Date of Award

12-11-2025

Degree Type

Thesis

Degree Name

Master of Science (M.S.)

Department

Electrical & Computer Engineering

First Advisor

Kamrul Hasan

Abstract

The growing scale and complexity of Internet-of-Things (IoT) edge networks complicate anomaly detection, particularly in identifying sophisticated Distributed Denial of Service (DDoS) attacks and zero-day behaviors under highly dynamic and imbalanced traffic conditions. This thesis proposes SD-CGAN, a Conditional Generative Adverserial Network optimzied with Sinkhorn Divergence as a geometry-aware one-class framework for robust IoT anomaly detection. SD-CGAN trains solely on benign traffic flows to learn a stable representation of normal traffic. To address class imbalance and improve the variety of the sample, we combine SD-CGAN with CTGAN-based synthetic data augmentation. Replacing the adversarial objective function with Sinkhorn Divergence yields smooth gradients, supports non-overlapping distributions, improves convergence and reduces mode collapse, while still remaining lightweight for IoT edge devices. We evaluate the model on exploitative subsets of CICDDoS2019 and compare against deep learning and GAN models. SD-CGAN achieves a strong accuracy of over 98% while remaining computationally efficient on CPU-only hardware (training in seconds and sub-second per sample scoring), which supports edge deployment. The main contributions of this thesis are: (i) a Sinkhorn Divergence-enhanced conditional GAN for stable geometry aware training; (ii) unsupervised, benign only anomaly scoring procedur suited for zero day detection; (iii) CTGAN integration to address data imbalance, validated by distributional analyses; and (iv) an end-to-end reproducible pipeline aligned with resource constraints at the edge. Together, these results show that SD-CGAN is a practical and accurate approach for protecting IoT edge networks from evolving DDoS attacks.

Download

Share

COinS