna leave you, na, na, na. even if i’m not here to stay. i still want your heart. your heart for takeaway, yeah, yeah, yeah-yeah. your heart for takeaway, yeah, yeah, yeah-yea. Done. Observation 1. Since the given seed/key phrase, has a repetitive phrase, the model can pin on nuance and use the same phrase and device to give the lyric a more poetic edge. Seed: “ m okay when thinkin’ ‘bout you. you were right when you said it, keep my cards to my chest and i neve”. Generated lyrics: r let no one know me, i pushed you out. now i get that you needed just a little more love. and to just see me wide open and breaking down. when no one’s there to listen, i will hear your pain. i know thatz i was less than, but you know people change. so tell me, what’s the difference if we still feel the same? it takes a little distance to make you see the way. i go from okay, i’m okay to i’m not okay fr. Done. Observation 2. Here the given seed/key phrase, has no repetitive phrase, yet the model can use the phrase “I’m okay” and use the ‘repetitive’ device to achieve the poetic flair”.

Without dropout layers: As compared to the former implementation with the dropout layers, here the ‘repetitive’ element only appears about 25% of the time, whereas in the former it was around 66.67%. The lyrics generated are otherwise indistinguishable from the former implementation, even with a higher overfitting

382

A. Dhandapani et al.

and validation loss. Furthermore, on performing an opening poll with peers, this set of results proved to be far more acceptable as song lyrics, with an edge of over 80%. Seed: “ ause he’s the reason for the teardrops on my guitar the only thing that keeps me wishing on a wishin” Generated Lyrics: g star he’s the song in the car i keep singing, don’t know why i do… drew walks by me can’t he tell that i can’t breathe? and there he goes, so perfectly the kind of flawless i wish i could be she’d better hood him tight, give him all her love look in those beautiful eyes and know she’s lucky ‘cau Done

Seed: “to me i can’t help but wish you took me with you and this is when the feeling sinks in i don’t wanna” Generated Lyrics: miss you like this come back be here, come back be here i guess you’re in london today i don’t wanna need you this way come back be here, come back be here this is falling in love in the cruelest way this is falling for you and you are worlds away nawhh you’re aooadu and thating inte me she break t Done

Seed: “and you said, "i never regretted the day that i called you mine" can i call you mine? ooh, ah, ooh”. Generated Lyrics: can i call you mine? ooh, ah, ooh. can i call you mine? ooh, ah, ooh. can i call you mine? do you mean, do you mean what you say? what you said, now you can’t take away. you’re my gospel, but i’m losing faith, losing faith. do you mean, do you mean what you say? take a minute, do you need to stop and think? what we have, no, we can’t throw away, throw away. show me that you mean it, ayy. show me that you mean it. do you really mean it? everything happens for a reason. show me that you mean it do you, d Done.

Lyrics Generation Using LSTM and RNN

383

4.2 Inclusion/Exclusion of the Dropout Layers As evident from the training versus testing loss values, the model appears to be overfitting the data. The conventional response is to try to mitigate this by the addition of dropout layers. A model trained with zero dropout layers served as the control for this test. Three possible configurations of the dropout layer(s) were tested (Figs. 5, 6, 7 and 8): • 1st dropout implementation (just after dense layer)—general implementation • 2nd dropout implementation, after every hidden layer. • 3rd dropout implementation, after hidden layers, just before the output dense layer. Through multiple runs, it was found that although the 3rd implementation did successfully reduce the overfitting to a certain extent, the difference between the Fig. 5 (Control) Comparison of validation loss versus testing loss in bidirectional LSTM implementation without Dropout layers. Training loss—0.0420 Validation loss—6.0612

Fig. 6 (Experiment 1) Comparison of validation loss versus testing loss in 1st dropout implementation (just after dense layer). Training loss—3.5126 Validation loss—4.1771

384 Fig. 7 (Experiment 2) Comparison of validation loss versus testing loss in 2nd dropout implementation, after every hidden layer: Training loss—0.0655 Validation loss—5.7330

Fig. 8 (Experiment 3) Comparison of validation loss versus testing loss in 3rd dropout implementation, after hidden layers, just before the output dense layer: Training loss—0.0427 Validation loss—6.6893

Fig. 9 Comparison of loss in bidirectional, 1st, 2nd, and 3rd dropout implementation

