Information and Communication Technologies: International Conference, ICT 2010, Kochi, Kerala, India, September 7-9, 2010, Proceedings (Communications in Computer and Information Science, 101) 3642157653, 9783642157653

Citation preview

Communications in Computer and Information Science

101

Vinu V Das R. Vijaykumar Srinivasa K. G. Hatim A. Aboalsamh Mohammad Hammoudeh Vahid Salmani Dinesh Kumar Tyagi Anjali Mohapatra Bharatheesh Jaysimha Eliathamby Ambikairajah Jon Blackledge (Eds.)

Information and Communication Technologies International Conference, ICT 2010 Kochi, Kerala, India, September 7-9, 2010 Proceedings

13

Main Editors Vinu V Das Engineers Network, Trivandrum, Kerala, India E-mail: [email protected] R. Vijaykumar NSS College of Engineering, Palakkadu, India E-mail: [email protected]

Library of Congress Control Number: 2010934002 CR Subject Classification (1998): H.3, H.4, F.1, C.2, H.2.8, J.3 ISSN ISBN-10 ISBN-13

1865-0929 3-642-15765-3 Springer Berlin Heidelberg New York 978-3-642-15765-3 Springer Berlin Heidelberg New York

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. springer.com © Springer-Verlag Berlin Heidelberg 2010 Printed in Germany Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India Printed on acid-free paper 06/3180 543210

Volume Editors

Editor-in-Chief Vinu V Das Engineers Network, India

Editor R Vijaykumar NSS College of Engineering Palakkadu, India

Associate Editors Srinivasa K G Hatim A Aboalsamh Mohammad Hammoudeh Vahid Salmani Dinesh Kumar Tyagi Anjali Mohapatra Bharatheesh Jaysimha Eliathamby Ambikairajah Jon Blackledge

M. S. Ramaiah Institute of Technology, India King Saud University, Saudi Arabia University of Wolverhampton, UK University of California, USA BITS-Pilani, India IIIT-Bhubaneswar, India ABIBA Systems, India University of New South Wales, Australia Dublin Institute of Technology, Ireland

Preface

The International Conference on Advances in Information and Communication Technologies, ICT, is organized every year with the aim of promoting research and developmental activities in computer science, information technology, computational engineering, and communication. Another goal is to promote scientific information interchange between researchers, developers, engineers, students, and practitioners working in India and abroad. The conference provides an ideal platform for people to share views and experiences in information technology, telecommunication, computing techniques and related areas. ICT 2010 was organized by the Association of Computer Electronics and Electrical Engineers (ACEEE), during September 7–9, 2010 in Kochi, Kerala, India. This volume contains 124 papers selected through a rigorous reviewing process. The contributions reflect the richness of research on topics within the scope of the conference and represent several important developments, specifically focused on theoretical foundations and methods for information processing and communications. Organizing a conference like this one is not possible without the assistance and continuous support of many people and institutions. Particularly I would like to express my gratitude to the organizers of sessions on dedicated topics that took place during the conference. I also thank Stefan Goeller, Janahanlal Stephen, R Vijay Kumar, and Nessy Thankachan for their constant support and guidance. I would like to express my gratitude to Springer’s LNCS-CCIS editorial team, especially Leonie Kunz, for producing such a wonderful proceedings book.

July 2010

Vinu V Das

ICT 2010 – Organization

Technical Chair Pascal Lorenz Abdelhamid Mellouk Joel Rodrigues

University of Haute Alsace, France University of Paris Est, France University of Beira Interior, Portugal

Technical Co-chair R Vijayakumar Ford Lumban Gaol William M. Mongan

NSS College of Engineering, India University of Indonesia Drexel University, USA

Organizing Chair Janahan Lal PM Thankachan

Viswajyuothi College of Engineering and Technology, India Mar Ivanious, India

Organizing Co-chair Srinivasa K G Hatim A Aboalsamh

M. S. Ramaiah Institute of Technology, India King Saud University, Saudi Arabia

General Chair Vinu V Das

Tabor Engineers Network P. Ltd., India

Publicity Chair Mohammad Hammoudeh

University of Wolverhampton, UK

Publication Chair Mohamed Jamaludeen

SRM University, India

X

Organization

Advisory Committee Sudarshan TSB Sumeet Dua Nirwan Ansari

BITS Pilani, India Louisiana Tech University, USA New Jersey Institute of Technology, USA

Programm Committee Shelly Sachdeva Pradheep Kumar K Rupa Ashutosh Fadnavis Muhammad Nubli Zhenyu Y Angz Keivan Navi Anil K Jacob B. Kannan Liviu Vladutu Malabika Basu Pritimoy Sanyal J.Arunadevi Vahid Salmani Dinesh Kumar Tyagi Anjali Mohapatra Bharatheesh Jaysimha Eliathamby Ambikairajah Jon Blackledge Chandan Debasish Kundu Jims Marchang R Sanjeev Kunte Rama Shankar Yadav Smriti Agrawal Vandana Bhattacherjee R.D. Sudhaker Samuel Amitabha Sinha

Jaypee Institute of Information and Technology University, India SEEE, India Yeshwantrao Chavan College of Engineering, India University Malaysia, Pahang, Malaysia Florida International University, USA Shahid Beheshti University, Iran MES College of Engineering, India CUSAT, India Dublin City University, Ireland Dublin Institute of Technology, Ireland West Bengal University of Technology, India Thiagarajar School of Management, India University of California, USA BITS-Pilani, India IIIT-Bhubaneswar, India ABIBA Systems, India University of New South Wales, Australia Dublin Institute of Technology, Ireland IIT Kharagpur, India IIT Kharagpur, India NIT Silchar, India J N N College of Engineering, India MNNIT, India MNNIT, India BITS Mesra, India S J College of Engineering, India West Bengal University of Technology, India

Table of Contents

Full Paper Design and Simulation of Security Sub-layer of WMAN IEEE 802.16 Standard (Wi-Max Compliant) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rajat Sheel Jain, Neeraj Kumar, and Brijesh Kumar

1

Speech Recognition of Assamese Numerals Using Combinations of LPC - Features and Heterogenous ANNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Manash Pratim Sarma and Kandarpa Kumar Sarma

8

Design and Modeling of Power Efficient, High Performance 32-bit ALU through Advanced HDL Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . K. Dhanumjaya, G. Kiran Kumar, M.N. Giriprasad, and M. Raja Reddy A New Multi-language Encryption Technique for MANET . . . . . . . . . . . . Prasenjit Choudhury, Rajasekhar Gaddam, Rajesh Babu Parisi, Manohar Babu Dasari, and Satyanarayana Vuppala

13

22

Bi-variate Polynomial Approximation of Fuzzy Controller Using Genetic Algorithm for Trajectory Control of PUMA560 . . . . . . . . . . . . . . . A. Mona Subramaniam, A. Manju, and Madhav J. Nigam

29

Protecting Digital Images Using DTCWT-DCT . . . . . . . . . . . . . . . . . . . . . . K. Ramani, E.V. Prasad, and S. Varadarajan

36

A Novel High Speed Dynamic Comparator Using Positive Feedback with Low Power Dissipation and Low Offset . . . . . . . . . . . . . . . . . . . . . . . . . Silpakesav Velagaleti, Pavankumar Gorpuni, and K.K. Mahapatra

45

VLSI Design of Four Quadrant Analog Voltage-Mode Multiplier and Its Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ankita Tijare and Pravin Dakhole

50

An Optimizing Technique for MDGT Using DRSA Algorithm Association with IP Traceback Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . S. Karthik, V.P. Arunachalam, T. Ravichandran, and M.L. Valarmathi Designing a Promotor for a Novel Target Site Identified in Caspases for Initiating Apoptosis in Cancer Cells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mala S. Kumar, K.L. Lainu, V. Aghila, Dhanya Purushothaman, K. Varun Gopal, P.K. Krishnan Namboori, and Vrinda Harishankar

55

62

XII

Table of Contents

Learning Classifier Systems Approach for Automated Discovery of Hierarchical Censored Production Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . Suraiya Jabin

68

Feature Based Watermarking Algorithm by Adopting Arnold Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . S.S. Sujatha and M. Mohamed Sathik

78

Inference of Gene Networks from Microarray Data through a Phenomic Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rio G.L. D’Souza, K. Chandra Sekaran, and A. Kandasamy

83

A Novel Lattice Based Research Frame Work for Identifying Web User’s Behavior with Web Usage Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . V.V.R. Maheswara Rao and V. Valli Kumari

90

Design and Analysis of Specification Based Ids for Wireless Networks Using Soft Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Vydeki Dharmar and K. Jayanthy

100

A New Image Content-Based Authenticity Verification Procedure for Wireless Image Authentication Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V. Lokanadham Naidu, K. Ramani, D. Ganesh, Sk. Munwar, and P. Basha

108

Enhanced Substitution-Diffusion Based Image Cipher Using Improved Chaotic Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . I. Shatheesh Sam, P. Devaraj, and R.S. Bhuvaneswaran

116

Network Forensic Analysis by Correlation of Attacks with Network Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Atul Kant Kaushik, Emmanuel S. Pilli, and R.C. Joshi

124

Robust and Real Time Data Delivery in Wireless Sensor Networks . . . . . Deepali Virmani and Satbir Jain

129

Multiple QoS Guided Heuristic for Independent Task Scheduling in Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sameer Singh Chauhan and R.C. Joshi

136

A Framework for Network Forensic Analysis . . . . . . . . . . . . . . . . . . . . . . . . . Emmanuel S. Pilli, Ramesh C. Joshi, and Rajdeep Niyogi

142

A New Trust Model Based on Time Series Prediction and Markov Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sarangthem Ibotombi Singh and Smriti Kumar Sinha

148

A Novel J2ME Service for Mining Incremental Patterns in Mobile Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ashutosh K. Dubey and Shishir K. Shandilya

157

Table of Contents

XIII

CDPN: Communicating Dynamic Petri Net for Adaptive Multimedia Presentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.P. Sarath Chandar, S. Arun Balaji, G. Venkatesh, and Susan Elias

165

Developing a Web Recommendation System Based on Closed Sequential Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Utpala Niranjan, R.B.V. Subramanyam, and V. Khanaa

171

Nearest Neighbour Classification for Trajectory Data . . . . . . . . . . . . . . . . . Lokesh K. Sharma, Om Prakash Vyas, Simon Schieder, and Ajaya K. Akasapu

180

Security-Aware Efficient Route Discovery for DSR in MANET . . . . . . . . . Sanjeev Rana and Anil Kapil

186

Regression Modeling Technique on Data Mining for Prediction of CRM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Manisha Rathi

195

Congestion Games in Wireless Channels with Multipacket Reception Capability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Debarshi Kumar Sanyal, Sandip Chakraborty, Matangini Chattopadhyay, and Samiran Chattopadhyay

201

Secure and Revocable Multibiometric Templates Using Fuzzy Vault for Fingerprint and Iris . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V.S. Meenakshi and G. Padmavathi

206

High Speed Cache Design Using Multi-diameter CNFET at 32nm Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Aminul Islam and Mohd. Hasan

215

Dynamic Load Balancer Algorithm for the Computational Grid Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rajkumar Rajavel, Thamarai Selvi Somasundaram, and Kannan Govindarajan

223

Instance-Based Classification of Streaming Data Using Emerging Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mohd. Amir and Durga Toshniwal

228

Modified Go-Left Balls and Bins Algorithm for Server Load Balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Prasun Banerjee, Stephan D’Costa, and Sukriti Bhattacharya

237

Three Layered Adaptation Model for Context Aware E-Learning . . . . . . . Minu M. Das, Manju Bhaskar, and T. Chithralekha

243

XIV

Table of Contents

AFDEP: Agreement Based CH Failure Detection and Election Protocol for a WSN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Amarjeet Kaur and T.P. Sharma

249

Problem Area Identification with Secure Data Aggregation in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Paresh Solanki, Gaurang Raval, and Srikant Pradhan

258

Operational Transconductance Amplifier Based Two-Stage Differential Charge Amplifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dinesh B. Bhoyar and Bharati Y. Masram

267

Graceful Degradation in Performance of WaveScalar Architecture . . . . . . Neha Sharma and Kumar Sambhav Pandey

274

ICT 2010 – Short Paper Dual Tree Complex Wavelet Transform Based Video Object Tracking . . . Manish Khare, Tushar Patnaik, and Ashish Khare

281

Design of New Indexing Techniques Based on Ontology for Information Retrieval Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . K. Saruladha, G. Aghila, and Sathish Kumar Penchala

287

Identifying the Attack Source by IP Traceback . . . . . . . . . . . . . . . . . . . . . . . K.C. Nalavade and B.B. Meshram

292

An Approach towards Secure and Multihop Time Synchronization in Wireless Sensor Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Arun Kumar Tripathi, Ajay Agarwal, and Yashpal Singh

297

Improving Dynamic Difficulty Adjustment to Enhance Player Experience in Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. Joy James Prabhu

303

A Novel Method for Cross-Language Retrieval of Chunks Using Monolingual and Bilingual Corpora . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tayebeh Mosavi Miangah and Amin Nezarat

307

Application of Kohonan SOM in Prediction . . . . . . . . . . . . . . . . . . . . . . . . . Sathya Ramadass and Annamma Abhraham

313

Evaluation of the Role of Low Level and High Level Features in Content Based Medical Image Retrieval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . K.S. Arun and K.S. Sarath

319

A Reinforcement Learning Approach for Price Offer in Supplier Selection Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Vali Derhami, Mohammad Ali Saadatjoo, and Fatemeh Saadatjoo

326

Table of Contents

XV

Generation of k-ary and (k,m)-ary Trees in A-order Using z-Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . N.A. Ashrafi Payaman

330

Wavelet and Hadamard Transforms for Image Retrieval Using Color Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sanjay N. Talbar and Satishkumar L. Varma

336

A Rough Set Integrated Fuzzy C-Means Algorithm for Color Image Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Byomkesh Mandal and Balaram Bhattacharyya

339

Local Monitoring Based Reputation System with Alert to Mitigate the Misbehaving Nodes in Mobile Ad Hoc Networks . . . . . . . . . . . . . . . . . . . . . K. Gopalakrishnan and V. Rhymend Uthariaraj

344

An Energy Efficient Cluster Based Broadcast Protocol for Mobile Ad Hoc Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . G. Kalpana and M. Punithavalli

350

A Hybridized Graph Mining Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sadhana Priyadarshini and Debahuti Mishra

356

Dynamic Contract Generation and Monitoring for B2B Applications with Composite Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kanchana Rajaram and S. Usha Kiruthika

362

Synchronization of Authorization Flow with Work Object Flow in a Document Production Workflow Using XACML and BPEL . . . . . . . . . . . . Subrata Sinha, Smriti Kumar Sinha, and Bipul Syam Purkayastha

365

Virtual Nodes for Self Stabilization in Wireless Sensor Networks . . . . . . . Deepali Virmani and Satbir Jain

371

Small Square Microstrip Antenna . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . L. Lolit Kumar Singh, Bhaskar Gupta, and Partha P. Sarkar

376

Extraction of Optimal Biclusters from Gene Expression Data . . . . . . . . . . J. Bagyamani, K. Thangavel, and R. Rathipriya

380

Analysis of Data Warehouse Quality Metrics Using LR . . . . . . . . . . . . . . . Rolly Gupta and Anjana Gosain

384

Similar - Dissimilar Victor Measure Analysis to Improve Image Knowledge Discovery Capacity of SOM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . N. Chenthalir Indra and E. RamaRaj

389

Cluster-Base Directional Rumor Routing in Wireless Sensor Network . . . Parvin Eftekhari, Hamid Shokrzadeh, and Abolfazl Toroghi Haghighat

394

XVI

Table of Contents

On the Calculation of Coldness in Iowa, a North Central Region, United States: A Summary on XML Based Scheme . . . . . . . . . . . . . . . . . . . Sugam Sharma, Shashi Gadia, and S.B. Goyal Development of a Three Layer Laminate for Better Electromagnetic Compatibility Performance at X-Band . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C. Dharma Raj, G. Sasibhushana Rao, P.V.Y. Jayasree, B. Srinu, and P. Lakshman

400

406

Pre-Confirmation Neural Network for Reducing the Region of Interest in an Image for Face Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. Femina Abdulkader and Ajit Joseph

411

Low Voltage Low Power Op Amp with Gain Boosting and Frequency Compensation Technique for Battery Powered Applications . . . . . . . . . . . K. Sarangam and Hameed Zohaib Samad

417

Performance of Clustering in Mobile Domain . . . . . . . . . . . . . . . . . . . . . . . . Soumen Kanrar and Aroop Mukherjee

423

Morphological Analyzer for Telugu Using Support Vector Machine . . . . . G. Sai Kiranmai, K. Mallika, M. Anand Kumar, V. Dhanalakshmi, and K.P. Soman

430

Visualization of State Transition Systems in a Parallel Environment . . . . Subbu Ramanathan, Haresh Suresh, Amog Rajenderan, and Susan Elias

434

Facilitating Efficient Integrated Semantic Web Search with Visualization and Data Mining Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . S.K. Jayanthi and S. Prema Image Object Classification Using Scale Invariant Feature Transform Descriptor with Support Vector Machine Classifier with Histogram Intersection Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Biplab Banerjee, Tanusree Bhattacharjee, and Nirmalya Chowdhury

437

443

Palmprint Recognition System Using Zernike Moments Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . P. Esther Rani and R. Shanmuga Lakshmi

449

A QOS Framework for Mobile Ad-Hoc in Large Scale Networks . . . . . . . . A. Boomaranimalany and R.M. Chandrasekaran

455

An Interactive Content Based Image Retrieval Method Integrating Intersection Kernel Based Support Vector Machine and Histogram Intersection Based Similarity Measure for Nearest Neighbor Ranking . . . Tanusree Bhattacharjee, Biplab Banerjee, and Nirmalya Chowdhury

458

Table of Contents

XVII

ICT 2010 – Poster Paper Analysis and Prediction of Blocking Probability in a Banyan Based ATM Switch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . R. Pugazendi and K. Duraiswamy

463

Performance Evaluation of QoS Aware Routing in OLSR (Optimized Link State Routing Protocol) Using Genetic Algorithm . . . . . . . . . . . . . . . M. Pushpavalli and A.M. Natarajan

469

Handloom Silk Fabric Defect Detection Using First Order Statistical Features on a NIOS II Processor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . M.E. Paramasivam and R.S. Sabeenian

475

Performance Modeling of MANET Routing Protocols with Multiple Mode Wormhole Attacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yogesh Chaba, Yudhvir Singh, Kanwar Preet Singh, and Prabha Rani

478

Mining a Ubiquitous Time and Attendance Application Schema Using Oracle Data Miner: A Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Binu Jacob and K.V. Promod

484

4T Carry Look Ahead Adder Design Using MIFG . . . . . . . . . . . . . . . . . . . . P.H.ST. Murthy, L. Madan Mohan, V. Sreenivasa Rao, and V. Malleswara Rao

490

Microcontroller Based Monitoring and Control of Greenhouse Enivironment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Gaytri Gupta

495

Hand Written Text to Digital Text Conversion Using Radon Transform and Back Propagation Network (RTBPN) . . . . . . . . . . . . . . . . . . . . . . . . . . . R.S. Sabeenian and M. Vidhya

498

Resource Allocation and Multicast Routing Protocol for Optical WDM Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . N. Kaliammal and G. Gurusamy

501

An Approximate Algorithm for Solving Dynamic Facility Layout Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Surya Prakash Singh

504

Homology Modeling and Protein Ligand Interaction to Identify Potential Inhibitor for E1 Protein of Chikungunya . . . . . . . . . . . . . . . . . . . C.S. Vasavi, Saptharshi, R. Radhika Devi, Lakshmi Anand, Megha. P. Varma, and P.K. Krishnan Namboori Tied Mixture Modeling in Hindi Speech Recognition System . . . . . . . . . . . R.K. Aggarwal and M. Dave

510

514

XVIII

Table of Contents

Propagation Delay Variation due to Process Induced Threshold Voltage Variation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Krishan Gopal Verma, Brajesh Kumar Kaushik, and Raghuvir Singh

520

Biomedical Image Coding Using Dual Tree Discrete Wavelet Transform and Iterative Projection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sanjay N. Talbar and Anil K. Deshmane

525

Privacy-Preserving Na¨ıve Bayes Classification Using Trusted Third Party and Offset Computation over Distributed Databases . . . . . . . . . . . . B.N. Keshavamurthy, Mitesh Sharma, and Durga Toshniwal

