Date Approved
7-6-2026
Embargo Period
7-6-2027
Document Type
Thesis
Degree Name
M.S. Computer Science
Department
Computer Science
College
College of Science & Mathematics
Advisor
Silvija Kokalj-Filipovic, Ph.D.
Committee Member 1
Andrea Lobo, Ph.D.
Committee Member 2
Patrick McKee, M.S.
Disciplines
Computer Sciences | Physical Sciences and Mathematics
Abstract
Adversarial attacks pose a major vulnerability for deep neural networks, threatening the reliability of machine learning systems deployed in real-world environments such as Next Generation (NextG) wireless networks and AI-driven sensing platforms. This thesis investigates Discrete-Space Variational Information Bottleneck (DSVIB) models to improve adversarial robustness in signal classification. While prior work has explored continuous space Variational Information Bottleneck (VIB), the potential benefits of discrete latent representations remain largely unexamined. To address this gap, we introduce discrete bottleneck architectures based on vector quantization, which compress input signals into finite codebooks that suppress adversarial perturbations while preserving task-relevant information. We develop and evaluate two DSVIB frameworks: a Vector Quantized Variational Autoencoder (VQVAE) that preprocesses signals by filtering adversarial noise prior to classification, and a Vector Quantized Classifier (VQC) that embeds the discrete bottleneck directly within the classifier for end-to-end robust learning. Both models are trained using information-theoretic VIB objectives. Across two signal domains, Radio Frequency (RF) modulation classification (Torchsig), and image classification (MNIST), discrete bottleneck architectures consistently outperform baseline classifiers and continuous latent-space models under a range of white-box attacks. These include gradient-based, saliency-based, and adaptive methods, as well as novel phase-preserving variants designed for RF signals. This work establishes discrete information bottlenecks as an effective and generalizable defense strategy for signal classification and highlights their potential to enhance the security and reliability of machine learning systems.
Recommended Citation
Garuso, Attanasia, "ADVERSARIAL ROBUSTNESS IN SIGNAL CLASSIFICATION THROUGH VECTOR QUANTIZED INFORMATION BOTTLENECK" (2026). Theses and Dissertations. 3558.
https://rdw.rowan.edu/etd/3558
