Electronic Theses and Dissertations

Title

Use of Linear and Nonlinear Classifiers against Adversarial Learning and Influence

Identifier

101

Author

Sudip Saha

Date

2010

Date of Award

7-26-2010

Document Type

Thesis (Access Restricted)

Degree Name

Master of Science

Major

Computer Science

Committee Chair

Dipankar Dasgupta

Committee Member

King-Ip Lin

Committee Member

Vinhthuy Phan

Abstract

Machine learning based classification techniques are being used in a growing number of security applications for anomaly detection related purposes. This is due to the adjusting capability and novel anomaly detection ability of machine learning. Previous works have shown that very simple classification techniques like linear classifiers can easily be learned and attacked by the adversary. This work addresses the adversarial learning possibility of nonlinear classifiers. It has been illustrated in this work that, nonlinear classifiers in general cannot be learned. However, special cases of nonlinearity have been addressed in which adversarial learning has been considered with some heuristic approaches.Another kind of attack on machine learning systems is the adversarial influence on learning. The adversary may change the distribution of data that is used in learning. In this work, the performance of a particular learning system has been studied against such kind of influence.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to the local University of Memphis Electronic Theses & dissertation (ETD) Repository.

This document is currently not available here.

Share

COinS