Date Approved

9-14-2022

Embargo Period

9-19-2022

Document Type

Thesis

Degree Name

M.S. Electrical and Computer Engineering

Department

Electrical and Computer Engineering

College

Henry M. Rowan College of Engineering

Advisor

Robi Polikar, Ph.D.

Committee Member 1

Umashanger Thayasivam, Ph.D.

Committee Member 2

Ravi Ramachandran, Ph.D.

Committee Member 3

Ghulam Rasool, Ph.D.

Keywords

Adversarial Defense, Adversarial Machine Learning, Cyberattacks

Subject(s)

Cyberinfrastructure--Security measures; Machine learning

Disciplines

Electrical and Computer Engineering

Abstract

Machine learning models are increasingly employed in making critical decisions across a wide array of applications. As our dependence on these models increases, it is vital to recognize their vulnerability to malicious attacks from determined adversaries. In response to these adversarial attacks, new defensive mechanisms have been developed to ensure the security of machine learning models and the accuracy of the decisions they make. However, many of these mechanisms are reactionary, designed to defend specific models against a known specific attack or family of attacks. This reactionary approach does not generalize to future "yet to be developed" attacks. In this work, we developed Broad Spectrum Defense (BSD) as a defensive mechanism to secure any model against a wide range of attacks. BSD is not reactionary, and unlike most other approaches, it does not train its detectors using adversarial data, hence removing an inherent bias present in other defenses that rely on having access to adversarial data. An extensive set of experiments showed that BSD outperforms existing detector-based methods such as MagNet and Feature Squeezing. We believe BSD will inspire a new direction in adversarial machine learning to create a robust defense capable of generalizing to existing and future attacks.

Share

COinS