529

Extraction of Pose Invariant Facial Features . . . . . . . . . . . . . . . . . . . . . . . . . Singh R. Kavita, Zaveri A. Mukeshl, and Raghuwanshi M. Mukesh

535

On the Segmentation of Multiple Touched Cursive Characters: A Heuristic Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tanzila Saba, Ghazali Sulong, Shafry Rahim, and Amjad Rehman

540

Context Representation and Management in a Pervasive Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B. Vanathi and V. Rhymend Uthariaraj

543

A Comparative Study and Choice of an Appropriate Kernel for Support Vector Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . R. Sangeetha and B. Kalpana

549

Color Image Restoration Method for Gaussian Noise Removal . . . . . . . . . J. Harikiran and R. Usha Rani

554

A Decision Tree Approach for Design Patterns Detection by Subgraph Isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Akshara Pande, Manjari Gupta, and A.K. Tripathi

561

Realisation of Various EBG Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B. Bhuvaneswari and K. Malathi

565

Identification of Melanoma (Skin Cancer) Proteins through Support Vector Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Babita Rathore, Sandeep K. Kushwaha, and Madhvi Shakya

571

BPNN and Lifting Wavelet Based Image Compression . . . . . . . . . . . . . . . . Renu Singh, Swanirbhar Majumder, U. Bhattacharjee, and A. Dinamani Singh

576

Combined Off-Line Signature Verification Using Neural Networks . . . . . . D.R. Shashi Kumar, R. Ravi Kumar, K.B. Raja, R.K. Chhotaray, and Sabyasachi Pattanaik

580

Table of Contents

Distribution of Continuous Queries over Data Aggregators in Dynamic Data Dissemination Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mahesh Gadiraju and V. Valli Kumari Formal Verification of IEEE802.16m PKMv3 Protocol Using CasperFDR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . K.V. Krishnam Raju, V. Valli Kumari, N. Sandeep Varma, and K.V.S.V.N. Raju Digital Image Steganography Based on Combination of DCT and DWT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Vijay Kumar and Dinesh Kumar Mobile Robot Pose Estimation Based on Particle Filters for Multi-dimensional State Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . J. Divya Udayan, T. Gireesh Kumar, Roshy M. John, K.J. Poornaselvan, and S.A. Lakshmanan

XIX

584

590

596

602

Neural Networks Based Detection of Purpose Data in Text . . . . . . . . . . . . P. Kiran Mayee, Rajeev Sangal, and Soma Paul

606

Comparative Performance Analysis of QoS-Aware Routing on DSDV, AODV and DSR Protocols in MANETs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rajneesh Gujral and Anil Kapil

610

Steering Control of an Automated Vehicle Using Touch Screen with Simulation Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Apeksha V. Sakhare, V.M. Thakare, and R.V. Dharaskar

616

A Low Cost GPS Based Vehicle Collision Avoidance System . . . . . . . . . . . Samarth Borker and R.B. Lohani

619

Human Skin Region Identification Using Fusion Technique . . . . . . . . . . . . Vijayanandh Rajamanickam and Balakrishnan Ganesan

622

Avoidance of Electromagnetic Interference in Modem with Minimal Bit Error Rate Using Tele Typewriter Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . C. Annadurai, D. MuthuKumaran, C. Charithartha Reddy, and M. Kanagasabapathy

626

Retracted: A State Variables Analysis for Emerging Nanoelectronic Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . K. Vaitheki and R. Tamijetchelvy Unconditional Steganalysis of JPEG and BMP Images and Its Performance Analysis Using Support Vector Machine . . . . . . . . . . . . . . . . . P.P. Amritha, Anoj Madathil, and T. Gireesh Kumar

632

638

XX

Table of Contents

ASSR Fair Load Distribution Using Efficiency Division Factor with Greedy Booster Approach for MANET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ahmad Anzaar, Husain Shahnawaz, Chand Mukesh, S.C. Gupta, and R. Gowri

641

Secret Sharing Scheme for Image Encryption Using New Transformation Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. Kalai Selvi and M. Mohamed Sathik

647

Measuring the Reusability Level of Software Packages Using Reusability Testing Scripts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . R. Kamalraj, A. Rajiv Kannan, P. Ranjani, and R. Hemarani

651

A New Threshold Calculation Approach in the Performance Enhancement of Spread Spectrum System Using Double Density Discrete Wavelet Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Arunarasi Jayaraman and Indumathy Pushpam

654

Recent Trends in Superscalar Architecture to Exploit More Instruction Level Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ritu Baniwal and Kumar Sambhav Pandey

660

A Neural Network Based Solution to Color Image Restoration Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Satyadhyan Chickerur and M. Aswatha Kumar

666

Spline Biorthogonal Wavelet Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . T. Arathi, K.P. Soman, and Latha Parameshwaran

672

Securing Password File Using Bezier Curves . . . . . . . . . . . . . . . . . . . . . . . . . Sunil Khurana and Sunil Kumar Khatri

675

Reading a 4-Bit SLC NAND Flash Memory in 180nm Technology . . . . . . Nilesh Shah, Rasika Dhavse, and Anand Darji

681

Keystroke Dynamics Authentication Using Neural Network Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Venkateswaran Shanmugapriya and Ganapathi Padmavathi

686

Moving Object Tracking Using Object Segmentation . . . . . . . . . . . . . . . . . Sanjay Singh, Srinivasa Murali Dunga, A.S. Mandal, Chandra Shekhar, and Anil Vohra

691

Mobile Health Care System for Patient Monitoring . . . . . . . . . . . . . . . . . . . Titir Santra

695

Erratum A State Variables Analysis for Emerging Nanoelectronic Devices . . . . . . . K. Vaitheki and R. Tamijetchelvy

E1

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

701

Design and Simulation of Security Sub-layer of WMAN IEEE 802.16 Standard (Wi-Max Compliant) Rajat Sheel Jain, Neeraj Kumar, and Brijesh Kumar Information Technology, Lingaya’s University, Faridabad, India [email protected] [email protected] [email protected]

Abstract. The Data could be provided from the MAC Layer to the Security Sub Layer that is going to be designed & simulated. In this paper the security sublayer will provide subscribers with total privacy across the fixed broadband wireless network. This is done by encrypting connections between Subscriber Station (SS) and Base Station (BS). Also, strong protection from theft of service is provided through the BS protection to data transport services, by enforcing encryption of the associated service flows across the network. The data when entered into the security sub layer first will be certified. After the certification of the data, it could be encrypted through techniques like Data Encryption Standard (DES), Advanced Encryption Standard (AES) that we will call as the Data Encryption. After the encryption, the data is again encrypted through Secure Hash Algorithm (SHA-1) which is used for the message digest of that encrypted data. This operation will take place in the Integrity Unit (part). Keywords: Advanced Encryption Standards, Data Encryption Algorithm, RSA Standards, Wireless Security.

1 Introduction The paper is about to design a security Sub-Layer, defined in Data Link Layer of IEEE 802.16-2004 [1] standard. This standard is defined for Wireless Metropolitan Area Networks (WMAN). It is an alternative technology to provide the broadband technology to the end users. IEEE 802.16 comprises a set of standards, the principal ones are: (a) IEEE 802.16, for the fixed users, working in frequency bands between 10 and 66 GHz. (b) IEEE 802.16e, for mobile users, working in frequency bands between 2 and 6 GHZ. The difference between nomadic and mobile users is: nomadic users do not require to be connected to the network while they are moving. While mobile users usually working with hand-held devices like mobile phones, pocket PC that remain always connected to the network, even when they are moving. Therefore, mobile users require the network to be able to perform hand-over (or hand-off) operation i.e. connections must be ‘kept on’ while users move from an area covered by a certain base station to another area covered by another base station. V. V Das, R. Vijaykumar et al. (Eds.): ICT 2010, CCIS 101, pp. 1–7, 2010. © Springer-Verlag Berlin Heidelberg 2010

2

R.S. Jain, N. Kumar, and B. Kumar

Fig. 1. Comparison between IEEE 802.16 Families of Standards

2 Various Encryption Techniques There are different cryptographic algorithms that are used in our framework to analyse them. In our crypto processor implementation, the dedicated crypto block of the crypto processor permits fast execution of encryption, decryption, and key scheduling operations for RSA and IDEA private Key. The 64-bit plain text data streams entered into the 64-bit input register are encrypted with a proper key and control signals based on the RSA and IDEA Algorithms After plain text data streams are encrypted, the 64-bit cipher texts are output to the 64-bit output register. The decryption process is the same as the encryption process except for the control signals. We use the crypto algorithms like DES, AES, SHA-1 and RSA that are going to be explained below one by one. 2.1 Data Encryption Standards (DES) The Data Encryption Standard (DES) is the most well-known symmetric-key block cipher. It is defined by the American standard FIPS 46–2. The design of DES is related to two general concepts: Product ciphers and Feistel ciphers. Each involves iterating a common sequence or round of operations. The basic idea of a product cipher is to build a complex encryption function by composing several simple operations which offer complementary, but individually insufficient protection. Basic operations include transpositions, translations (e.g., XOR) and linear transformations, arithmetic operations, modular multiplication, and simple substitutions. DES is a Feistel cipher which processes plaintext blocks of n = 64 bits, producing 64-bit cipher text blocks. The effective size of the secret key K is k=56 bits; the input key K is specified as a 64-bit key, 8 bits of which (bits 8; 16; : : : ; 64) may be used as parity bits. A widely held belief is that the parity bits were introduced to reduce the effective key size from 64 to 56 bits, to intentionally reduce the cost of exhaustive key search by a factor of 256.

Design and Simulation of Security Sub-layer of WMAN IEEE 802.16 Standard

3

2.2 RSA Public-Key Encryption The RSA cryptosystem is used to provide both secrecy and digital signatures and its security is based on the intractability of the integer factorization problem. In public key cryptosystems, everyone has two related complementary keys, a publicly revealed key and a secret key. Each key unlocks the code that the other key makes. Knowing the public key does not help you deduce the corresponding secret key. The public key can be published and widely disseminated across a communications network. This protocol provides privacy without the need for the same kind of secure channels that a conventional cryptosystem requires. Anyone can use a recipient's public key to encrypt a message to that person, and that recipient uses her own corresponding secret key to decrypt that message. No one but the recipient can decrypt it, because no one else has access to that secret key. Not even the person who encrypted the message can decrypt it. Message authentication is also provided. The sender's own secret key can be used to encrypt a message, thereby ``signing'' it. This creates a digital signature of a message, which the recipient (or anyone else) can check by using the sender's public key to decrypt it. This proves that the sender was the true originator of the message, and that the message has not been subsequently altered by anyone else, because the sender alone possesses the secret key that made that signature. Forgery of a signed message is infeasible, and the sender cannot later disown his signature. Algorithm: Key Generation for RSA Public-Key Encryption Summary: Each entity creates an RSA public key and a corresponding private key. Each entity A should do the following: 1. Generate two large random (and distinct) primes p and q, each roughly the same size. 2. Compute n = pq and _ = (p − 1)(q − 1). 3. Select a random integer e, 1 < e < phi, such that gcd (phi (n), e) = 1. 4. Use the extended Euclidean algorithm to compute the unique integer d, 1 < d LUBR) Resource is overloaded Use Algorithm 3 else if ( resource Load Cost < LLBR ) Resource is least loaded else Resource is moderately load End

Algorithm 2. Job Migration (JM) Algorithm Begin Get the list of overloaded resource Get the list of least loaded resource for ( all the overloaded resource) While (overloaded resource load cost > LUBR) Take the one job from resource queue Compute the completion time of job CTJK in each resource for ( all least loaded resource ) Choose the best resource BR(JK) having minimum CTJ Then migrate the job to the resource End

In-order to provide better solution, it is necessary to estimate the Completion Time of Job (CTJ) in remote resource before the job is migrated to remote location as follows, CTJK (Ri) = ET (Jk) + DTT (Jk) + CTWJ (Ri)

(9)

Where CTJK (Ri) represents the completion time of job Jk at resource Ri, DTT (Jk) denotes the Data Transfer Time of the job Jk to the resource Ri and CTWJ (Ri) denotes the Completion Time of Waiting Job at resource Ri. In the JM algorithm, it calculates CTJ in the remote resource and chooses the resource having minimum CTJ as, BR (JK)= min ( CTJK (R1), ....., CTJK (Rm) ) Where BR(JK) represent the best resource for migrating job JK.

(10)

Dynamic Load Balancer Algorithm for the Computational Grid Environment

227

3 Experimental Results and Performance Evaluation In the result phase we have simulated the result by exploiting ten resources and hundreds of jobs. The performance of the proposed Load Balancer algorithm using Boundary value approach works better than the Normal Load Balancer with respect to the load of the resource and waiting time of job as shown in the Fig 2. L o a d V s W a i t i n g T i me o f j o b

Wai t i ng T i me

N o r ma l Loa d B a l a nc e r A l go r i t h m

10 5

of j ob 0 0.16

0. 16 Load

0. 243

Loa d B a l a nc e r A l go r i t h m usi ng B o unda r y Va l ue

Fig. 2. Performance Evaluation of Normal Vs Proposed Load Balancer Algorithm

4 Conclusion Our proposed LB algorithm is evaluated with simulation and traces from real time grid environment in CARE laboratory. The proposed LB algorithm using Boundary value approach works better than Normal LB algorithm. The result obtained with performance evaluation can balance the load and increase the utilization of the resource, which are idle or least loaded.

References 1. Foster, I., Kesselman, C.: The Grid: Blueprint for a future Computing Infrastructure. Morgan Kaufmann, San Francisco (1999) 2. Coulouris, G., Dolimore, J., Kindberg, T.: Distributed Systems: Concepts and Design. Addison-Wesley Longman, Amsterdam (1994) 3. Somasundaram, T.S., Amarnath, B.R., Kumar, R., Balakrishnan, P., Rajandar, K., Rajiv, R., Kannan, G., Rajesh Britto, G., Mahendran, E., Madusudhanan, B.: Care Resource Broker: A Framework for scheduling and supporting Virtual Resource Management. Journal: Future Generation Computer System (2010) 4. Xiangchun, H., Duanjun, C., Jing, C.: One centralized Scheduling pattern for Dynamic Load Balance in Grid. In: IEEE International Forum on Information Technology and Applications (2009) 5. Lu, B., Zhang, H.: Grid Load Balancing Scheduling Algorithm based on Statistics Thinking. In: 9th IEEE International Conference for Young Computer Scientists (2008) 6. Gondhi, N.K., Durgesh Pant: An Evolutionary Approach for Scalable Load Balancing in Cluster Computing. In: IEEE International Advance Computing Conference (2009) 7. Shah, R., Veeravalli, B., Mistra, M.: On the Design of Adaptive and Decentralized LoadBalancing Algorithms with Load Estimation for Computational Grid Environments. IEEE Transactions on Parallel and Distributed Systems (December 2007)

Instance-Based Classification of Streaming Data Using Emerging Patterns Mohd. Amir and Durga Toshniwal Department of Electronics and Computer Engineering IIT Roorkee, Roorkee, India [email protected], [email protected]

Abstract. Classification of Streaming Data has been recently recognized as an important research area. It is different from conventional techniques of classification because we prefer to have a single pass over each data item. Moreover, unlike conventional classification, the true labels of the data are not obtained immediately during the training process. This paper proposes ILEP, a novel instance-based technique for classification of streaming data with a modifiable reference set based on the concept of Emerging Patterns. Emerging Patterns (EPs) have been successfully used to catch important data items for addition to the reference set, hence resulting in an increase in classification accuracy as well as restricting the size of the reference set. Keywords: Instance based, classification, emerging patterns, k nearest neighbour, streaming data, single pass classification, streaming data, ILEP.

1 Introduction Data streams are applications in which the data is modelled best not as persistent relations but rather as continuous flow of data sets. This class of data-intensive application has been of wide usage and recognition. In this paper, we address the problem of classification of streaming data. In conventional data sets, we assume the availability of a large labelled training set and the training procedure is error driven because the correct label can be obtained as soon as the classification is done. But in case of streaming data classification problem, we assume that in the beginning, we do not have any training set or we have a very small training set. Moreover, the test data set comes in the form of data streams to the system and the true labels are obtained only after certain time lag. We derive our new classification algorithm from Instance-based (IB) classifiers [1]. The main step in designing an IB classifier is the maintenance of the reference set, which stores only some of the data items seen so far. The idea is to store only those data items in the reference set, that seem to be key to accurate classification of newer data items. Moreover, the size of the reference set is also an important factor, because it determines the speed of classification. The smaller the size of reference set, the faster is the classification. We have implemented the classical IB classifiers with modifiable reference set and used an EP miner, developed in [2], to extract emerging V. V Das, R. Vijaykumar et al. (Eds.): ICT 2010, CCIS 101, pp. 228–236, 2010. © Springer-Verlag Berlin Heidelberg 2010

Instance-Based Classification of Streaming Data Using Emerging Patterns

229

patterns from the reference set. We then use the obtained emerging patterns to selectively update the reference set with the streaming data, hence we name it ILEP – Instance-based Learning using Emerging Patterns. This updated reference set is in turn used for classifying future streaming data. The paper proceeds with a brief discussion of the background work that has been used by us. In section 2, we discuss the classical algorithms of Instance-based classifiers, namely IB1, IB2 and IB3. In section 3, we present the concept of emerging patterns and how they can be used for mining streaming data. This is followed by section 4, where we present the design of ILEP. Section 5 gives details about the experiments performed with our model and also presents a comparative study with the previous technique. The paper ends with concluding remarks in section 6.

2 Instance Based Classifiers For our purpose, we mainly focus on online nearest neighbour classification strategies (Instance-based learning). Instance based classifiers are used because they are the most suitable types of online classifiers for streaming data classification. There are three popular algorithms for instance based classification, namely IB1, IB2 and IB3. For our purpose, we employ IB2 and IB3 algorithms. This is because IB1 algorithm adds all data items seen till the instant to the reference set, which will not be possible in streaming data classification, due to huge storage requirements in this case. IB2 starts with an empty reference set S. Upon receiving a new object, it classifies it using the objects currently held in memory. If the classification is correct, the object is discarded. Conversely, if the classification is incorrect, the object is added into S. IB3 is an extension of IB2 that employs a "wait and see" evidence gathering method to determine which of the saved instances are expected to perform well during classification. IB3 maintains a classification record (i.e., the number of correct and incorrect classification attempts) with each saved instance. A classification record summarizes an instance's classification performance on subsequently presented training instances and suggests how it will perform in the future. IB3 employs a significance test to determine which instances are good classifiers and which ones are believed to be noisy. IB3 accepts an instance if its classification accuracy is significantly greater than its class's observed frequency and removes the instance from the concept description if its accuracy is significantly less. Kuncheva et al. in [3] have proposed a classification strategy for streaming data by making modifications in IB2 and IB3 algorithms. Due to unavailability of true labels at the time of classification, the additions to the reference set have been made based on a simple heuristic i.e. whether the two nearest neighbours of the data item in the reference set have the same label. If not, the classification is assumed to be wrong and the data item is added to the reference set with the label of the nearest neighbour. This technique resulted in good accuracy of classification, but still led to increase in the size of reference set, which was seen to be marginal in large data sets. Although no name was given to this scheme, for our scheme, we will refer to it as the Simple IB Scheme. We will perform comparisons of this technique with ours in section 5.

230

M. Amir and D. Toshniwal

3 Emerging Patterns EPs (Emerging Patterns) are a new kind of patterns introduced recently [4]. They have been proved to have a great impact in many applications. EPs can capture significant changes between datasets. They are defined as item sets whose supports increase significantly from one class to another. The discriminating power of EPs can be measured by their growth rates. The growth rate of an EP is the ratio of its support in a certain class over that in another class. Usually the discriminating power of an EP is proportional to its growth rate. The process of finding out the emerging patterns is based on mining the strongest EPs from strongest instances in the reference set. The set of EPs is updated according to the strength of EPs and data instances. If gr(e) is the growth rate of a pattern from one class to another and s(e) is the support of that patter, then the strength of an EP e, strg(e), is defined as follows. ( )

( ) ( )

( )

(1)

The strength of an EP is proportional to both its growth rate (discriminating power) and support. Notice that if an EP has a high growth rate and a low support its strength might be low. In addition, if it has a low growth rate and a high support its strength might also be low. The strength of an instance, I, is defined by a fitness function as follows. ( )

∑

( )

(2)

The fitness function of a data instance can be measured by the average support of the attribute values in this instance. Suppose that we have an instance i {a1, a2, a3, ... an}. We first find the supports of all the attribute values (from a1 to an). We then average these supports to obtain a measure that tells how good the instance is. For our purpose, we use the Emerging Patterns Discovery application developed by Roman Podraza and Mariusz Kalinowski [2].

