A* ATTACK: A NOVEL PATH-FINDING APPROACH TO ADVERSARIAL EXAMPLES

Loading...
Thumbnail Image
Authors
Clark, Christopher D.
Subjects
deep neural network
DNN
artificial intelligence
A*
A* search
A-Star
adversarial example
MLaaS
machine learning as a service
black-box attack
Advisors
Barton, Armon C.
Date of Issue
2023-09
Date
Publisher
Monterey, CA; Naval Postgraduate School
Language
Abstract
This paper presents a novel approach to exploiting a key vulnerability of deep neural networks (DNNs) to adversarial examples with a focus on the black-box machine learning as a service (MLaaS) environment. We introduce A* Attack, a unique adversarial example attack that leverages the A* Search algorithm to find adversarial perturbations. This innovative approach is designed to overcome the challenges of both excessive model queries in decision- and score-based attacks and the limitations of transferability from white-box attacks. The A* Attack demonstrates competitive performance in the white-box setting and sets a new standard in the decision-based black-box setting, achieving high attack success rates with minimal queries. This represents a significant advancement in the field, offering a new approach to the black-box attack method. This paper provides a competitive evaluation of the A* Attack on CIFAR10 and ImageNet, comparing its performance against other leading attacks and defenses.
Type
Thesis
Description
Department
Computer Science (CS)
Organization
Identifiers
NPS Report Number
Sponsors
Funder
Format
Citation
Distribution Statement
Approved for public release. Distribution is unlimited.
Rights
This publication is a work of the U.S. Government as defined in Title 17, United States Code, Section 101. Copyright protection is not available for this work in the United States.
Collections