Adversarial Attack Detection and Defense in Graph Alignment and Text-to-Image Generation

Abstract

Recent advances in machine learning have highlighted critical vulnerabilities in graph matching models and text-to-image diffusion models (T2I DMs), where adversarial attacks can significantly compromise system performance while remaining imperceptible to users. This dissertation addresses the dual challenges of developing effective adversarial attacks and robust defense mechanisms across two domains: graph matching systems (including network alignment and cross-lingual entity alignment in knowledge graphs) and text-to-image generation models. Our research tackles fundamental issues in adversarial machine learning: generating effective attacks while ensuring imperceptibility, and developing defenses that maintain system performance. We identify and solve gradient vanishing issues in iterative attack methods and address the challenge of defending against adversarial perturbations without compromising matching or generation quality.