4 Design of the ILEP Scheme The basic design of the ILEP has been shown in Fig. 1. The design includes an implementation of the instance based classifier – the IB classifier, two sets of data namely emerging patterns set (contained in the EP Filter) and the reference set and a decision block that decides whether to add a data to the reference set or not. The input to the system includes a stream. At step t in the delayed labeling scenario, we have a reference set St and a new unlabelled object xt+1. After predicting the label for xt+1, we receive the label of the object that arrived τ steps earlier, xt−τ+1. Objects from xt−τ+2 to xt+1 are still without labels.

Instance-Based Classification of Streaming Data Using Emerging Patterns

231

To output

Streaming Data

Delayed Labels

IB Classifier

Reference Set

EP Filter

Incorrect?

Predictor Discard

Correct?

Add Data to reference set

Fig. 1. Design of the ILEP scheme

Using the techniques of EP mining discussed before, we first mine the EPs from the initial reference set and maintain an Emerging Patterns Set in the EP Filter. At any instant, as a data point is received as input, it is instantly classified by the IB classifier and the result is sent to the output. At the same time, EP Filter matches this data point with the mined EP Set. If the data point satisfied a strong EP i.e. an EP with large value of growth and high support of target class, then this data item is passed to the decision making system (else it is filtered out), which decides whether to add that item to the reference set or not. The versions of ILEP that uses IB2 is called ILEP2 while one that uses IB3 is called ILEP3 There are essentially two areas of concern where we have to focus for accurate classification – addition of data items to the reference set and EP set extraction. 4.1 Maintenance of Reference Set In ILEP, the reference set forms the backbone of the IB classifier. At the same time, it is also used to mine EPs to be used by the EP Filter. Then, it is used by the predictor component to decide whether to add the newer data items or not. The reference set is non-static, meaning that it will be continuously modified during the course of classification. While sometimes, it will be modified by the IB classifier’s error driven correction mechanism, sometimes, it will also be modified by the output data items from the Predictor. Here modification in the former case implies addition or deletion of the data item, while in the latter case, it implies only addition to the reference set.

232

M. Amir and D. Toshniwal

Once we have a data item that satisfies a strong EP, we may need to add that item to the reference set. For addition of a data item to the reference set, we need to know its true label. But in the problem statement, it is made clear that there is no way of knowing the true label of data item before a certain time lag. First we make sure if the data item is important enough to be added to the reference set. In the simple IB scheme, an intuitive heuristic method was proposed for deciding whether to add the data item to the reference set. This heuristic checked the first two nearest neighbours of the item in the reference set and if they were dissimilar, the item was added to the reference set. We improve upon this heuristic and add more conditions to the decision maker, emphasizing the fact that one of our main focus is to keep the reference set under limit. The two conditions that we add to the decision maker are: firstly, whether the target class predicted by the EP miner and the target class predicted by the IB classifier match and secondly, whether label of the nearest neighbour and the target class of the EP miner are same. If so, we conclude that the data item is a strong distinguishing point and its addition to the reference set will improve the classification accuracy of the IB classifier. Hence we add such a data item to the reference set, with its label same as that predicted by the IB classifier. It must be noted at this point that such an addition may lead to addition of a data item with wrong label. 4.2 Mining of Emerging Patterns The procedure for mining of the EPs has already been discussed before. The EP set consists of only few top EPs from the total mined EPs, with maximum growth rates. But the main issue is that we cannot classify the complete set of streaming data using one single EP set that was mined in the beginning as in that case, any recent changes in the concept description will not be captured by the EP set and hence the decision maker will rely upon the old EP set even for newer data items. Hence we need to have a progressive EP set. For this, we need continuous EP mining of streaming data. But mining EPs continuously is neither feasible (as it is a very intensive process and cannot be used in case of streaming data) nor required as small addition in the reference set will not make much difference in the EP set. A simple yet effective strategy would be to time the mining of EPs at regular intervals, so that the decision making system is constantly updated. A more active strategy would be based upon the number of modifications in the reference set. As part of this strategy, we maintain a modification counter initialized to zero. Every time a data item is added or deleted from the reference set, we increment this counter. Every time this counter reaches a certain overflow limit, we update the EP set by mining the EPs again and reset the counter. The overflow limit will be called EP update counter limit. This strategy is much better than the fixed update frequency method because in that case, we have to initialize the frequency in advance, which will depend upon the type of data set. Hence we need to have some knowledge about the data set we are working on. But in this approach, the strategy decides the value of the update frequency by itself. We just need to initialize the counter limit, which can be same for all type of data sets.

Instance-Based Classification of Streaming Data Using Emerging Patterns

233

The complete classification procedure can be summarized in following steps: 1. 2.

3. 4. 5. 6.

As a new data stream block is received, it is classified using the IB classifier and the result is outputted. Proceed to step 2. An EP Mining system, that extracts the emerging patterns from the reference set of the IB classifier, is used to filter out data items matching the emerging patterns with high growth rate. A decision making system, which uses various heuristics proposed previously, adds some of the data items from step 2 to the reference set. If modification counter reaches the EP update counter limit, mine EPs again and reset the counter. On receipt of the delayed class labels of already received data items, the corresponding items of the reference set are updated accordingly On arrival of newer streaming data, continue again from step 1.

5 Experiments Due to unavailability of real time streaming data, we carried out our experiments on streaming data classification techniques by using data sets chiefly from UCI repository [6]. We simulated the streaming process of data items for classification using data sets given in table I. Table 1. Data sets used Data Sets

Features

Classes

Objects

Iris Wine Sonar Glass Ionosphere Liver Page Block Vehicle Letters

4 13 60 10 34 6 10 18 16

3 3 2 7 2 2 5 3 26

120 148 178 171 316 293 5409 752 17619

The determining factors for effective strategy will be accuracy of the classification as well as the final percentage change in the size of the reference set. We mainly conducted three experiments to compare our results. First we obtained accuracy results for the simple IB scheme (using IB3 algorithm, as according to [3], IB2 does not perform well with the ‘Do Nothing’ approach that we have used in our work). Then we obtain accuracy results for ILEP2 and then ILEP3 algorithm. The results have been compared as bar charts in figure 2. Similarly, we obtain results for change in size of reference set in these three cases. These results have been compared in Figure 3. The value of K for the IB classification (K – nearest neighbour) is 20% of

234

M. Amir and D. Toshniwal

the initial size of reference set. The EP update counter limit is set to 15 and delay in arrival of true labels is 10. The percentage of top EPs considered by the decision maker is set to 30.

Accuracies

100 80 60 40 20 0

Data Sets Accuracy using Simple IB technique

Accuracy using ILEP2

Accuracy using ILEP3 Fig. 2. Comparing classification accuracies using the simple IB scheme, ILEP2 and ILEP3

The first and most important result is the effectiveness of using emerging patterns technique in classification of streaming data. Figure 2 compares the accuracies obtained from the classification experiments using the simple IB scheme (IB3), ILEP2 and ILEP3. We observe that using ILEP, we could not gain much increase in the accuracy level. In fact, in many cases, the classifier of simple IB scheme was found to be more accurate. Also, while comparing IB2 and IB3, IB2 is found to have an edge over IB3, but the difference is not very high. But, from the point of view of reference set size, our technique definitely stands ahead, as is clear from figure 3, which shows the percentage increase in the reference set size of various data sets after the final classification in the three experiments that we did. It should be noted that in smaller data sets as those of Iris, wine, sonar etc., the increase in reference set size is almost zero. This, along with the fact that there classification accuracy was good (using any technique) resonates well with the fact that reference set size increases on addition of data items when there classification is predicted incorrectly. The advantage of using our technique becomes clear from the results of large data sets like Letters, where the increase in reference set size using EP miner is almost zero whereas that using technique of [3] is around 1500%. The simple IB scheme ends up with good accuracy but with a very huge final reference set size due to large number of testing objects. It can be easily predicted that this technique will keep on increasing its reference set size as more and more objects are tested by it and hence is very unsuitable for large data sets, unlike ILEP.

% Increase in reference set sizes

Instance-Based Classification of Streaming Data Using Emerging Patterns

235

1600 1400 1200 1000 800 600 400 200 0 -200

Data Sets % increase in reference set size using Simple IB technique % increase in reference set size using ILEP2 % increase in reference set size using ILEP3 Fig. 3. Comparing increase in reference set sizes using simple IB scheme, ILEP2 and ILEP3

Figure 3 also shows comparisons between classifications obtained from ILEP2 and ILEP3 algorithms. As is clear, while the reference set size increases in case of ILEP2 algorithm for all data sets, ILEP3 is able to do an almost equally good classification with a shorter reference set. This result is of great importance, because in case of streaming data sets, the amount of data input is huge, and if we do not limit the size of reference set, the classification efficiency will be very low, resulting in slow down of the process, which may be disastrous. Therefore, we would say that if the accuracy of the system can be slightly compromised, we will always choose the EP miner with IB3 method for classification. But if the accuracy of the system is important, then we will have to compare the two algorithms upon the data set that is going to be used.

6 Conclusion This paper implemented a novel technique that integrates instance-based classifiers with EP miner for classification of streaming data. The ILEP2 and ILEP3 classifiers were applied on various data sets of the UCI repository. The accuracy results of the implemented classifiers were found to be similar to those of the simple IB scheme, but a significant gain was obtained in terms of the final reference set size. While in the simple IB scheme, the size of reference set drastically increases and goes to values as high as 1000% to 1500%, in ILEP, the reference set remains almost of the same size in case of ILEP2 algorithm and decreases in size in case of ILEP3 algorithm. It was found that both ILEP2 and ILEP3 can be of good use depending upon the type of

236

M. Amir and D. Toshniwal

data sets used. While ILEP2 is ahead of ILEP3 in terms of accuracy, ILEP3 is better than ILEP2 when it comes to restricting the size of reference set. Although the implementation for the proposed work has been done on a centralized system, the scalability of the design becomes more apparent, when a distributed system is used for the purpose of classification of data streams. Distributed systems are getting more and more importance in this age of overwhelming data storage. Since streaming data is normally associated with huge amount of data, which is very difficult to be handled with a centralized system, the use of distributed systems for classification purpose becomes all the more obvious. The most important thing about our proposed scheme that makes it suitable for distributed systems is that our scheme is easily decomposable into distributed and self dependent components that can be easily implemented on any platform. There are many aspects that have been left untouched for future works. We have not considered the time taken by the online classification for each data item. Moreover, the EP mining system that we used in our work is meant for conventional data. More work needs to be done so as to develop a system for mining EPs from streaming data systems.

References [1] Podraza, R., Walkiewicz, M., Dominik, A.: Credibility Coefficients in ARES Rough Set Exploration System. In: Ślęzak, D., Yao, J., Peters, J.F., Ziarko, W.P., Hu, X. (eds.) RSFDGrC 2005. LNCS (LNAI), vol. 3642, pp. 29–38. Springer, Heidelberg (2005) [2] Kuncheva, L.I., S´anchez, J.S.: Nearest Neighbour Classifiers for Streaming Data with Delayed Labelling. In: 8th IEEE International Conference on Data Mining, Pisa, Italy, pp. 869–874 (2008) [3] Dong, G., Li, J.: Efficient Mining of Emerging Patterns: Discovering Trends and Differences. In: International Conference on Knowledge Discovery and Data Mining (KDD 1999), San Diego, CA, USA, pp. 43-52 (1999) [4] Nakayama, H., Yoshii, K.: Active forgetting in machine learning and its application to financial problems. In: International Joint Conference on Neural Networks, Como, Italy, pp. 123–128 (2000) [5] UCI Machine learning repository, http://archive.ics.uci.edu/ml/ [6] Siebert, J.P.: Turing Institute Research Memorandum TIRM-87-018: Vehicle Recognition Using Rule Based Methods (March 1987) [7] Seeger, M.: Learning with labeled and unlabeled data. Technical Report, University of Edinburgh, UK (2002) [8] Wilson, D.R., Martinez, T.R.: Reduction techniques for instance-based learning algorithms. Machine Learning 38(3), 257–286 (2000) [9] Domingos, P., Hulten, G.: Mining High-Speed Data Streams. In: 6th International Conference on Knowledge Discovery and Data Mining, pp. 71–80. Association for Computing Machinery, Boston (2000)

Modified Go-Left Balls and Bins Algorithm for Server Load Balancing Prasun Banerjee, Stephan D’Costa, and Sukriti Bhattacharya Department of Computer Science & Engineering University of Calcutta , West Bengal, India {banerjee.prasun,dcosta.stephan,bhattacharya.sukriti}@gmail.com http://www.caluniv.ac.in

Abstract. This paper proposes a modified version of Go-left Balls & Bins algorithm for server load balancing using K-Partite property of a graph. The previous algorithms all had to keep knowledge of the past load distribution information while distributing new load -an issue that is virtually impossible in real life as it itself congests the server with load. But ours algorithm is random in nature and hence sheds this overhead; this is also quite realistic in nature and close to the implementation domain. Keywords: Server Load balancing, Go-left Balls & bins, K-partite graph.

1 Introduction Internet server programs supporting mission-critical applications such as financial transactions, database access, corporate intranets, and other key functions must run 24 hours a day, seven days a week. Network servers are now frequently used to host ERP, e-commerce and a myriad of other applications. The foundation of these sites the e-business infrastructure - is expected to provide high performance, high availability, and secure and scalable solutions to support all applications at all times. However, the availability of these applications is often threatened by network overloads as well as server and application failures. Resource utilization is often out of balance, resulting in the low-performance resources being overloaded with requests while the high-performance resources remain idle. Server load balancing [1][2][4] is a widely adopted solution to performance and availability problems. Load balancing is a technique used for distributing service requests evenly among servers that offer the same service. Load refers to a number assigned to a service request based on the amount of time required to execute that service. Loads are assigned to services so that the system can understand the relationship between requests. To keep track of the amount of work, or total load, being performed by each server in a configuration, the administrator assigns a load factor to every service and service request. A load factor is a number indicating the amount of time needed to execute a service or a request. On the basis of these numbers, statistics are generated for each server and maintained on the bulletin board on each machine. Each bulletin board keeps track of the cumulative load associated with each server, so that when all servers are busy, the system can V. V Das, R. Vijaykumar et al. (Eds.): ICT 2010, CCIS 101, pp. 237–242, 2010. © Springer-Verlag Berlin Heidelberg 2010

238

P. Banerjee, S. D’Costa, and S. Bhattacharya

select the one with the lightest load. The two major categories for load-balancing algorithms are static, and dynamic. Static load balancing algorithms allocate the tasks of a parallel program to workstations based on either the load at the time nodes are allocated to some task, or based on an average load of workstation cluster. Dynamic load balancing algorithms make changes to the distribution of work among workstations at run-time; they use current or recent load information when making distribution decisions. As a result, dynamic load balancing algorithms can provide a significant improvement in performance over static algorithms. There are three major parameters which usually define the strategy a specific load balancing algorithm will employ. These three parameters answer three important questions: Who makes the load balancing decision? What information is used to make the load balancing decision? and Where the load balancing decision is made? Based on these three questions can categorize the load balancing strategies as – Sender-Initiated vs. Receiver-Initiated Strategies: The first question is answered based on whether a sender-initiated or receiver-initiated policy is employed. In sender initiated policies, congested nodes attempt to move work to lightly-loaded nodes. In receiver-initiated policies, lightly-loaded nodes look for heavily-loaded nodes from which work may be received. – Global vs. Local Strategies: Global or local policies answer the second question. In global policies, the load balancer uses the performance profiles of all available workstations. In local policies workstations are partitioned into different groups. The benefit in a local scheme is that performance profile information is only exchanged within the group. – Centralized vs. Distributed Strategies: A load balancer is categorized as either centralized or distributed, both of which define where load balancing decisions are made. In a centralized scheme, the load balancer is located on one master workstation node and all decisions are made there. In a distributed scheme, the load balancer is replicated on all workstations. In this paper we state and improve the limitations of traditional go-left balls and bins algorithm and propose a centralized, global and dynamic version of it, using K-partite property of the graph. The paper is structured as follows. Section 2 illustrates traditional ’Balls & Bins’ go-left algorithm and its limitations. In section 3, we propose our algorithm, the modified version of ’Balls & Bins’ go-left using K-partite graph property. In section 4, we show the results. In section 5, we conclude by discussing the main advantages of our algorithm.

2 Go-Left 'Balls and Bins' Suppose n balls have to be assigned to n bins, where each ball has to be placed without knowledge about the distribution of previously placed balls [3][4]. The goal is to achieve an allocation that is as even as possible so that no bin gets much more balls than the average. In [1] it is shown that a non-uniform and possibly dependent choice

Modified Go-left Balls and Bins Algorithm for Server Load Balancing

239

of the locations for a ball can improve the load balancing. The algorithm can be stated as follows, For each new incoming unit load (call it a ball) perform the following: Step 1: Divide the graph (of nodes) into 2 partitions. Step 2: Select a node (call it a bin) randomly from the first partition; Step 3: And choose a 2nd bin from the second partition depending on the 1st choice. Step 4: Then compare the load (number of balls) of both the bins and places the ball in the bin with lesser load. In case of a tie it goes left (quite in a biased way) and places the ball into the 1st bin, i.e., go-left. This algorithm is called dependent, because every node from the 2nd partition is chosen depending on the choice of node of the 1st partition. This algorithm is called nonuniform for a biased tie-breaker mechanism. The above algorithm lacks the following points: • • •

We need a uniform node numbering; based on which the nodes of the 2nd partition can be chosen as a function of the nodes chosen from the 1st partition. The algorithm mentioned above is silent about how to make partitions; it somehow makes that partitions randomly or assume that is already done. Though it makes partitions it completely ignores the concept of clustering network; an idea that can be incorporated depending on node’s connectivity, topology and geographical locations.

3 K-partite Go-Left 'Balls and Bins' Algorithm Here we look forward to use the centralized-global-dynamic algorithm by making use of previous Go-left balls and bins algorithm. K-partite Go-left Algorithm has been proposed to distribute the jobs among several nodes for processing. Thus increasing the system performance. Asymmetry often helps to distribute load and still brings good performance that we don’t usually expect it shows that performance gets a hike if we distribute it rather nonuniformly and in a dependent fashion. Therefore, we tackle the limitations stated in the last section as follows: 1. We are using a uniform numbering system of the form (Partition Id:node Id). This form of numbering easily finds its’ slimily in the network address of the form (Host Id:net Id) 2. Here we make partitioning explicitly by using K-Partite property of a graph. 3. The K-partite nature of the graph account for its connectivity and topologyan issue that is considered when clustering a network. With these modifications our algorithm is stated in the following. The algorithm starts with a connected random graph G with n nodes, each node has a unique node index associated with it. The load associated to each node index is set to zero in the

240

P. Banerjee, S. D’Costa, and S. Bhattacharya

initialization phase. Then G is divided into K parts depending on its k-partite property [2][3][5]. Each partition has n unique partition index associated. The incoming loads are distributed in the following way. Modified K-partite Go-left Balls & Bins Algorithm: 1: 2: 3:

4:

Generate random graph (We don’t consider the graph with disconnected components). Initialize it with Load 0 for all the nodes. Partition the graph into K parts depending on its k-partite property: 3.1: Take the first node randomly and assign it the partition index = 0 and the node index = 0. 3.2: For every other node that are still not member of any partition 3.2.1: For every other partition 3.2.1.1: Check whether the new node is connected to any nodes of this partition. 3.2.1.2: If it is not, let it be a member of this partition 3.2.2: Otherwise assign it to a new partition(with new partition index). Distribute the load between the partition using a modified version of the goleft algorithm: 4.1: For each unit load 4.1.1: Take a random node from the first partition. 4.1.2: For each remaining partition. 4.1.2.1: Choose a node Newnodeindex = f(Oldnodeindex) Function f can be defined as f(Oldnodeindex) = counttotal_node (partion index) + (Oldnodeindex)2% countnumber_of_partitions

(

counttotal_node countnumber_of_partitions

)

4.1.2.2: If this node has a load less than the previous node assign it a unit load. 4.1.2.3: In case of a tie; always go left.