A. Dhandapani et al.

Lyrics Generation Using LSTM and RNN

385

testing and training loss was still worthwhile. But the difference arises in the lyrics generated. The model with a dropout layer often got caught in endless loops of certain words and phrases. The final lyrics generated were also not of sufficiently high quality. Whereas the model, without any dropout layer, having a more distinct difference between training and testing loss, was able to generate higher quality lyrics at every iteration. Further Comparisons: The traditional implementation of the dropout layers in a neural network places it just after the dense Layer. As seen in the following graphs, this becomes quite evident even with this system. Over subsequent changes in the number of epochs spent in training the model, the graphs of training and validation loss show fairly constant values for the traditional implementation. (dropout 1). The dropout layer positioned just before the dense layer (dropout-3) produces training and validation loss closely resembling that of the control (bidirectional) curve. This shows that having the dropout layer at this position contributes next to nothing toward the goal of countering the overfitting issue. When the dropout layer is positioned after every hidden layer (dropout-2), the approach worked best particularly for the lower values of epochs, up to around 12. Till this point, it was not only providing stable loss values for both training and validation sets but lower loss values. However, after the critical point of 12 epochs, the values showed a drastic change assuming a sigmoid structure just like the control (bidirectional) curve but heavily skewed toward the right, i.e., higher epoch values, and finally merging with the control and dropout-3 curves at around epoch 25. Thus, it was evident that for an overall controlled state, the traditional approach proved to be the best; however at lower epoch values, positioning similar dropout layers between hidden layers also serves as a viable alternative (Figs. 9 and 10). Fig. 10 Comparison of validation loss in bidirectional, 1st, 2nd, and 3rd dropout implementation

386

A. Dhandapani et al.

5 Discussions The phenomenon of overfitting (inferring bias) draws close parallels to human creativity. As social being, humans have some likes or dislikes and thus always tend to have a bias toward the things they admire. This is reflected in the artist’s songwriting as well, where the degree of bias is shown toward certain emotions. The overfitting, in theory, is somewhat able to simulate that creative bias, resulting in far more ‘human-like’ lyrics being generated. According to the over-fitted brain theory, organisms’ brains face the same difficulty of overfitting to their daily stimulus distribution, resulting in poor generalization and overfitting. The brain may recover the generalizability of its perceptual and cognitive capacities and improve task performance by imagining out-of-distribution sensory stimuli every night [24]. Similarly, it is possible that the overfitting problem faced by the model can be generalized using intermittent and finely tuned ‘noise-injections’, in the form of noisy or corrupted, or faulty data. This is a common method implemented in handing overfitting in the case of DNNs.

6 Conclusion and Future Work The lyric generator proposed by this study is significant and interesting as it enables a dive deeper into NLP and neural networks to see what kind of abstract applications they might have. In this day and age, when everything is becoming automated and enhanced due to machine learning, the results of this study seem to align with the idea that the applications of NLP and data science can extend to any field imaginable. It is fascinating to see how interdisciplinary data science truly is and how it can bridge the gap between STEM and other distinct fields (in this case, music). There is a plethora of ways to enhance and build on this concept in the future. First and foremost, several of the difficulties highlighted in the preceding sections, such as the overfitting problem, may certainly be addressed, by noise injections. Despite the usage of GPU-accelerated neurons, the training time was fairly long, and this has to be addressed in the future too. Fine-tuning parameters is one course of action that may be taken to see whether model performance can be improved, both from the LSTM and BERT sides of the application. This concept might be extended to different literary styles and mediums, such as poetry, and with the correct dataset, the lyrics generator may evolve into a poetry generator. Adding further layers or establishing a concurrent LSTM to add pitch and rhythm to the lyrics is one avenue that can also be considered. As a result of which, one might be able to create a more comprehensive music-generating engine.

