AttackGNN: Red-Teaming GNNs in Hardware Security Using Reinforcement Learning

Explore a 15-minute conference presentation from USENIX Security '24 that introduces AttackGNN, a pioneering red-team attack methodology targeting Graph Neural Network (GNN) applications in hardware security. Learn how researchers from Texas A&M University and the University of Delaware developed a novel reinforcement learning agent to generate adversarial circuit examples that successfully challenge GNN-based security techniques. Discover how this approach addresses challenges in effectiveness, scalability, and generality while targeting five GNN-based techniques across four critical hardware security domains: intellectual property piracy, hardware Trojan detection/localization, reverse engineering, and hardware obfuscation. Understand the implications of achieving a 100% success rate in generating adversarial circuits that consistently fool GNN-based defenses, highlighting potential vulnerabilities in current hardware security implementations.