4 Results L

The optimum efficiency of each node is calculated by n %, where L is the total load to be distributed and n is the total number of nodes. The result of our algorithm (the actual efficiency) is plotted across X and Y-axis along with the optimum efficiency. The X-axis denotes total number of nodes considered each execution (from 10 to 100), while the corresponding Y-axis shows the maximum amount of load a node can support (the actual efficiency). The black curve denotes efficiency of our algorithm where the gray curve denotes optimal efficiency. Here we run our algorithm for 1024 loads.

Modified Go-left Balls and Bins Algorithm for Server Load Balancing

241

Fig. 1. Performance graph for 1024 loads

5 Discussion This proposed K-partite-go-left algorithm finds its use in the load balancer software used to do server side load balancing. In actual world there can be some CPU overhead for partitioning the network, re-partitioning the network once a new node is added or removed, balancing the load, re-balancing the load when new bulk of loads come. The complexity of load distribution algorithm is O(n2) and the complexity of partitioning is O(n3); and this is an overhead of this algorithm. But this algorithm has the modularity to support the dynamic load and dynamic node availability. In case a new node is added or deleted, we only need to run the partition routine on the modified adjacency matrix. In case a new load comes, we need to run the distribute routine. Additionally for managing a new bulk of loads together, we can actually run the load balancer in multiple synchronized threads. Our algorithm is actually not an optimal choice algorithm as it is nonuniform and dependent in nature in choosing the right bin. But it pays much lesser cost in building this near optimal algorithm. Because this algorithm uses the biased go left method, it removes the load oscillation. Acknowledgements. The authors would like to thank Mr. Sanjit Setua for his continuous support and helpful suggestions.

References 1. Vocking, B.: How Asymmetry Helps Load Balancing. In: Proceedings of the 40th Annual Symposium on Foundations of Computer Science. IEEE, Los Alamitos (1999)

242

P. Banerjee, S. D’Costa, and S. Bhattacharya

2. Long, B., Wu, X., Zhang, Z.M., Yu, P.S.: Unsupervised learning on k-partite graphs. In: Proceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, New York (2006) 3. Harary, E.: Graph Theory. Addison-Wesley, Reading 4. Mitzenmacher, M.: The Power of Two Choices in Randomized Load Balancing. PhD thesis, University of California at Berkeley (1996) 5. Knuth, D.E.: The Art of Computer Programming, vol. 3. Addison Wesley, Reading (1998)

Three Layered Adaptation Model for Context Aware E-Learning Minu M. Das, Manju Bhaskar, and T. Chithralekha Pondicherry Central University, Pondicherry, India [email protected], [email protected], [email protected]

Abstract. Current context aware e-learning system lacks in providing highly customized information to the learner, which considers context parameters discretely and there is no complete standardized set of learner contexts. The proposed three layered adaptation model gives all most all learner contexts in a standardized way, which improves the efficiency of learning process. The design solution provides a three layered architecture based on learner’s characteristics, which is divided into three layers such as conceptual layer, logical layer and physical layer which helps to improve the efficiency of learning process. Keywords: e-learning, context aware e-learning, layered architecture, adaptation model, learning path, intention, preference.

1 Introduction E-learning is aimed at supporting learners in fulfilling a specific learning need with in a specific situation through the use of information and communication technology. Unfortunately, current e-learning systems do generally not consider some important characteristics capable of providing a complete context aware e-learning system to the learner. Existing e-learning systems do not pay attention to the learner’s complete characteristics, thus all students are given the same materials and activities. Nor do didactic materials offer, due to the access possibilities to different devices such as PDA, mobile, PC, Laptops and so on, in an efficient way. Context aware e-learning systems select or filter the learning resources in order to make the e-learning content more relevant and suitable for the learner in his/ her situation. However, looking for a stronger learning personalization, the design of courses to ensure the right knowledge acquisition by the student, in his own way, taking in to account of learners characteristics, preferences, intentions and device used. Thus a standardization of contexts of learner needed for efficient e-learning process. The proposed three layered adaptation model for Context aware e-learning gives the standardized contexts of learner in all aspects. The whole paper is organized as follows: Section 2 describes the existing works in context aware e-learning. The Need for standardized Context aware e-learning is described in section 3. Section 4 illustrates the three layered adaptation model for context aware e-learning. The case study and system snapshots are described in section 5. Finally section 6 concludes the paper. V. V Das, R. Vijaykumar et al. (Eds.): ICT 2010, CCIS 101, pp. 243–248, 2010. © Springer-Verlag Berlin Heidelberg 2010

244

M.M. Das, M. Bhaskar, and T. Chithralekha

2 Existing Works in Context Aware E-Learning In the existing works contextualization are achieved by delivering appropriate material, or by considering the preference or intentions of the learner, or by giving the learning material in suitable media. The details of each category of contextualization are described in the following section. 2.1 Learning Paths Learning path defines the sequence of learning activities that is carried out by learner going through the learning units in the e-learning system. Learning unit is an abstract representation of a course, a lesson, a workshop, or any other formal or in-formal learning or teaching event [17]. Context aware learning path generation is about creating a learner experience that purposely adjusts to various conditions such as personal characteristics, pedagogical knowledge and the learner interactions over the period of time with the intentions to increase success criteria. 2.2 Learning Object Model The design of the standardized context model requires a flexible learning object model. That is, the learning object structure should not be static. The structure of the learning object will change according the learner’s preferences and intentions. This requires that the learning object is structured in terms of different levels of abstractions as given below in Table 1. That is, the same learning object is available in the form of a concept, detailed concept, example, case study, simulation and demonstration. Each of these corresponds to the various abstraction of the same learning object. Every abstraction would be available in different media types such as text, audio, video, animation etc. When a learner whose learning preference is learning by case study approaches the E-Learning system with the intention of preparing for interview, the different learning object abstraction chunks chosen to constitute the learning object structure and the sequencing order of these abstractions while presenting to the learner is as shown below Simple Concept

Case Study

Example.

Thus, for the above mentioned scenario, the learning object is structured with three abstractions – simple concept, case study and example. This structure is dynamically determined based on the learning preference and intention of the learner. Formalized way of representing the learning object model for different intentions of the learner is given in Table 2. This shows the different learner’s intention and the corresponding sequencing of the learning object abstractions for each of these intentions.

Three Layered Adaptation Model for Context Aware E-Learning

245

Table 1. Learner’s Preferences and Intentions in Existing System ontext Parameters Considered Learning Preference

Learning Intention

Sub Context Parameters Conceptual Example-Oriented Case Study Simulation Demonstration Research Survey/ Overview Quick Reference Basic Introduction Project Assignment Seminar

Existing Systems [1],[2],[3],[4],[5],[6],[7] [9],[10],[11],[12],[13], [15],[16],[18],[19]

[6],[7],[14],[18],[19]

Table 2. Sequencing of contents based on learner’s preferences and Intentions Learner’s Intentions Research Survey

Learning chunk abstraction constituting the learning object Simple Concept, Detailed Concept, Example, Case Study, Demonstration, Simulation Detailed Concept, Example, Case Study

Quick Reference Basic Introduction Project

Simple Concept, Example, Case Study Simple Concept, Example Detailed Concept, Example, Case Study, Simulation, Demonstration

Seminar Assignment

Detailed Concept, Example, Case Study, Demonstration Detailed Concept, Example, Case Study

3 Need for the Proposed Context Model Existing context aware e-learning systems are using few parameters which are described in section 2. There is no such context aware e-learning system which in cooperates all of the parameters which are described in section 2. So in order to include all the parameters and to give a standardized form for learner characteristics, a three layered adaptation model is proposed. Proposed system is using a three layered adaptation model, which helps to include all the learning characteristics. The following section 4 describes about the three layered adaptation model for context aware e-learning.

4 Three Layered Adaptation Model for Context Aware E-Learning Three layered adaptation model gives a standardized model which completely capture the learner’s characteristics. The adaptation model divides all learner contexts in three

246

M.M. Das, M. Bhaskar, and T. Chithralekha

layers so that it helps efficient personalization in the learning process. Fig. 1. shows the three layered model for context aware e-learning system, in which each layer taken care of specific characteristics of learner. Three layers in the model are given below.

Physical Layer

Logical Layer

Adaptivity Increases

4.1 Conceptual Layer 4.2 Logical Layer 4.3 Physical Layer

Conceptual Layer

Fig. 1. Three Layered Adaptation Model for Context Aware E-learning

4.1 Layer 1: Conceptual Layer Conceptual layer deals with learning path generation. The learning path generation is important because different learners may have different characteristics, prior knowledge or motivation or needs [8]. This diversity commonly requires the presentation of different information to different learners in different formats [19]. That is why conceptual layer is very important which consider various aspects of individual students when presenting information and or practice opportunities in order to make the learning process as effective, efficient and motivating as possible. The parameters helps for providing adaptivity in conceptual layer given [8] are prior knowledge and prior skills, Learning capabilities, Learning preferences, Performance level and knowledge state, Interest, Personal circumstances, Motivation.

Three Layered Adaptation Model for Context Aware E-Learning

247

4.2 Layer 2: Logical Layer The second layer, logical layer mainly concentrates the learner’s preferences and learner’s intentions. Each learner has his/her specific preference and intentions for learning. The values for learner’s preferences which are considered for adaptation are Concept, Detailed Concept, Example, Case study, Simulation and Demonstration of a particular learning material. The intention of the learner can be research, survey work, interview purpose, assignment work, project or just to learn the concept. Based on the intentions and preferences the sequencing of the contents will differ. 4.3 Layer 3: Physical Layer The third layer, physical layer provides information about the device used by the learner for learning. The device can be mobile, PDA, laptop, PC and so on. According to the intention, preferences and media preferred by the learner, physical layer delivers appropriate learning material in specified media. Media can be audio, video , text or animation.

5 Case Study: Gurudev To explain how to perform the three layered adaptation process in e-learning system, the following section briefly explains the procedure. The experimental system is named as Gurudev and computer network subject is taken as example.. Based on the conceptual layer, logical layer and physical layer, context adaptive learning scheme is generated. Fig. 3. shows the generated learning scheme which is given to the server to get the personalized learning materials. The server is acted as learning content management system which dynamically composes the learning objects and is given to the learner which is given in Fig 4.. The snapshots of implemented system are given below.

Fig. 3. The generated Adaptive Learning Scheme Fig. 4. Dynamically composed learning object

6 Conclusion Three layered adaptation model for context aware e-learning system provides three layers of adaptation which helps for achieving personalization. The three layered adaptation model considers all most all learner characteristics in a standardized way. So the model helps to improve the e-learning process in an efficient manner.

248

M.M. Das, M. Bhaskar, and T. Chithralekha

References 1. Chen, P., Meng, A., Zhao, C.: Constructing Adaptive Individual Learning Environment Based on Multi- Agent System. In: IEEE International Conference on Computational Intelligence and Security Workshop, pp. 374–377 (2007) 2. Dietze, S., Gugliotta, A., Domingue, J.: Addressing Context-Awareness and Standards Interoperability in E-Learning: A Service-oriented Framework based on IRS III 3. Jeongwoo, K., Fumihiko, M., Teruko, M., Eric, N., Masahiko, T., Ichiro, A.: ContextAware Dialog Strategies for Multimodal Mobile Dialog Systems. J. AAAI (2006) 4. Jose, M.M., Juan, A.O., Luis, G., Francisco, V.: Creating adaptive learning paths using Ant Colony Optimization and Bayesian Networks. In: IEEE International Joint Conference on Neural Networks, pp. 3834–3839 (2008) 5. Jovanovic, J., Gasevic, D., Knight, C., Richards, G.: Ontologies for Effective Use of Contextin e-Learning Settings. Educational Technology & Society 10(3), 47–59 (2007) 6. Howe, D.: Free online dictionary of computing, Imperial College Department of Computing London, UK (2006) 7. Kareal, F., Klema, J.: Adaptivity in e-learning. Current Developments in TechnologyAssisted Education, 260–265 (2006) 8. Kawanishi, N., Jin, K.S., Si, H., Kawahara, Y., Morikawa, H.: Building Context-Aware Applications and Probe Space Infrastructure. In: IEEE International Symposium on Intelligent Signal Processing and Communications, pp. 103–106 (2006) 9. KounTem, S., HsinTe, C.: The Study of Using Sure Stream to Construct Ubiquitous Learning Environment. In: 2008 IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing, pp. 534–548 (2008) 10. Lanzilotti, R., Ardito, C., Costabile, M.F., De, A.A.: eLSE Methodology: a Systematic Approach to the eLearning Systems Evaluation. Educational Technology & Society 9(4), 42–53 (2006) 11. Limongelli, C., Sciarrone, F., Temperini, M., Vaste, G.: Adaptive Learning with the LSPlan System: a Field Evaluation. IEEE Transaction on Learning Technologies (2008) 12. Srimathi, H., Srivatsa, S.K.: Identification of ontology based learning object using instructional design (2008) 13. Stockley, D.: E-learning Definition and Explanation (Elearning, Online Training, Online Learning) (2003) 14. Sun Microsystems E-learning Framework 15. Thyagharajan, K.K., Ratnamanjari, N.: Adaptive Content Creation for Personalized e-Learning Using Web Services. J. Applied Sciences Research 3(9), 828–836 (2007) 16. Tuparova, D., Tuparo, G.: Learning paths in open source e-learning environments. Current Development in Technologies-Assisted Education, 1565–1569 (2006) 17. Wang, M., Ci, L., Zhan, P., Xu, Y.: Applying Wireless Sensor Networks to ContextAwareness in Ubiquitous Learning. In: IEEE Third International Conference on Natural Computation, vol. 5, pp. 791–795 (2007) 18. Yang, S.J.H.: Context Aware Ubiquitous Learning Enviornments For Peer to Peer Collaborative Learning. Educational Technology & society 9(1), 188–201 (2006) 19. Zajac, M.: Using Learning Styles to Personalize Online Learning. J. Campus- Wide Information System 26(3), 256–265 (2009)

AFDEP: Agreement Based CH Failure Detection and Election Protocol for a WSN Amarjeet Kaur and T.P. Sharma Computer Science And Engineering Department, National Institute of Technology, Hamirpur – 177005 India [email protected], [email protected]

Abstract. In this paper, we propose an agreement-based fault detection and recovery protocol for cluster head (CH) in wireless sensor networks (WSNs). The aim of protocol is to accurately detect CH failure to avoid unnecessary energy consumption caused by a mistaken detection process. For this, it allows each cluster member to detect its CH failure independently. Cluster members employ distributed agreement protocol to reach an agreement on failure of the CH among multiple cluster members. The detection process runs concurrently with normal network operation by periodically performing a distributed detection process at each cluster member. To reduce energy consumption, it makes use of heartbeat messages sent periodically by a CH for fault detection. Our algorithm would provide high detection accuracy because of agreement protocol.

Keywords: Wireless Sensor Network, Clustering, Fault detection, Agreement protocol, Detection accuracy.

1 Introduction Wireless sensor networks (WSNs) consist of hundreds and even thousands of small tiny devices called sensor nodes distributed autonomously to monitor physical /environmental conditions, infrastructure protection, battlefield awareness etc. Each sensor node has sensing, computation, and wireless communication capabilities [1]. Sensor nodes sense the data and send it to base station (BS). Sensor nodes have energy constraint. Sensor nodes are often left unattended which makes it difficult or impossible to re-charge or replace their batteries. The cost of transmitting information is much higher than computation and hence it is necessary to reduce the number of transmissions. Clustering is an effective way to reduce number of transmission and prolongs network lifetime. There are number of clustering-based routing protocols proposed in literature for WSNs [2-11]. These protocols improve energy consumption and performance when compared to flat large-scale WSNs. Sensor nodes are prone to failure due to harsh environment. The failure of a sensor node affects the normal operation of a WSN [12].The failure of a CH makes situation even worse. In literature, number of authors have proposed fault tolerant protocols [13-16]. In this paper, we propose a fault tolerant protocol for WSN, which is based on agreement protocol. V. V Das, R. Vijaykumar et al. (Eds.): ICT 2010, CCIS 101, pp. 249–257, 2010. © Springer-Verlag Berlin Heidelberg 2010

250

A. Kaur and T.P. Sharma

2 Related Work Clustering is an effective way for improving the energy efficiency and prolonging the network lifetime of WSNs. The CH failure causes the connectivity and data loss within cluster. It also disconnects cluster members from rest of the network. Hence, it is crucial to detect and recover the CH failure to maintain normal operation of cluster and network as a whole. In [17], an agreement-based fault detection mechanism is proposed for detecting CH failures in clustered Underwater Sensor Networks (UWSNs). In this, each cluster member is allowed to independently detect the fault status of its CH and at the same time a distributed agreement protocol is employed to reach an agreement on the fault status of the CH among multiple cluster members. It runs parallel with normal network operation by periodically performing a distributed detection process at each cluster member. It provides high detection accuracy under high packet loss rates in the harsh underwater environment. FTEP [18] is a dynamic and distributed CH election algorithm with fault tolerance capabilities based upon two-level clustering scheme. If energy level of current CH falls below a threshold value or any CH fails to communicate with cluster members then election process is started which is based on residual energy of sensor nodes. This election process appoints a CH and a back up node to handle CH failure. It has a single point (back up node) to detect failure which may itself be disastrous. In cellular approach to fault detection and recovery [19], network is partitioned into a virtual grid of cells, where each cell consists of a group of sensor nodes. A cell manager and a secondary manager are chosen in each cell to perform fault management tasks. Secondary manager works as back up node which will take control of the cell when cell manager fails to operate. This protocol handles only those failures which are caused due to energy depletion. In [20], a method to recover from a gateway fault is proposed, which is based on agreement protocol. A periodic status updates are sent through inter-gateway communication. The status updates can be missed due to link failures between two sensor nodes, hence a consensus has to be reached by all gateways before recovery commences. When a gateway is identified as completely failed all the sensor nodes in its cluster are recovered. Venkataraman algorithm [21], proposed a failure detection and recovery mechanism which is also based on energy exhaustion. It focused on sensor node notifying its neighboring sensor nodes before it completely shuts down due to energy exhaustion.

3 System Model 3.1 Assumptions Here we make some assumptions as following: • Nodes failure due to energy depletion or any hardware or software problem. • All nodes are homogenous, immobile and have limited energy [22]. And initially have same amount of energy. • Each node has fixed number of transmission power level.

AFDEP: Agreement Based CH Failure Detection and Election Protocol for a WSN

251

• Transmission power is uniform across network. • Every node knows its current energy level [18]. • A message sent by a node is received correctly with in a finite time by all nodes in the cluster. • The clusters are static i.e. are formed at the start of the network. After that CH rotates. • Nodes know about their location [18]. • The BS is fixed and not located between sensor nodes [22]. 3.2 Network Model Fig. 1(a) shows the network model used. Various symbols and terms used are shown in Table 1. All sensor nodes are homogeneous, which have two transmission modes i.e. high power transmission mode for communication between CHs and BS and low power transmission mode for communication between cluster members and CH. The distribution of sensor nodes is uniform throughout the environment. Communication medium is radio links. Links between two sensor nodes is considered bidirectional. There is only single channel for communication between sensor nodes.