Lyrics Generation Using LSTM and RNN

387

References 1. History of music. http://en.wikipedia.org/wiki/History_of_music [Accessed: 18th April 2014] 2. Ali SO, Peynircioglu ZF (2006) Songs and emotions: are lyrics and melodies equal partners? Psychol Music 34(4):511–534. 2006, publisher: SAGE Publications Ltd 3. Mahedero JP, MartÍnez Á, Cano P, Koppenberger M, Gouyon F (2005) Natural language processing of lyrics. 475–478. https://doi.org/10.1145/1101149.1101255 4. Settles B (2010) Computational creativity tools for songwriters 49–57 5. Gervás P (2013) Computational modelling of poetry generation, Corpus ID: 35020911 6. Logan B, Kositsky A, Moreno P (2004) Semantic analysis of song lyrics 2:827–830. https:// doi.org/10.1109/ICME.2004.1394328 7. Mayer R, Neumayer R, Rauber A (2008) Rhyme and style features for musical genre classification by song lyrics. In: ISMIR 2008 - 9th international conference on music information retrieval 337–342 8. Nguyen H, Sa B (2009) Rap lyrics generator. https://nlp.stanford.edu/courses/cs224n/2009/fp/ 5.pdf 9. Pudaruth S, Amourdon S, Anseline J (2014) Automated generation of song lyrics using CFGs 613–616. https://doi.org/10.1109/IC3.2014.6897243 10. Santhoopa J (2020) Sequence models and recurrent neural networks (RNNs), Jul 27, 2020. https://towardsdatascience.com/sequence-models-and-recurrent-neural-networks-rnns62cadeb4f1e1 11. Laskar MTR (2021). York University utilizing the transformer architecture for question answering. https://doi.org/10.13140/RG.2.2.22446.23364 12. SuperDataScience Team (2018) Recurrent neural networks (RNN) - the vanishing gradient problem. https://www.superdatascience.com/blogs/recurrent-neural-networks-rnn-the-vanish ing-gradient-problem 13. Gupta TK, Raza K (2019)Chapter 7 - optimization of ANN architecture: a review on natureinspired techniques. Dey N, Borra S, Ashour AS, Shi F (eds) Machine learning in bio-signal analysis and diagnostic imaging. Academic, pp 159–182, ISBN 9780128160862. https://doi. org/10.1016/B978-0-12-816086-2.00007-2 14. DiPietro R, Hager GD (2020) Chapter 21 - Deep learning: RNNs and LSTM. In: Kevin Zhou S, Rueckert D, Fichtinger G (eds) The Elsevier and MICCAI society book series, handbook of medical image computing and computer assisted intervention. Academic, pp 503–519. ISBN 9780128161760. https://doi.org/10.1016/B978-0-12-816176-0.00026-0 15. Nicholson C (2020) A beginner’s guide to LSTMs and recurrent neural networks. https://wiki. pathmind.com/lstm 16. Singh J (2020) Activation function breakthrough. inblog.in/ACTIVATION-FUNCTIONBREAKTHROUGH-VOyvxhTELU 17. Ramasamy L, Kaliappan J, Prem V, Nonghuloo M (2020) Analyses and modeling of deep learning neural networks for sequence-to-sequence translation. Int J Adv Sci Technol 3152– 3159 18. Phi M (2018) Illustrated guide to LSTM’s and GRU’s: a step-by-step explanation, Sep 24, 2018. https://towardsdatascience.com/illustrated-guide-to-lstms-and-gru-s-a-step-by-step-exp lanation-44e9eb85bf21 19. Shah D (2021) Song lyrics dataset, version 5. Retrieved October 12, 2021, from https://www. kaggle.com/deepshah16/song-lyrics-dataset 20. Brownlee J (2018) A gentle introduction to dropout for regularizing deep neural networks, December 3, 2018. https://machinelearningmastery.com/dropout-for-regularizing-deep-neu ral-networks/ 21. Horev R (2018) BERT explained: state of the art language model for NLP, Nov 10, 2018. https://towardsdatascience.com/bert-explained-state-of-the-art-language-model-fornlp-f8b21a9b6270 22. Aggarwal R (2019) Bi LSTM, Jul 4, 2019. https://medium.com/@raghavaggarwal0089/bilstm-bc3d68da8bd0

388

A. Dhandapani et al.

23. Hoel E (2021) The overfitted brain: dreams evolved to assist generalization. Patterns 2(5):100244, ISSN 2666–3899. https://doi.org/10.1016/j.patter.2021.100244 24. Yu Y, Srivastava A, Canales S (2021) Conditional LSTM-GAN for melody generation from lyrics. ACM Trans Multimedia Comput Commun 1, Article 35 (February 2021):20. https:// doi.org/10.1145/3424116 25. Briot J-P, Pachet F (2017) Music generation by deep learning—Challenges and directions. Accessed from http://arxiv.org/abs/1712.04371 26. Potash P, Romanov A, Rumshisky A (2015) GhostWriter: using an LSTM for automatic rap lyric generation 1919–1924. https://doi.org/10.18653/v1/D15-1221 27. Enrique A (2018) Word-level LSTM text generator. Creating automatic song lyrics with neural networks, June 2018. https://medium.com/coinmonks/word-level-lstm-text-generatorcreating-automatic-song-lyrics-with-neural-networks-b8a1617104fb. Accessed 27 Feb 2019 28. Tong Y, Liu YuLing, Wang J, Xin G (2019) Text steganography on RNN-generated lyrics. Math Biosci Eng 16(5):5451–5463. https://doi.org/10.3934/mbe.2019271 29. Aarthi D, Viswanathan V, Nandhini B, Ilakiyaselvan N (2019) Question classification using a rule based model. Int J Innov Technol Explor Eng 9(1)