TROPICAL GEOMETRY IN NEURAL NETWORKS: A NOVEL DEFENSE AGAINST ADVERSARIAL ATTACK
Loading...
Authors
Pasque, Kurt R.
Subjects
Department of Defense
DoD
Tropical Convolutional Neural Network
TCNN
artificial intelligence
AI
neural networks
image classification
adversarial attacks
tropical activation layers
tropical geometry
robustness
model structure
rectified linear unit
ReLU
rectified linear unit activation
ReLU activation
neural network design
image classification
adversarial manipulations
DoD
Tropical Convolutional Neural Network
TCNN
artificial intelligence
AI
neural networks
image classification
adversarial attacks
tropical activation layers
tropical geometry
robustness
model structure
rectified linear unit
ReLU
rectified linear unit activation
ReLU activation
neural network design
image classification
adversarial manipulations
Advisors
Yoshida, Ruriko
Huang, Jefferson
Date of Issue
2024-09
Date
Publisher
Monterey, CA; Naval Postgraduate School
Language
Abstract
Neural networks, though powerful, are highly susceptible to adversarial attacks, which can severely impact their reliability and performance in critical applications, especially within the Department of Defense (DoD). This thesis employs Tropical Convolutional Neural Networks (TCNN), a novel technique leveraging tropical geometry to enhance neural network robustness against adversarial attacks. By reconfiguring neural networks to use a tropical embedding layer as the final layer, this approach significantly improves resistance to attacked input data without compromising model accuracy, complexity, or training time. The research includes comprehensive experiments on image classification tasks across multiple datasets, demonstrating that TCNNs maintain higher accuracy under adversarial conditions compared to identically structured traditional models. This work not only contributes to the advancement of secure artificial intelligence (AI) applications in defense but also opens new avenues for future research in neural network robustness.
Type
Thesis
Description
Series/Report No
Department
Organization
Identifiers
NPS Report Number
Sponsors
Funder
Format
Citation
Distribution Statement
Distribution Statement A. 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.