Sensor node Cluster Head

Fig. 1. Network Model Table 1. Notions used to explain protocol Symbol

′ ′′

Meaning of symbol Distance that message travels Number of bits in the message Energy dissipated in transmitter electronics per bit (taken to be 50nJ/bit) Energy dissipated in transmitter amplifier (taken to be 50nJ/bit) Energy dissipated in receiver electronics per bit (taken to be 50nJ/bit) Energy consumed in transmission Energy consumed in receiving Location of node Current energy of node Energy level at which sensor node can participant in election of Energy level at which current starts election process Energy level up to which election process must be completed Candidate set

252

A. Kaur and T.P. Sharma

During the network deployment, all the sensor nodes are assigned same initial energy value. All sensor nodes are assumed to know their geographical [23]. We assume that clusters may overlap during election procedure so that every sensor node comes under at least one cluster. Initially, some sensor nodes are randomly selected as CHs and they announce their energy levels and location information. These CHs start working in high power transmission mode while other regular sensor nodes work in low power transmission mode. 3.3 Sensor Node’s Energy Model A sensor node consists of sensors, analog signal conditioning, data conversion circuitry, digital signal processing and a radio link [3]. Each component of sensor node consumes energy for sending and receiving data. The following energy consumption model shows the energy consumed by components of sensor node as shown in Fig 1 (b) (redrawn from [3]). Assuming 1/ path loss, the energy consumption on each sensor node is: (1) (2) to send bits; where According to eq. 1, the transmitter unit consumes energy is the energy consumed by transmitter electronics per bit and is the energy used by amplifier per bit. According to eq. 2, the receiving unit consumes energy is the energy used by receiver electronics per bit. to receive bits, where Table 1 summarizes the meaning of each term and its typical value. The values for ′ , and are updated during each election process. Typically, value of for next election round is set to the average value of the energy levels of all ′ is set according to candidate nodes during current election round. The values of ′′ ′ . The values of is set according to as follows: ′′

- (energy consumption during election process + energy consumption in data transmission during that period).

4 AFDEP Protocol 4.1 Setup Phase Clusters are formed only once during the setup phase before the network starts to run. Initially, some sensor nodes are randomly selected as a CH, because energy of each sensor node is equal in amount. CHs send advertisement messages that contain energy and location information of CHs to neighboring sensor nodes. Each sensor node that listen to this advertisement message responds with a return message comprising its residual energy and location. However, a sensor node may be in the range of multiple CHs, but finally it must be associated with a single CH. If any sensor node falls within the overlapping region of more than one CHs, it decides its association to a CH by calculating the value of e/d (energy/distance). CH, has maximum e/d value is selected as final CH for that sensor node. If more than one CHs yields same

AFDEP: Agreement Based CH Failure Detection and Election Protocol for a WSN

253

maximum e/d value, then any of them is randomly selected. If a sensor node does not fall within the range of any CH, it declares itself as a CH and gets activated in high power transmission mode. When clusters are established, the CHs collect the data from cluster members, perform local data aggregation and communicate with the BS. 4.2 Steady State Phase Once cluster is formed, CH creates a TDMA schedule for cluster members and sends it to them. Sensor nodes sense data and send it to CH according to TDMA schedule. ) equals This process continues for all clusters until CH’s current energy level ( ′

to or less than . Then CH starts election process of new CH for next round. CH for next round in low power transmission mode, which is average broadcasts energy of those cluster members who participated in last election process. All sensor ). nodes within cluster listen message and compare with their current energy level ( Sensor node which have greater than or equal to , marks itself as a participant for election process. All participant sensor nodes broadcast their and location ( ) in low transmission mode. All participant sensor nodes can listen to each other because all sensor nodes are within low power transmission range of each and of each other. Because of this, all participant sensor nodes know about other. Hence, each participant sensor node is aware about higher energy participant upgrades itself as sensor node. The participant sensor node with highest value of CH and gets activated in high power mode; where as sensor node with second highest and of all energy upgrades itself as back up CH. New CH receives participant sensor nodes during election process, it calculates average of all and gets value of , which is used for next round. Both new CH and back up node know . All participant sensor nodes mark themselves as non-participant the value of sensor nodes again. The previous CH also starts working in low power mode. Failure Detection. The detection process runs parallel with normal network operation by periodically performing a distributed detection process at each cluster member. For failure detection mechanism each cluster member maintains a status vector. In status vector each bit corresponds to a cluster member. Initially all bits of are set to zero of status vector on each sensor node. A bit in the vector is set once its corresponding cluster member detects that CH has failed. CH of each cluster periodically sends a hello message (i.e. notification that CH is alive) to cluster members. Cluster member, who does not listen hello message, sets its corresponding bit as one in status vector and locally decides that CH has failed and broadcasts data plus status vector. Other cluster members also listen this message. They extract status vector from message and merge it with own status vector and this process continuous up to the end of the TDMA schedule. At the end of the TDMA frame, cluster members reach on an agreement about failure of CH. If all bits of status vector are set then it is decided that CH has failed. Failure Recovery. By using agreement protocol when cluster members confirm about CH failure then cluster member who has last slot in TDMA schedule informs to back up node about failure. Back up node elects itself as a CH and sends an advertisement message in high power transmission mode. It keeps on working as CH till its residual

254

A. Kaur and T.P. Sharma

energy level reaches a critical limit or it fails. New back up node is required for new in low CH, so CH start election process for new back up node with sending power transmission mode. Back up node election process is similar to election process of CH.

5 Performance Evaluation 5.1 Simulation Environment In this section, we evaluate the performance of our proposed AFDEP protocol. We used OMNET-4.0 [24] as simulator and same radio model as discussed in. The basic simulation parameters are given in Table 2.

In order to check the performance of AFDEP protocol, we take following metrics/ clustering attributes: Network lifetime. This metric gives the time up to which a network remains alive. It shows number of rounds (including fault tolerance) up to which network remains alive for different number of nodes in network. One round consists of an operation of network from sensing the phenomenon to receiving data at sink node including election process and fault handling if any. Detection Accuracy. It shows how accurately nodes can detect faults. The detection accuracy is defined by the probability of false alarm, which is the probability that an operational CH is mistakenly detected as a faulty one. Detection accuracy performance is measured under different packets loss rates and cluster sizes. Table 2. Experiment Parameter Parameter Area of sensor field Sink position Initial energy per node Path loss exponent

Value 100×100 m2 At origin (0,0) 1J 2 50 nJ/bit 100 pJ/bit/m2 50 nJ/bit Size of data packet 500 bits Size of control packet 20 bits Sensing Interval 0.5 s High transmission range 60 m Low transmission range 20 m No of Nodes 300 Cluster Size 10, 20, 30

5.2 Simulation Results and Discussion To find out more reliable and accurate results, we executed AFDEP protocol with different number of nodes, number of times and failure frequency.

AFDEP: Agreement Based CH Failure Detection and Election Protocol for a WSN

255

Network Lifetime. It can be observed form Fig. 2 that as the number of nodes increases, network lifetime increases. But after certain number of nodes, the network life time starts decreasing due to more overhead of cluster maintenance. When number of nodes are 40, network is alive up to 900 rounds. AFDEP FTEP

1600 1400

Number of Rounds

1200 1000 800 600 400 200 0 0

50

100

150

200

250

300

Number of Nodes

Fig. 2. Network Lifetime

Detection Accuracy. From Fig. 3, we can observe the effects of the packet loss rate on detection accuracy for different cluster size. For simulation, we consider the packet loss rate range from 0.2 to 0.4. It can be observed that with the increase of the packet loss rate the probability of false alarm positive increases, which leads to lower detection accuracy. A larger number of sensor nodes lead to a smaller probability of false alarm positive, i.e., higher detection accuracy. As expected, AFDEP can achieve high detection accuracy. 0.1

Probability of False Alarm Positive

0.01

n=20 n=15 n=10

1E-3

1E-4

1E-5

1E-6

1E-7

1E-8 0.20

0.25

0.30

0.35

0.40

Packet Loss Rate

Fig. 3. Detection Accuracy

5 Conclusion AFDEP is agreement-based fault detection and recovery protocol for faulty CH in WSNs. AFDEP periodically checks for CH failure. This detection process runs parallel

256

A. Kaur and T.P. Sharma

with network operation. It provides high accuracy, because it allows each cluster member to detect its faulty CH independently. It employs a distributed agreement protocol to reach an agreement on the failure of CH among multiple cluster members. In order to recover from faulty CH, back up node is elected as new CH and new back up node is elected locally. Election of CH and back up node is based on residual energy of sensor nodes. Simulation results show, AFDEP achieves high detection accuracy in harsh environment.

References 1. Akyildiz, I.F., et al.: Wireless sensor networks: a survey. Computer Networks 38, 393–422 (2002) 2. Akkaya, K., Younis, M.: A survey of routing protocols in wireless sensor networks. Elsevier Ad Hoc Network 3/3, 325–349 (2005) 3. Heinzelman, W., Chandrakasan, A., Balakrishnan, H.: Energy-efficient communication protocol for wireless sensor networks. In: Proceeding of the Hawaii International Conference System Sciences, Hawaii (2000) 4. Manjeshwar, A., Agrawal, D.P.: TEEN: a protocol for enhanced efficiency in wireless sensor networks. In: Proceedings of the 1st International Workshop on Parallel and Distributed Computing Issues in Wireless Networks and Mobile Computing, San Francisco, CA (2001) 5. Manjeshwar, A., Agrawal, D.P.: APTEEN: a hybrid protocol for efficient routing and comprehensive information retrieval in wireless sensor networks. In: Proceedings of the 2nd International Workshop on Parallel and Distributed Computing Issues in Wireless Networks and Mobile computing, Ft. Lauderdale, FL (2002) 6. Younis, O., Fahmy, S.: HEED: A Hybrid, Energy-Efficient, Distributed Clustering Approach for Ad Hoc Sensor Networks. IEEE Transaction on Mobile Computing 3(4) (2004) 7. Hussain, S., Matin, A.W.: Hierarchical Cluster-based Routing in Wireless Sensor Networks. In: Proceeding of 5th Intl. Conf. on Information Processing in Sensor Network (IPSN 2006), USA (2006) 8. Banerjee, S., Khuller, S.: A clustering scheme for hierarchical control in multi-hop wireless networks. In: Proceeding of IEEE INFOCOM, Anchorage, Alaska, USA, vol. 2, pp. 1028–1037 (2001) 9. Sajjanhar, U., Mitra, P.: Distributive energy efficient adaptive clustering protocol for wireless sensor networks. In: Proceeding of International Conference on Mobile Data Management (MDM 2007), Mannheim, Germany (2007) 10. Moussaoui, O., Naimi, M.: A distributed energy aware routing protocol for wireless sensor networks. In: Proceeding of ACM PE-WASUN 2005, Montreal, Quebec, Canada, pp. 34– 40 (2005) 11. Shah, R.C., Rabaey, J.M.: Energy aware routing for low energy ad-hoc sensor networks. In: Proceeding of IEEE Wireless Communication and Networking Conf (WCNC), Orlando, pp. 1–5 (2002) 12. Jiang, P.: A New Method for Node Fault Detection in Wireless Sensor Networks 12821294 (2009), http://www.mdpi.com/journal/sensors 13. Souza, L.M., Vogt, H., Beigl, M.: A survey on Fault Tolerance in Wireless Sensor Networks, http://www.cobis-online.de/

AFDEP: Agreement Based CH Failure Detection and Election Protocol for a WSN

257

14. Yu, M., Mokhtar, H., Merabti, M.: A Survey on Fault Management in Wireless Sensor Networks. Computer Networks (2007) 15. Lee, M.H., Choi, Y.H.: Fault detection of wireless sensor networks. Elsevier Computer Communications 31, 3469–3475 (2008) 16. Jiang, P.: A New Method for Node Fault Detection in Wireless Sensor Networks, pp. 1282–1294 (Feburary 2009), http://www.mdpi.com/journal/sensors 17. Wang, P., Zheng, J., Li, C.: An Agreement-Based Fault Detection Mechanism for Under Water Sensor Networks. In: Proceeding Global Telecommunications Conference, GLOBECOM 2007. IEEE, Los Alamitos (2007) 18. Bansal, N., Sharma, T.P., Misra, M., Joshi, R.C.: FTEP: A Fault Tolerant Election Protocol for Multi-level Clustering in Homogeneous Wireless Sensor Networks. In: Proceeding 16th IEEE International Conference on Networks, ICON 2008 (2008) 19. Asim, M., Mokhtar, H., Merabti, M.: A cellular approach to fault detection and recovery in wireless sensor networks. In: Third International Conference on Sensor Technologies and Applications, SENSORCOMM 2009, 18-23, pp. 352–357 (2009) 20. Venkataraman, G., Emmanuel, S., Thambipillai, S.: Energy-efficient cluster-based scheme for failure management in sensor networks. IET Commun. 2(4), 528–537 (2008) 21. Venkataraman, G., Emmanuel, S., Thambipillai, S.: A Cluster-Based Approach to Fault Detection and Recovery in WSNs. In: IEEE ISWCS 2007 (2007) 22. Khadivi, A., Shiva, M.: FTPASC: A Fault Tolerant Power Aware Protocol with Static Clustering for Wireless Sensor Networks (2006) 23. Fan, K.W., Liu, S., Sinha, P.: On the Potential of Structure-free Data Aggregation in Sensor Networks. In: Proceedings IEEE Infocom (2006) 24. http://www.omnetpp.org/

Problem Area Identification with Secure Data Aggregation in Wireless Sensor Networks Paresh Solanki, Gaurang Raval, and Srikant Pradhan Institute of Technology, Nirma University, Ahmedabad, Gujarat, India {08mce018,gaurang.raval,snpradhan}@nirmauni.ac.in http://nirmauni.ac.in/it/

Abstract. The primary use of wireless sensor networks (WSNs) is to collect and process data. Most of the energy consumption is due to data transmission. Because of the unique properties of WSNs all raw data samples are not directly sent to the sink node instead data aggregation is preferred. Since sensor nodes are deployed in an open environment such as a battlefield or similar applications, data confidentiality and integrity are vital issues in such conditions, hence secure aggregation is required. End to end secure aggregation is less demanding compared to hop by hop secure aggregation so former is superior. When aggregation is performed on data, crucial information is lost which may be indicating alarming situation. This paper presents an idea to reduce the amount of information transmitted with retention of critical data so that the problem area could be identified. Privacy Homomorphism(PH) preserves the data characteristics even in the encrypted form. This paper is based on the PH technique which provides secure data aggregation without significant loss of individuality of data. Keywords: Data aggregation, convergecast, security, wireless sensor networks.

1 Introduction In wireless sensor networks, sensor nodes collect data from hostile environment and send it to sink node where it is processed, analyzed and used by the application. In these resource constrained networks, the general approach is to send the data jointly which is generated by different sensor nodes. While being forwarded towards the base station such in-network processing of data is known as data aggregation. When base station queries to the network, all nodes do not send their data to sink node directly but aggregator node collects data and responds to sink node. Data aggregation reduces the number of data transmissions thereby improving the bandwidth and energy utilization in the network but results in loss of individuality of reading which could be of important use. Because of peculiar characteristics of sensor network, security on data aggregation [2] is most crucial. There is a strong conflict between security and data aggregation protocols. Security protocols require sensor nodes to encrypt and authenticate any sensed data prior to its transmission and prefer data to be decrypted by the base station. On the other hand, data aggregation protocols prefer plain data to V. V Das, R. Vijaykumar et al. (Eds.): ICT 2010, CCIS 101, pp. 258–266, 2010. © Springer-Verlag Berlin Heidelberg 2010

Problem Area Identification with Secure Data Aggregation in WSNs

259

implement data aggregation at every intermediate node so that energy efficiency is maximized. Moreover, a data aggregation result in alterations in sensor data and therefore it is a challenging task to provide source and data authentication along with data aggregation. Due to these conflicting goals, data aggregation and security protocols must be designed together so that data aggregation can be performed without sacrificing security and individuality of data. This paper is based on secure data aggregation using cluster based approach for problem area identification. In the implementation Jist/SWANS simulator [10][11] was used. The basic code of heartbeat application was modified to implement the clustering strategy with dynamic selection of clusters. The energy model [13] was hooked in to the simulator. Separate application was developed for both plain aggregation and secure aggregation with PH method integration. Specific reason of selecting PH method is its ability to preserve individuality of data after encryption. It was assumed that nodes are aware of their location. Jist/SWANS significantly outperform ns2 and GloMoSim, both in time and space [12].

2 Related Works In wireless sensor network, there are so many challenges like how to improve lifetime of network, how to provide robustness to network and security issues. WSNs collect the data from sensor nodes, process it and send it to the base station. 70% [2] of energy consumption is due to data transmission. It is widely accepted that the energy consumed in one bit of data transfer can be used to perform a large number of arithmetic operations in the sensor processor. Moreover in a densely deployed sensor network the physical environment would produce very similar data in close-by sensor nodes and transmitting such data is more or less redundant. Therefore, all these facts trigger the concept of grouping of nodes such that data from a group can be combined together in an intelligent way and transmit only compact data. This process of grouping of sensor nodes in a densely deployed large-scale sensor network is known as clustering. One major goal of clustering is to allow in network pre-processing, assuming that cluster heads collect multiple data packets and relay only one aggregated packet [1]. To reduce the latency present in the tree-based aggregation, recent trend is to group sensor nodes into clusters so that data is aggregated with improved efficiency and low latency. Attackers may capture secret data as sensor network deployments are vulnerable, so secure aggregation is required. By using traditional symmetric key cryptography algorithms, it is not possible to achieve end-to-end confidentiality and in-network data aggregation together. If the application of symmetric key based cryptography algorithms is combined with data aggregation, then the messages must be encrypted hop-by-hop. Clearly, this is not an energy efficient way of performing secure data aggregation and it may result in considerable delay. Secondly due to resource constraints of sensor nodes, symmetric key cryptography is preferable over asymmetric key cryptography [4]. In addition, this process requires neighboring data aggregators to share secret keys for decryption and encryption. Hop by hop secure data aggregation is highly resource consuming because data aggregator nodes first decrypt the data then aggregate it and again encrypt it. So end to end secure encrypted

260

P. Solanki, G. Raval, and S. Pradhan

data aggregation is preferred because aggregation process is done on encrypted data[3][5]. In order to achieve end-to-end data confidentiality and data aggregation together without requiring secret key sharing among data aggregators, PH based cryptography becomes obvious choice. Sensor nodes share a common symmetric key with the base station that is kept hidden from intermediate aggregators. Currently different schemes are available for end to end secure encrypted data aggregation but needs some more attention. Few enhancements cannot be ruled out in existing secure data aggregation methods to solve the issue of problem area identification.

3 Secure Data Aggregation Aggregated WSNs provide better power conservation and efficient use of communication channels but also introduce additional security concerns. Most existing schemes for data aggregation are subject to attack. Because of this, the need for secure data aggregation is raised and its importance needs to be highlighted [6]. Hop-by-hop secure data aggregation increases the computational demand at the inner nodes (aggregator) a lot though they are the most important ones and should save on energy as much as possible. Thus it would be desirable to process data without having to decrypt it while preserving the content. The aggregating node does not necessarily need to interpret the data; it only has to be able to work with it. A concept which meets the above requirements is called Privacy Homomorphism and has been introduced by Rivest, Adelman and Dertouzos. This PH method is used in end-to-end secure data aggregation. 3.1 Privacy Homomorphism PH is an encryption function which allows operations like additions or multiplications on the encrypted data. The result will yield an encrypted codeword which is similar to the codeword that would be obtained by applying the operation on the cleartext first and encrypting the result afterwards. Additions or multiplications are of particular interest in this context. An instance of such method was suggested by Domingo-Ferrer in a provably secure additive and multiplicative privacy homomorphism. PH is an encryption transformation that allows direct computation on encrypted data [7][8][9]. It is a symmetric encryption scheme which uses the same key for encryption and decryption. Let Q and R denote two rings, + denote addition and x denote multiplication on both. Let K be the keyspace. Following is the encryption transformation E: K x Q -> R and the corresponding decryption transformation is D : K x R -> Q. Given a, b є Q and k є K we term additively homomorphic and multiplicatively homomorphic. a + b = Dk (Ek (a) + Ek (b))

(1)

a * b = Dk (Ek (a) * Ek (b))

(2)

RSA is a multiplicative PH, while Domingo-Ferrer presented an additive and multiplicative PH which is a symmetric scheme and secures against chosen cipher text attacks. Asymmetric PH is not acceptable in the context of WSNs due to execution times.

