Masters Theses
Keywords and Phrases
Adversarial attack; Attack generator; Genetic algorithm; MAT; Multi-objective optimization; Neural network
Abstract
"Vulnerability to adversarial attacks is a recognized deficiency of not only deep neural networks (DNNs) but also multi-task deep neural networks (MT-DNNs) that attracted much attention in the past few years. To the best of our knowledge, all multi-task deep neural network adversarial attacks currently present in the literature are non-targeted attacks that use gradient descent to optimize a single loss function generated by aggregating all loss functions into one. On the contrary, targeted attacks are sometimes preferred since they give more control over the attack. Hence, this paper proposes a novel targeted multi-objective adversarial ATtack (MAT) based on genetic algorithms (GA)s that can create an adversarial image capable of affecting only targeted loss functions of the MT-DNN system. MAT is trained on the Taskonomy dataset using a novel training algorithm GAMAT that consists of five specific stages. The superiority of the proposed attack is demonstrated in terms of the fitness-distance metric"--Abstract, p. iv
Advisor(s)
Nadendla, V. Sriram Siddhardh
Committee Member(s)
Madria, Sanjay Kumar
Tripathy, Ardhendu S.
Department(s)
Computer Science
Degree Name
M.S. in Computer Science
Publisher
Missouri University of Science and Technology
Publication Date
Spring 2023
Pagination
ix, 64 pages
Note about bibliography
Includes_bibliographical_references_(pages 59-63)
Rights
© 2023 Nikola Andric, All Rights Reserved
Document Type
Thesis - Open Access
File Type
text
Language
English
Thesis Number
T 12290
Electronic OCLC #
1426046764
Recommended Citation
Andric, Nikola, "MAT: Genetic Algorithms Based Multi-Objective Adversarial Attack on Multi-Task Deep Neural Networks" (2023). Masters Theses. 8161.
https://scholarsmine.mst.edu/masters_theses/8161