Problem Area Identification with Secure Data Aggregation in WSNs

261

4 Implementation of Secure Data Aggregation The algorithm requires the same secret key for encryption and decryption. The aggregation is performed with a key that can be publicly known, i.e., the aggregator nodes do not need to decrypt the encrypted messages. However, it is required that the same secret key is applied on every node in the network that needs to encrypt data. For very secure parameter combinations (d > 100), the messages become very big. However, with reasonable parameters it also fits the needs of constrained devices. Parameter settings, encryption and decryption algorithm process is shown below: Parameter Settings: 1. A large integer p which holds the following two properties: (a) It should consist of a large number of divisors. p simply is a product of integers with repeatedly multiplying prime numbers. (b) The large p should be chosen such that many integers c can be found for which an cinv exists so that c x cinv mod p = 1 where c1, the energy of the final reconstructed image is closer to the energy of the original. This improves the reconstructed image’s PSNR, but at the expense of more non-zero coefficients. We consider the number of non-zero coefficients to be a measure of efficiency because zero coefficients are very efficient to code, and so bit rate tends to be proportional to the number of non-zero coefficients. For implementation of the method we used k=1.6 for better results. Initially we consider a large threshold =64 and decreasing it at each iteration until our target threshold, 32 is reached. A large starting threshold causes many coefficients to be eliminated initially. With the right balance between the energy gain k and the amount by which the threshold decreases each iteration, most insignificant coefficients remain significant, while the signal’s energy is maintained and image reconstruction improves. The control system, where the initial threshold 64 is reduced by 1 each iteration until =32, performs best. Comparing the efficiency–distortion data for this system with the data corresponding to a threshold reduction of 4/ iteration, the system with the slower threshold reduction rate produces a DDWT signal with fewer non-zero coefficients but the same reconstruction image quality. For the experiments test images are decomposed up to 2 levels. With respect to DDWT, there is no strong constraint on the filters for the first level decomposition. We choose the CDF 9/7 filter bank since this filter bank shows excellent performance, and is widely used in image coding. The 6-tap Qshift filters [6] are used for the remaining stages of DDWT. Table 1 shows the results for CT_brain image (256 256). Initially PSNR is calculated for DWT and DDWT reconstructed image. Then IP-NS algorithm is applied to the DWT and DDWT and PSNR is calculated for the same. Finally the result for DWT, DWT with IP-NS, DDWT, DDWT with IP-NS is displayed. The results show that the reconstruction quality of the DDWT with IP-NS is better than DDWT without noise shaping, DWT and DWT with IP-NS since the PSNR value for DDWT with noise shaping is greater than other methods. Visual quality for CT_brain is shown in Fig 2.

528

S.N. Talbar and A.K. Deshmane Table 1. PSNR results for DDWT, DDWT with IP_NS, DWT and DWT with IP_NS

64

33.41

PSNR for DDWT with IPNS 37.90

56

37.48

43.57

42.02

52

38.10

44.79

43.11

44

38.69

43.27

42.89

40

38.91

44.52

43.96

36 32

39.22 39.51

44.84 44.17

43.04 40.12

Thres hold

Fig. 2. Visual Quality for CT brain image (Threshold=32)

PSNR for DDWT

PSNR for DWT with IP-NS 36.32

5 Conclusion The threshold reduction rate affects the performance results. With decreasing threshold at each iteration PSNR value increases that means with decreasing threshold the quality of image increases. And we get best results at 32 . The choice of gain factor k affects the noise components in the system. It is always better to choose k greater than one. When, 1 the energy of the final reconstructed image is closer to the energy of the original. This improves the reconstructed image PSNR, but at the expense of more non-zero coefficients and we obtain best results at k=1.6.

References 1. Reeves, T.H., Kingsbury, N.G.: Overcomplete image coding using iterative projectionbased noise shaping. In: Proc. Int. Conf. Image Pro., NY, pp. 597–600 (2002) 2. Yang, J., Wang, Y., Xu, W., Dai, Q.: Image coding using Dual Tree Discrete Wavelet Transform. IEEE Tras.on Image Processing 17(9) (September 2008) 3. Nguyen, T.T., Oraintara, S.: Multiresolution direction filter banks: Theory, design and application. IEEE Trans. Signal Process. 53(10), 3895–3905 (2005) 4. Kingsbury, N.G., Reeves, T.H.: Redundant representation with complex wavelets: How to achieve sparsity. In: Proc. Int. Conf. Image Process, Barcelona (September 2003) 5. Selesnick, I.W., Baraniuk, R.G., Kingsbury, N.G.: The dual-tree complex wavelet transform. IEEE Signal Process. Mag. 22, 123–151 (2005)

Privacy-Preserving Naïve Bayes Classification Using Trusted Third Party and Offset Computation over Distributed Databases B.N. Keshavamurthy, Mitesh Sharma, and Durga Toshniwal Department of Electronics & Computer Engineering, Indian Institute of Technology, Roorkee, Uttarakhand, India [email protected], [email protected], [email protected]

Abstract. Privacy-preservation in distributed databases is an important area of research in recent years. In a typical scenario, multiple parties may wish to collaborate to extract interesting global information such as class labels without revealing their respective data to each other. This may be particularly useful in applications such as car selling units, medical research etc. In the proposed work, we aim to develop a global classification model based on the Naïve Bayes classification scheme. The Naïve Bayes classification has been used because of its simplicity and high efficiency. For privacy-preservation of the data, the concept of trusted third party with two offsets has been used. The data is first anonymized at local party end and then the aggregation and global classification is done at the trusted third party. The proposed algorithms address various types of fragmentation schemes such as horizontal, vertical and arbitrary distribution. Keywords: privacy-preservation, distributed database, partitions.

1 Introduction Due to the advancement of computing and storage technology in recent times, digital data can now be easily collected. It is very difficult to analyze the entire data manually. Thus a lot of work is going on for mining and analyzing such data. In many real world applications such as hospitals, direct marketing, design-firms and university databases, data is distributed across different sources. The distributed database consists of horizontal, vertical or arbitrary fragments. In case of horizontal fragmentation, each site has complete information on a distinct set of entities. An integrated dataset consists of the union of these datasets. In case of vertical fragments each site has partial information on the same set of entities. An integrated dataset would be produced by joining the data from the different sites. Arbitrary fragmentation is a hybrid of previous two. Normally, the data sources of distributed database scenarios are willing to go for the global instance of their integrated data for their mutual benefits. Thus preserving privacy of personal sensitive data is a non-trivial issue. V. V Das, R. Vijaykumar et al. (Eds.): ICT 2010, CCIS 101, pp. 529–534, 2010. © Springer-Verlag Berlin Heidelberg 2010

530

B.N. Keshavamurthy, M. Sharma, and D. Toshniwal

Initially, for privacy-preserving data mining, randomization methods were used [2] [3]. In [4] [5] [6], the usage of the approach for classification was discussed. A number of other techniques [6] [7] have also been proposed for privacy preservation which work on different classifiers. The work in [8] [9] describes the methods for improving the effectiveness of classification. In [9] it proposes a method which eliminates the privacy breach and increase utility of the released database. In case of distributed environment, the most widely used technique in privacy preservation mining is secure sum computation [10]. Naive Bayes approach for classification is described in [1]. A few of research papers have discussed the privacy preserving mining across distributed databases. The key research gap with the existing methods of computation is that the global pattern computation is done at one of the data source itself. Our paper addresses this issue effectively by using a trusted third party who is responsible for computing result for aggregate classification data of the collaborating parties. Secondly all the class types need not be present at every collaborating party. This issue is addressed by providing two offsets, one for the newly initiating class instances at each collaborating par and other for the class instances which are initiated by predecessor. The rest of the paper is organized as follows: Section 2 briefs about related research work followed by proposed work, privacy-preserving Naïve Bayes classification using trusted third party computation with two offset over distributed databases. Section 3 gives experimental results. Section 4 includes conclusion.

2 Proposed Work The Naïve Bayes algorithm for different distribution scenarios such as horizontal, vertical and arbitrary is proposed for distributed databases. 2.1 Horizontal Partitioning Scenario Here each party locally computes the local count of instances. The global count is given by the sum of the local counts. To count global values by summing all local values we use modified secure sum algorithm which sends the sum to the trusted third party. The following steps are followed in computing global class instances for the collaborating parties using trusted third party computation: 2.1.1 At Trusted Third Party Following steps were followed: 1. The collaborating parties interested in global pattern computation send their request to trusted third party. 2. Trusted third party sends offset and address to collaborating parties to perturb his class instances and send perturbed class instances to logical adjacent one. 3. Finally after receiving integrated class instances from the last collaborating party, trusted third party will compute the integrated class instances by subtracting his random values form the class instance values obtained from last party and will get actual class instances and same will be conveyed to collaborating parties

Privacy-Preserving Naïve Bayes Classification Using Trusted Third Party

531

2.1.2 At Collaborating Parties The following steps were followed: 1. Third party will compute the integrated class instances by subtracting his random values form the class instance values obtained from last party and will get actual class instances and same will be conveyed to collaborating parties. 2. The randomly selected initial party receives one offset called newoffset because instance of class items starts from here. He sends perturbed values its logical neighbor 3. For all data sources except the initial collaborator. If There is no new instance of classes in compare with old list of class instances send by previous collaborator then collaborator needs only one offset oldoffset to perturb his class instances, then add his perturbed values to the values received from his predecessor and then send the resulting instance of class to to logical successor else collaborator needs oldoffset for class instances already initiated by its predecessor and newoffset for class instances initiates form the here and then send the resulting instance of class to logical successor. 4. Finally, the logical last collaborating party after performing step c operation it will sends the class instances to the trusted third party. The necessary algorithms are given follows: Here each party locally computes the local count of instances. The global count is given by the sum of the local counts. To count global values by summing all local values we use modified secure sum algorithm with trusted third party.

Algorithm Assumptions: n parties, r class values, z attribute values, jth attribute contain lj different values ,S– Supervisor/Trusted Third Party and P – Parties/collaborating parties, C il.r = no of instances with party, Pi having classes r and attribute values l and N ir = no of instances with party Pi having classes r. At P:

For all class values y do For all z, Party Pi locally computes ciyz Party Pi locally computes niy EndFor Encrypt values and send to third party

At S: Receive values and decrypt them;

From all parties, we can get all possible C1l.r and N1r Parties calculate the required probabilities from global C1l.r and N1r , on that basis will predict the class. Fig. 1. Algorithm for Horizontal Scenario

532

B.N. Keshavamurthy, M. Sharma, and D. Toshniwal

2.2 Vertical Partitioning Scenario Each party computes nac and nc, and send them to the trusted third party. Here we have been making use of encryption and decryption techniques correspondingly at collaborating and trusted third party to preserve the privacy of import sensitive information. The following steps are used to compute the aggregate values of the collaborating parties: 2.2.1 At Trusted Third Party 1. Trusted third party will decrypt all the values and integrate them to get global instances of all required values, and used to predict classes. 2. Send the integrated values to all the collaborating parties. 2.2.2 At Collaborating Parties 1. The collaborating parties interested in global pattern computation send their encrypted values to the third party. 2. After receiving the integrated values from trusted third party, the actual values will be interpreted using private key. The corresponding algorithm is given in Fig. 2 is as follows: In nominal attributes, each party calculates their local instances of nac and na, of the attribute values they have. As each party have different attributes, so no parties have same value of instance of attribute and class. Hence there is no need to calculate the sum of values. At a particulate timestamp, we calculate the local values of nac and nc, and send them to the trusted third party. Algorithm Assumptions: n parties, r class values, z attribute values, jth attribute contain lj different values ,S– Supervisor/Trusted Third Party and P – Parties/collaborating parties, C il.r = no of

instances with party, Pi having classes r and attribute values l and N ir = no of instances with party Pi havig classes r. At P:

For all class values y do For all z, Party Pi locally computes ciyz Party Pi locally computes niy EndFor Encrypt values and send to third party

At S: Receive values and decrypt them;

From all parties, we can get all possible C1l.r and N1r Parties calculate the required probabilities from global C1l.r and N1r , on that basis will predict the class.

Fig. 2. Algorithm for Vertical Fragmentation

Privacy-Preserving Naïve Bayes Classification Using Trusted Third Party

533

2.3 Arbitrary Partitioning Scenario

It is an inter mix of the vertical and the horizontal partition. In arbitrary fragmentation over distribute databases; the trusted third party well in advance knows the collaborating parties who are belong to horizontal and vertical partition. To compute the global class instances the following steps were used: 1. If the collaborator is horizontally distributed, follow the procedure mentioned in section 2.1, to collect, integrate, compute global instances of class values and disseminate the global class instances for to the different collaborating parties. 2. If the collaborator is vertically fragmented, follow the procedure mentioned in section 2.2 to collect, integrate, compute the global integrated values from different collaborating parties and finally at collaborating parties the actual values of the class instances are found using the private key.

3 Results The dataset used for the purpose of experimentation is car-evaluation [11]. The algorithm was applied on non-distributed database and the percentage accuracy was obtained. The percentage accuracy for the distributed scenario should intuitively be less than or equal to non-distributed scenario. The accuracy is coming out to be same as that to non-distributed scenario which is best case result. The analysis results of different partitions of distributed databases are as follows in table 1: Table 1. Classification of distributed and non-distributed databases

S. No.

Description

1

Classification of data is at single party (No distribution)

Number of Parties

Total Number of Records

% Accuracy

1

1728 Records Party1: 500 Records Party2: 500 Records Party3: 728 Records

85

Classification of data distributed in horizontal scenario

2

Classification of data distributed in vertical scenario

3

3

4

Classification of data distributed in arbitrary scenario

2

3

Party1: 2 Attributes Party2: 2 Attributes Party3 : 3 Attributes Party1: 500 Records Party2: 4 Attributes Party3: 3 Attributes

85

85

85

4 Conclusion In our proposed work, we have proposed a set of algorithms for classifying data using Naïve Bayes from a group of collaborative parties without breaching privacy. The non-distribution and various distribution scenarios such as horizontal, vertical and arbitrary scenarios are compared and their accuracy is calculated on the data. The accuracy comes out to be the same showing that the algorithm is giving best case

534

B.N. Keshavamurthy, M. Sharma, and D. Toshniwal

results. Privacy is also preserved using privacy preservation techniques such as offset computation and encryption. The third party concept is also introduced to calculate global classification results with privacy preservation. In this case, Naïve Bayes algorithm is applied to static data but the algorithm can also be extended for dynamic databases in future work. Also the algorithm can be modified for numeric data to widen its scope.

References 1. Vaidya, J., Kantarcıoğlu, M., Clifton, C.: Privacy-Preserving Naïve Bayes Cassification. International Journal on Very Large Data Bases 17(4), 879–898 (2008) 2. Leew, C.K., Choi, U.J., Liew, C.J.: A data distortion by probability distribution. ACM TODS, 395–411 (1985) 3. Warner, S.L.: Randomized Response: A survey technique for eliminating evasive answer bias. Journal of American Statistical Association, 63–69 (1965) 4. Agarwal, R., Srikanth, R.: Privacy–preserving data mining. In: Proceedings of the ACM SIGMOD conference (2005) 5. Agarwal, D., Agarwal, C.C.: On the design and Quantification of Privacy–Preserving Data Mining Algorithms. In: ACM PODS Conference, pp. 1224–1236 (2002) 6. Zhang, P., Tong, Y., Tang, D.: Privacy–Preserving Naïve Bayes Classifier. In: Li, X., Wang, S., Dong, Z.Y. (eds.) ADMA 2005. LNCS (LNAI), vol. 3584, pp. 744–752. Springer, Heidelberg (2005) 7. Zhu, Y., Liu, L.: Optimal Randomization for Privacy–Preserving Data Mining. In: KDD ACM KDD Conference (2004) 8. Gambs, S., Kegl, B., Aimeur, E.: Privacy –Preserving Boosting, Journal (to appear) 9. Poovammal, E., Poonavaikko: An Improved Method for Privacy Preserving Data Mining. In: IEEE IACC Conference, Patiala, India, pp. 1453–1458 (2009) 10. Yao, A.C.: Protocol for secure sum computations. In: Proc. IEEE Foundations of Computer Science, pp. 160–164 (1982) 11. Bohanec, M., Zupan, B.: UCI Machine Learning Repository (1997), http://archive.ics.uci.edu/ml/datasets/Car+Evaluation

Extraction of Pose Invariant Facial Features Singh R. Kavita1, Zaveri A. Mukeshl2, and Raghuwanshi M. Mukesh3 1

Department of Computer Technology, YCCE, Nagpur, 441 110, India Computer Engineering Department, S.V.National Institute of Technology, Surat, 329507, India 3 NYSS College of Engineering and Research, Nagpur, 441 110, India [email protected], [email protected], [email protected] 2

Abstract. In this paper, we describe a method for extraction of facial features of 2D still faces with variations in view in a certain viewing angle range. The images we have considered vary beyond left 30 degrees to right 30 degree out of plane rotation. Our technique applies skin separation and corner detection for extraction of features of faces in different poses. Just detecting the location of two facial points namely the corner of eyes and location of nose tip, the other features will be derived from them automatically; thus saving the time during the feature extraction. Keywords: feature extraction, skin segmentation, corner detection.

1 Introduction Pattern recognition finds its application in many face processing techniques such as face recognition. In general, any face recognition system consists of three different modules; viz: face detection, feature extraction and recognition [1]. Face recognition on a broader spectrum uses mainly the following techniques: facial geometry, based on geometrical characteristics of the face; skin pattern, based on visual skin pattern and lastly facial thermogram, based on an infrared signals to map the face. Recognizing someone based on facial geometry makes human recognition a more automated process. Based on this fact, feature extraction plays a significant role. So, it becomes important to extract the prominent features in order to perceive human faces. In the last so many years, the approaches proposed for feature extraction can be broadly categorized as (i) global approach [2],[3],[4]; (ii) template based approach [5],[6]; (iii) local feature based approach [7],[8] and (iv) geometry based approach[9],[10],[11]. In addition, many approaches [12], [13] used eyes as the only facial features for the initialization of faces for any face recognition technique (FRT). All the approaches cited above suffers from a major drawback and that is, their decreasing accuracy as the subject face has pose angles more than 60 degrees or so. However, in general for faces which are rotated out of plane where both eyes are not visible, techniques may require more features than just the eyes. V. V Das, R. Vijaykumar et al. (Eds.): ICT 2010, CCIS 101, pp. 535–539, 2010. © Springer-Verlag Berlin Heidelberg 2010

536

S.R. Kavita, Z.A. Mukeshl, and R.M. Mukesh

Our proposed technique takes an account of the geometry based features which makes use of skin separation and corner detection for extraction of feature points on faces in different poses. Just detecting the location of two facial points namely the corner of eyes and location of nose tip, the other features will be derived from them automatically; thus saving the time during the feature extraction. The paper is organized as follows: in Sect. 2, discuss about the proposed methodology for feature extraction. Experimental results are presented in Sect. 3. Finally, a conclusion is given.

2 Our Approach for Feature Points Extraction Faces which are out of plane something other than eye balls is required to localize the face before extraction of features. Base on this fact, we have considered nose as the localizing component on face. Moreover, the nose profile can be characterized as the set of points with the highest symmetry and high luminance values; therefore we can identify the nose tip as the point that lies on the nose profile, above the nose baseline, and that corresponds to the brightest gray level. These considerations allow localizing the nose tip robustly shown in Fig 1.

Fig. 1. Localized Nose tip

As we know the geometry-based technique works efficiently on the frontal faces, however; we take an opportunity to extend this approach for feature extraction for faces under variations in pose. The overall methodology for feature extraction is described in Fig 2. Selecting nose tip

the

Image Normalization

Skin segmentation

Input Image Corner Detection

Feature Vector

Fig. 2. Our approach to feature extraction

In order to detect facial feature candidates properly, the unnecessary information in a face image must be removed in advance. So in our approach, the first stage is

Extraction of Pose Invariant Facial Features

537

cropping the face area as soon as the image is read from the database and the remaining part of preprocessing is performed on that cropped images. In this respect, when an image is selected we first localize the nose tip as mentioned earlier. Based on the threshold value computed from the located nose tip, a frame is designed for cropping the faces. Accordingly we design the three different efficient frames for the frontal, left oriented and right oriented images. Cropped faces from the images are shown in Fig 3.

Fig. 3. Sample images from database top row (cropped frontal faces), middle row (cropped left oriented faces), and bottom row (cropped right oriented faces).

To the cropped images we further employ segmented skin region candidates that satisfy the homogeneity property of the human skin. If corner detection is directly applied to cropped images the fiducially points on the face are not predominant. Once the position of the extracted points as in red in Fig 4.b is located, features are inferred on the basis of simple geometrical consideration. The features that we have considered are eye_lip_distance, eye_nose_distance, nose_lip_distance, angle_eye_lip, angle_eye_nose and angle_nose_lip.

(a)

(b)

Fig. 4. (a) Skin segmented faces, (b) Corner detected faces

3 Experimentation and Results In our experiment, we have used subset of Indian database [14]. This database is a rich set of 611 colored images of size 567 X 334 with different expressions, varying illumination and varying orientation. We have worked with 183 images of 61 different subjects (22 female and 39 male). For each subject we have selected, randomly, three images irrespective of illumination or expression consideration. For feature extraction we have taken into consideration three images per subject (one frontal, one right oriented and one left oriented). Our algorithm is not restricted for rotation range to be in between +30 and –30 degrees. We have considered even those faces where only single eye is visible. The features are extracted from cropped images of dimension 301 × 351 which contains only the faces, using the procedure explained in Fig 2.

538

S.R. Kavita, Z.A. Mukeshl, and R.M. Mukesh

Once the features have been extracted with our proposed method and PCA, we have analyzed that the features extracted from PCA varies from 0 to some greater range. Moreover, features are more consistent in case of our approach between specific ranges, except for few images. It implies that features we have considered are almost consistent for variations in poses of faces. For the faces from each orientation, the correct features were extracted in 98 % of the faces. However, for 2% of faces features were not extracted correctly due effect of illumination. For testing, we have selected any image randomly from the test database. While deciding that test image should go to which cluster, a simple tactic is used and that is finding nose tip as the pixel under a threshold in the test image so that it can be said to be at right or left of the origin. For the selected pixel within the threshold limit, the image has been considered to have a frontal face. In this case, the pixel corresponding to tip of the nose has been considered as it gave clear idea of the face angle ( since the only information required for clustering is whether the pixel is located left or right of the origin, if none, it is frontal face.) Mathematically, the above explained method can be represented as, , 40 ,,

,

40 40

40

where, It =test image under consideration, (µ= threshold value, selected X-coordinate320). Now, when the system knows the clusters and has classified the test image as native of one of those, the subsequent tasks are to select and match feature vectors and produce the correct match accurately. For this, any of the classical methods can be used but here we have concentrated on Euclidean distance classifier as an ease of implementation without any breach in accuracy. Hence the feature vectors contain only those points which are common to all the clusters. The results of the experiments show quite a fascinating success rate. The extracted features have been used to classify 183 images with variation in pose and we achieved an 85% success rate on an average with equation. We have also evaluated the performance of feature extraction algorithm against the established PCA algorithm as shown in Table 1. Table 1. Comparative result of two approaches

PCA

Number of images 183

Error rate 0.37

Success percentage 63

Our

183

0.15

85

Database

Approach

The overall results however are very encouraging and hold good prospects of further research. The other factor which decides the efficiency of the algorithm is time taken to find a match for the subject image. Although the results below indicate time

Extraction of Pose Invariant Facial Features

539

efficiency for only near hundred images which gave correct matches and about four dozen images which produced wrong match or no match at all, they will vary only slightly for an increase in database.

4 Conclusion and Future Work Although the results presented in this paper represent an initial analysis of the pose invariant geometry based approach, we are extremely encouraged that our approach is competitive with current algorithms. The experiment result shows that the system can work in the different orientation and expression.

References [1] Zhao, W., Chellappa, R., Phillips, P., Rosenfeld, A.: Face recognition: A literature survey. ACM, Computing Surveys 35(4), 399–458 (2003) [2] Turk, M., Pentland, A.: Eigenfaces for Recognition. Cognitive Neuroscience 3(1), 71–96 (1991) [3] Zhao, W., Krishnaswamy, A., Chellappa, R., Swets, D.L., Weng, J.: Discriminant analysis of principal components for face recognition. In: face recognition:From Theory to applications, pp. 73–85. Springer, Heidelberg (1998) [4] Belhumeur, P.N., Hespanha, J.P., Kriegman, D.J.: Eigenfaces vs. Fisherfaces: recognition using class specific linear projection. IEEE Trans. Pattern Anal. Mach. Intell. 19(7), 711– 720 (1997) [5] Yuille, A., Cohen, D., Hallinan, P.: Facial feature extraction from faces using deformable templates. In: Proc. IEEE Computer Soc. Conf. on Computer Vision and Pattern Recognition, pp. 104–109 (1989) [6] Feris, R.S., De Campos, T.E., Cesar Jr., R.M.: Optimization Techniques 1973. LNCS (LNAI), vol. 4, pp. 127–135 (2000) [7] Manjunath, B.S., Chellappa, R., Von Der Malsburg, C.: A Feature Based approach to Face Recognition. In: Proc. of IEEE Computer society Conference on Computer Vision and Pattern Recognition, pp. 373–378 (1992) [8] Okada, K., Steffens, J., Maurer, T., Hong, H., Elagin, E., Neven, H., von der Malsburg, C.: The Bochum/U. SC Face Recognition System and How it Fared in the FERET phase III Test In face Recognition:from Theory to applications. Springer, Heidelberg (1998) [9] Kanade, T.: Computer Recognition of Human faces. Basel and Stuttgart, Birkhauser 19(7), 721–732 (1997) [10] Gu, H., Su, G., Du, C.: Feature Points Extraction from Faces. Image and Vision Computing NZ, 154–158 (2003) [11] IoanNou, S., Caridakis, G., Karpouzis, K., Kollias, S.: Robust Feature Detection for Facial Expression Recognition. EURASIP Journal on Image and Video Processing, vol. 2007, Article ID 29081 (2007) [12] Kapoor, A., Picard, R.W.: Real-Time, Fully Automatic Upper Facial Feature Tracking. In: Proceedings from 5th International Conference on Automatic Face and Gesture Recognition, pp. 10–15 (2002) [13] Gourier, N., James, D.H., Crowley, L.: Facial Features Detection Robust to Pose, Illumination and Identity. In: SMC 2004, pp. 617–622 (2004) [14] http://viswww.cs.umass.edu/~vidit/IndianFaceDatabase/

On the Segmentation of Multiple Touched Cursive Characters: A Heuristic Approach Tanzila Saba1, Ghazali Sulong2, Shafry Rahim2, and Amjad Rehman1 1

Department of Computer Science B.Z.University Multan Pakistan 2 Graphics and Multimedia Department FSKSM UTM

Abstract. Heuristics are based on the experiences and solves problems approximately that cannot be solved exactly. In handwritten documents recognition, the most difficult phase is touched character segmentation as incorrectly segmented characters cannot be recognized correctly. Accordingly, this paper presents a heuristic approach for multiple touched cursive characters. Initially, a possible segmentation zone is detected using peak to valley function. Next, greedy best search algorithm is implemented in the possible segmentation zone for touched characters segmentation. Experimental results on a test set extracted from the IAM benchmark database exhibit high segmentation accuracy up to 91.63%. Moreover, proposed approach is very fast and can handle multiple cursive touching characters. Keywords: OCR, touched characters segmentation, heuristics, peak to valley function, greedy best search.

1 Introduction Inventions of modern technologies have brought significant changes in cursive handwriting. Touched cursive character segmentation is a current issue for optical character recognition (OCR) systems and is the main cause of low accuracy. Although literature is replete with many character segmentation techniques, their feasibility is for machine printing, hand printing and well written cursive scripts only. All these techniques failed for touched cursive character segmentation [1]. In this regard detailed review can be viewed in [2].

2 Touched Character Segmentation: Proposed Approach This section elaborates proposed strategy for multiple touched cursive characters segmentation using heuristics. The methodology comprises of preprocessing, possible segmentation zone detection and heuristic segmentation of cursive touched characters. Multiple touched cursive characters are extracted from IAM forms scanned in grayscale format at 300 dpi [3]. However, more than three touched cursive characters are unavailable. Therefore, for the sake of experiments, two and three touched cursive characters are considered. Prior to character segmentation, V. V Das, R. Vijaykumar et al. (Eds.): ICT 2010, CCIS 101, pp. 540–542, 2010. © Springer-Verlag Berlin Heidelberg 2010

On the Segmentation of Multiple Touched Cursive Characters: A Heuristic Approach

541

Fig. 1. Preprocessing results: Original sample, Thresholding, Core-region

digital images are binarized using Otsu method. Furthermore, the core-region is detected to avoid ascenders and descenders of the overlapped characters [4]. Fig 1 exhibits preprocessing results. Possible segmentation zone (PSZ) is defined as an area that most probably occupies the segmentation path between the touching boundaries of the characters. In this research, possible segmentation zone is detected using peak to valley function (see Fig. 2) V (l p ) − 2 *V ( x) + V (rp ) (1) Vp x = V ( x) + 1

In equation (1), V p x is the vertical projection function at

location.

l p is peak

position on the left side of x, rp is the peak position on the right side of x . The zone between l p and rp is the possible segmentation zone. Here, w is the width of image and x is determined heuristically as below. ⎧ w / 2, for two touched characters x=⎨ ⎩ w / 3, for threetouched characters

(2)

Fig. 2. Possible segmentation zone detection (PSZ)

Following possible segmentation zone detection, the next step is segmentation of touched characters. Accordingly, within PSZ, from upper baseline to lower baseline, numbers of transactions from background (white pixels) to foreground (black pixels) are counted in each column. The columns with transactions value 2 are termed as candidate segmentation columns and are stored in a queue. However, to locate the best segmentation column among the candidates, greedy best search algorithm is implemented [5]. The greedy best search algorithm evaluated all candidate segmentation columns and finally column with minimum stroke thickness (vertically) selected.

3 Results and Discussion In the proposed approach, PSZ is detected using peak to valley function. Finally, successful touched character segmentation is performed in the PSZ using greedy best

542

T. Saba et al.

search algorithm. For experiments, two hundred touched characters were extracted from IAM forms. Segmentation accuracy rate 91.63% obtained in this research. Segmentation rate is computed as below. (For segmentation results see Fig 4). % Segmentation accuracy = (correct segmentation/ total test words) * 100

Fig. 4. Touched character segmentation using proposed approach

The proposed approach remained successful for solving single and multipletouched characters segmentation problems. The promising results are due to the clustering algorithm that worked on the whole image rather than on specific features. In addition, proposed algorithm is independent of stroke’s thickness.

4 Conclusion This paper presented a new heuristic algorithm for cursive touched character segmentation. The algorithm handles multi touching characters successfully. Proposed algorithm searches within the search area (PSZ) that makes its search narrow and fast. The promising experimental results achieved up to 93.61%.

References [1] Kurniawan, F., Rehman, A., Dzulkifli, M., Mariyam, S.: Self Organizing Features Map with Improved Segmentation to Identify Touching of Adjacent Characters in Handwritten Words. In: IEEE Ninth International Conference on Hybrid Intelligent Systems, HIS 2009, China, pp. 475–480 (2010) [2] Saba, T., Sulong, G., Rehman, A.: A Survey on Methods and Strategies on Touched Characters Segmentation. International Journal of Research and Reviews in Computer Science 1(2), 103–114 (2010) [3] Marti, U., Bunke, H.: The IAM database: An English sentence database for off-line handwriting recognition. International Journal of Document Analysis and Recognition 15, 65–90 (2002) [4] Rehman, A., Dzulkifli, M., Sulong, G., Saba, T.: Simple and Effective Techniques for Core Zone Detection and Slant Correction in Script Recognition. In: The IEEE International Conference on Signal and Image Processing Applications (ICSIPA 2009), pp. 15–20 (2009) [5] Russell, S.J., Norvig, P.: Artificial Intelligence: A Modern Approach, 2nd edn., pp. 94–95. Prentice Hall, Upper Saddle River (2003), http://aima.cs.berkeley.edu/

Context Representation and Management in a Pervasive Environment B. Vanathi and V. Rhymend Uthariaraj Ramanujan Computing Centre, Anna University Chennai, Chennai, Tamil Nadu, India [email protected] and [email protected]

Abstract. Integrating computing and computing applications into surroundings instead of having computers as discrete objects is the objective of pervasive computing. Applications must adjust their behavior to every changing surroundings. Adjustment involves proper capture, management and reasoning of context. This paper proposes representation of context in a hierarchical form and storing of context data in an object relational database than an ordinary database .Semantic of the context is managed by ontology and context data is handled by object relational database. These two modeling elements are associated to each other by semantics relations build in the ontology. The separation of modeling elements loads only relevant context data into the reasoner therefore improving the performance of the reasoning process.

1 Introduction The continuing technical progress in computing and communication lead to an all encompassing use of networks and computing power called ubiquitous or pervasive computing. Pervasive computing system targets at constantly adapting their behavior in order to meet the needs of users within every changing physical, social, computing and communication context. Pervasive devices make ad-hoc connections among them and may be connected to different types of sensors to capture changes in the environment. Fig. 1. Shows the flow in the evolution chain from centralized computing to pervasive computing as presented by [1] [2].Context awareness is at the heart of pervasive computing problems. Context can be defined as an operational term whose definition depends on the intension for which it is collected and the interpretation of the operations involved on an entity at a particular time and space rather than the inherent characteristics of the entities and the operations themselves according to Dey & Winogards [3, 4].The complexity of such problems increases in multiplicative fashion rather than additive with the addition of new components into the chain. Pervasive Context aware computing has three major basic components: pervasive environment, Context management modeling and context-aware service. Pervasive environment is characterized by dynamicity, heterogeneity and ubiquity of users, devices and other computing resources, ad-hoc connection among the devices and existence of hardware and software sensors. Context management modeling deals with how context data is collected, organized, represented, stored and presented to the V. V Das, R. Vijaykumar et al. (Eds.): ICT 2010, CCIS 101, pp. 543–548, 2010. © Springer-Verlag Berlin Heidelberg 2010

544

B. Vanathi and V. Rhymend Uthariaraj

Fig. 1. Evolution Chain

reasoning module. Context aware service performs context reasoning and decisions about the actions to be triggered. In this paper special emphasize on context data management which can be further used in the development of collaborative context application is proposed. Proper modeling, specification and definition of context and its management are essential for efficient reasoning, interpretation, and utilization of context data.

2 Related Works in Context Management Modeling Data required for modeling are obtained from the applications using sensors .Sensors can be physical, virtual or logical sensors. After collecting the data from the application; it has to be represented in a suitable way for processing. Various context management and modeling approaches are introduced to present context for reasoning in different application area. Data from the sensors are presented using any of the following modeling approaches like key-value-pair modeling, Graphical modeling, Object Oriented modeling, logic based modeling, Mark up scheme modeling and Ontology modeling. Among all the modeling approaches ontology based context model is more suitable for context aware computing [2].Ontology is defined as explicit specification of a shared conceptualization [4].Context is modeled as concepts and facts using ontology. Some context aware systems that use this approach are discussed. CONON (CONtext Ontology) [5] is based on treatment of high-level implicit contexts that are derived from low-level explicit context. It supports interoperability of different devices. It defines generic concepts regarding context and provides extensibility for adding domain specific concepts. Context reasoning in pervasive environment is time-consuming but is feasible for non-time-critical applications. For time-critical applications the data size and rule complexity must be reduced. This is an infrastructure based environment. CoBrA-ONT [6] is a context management model that enables distributed agents to control the access to their personal information in context-aware environments. It is designed to overcome the interoperability of different devices. It is central part of CoBrA, broker-centric agent architecture in smart spaces. CoBrA-ONT assumes that there always exists a contextbroker server known by all the participants. It is infrastructure-centric and is not for pervasive computing. SOUPA (Standard Ontology for Ubiquitous and Pervasive Applications) [7] includes modular component vocabularies to represent intelligent agents with associated beliefs, desires and intension, time, space, events, user profiles, actions and policies for security and privacy. SOUPA is more comprehensive than CoBrA-ONT because it deals with more areas of pervasive computing and ontology

Context Representation and Management in a Pervasive Environment

545

can be reused. GAS Ontology [8] is ontology designed for collaboration among ubiquitous computing devices. The basic goal of this ontology is to provide a common language for the communication and collaboration among the heterogeneous devices that constitute these environments. The GAS Ontology also supports the service discovery mechanism that a ubiquitous computing environment requires.

3 Limitations of Ontology Based Context Management Context aware systems are based on ad-hoc models of context, which causes lack of the desired formality and expressiveness. Existing models do not separate processing of context semantics from processing and representation of context data and structure. Ontology representation tools are suitable for statically representing the knowledge in a domain. They are not designed for capturing and processing constantly changing information in dynamic environment in a scalable manner. Existing ontology languages and serialization formats are test based (xml/rdf/owl) and not designed for efficient query optimization, processing and retrieval of large context data. The main drawbacks of pure ontological approaches are low performance and data throughput.

4 Proposed Work The proposed context aware system has context acquisition layer, context middleware (representation layer, context management layer and decision layer) and application layer. Context acquisition layer gathers the context from the environment using sensors. Context representation layer represents context as entity relation hierarchy form. In the context management layer context is further classified as inactive context and active context. Predicates are used to decide the inactive context. For example context defined using predicates like ownedby are inactive and context defined using predicates like locatedIn are active. Inactive context are stored in Object relational database and active context are stored in Ontology. In the middleware, rules learned or rules derived from other rules are also maintained. Using the rules, relevant context from database and ontology are forwarded to the reasoning component. From reasoning layer appropriate context is sent to user in the application layer. 4.1 Context Representation Dey and Winogards define Context in terms of a statement that is made about the characteristics of the entities, their relationships and properties of the relationships [4]. Context can be personal, device, physical, activity, network, location etc. Personal entity provides contexts like person's identity, address, activity, location etc. Context can be represented as Entity, Hierarchy, Relation, Axiom and Metadata [9]. Hierarchy is a set of binary relations that form an inversely directed acyclic graph. Relation is union of set of binary relations. Relation can be either entity relation or attribute relation. Entity relation has a set of binary relations having either its domain or range from the set of entity. SubEntityOf relation, SubPropertyOf, domain, range etc is some relations used to link entities. Root of the hierarchy is a global entity called ContextEntity .Entities and relations are sources of context data. Relation can be generic or domain based. The flow of context is shown in Fig. 2.

546

B. Vanathi and V. Rhymend Uthariaraj

S ta rt C on te xt A cqu is ition (se n s o rs , b a d g e s, C a m e ra ,W iFi e tc)

C o n te xt re p res en ta tion

Is C on tex t in active

Y es

Sto re C o n text in O b ject R e latio n al D a tab a se(O R D B M S)

No Sto re A ctiv e C o n text in O n tolo g y R ep o s ito r

R u le s C a p tu re d / D e rive d

Se le ct A p p ro p riate C o n text u sin g re as o n e r an d se n t to th e u s er End

Fig. 2. Flow of Context

For example in a generic level, relation is defined as person isLocatedIn Location and in a domain level, relation is defined as Student isLocatedIn Class .Attribute relation is the set of binary relations defined from set of entities to set of literals. Axiom is the axiomatic relations. Few generic level axiomatic relations are sameAs, inverse, symmetric and transitive. Meta data are information about a defined relation instance. Information like time of occurrence, precision, source of data can be a part of Meta information. Consider an example, Student isLocatedIn Class at a given time t. 4.2 Storing Context Data in an Relational Database Context represented in entity relation hierarchy form is stored in a relational database using the algorithm steps mentioned below. The attributes CEntity stores name of context entities. Attribute that stores name of the entity one step above in the hierarchy isa relation. Layer stores whether an entity is in the generic or domain layer. Relation stores name of relations. Persistence of the relation stores whether an entity can be inactive or active. Values of relations with static persistence are stored in the persistent context repository and values with dynamic persistence are stored temporarily for immediate use in the field Persistence. ValueForm stores source of value as a context entity or as a Literal. An attribute that stores name of instances is EInstance. Value is an attribute that stores value of the relation after applied to the instance. Timestamp stores context time. Source stores source of context. Precision stores precision of the context .Structure of the relational table is shown below: Step 1: Entity table (Context Entity, direct hierarchy relation, Layer) Step 2: Relation table (Relation, Persistence)

Context Representation and Management in a Pervasive Environment

547

Step 3: Relation Instance table (Relation name, Context Entity, Value) Step 4: Entity Instance table (Entity Instance, Context Entity) Step 5: Context Instance table (Entity Instance, Relation, Value, Time Stamp, Context Source and Context Precision) 4.3 Advantages of ORDBMS and RDBMS Relational models provide standard interfaces and query optimization tools for managing large and distributed context database or receive and send notification on context changes. Relational models are not designed for semantic interpretation of data. Relational database alone cannot be used to represent context in a pervasive environment. For semantic interpretations, ontology is used along with relational database. The Table 1 below summarizes the appropriateness of the approaches in relation to the necessary features. All approaches have strong and weak sides with respect to features for context management modeling. Best of three worlds are combined to form a hybrid context management model. Ordbms approach is more suitable than rdbms approach. It ensures large storage capacity, quick access speed. Table 1. Comparison of rdbms, ordbms and ontology approach

Necessary Feature Semantic Support Ease of transaction Query optimization Reasoning Support Formality Scalability

Relational Approach No Yes Yes No Yes Yes

Ontology Approach Yes No No Yes Yes No

Object Relational Approach No Yes Yes No Yes Yes

Object relational database supports several storage units like collection list, arrays, types and UDTs (User defined data types) and most of them are represented as objects arrays. Ordbms have a massive scalability compared to relational approach and excellent manipulation power of object databases. It supports rich data types by adding a new object-oriented layer. The systems are initially implemented by storing the inactive context to a relational database and active context to an ontology. Then the response time to get the relevant time is noted. Further system is implemented by replacing the storage of inactive context to relational database by object relational database. Then appropriate service can be provided to the user using service discovery [10]. 4.4 Metrics Basic metrics used to evaluate the relational database and object relational database are throughput or response time, memory and/or storage usage/requirements. Throughput is the number of queries per time. Evaluation time can be wall-clock (real), Central Processing Unit (user), Input/output (system), server side versus client

548

B. Vanathi and V. Rhymend Uthariaraj

side. To compare the relational database and object relational database with respect to ontology complexity of complex columns (CCC) is considered.

5 Conclusion Context is represented using layered and directed graph. Layered organization helps to classify and tag context data as generic domain independent or as domain dependent. A combination of context model using ontology and object relational database is proposed. This paper focuses on context representation and storage of context. Reasoning and decision making of the context obtained from the context management are the future work.

References 1. Satyanarayanan, M.: Pervasive Computing Vision and Challenges. IEEE Personal Communications, 10–17 (2000) 2. Strang, T., LinnhoPopien, C.: A Context Modeling Survey. In: Davies, N., Mynatt, E.D., Siio, I. (eds.) UbiComp 2004. LNCS, vol. 3205. Springer, Heidelberg (2004) 3. Winograd, T.: Architectures for Context. Human-Computer Interaction 16(2-4), 401–419 (2001) 4. Dey, A.K., Abowd, G.D.: Towards a Better Understanding of Context and Context Awareness. In: Proceedings of the CHI Workshop on the What, Who, Where and How of Context- Awareness, The Hague, The Netherlands (2000) 5. Wang, X., Zhang, D., Gu, T., Pung, H.K.: Ontology Based Context Modeling and Reasoning using OWL, workshop on context modeling and reasoning. In: IEEE International Conference on Pervasive Computing and Communication, Orlando, Florida (2004) 6. Chen, H.: An Intelligent Broker Architecture for Pervasive Context-Aware Systems. PhD Thesis University of Maryland, Baltimore Country, USA (2004) 7. Chen, H., Perich, F., Finin, T., et al.: SOUPA: Standard Ontology for Ubiquitous and Pervasive Applications. In: International Conference on Mobile and Ubiquitous Systems: Networking And Services, Boston, USA (2004) 8. Christopoulou, E., Kameas, A.: GAS Ontology: ontology for collaboration among ubiquitous Computing devices. International Journal of Human-Computer Studies 62(5), 664–685 (2005) 9. Ejigu, D., Scuturi, M., Brunie, L.: An Ontology Based Approach to Context Modeling and Reasoning in Pervasive Computing. In: 5th IEEE International Conference on Pervasive Computing and Communications Workshops, pp. 14–19 (2007) 10. Vanathi, B., Rhymend Uthariaraj, V.: Ontology based service discovery for context aware computing. In: 1st IEEE International Conference on Advanced Computing. IEEE Computer Society, Chennai (2009)

A Comparative Study and Choice of an Appropriate Kernel for Support Vector Machines R. Sangeetha and B. Kalpana Department of Computer Science, Avinashilingam University for Women, Coimbatore [email protected], [email protected]

Abstract. Support Vector machine (SVM) has become an optimistic method for data mining and machine learning. The exploit of SVM gave rise to the development of a new class of theoretically refined learning machines, which uses a central concept of kernels and the associated reproducing kernel Hilbert space. The performance of SVM largely depends on the kernel. However, there is no premise about how to choose a good kernel function for a particular domain. This paper focuses in this issue i.e. the choice of the Kernel Function is studied empirically and optimal results are achieved for binary class SVMs. The performance of the Binary class SVM is illustrated by extensive experimental results. The experimental results of the datasets show that RBF Kernel or any other kernels is not always the best choice to achieve high generalization of classifier although it is often the default choice. Keywords: Support Vector Machine, Pattern Classification, Kernel Function, Support Vectors, Mercer Kernels.

1 Introduction Support Vector machines are based on Statistical Learning Theory developed by Vapnik [1] and designed originally for binary classification. The formulation embodies the Structural Risk Minimization principle has given by the authors in [5-6], which has been shown to be superior to traditional Empirical Risk Minimization principle employed by conventional neural networks. SRM minimizes an upper bound on the generalization error but ERM minimizes the error on the training data. Use of Kernel outlined in [3] enables the curse of dimensionality to be addressed and the solution implicitly contains Support Vectors. In training a SVM we need to select a Kernel and its parameters. There is no method to determine how to choose an appropriate Kernel and its parameters for a given dataset to achieve high generalization of classifier. Furthermore, the choice of the regularization parameter C is crucial to obtain good classification results. This paper is organized as follows. Section 2 gives the brief overview of support vector machines. Section 3 discusses the various types of kernels. This is followed by experimental results. Section 5 concludes the work. V. V Das, R. Vijaykumar et al. (Eds.): ICT 2010, CCIS 101, pp. 549–553, 2010. © Springer-Verlag Berlin Heidelberg 2010

550

R. Sangeetha and B. Kalpana

2 Support Vector Machines Traditional optimization algorithms such as Newton Method or Quasi-Newton Method cannot work any more due to the memory problem. A special property of SVM [2-4] is it simultaneously minimizes the empirical classification error and maximizes the geometric margin, called as Maximum Margin Classifiers. SVM maps input vector to a higher dimensional space where a maximal separating hyperplane is constructed. Two parallel hyperplanes are constructed on each side of the hyperplane that separates the data. One of the hyperplanes that maximizes the margin is named as the optimal separating hyperplane. Consider the problem of separating the set of training vectors belonging to binary classes or dichotomization (xi, yi), i = 1,…. l, xi ∈ Rn, yi ∈ {+1, −1}, where the Rn is the input space, xi is the feature vector and yi is the class label of xi. The separating hyperplanes are linear discriminating functions as follows, f ( x ) = wT x + b

(1)

where w is a weight vector and b is called the bias value.

3 Kernel Induced Feature Space SVM uses an efficient mathematical function for mapping the classification data called as kernel trick and a dot product given in [7] for mapping higher dimension. 3.1 Mercer Kernel In statistical learning theory, if kernels are positive definite then they satisfy Mercer’s condition [8] in Hilbert space as a dot product and called as Mercer Kernel. 3.1.1 Mercer Theorem [9] 2 Suppose that K ∈ L∞ ( χ ) is a symmetric real-valued kernel such that the integral operator T K : L 2 ( χ ) → L 2 ( χ )

TK f ( x) =

∫ K ( x , y ) f ( y ) dy

(2)

x

is positive,i.e.., for all f ∈ L2 (χ ) we have ,

∫ K ( x , y ) f ( x ) f ( y ) dxdy

≥ 0 . If any

kernel function satisfies equation (2), then the kernel K maps data in feature space. 3.2 Reproducing Kernel Hilbert Spaces A Hilbert Space H (an infinite dimensional linear space endowed with a dot product) is a Reproducing Kernel Hilbert Space (RKHS) [8] ,if the evaluation functional bounded i.e., there exists a M such that

A Comparative Study and Choice of an Appropriate Kernel for Support Vector Machines

Ft [ f ] = f ( t ) ≤ M f

H

∀f ∈ H .

551

(3)

4 Experimental Results In this section, we use the different type of kernels depicted in table 1 to three bench mark datasets (Liver Disorder, Iris, Heart) from UCI database repository. Our aim is to make the kernels generalized for every domain. In classification tasks Ten-fold cross validation is used for generalization performance. The table 2 suggests that the few kernels give good classification performance and low error rate. After comparing all the features of the kernels, the appropriate kernels for small binary class datasets are spline and polynomial which has minimum number of support vectors, minimum value as error rate and good classification rate. Table 1. Types of Kernels Kernels Linear

Function

K ( xi , x j ) = 1 + xi x j T

Polynomial

K ( xi , x j ) = (1 + xi x j ) p

Radial Basis Function (RBF) Gaussian RBF

K(xi , x j ) = exp(−γ xi − x j )

Exponential RBF Sigmoid Spline

T

2

K ( x i , x j ) = exp( −

K ( x i , x j ) = exp( −

xi − x j

2

2σ 2 xi − x j

2σ

2

)

)

K ( xi , x j ) = tanh(kxi x j − δ ) T

k spline (u , v ) = 1 + uv + ∫ (u − t )+ (v − t )+ dt 1

0

Anova Spline Additive

K ( u , v ) = 1 + k ( u 1 , v 1 ) + k ( u 2 , v 2 ) + k ( u 1 , v 1 ) k (u 2 , v 2 )

∑

K (u , v ) =

K i (u , v )

i

Tensor product

n

K (u , v ) = ∏ K m (u m , v m ) m =1

Generally kernel functions are classified as Translation invariant kernel and Rotation invariant Kernel. From the above study, it is proposed that hybrid kernels are obtained by combining more than one invariant kernel, which can find numerous applications in practice. It works as an efficient classifier while comparing with standard kernel classifier and gives high accuracy for classification. In the hybrid kernel function, the parameters can be easily modified for different domains and data can be classified with low execution time.

552

R. Sangeetha and B. Kalpana

Table 2. Comparing Kernels with datasets. (Support Vectors,Classification and Error Rate) (a) Iris dataset Kernel Function (Parameters) Polynomial RBF Exponential RBF Gaussian RBF Sigmoid Spline Anova Spline

C=10,p=1 C=50000, p=2 C=5, γ=5 C=50000, γ =2 C=10, σ =2 C=80000, σ =5 C=10, σ =2 C=5000, σ =5 C=10,k=0.5 δ=0 C=inf, k=2, δ=4 C=10 C=1000000 C=10,k=5 C=50000,k=10

Kernel Function (Parameters) Polynomial RBF Exponential RBF Gaussian RBF Sigmoid Spline Anova Spline

RBF Exponential RBF Gaussian RBF Sigmoid Spline Anova Spline

SV%

CR

ER

33 29 55 40 36 34 43 42 60 60 31 29 44 47

55.0 48.3 91.7 66.7 60.0 56.7 71.7 70.0 100 100 51.7 48.3 73.3 78.3

0.7500 0.7500 0.9835 0.9667 0.7500 0.9833 0.9667 1.0 0.7600 0.7600 0.7500 0.7700 0.9998 0.9997

0.2500 0.2500 0.0165 0.0333 0.2500 0.0167 0.0333 0.0 0.2400 0.2400 0.2500 0.2300 0.0002 0.0333

(b) Liver dataset SV SV%

C=10,p=5 C=500000,p=2 C=10, γ=0.05 C=1000000, γ =1 C=10, σ =2 C=100000, σ =5 C=10, σ =2 C=50, σ =5 C=5,k=1, δ=1 C=500000,k=5, δ=10 C=10 C=1000000 C=10,k=5 C=100,k=10

Kernel Function (Parameters) C=10,p=3 Polynomial

SV

59 54 62 54 61 58 75 74 56 56 55 54 56 56

54.6 50.0 57.4 50.0 56.5 53.7 69.4 68.5 51.9 51.9 50.9 50.0 51.9 51.9

(c) Heart dataset SV SV%

C=100, p=5 C=10, γ =2 C=1000000, γ=5 C=10, σ =2 C=1000000, σ =5 C=10, σ =2 C=1000000, σ =5 C=10,k=1, δ=3 C=100000, k=5, δ=10 C=10 C=1000000 C=10,k=2 C= 100000, k=10

73 72 115 70 85 70 75 70 138 70 71 70 75 74

52.9 52.2 83.3 50.7 61.6 50.7 54.3 50.7 100 50.7 51.4 50.7 54.3 53.6

CPU (s) 0.1 0.1 0.1 0.3 0.1 0.3 0.1 0.1 0.1 0.1 0.0 0.1 0.1 0.1

CR

ER

0.6852 0.6853 0.6852 0.6854 0.6852 0.6852 0.6857 0.6852 0.6852 0.6759 0.6759 0.6852 0.6667 0.6667

0.3148 0.3147 0.3148 0.3146 0.3148 0.3148 0.3143 0.3148 0.3148 0.3241 0.3241 0.3148 0.3333 0.3333

CPU (s) 0.3 0.3 0.4 0.4 0.3 0.3 0.3 0.2 0.1 0.2 0.3 0.2 0.1 0.2

CR

ER

CPU(s)

0.5580 0.5580 0.5581 0.5580 0.5581 0.5581 0.5000 0.6154 0.5714 0.6154 0.4286 0.7143 0.5714 0.3571

0.4420 0.4420 0.4419 0.4420 0.4419 0.4419 0.5000 0.3846 0.4286 0.3846 0.5714 0.2857 0.4286 0.6429

0.5 0.5 0.6 0.5 0.4 0.6 0.4 0.5 0.5 0.5 0.5 0.4 0.5 0.4

A Comparative Study and Choice of an Appropriate Kernel for Support Vector Machines

553

5 Conclusion The experimental results of the three datasets show that opting the Kernels randomly is not always the best choice to achieve high generalization of classifier. It illustrates that the crucial kernels effectively improves the classification performance and generializability by optimizing the result. We exhibit the dependency of classifier accuracy on the different Kernel Functions of the Binary-class SVM using different datasets. It will be interesting and practically more useful to determine some method for determining the Kernel Function and its parameters based on statistical properties of the given data. In this paper we propose a Hybrid kernel method for classification of data. A systematic methodology and optimization technique is needed for the construction of Hybrid kernel in SVM. Particle swarm optimizations, Genetic Algorithm are the areas which provide a feasible solution for the optimization problems. So, they could be applied as an optimization technique for constructing hybrid kernel function based on large margin learning theory. Therefore, the crisis on choosing the right kernels and the best systematic method of combining them would be our research work in future.

References 1. Vapnik, V.: An overview of statistical learning theory. IEEE Trans. on Neural Networks 10(5), 988–999 (1999) 2. Cristianini, N., Shawe-Taylor, J.: Introduction to Support Vector Machines. Cambridge University Press, Cambridge (2000) 3. Schölkopf, B., Smola, A.: Leaning with Kernels. MIT Press, Cambridge (2001) 4. Burges, C.J.C.: A tutorial on support vector machines for pattern recognition. Data Mining and Knowledge Discovery 2(2), 56–89 (1998) 5. Cortesand, C., Vapnik, V.: Support-Vector Networks. Machine Learning 20 (1995) 6. Gunn, S.R.: ” SVM for Classification and Regression”, Technical Report, Image Speech and intelligent system groups (1998) 7. Herbrich, R.: Learning kernel classifiers: theory and algorithms. MIT Press, Cambridge (December 2001) 8. Xia, G.-E., Shao, P.-J.: Factor Analysis Algorithm with Mercer Kernel. IEEE Second International Symposium on Intelligent Information Technology and Security Informatics (2009) 9. Schölkopf, B., Mika, S., Burges, C., et al.: Input Space versus Feature Space in KernelBased Methods. IEEE Transactions on Neural Networks (5), 1000–1017 (1999) 10. Hofmann, T., Schölkopf, B., Smola, A.J.: Kernel Methods in Machine Learning. The Annals of Statistics 36(3), 1171–1220 (2008)

Color Image Restoration Method for Gaussian Noise Removal J. Harikiran and R. Usha Rani Department of Information Technology, GIT, GITAM University, Visakhapatnam [email protected], [email protected]

Abstract. A new approach to the restoration of color images corrupted by Gaussian noise is presented. The proposed technique adopts a multipass processing approach that gradually reduces the noise in the color information components of the image. Two different models for data smoothing are proposed based on the different classes of noisy pixels. The subsequent algorithm for edge detection is designed to better appraise the noise cancellation behavior of our filter from the point of view of human perception. This method does not require any “a priori” knowledge about the amount of noise corruption. Experimental results show that the filtering performance of the proposed approach is very satisfactory and accurate edge maps are achieved even in the presence of highly corrupted data. Keywords: Gaussian noise, image enhancement, nonlinear filter, image restoration, image processing.

1 Introduction Image noise is the random variation of brightness or color information in images produced by the sensor and circuitry of a scanner or digital camera. The development of techniques for noise removal is of paramount importance for image-based measurement systems [1]. In-order to smooth out from noise many filtering architectures have been proposed in the literature [2]. Indeed, noise can significantly decrease the accuracy of very critical operations such as feature extraction and object recognition. The goal of the filtering action is to cancel noise while preserving the integrity of edge and detail information, non-linear approaches generally provide more satisfactory results than linear techniques. However, a common drawback of the practical use of these methods is the fact that they usually require some “a priori” knowledge about the amount of noise corruption. Unfortunately such information is not available in real time applications. In this paper, a new approach to filtering of Gaussian noise and edge detection in color images is presented. The filtering behavior is based on the classification of noisy pixels into two different (fuzzy) classes. 1) Pixels corrupted by noise with amplitude not two different from that of neighbors. (Type A pixels). 2) Pixels corrupted by noise with amplitude much larger than that of the neighbors. (Type B pixels). V. V Das, R. Vijaykumar et al. (Eds.): ICT 2010, CCIS 101, pp. 554–560, 2010. © Springer-Verlag Berlin Heidelberg 2010

Color Image Restoration Method for Gaussian Noise Removal

555

The noisy color image is processed by first converting it from RGB to YIQ domain. Then prefiltering is applied (for both Type A and Type B pixels) only to the color information (I and Q) components of the image and the filtered image is converted from YIQ into RGB domain resulting a noise free image. The images generated by our method have been obtained without taking into account the original uncorrupted data, in order to simulate a real application where only noise data is available. This is the key advantage of the proposed method. This paper is organized as follows. Section 2 presents conversion of color image from RGB to YIQ domain. Section 3 presents filtering architecture for Type A pixels, Section 4 presents filtering architecture for Type B pixels, Section 5 presents the edge detection algorithm, Section 6 presents the experimental results and finally Section 7 report conclusions.

2 Conversion from RGB to YIQ Domain The early approaches to color image processing are performed by processing each RGB components separately. A disadvantage of these methods is the loss of correlation between the color channels resulting in color shifts [4][5][6]. That is a substitution of a noisy pixel color through a new false color, which does not fit into the local neighbourhood. This means that a noisy pixel is replaced by another noisy pixel. In our work, the YIQ system is used to process color image. The principle advantage of this space in image processing is that the color information components (I and Q) are processed leaving the luminance component (Y) It will be a need to convert RGB to YIQ system for this purpose. The conversion from RGB to YIQ and YIQ to RGB [3] is given as follow: Y I Q

0.299 =

0.587

0.114

R

R

1

0.596 -0.274

-0.322

G

G

= 1

0.312

B

B

1

0.211

-0.523

Fig. 1. Conversion from RGB to YIQ

0.9563

0.6210

Y

-0.2721 -0.6474

I

-1.1070

Q

1.7046

Fig. 2. Conversion from YIQ to RGB

3 Type A Prefiltering Pixels corrupted by noise with amplitude not too different from that of the neighbors are Type A pixels. The filter, we are using in our approach is called “zed filter”. Let us suppose we deal with digitized images having L levels (color images). Let x(n) be the pixel luminance at location n=[n1,n2] in the noisy image. Let x1(n), x2(n) ……, xN(n) be the group of N=8 neighboring pixels that belong to a 3X3 window around x(n). The output y(n) of the filter is defined by the following relationship: y(n) = x(n) + Where

1 N

N

∑ ζ ( x (n), x(n)) i

i =1

(1)

556

J. Harikiran and R. Usha Rani

u–v

ζ

(u,v)=

|u – v| ≤ p

3 p− | u − v | sgm(u-v) p3p

(2)

and “p” is an integer ( 0