Edge AI-Powered Cyber-Physical Attack Mitigation: A Graph Reinforcement Learning Approach for Autonomous Threat Response

Abstract

Cyber-physical systems (CPS) at the edge require real-time autonomous defense mechanisms. We propose GRL-Shield, a graph reinforcement learning (GRL) framework that models CPS networks as dynamic graphs and autonomously mitigates multi-vector attacks. Using attention-based graph neural networks (GATs) and proximal policy optimization (PPO), GRL-Shield reduces false positives by 34% while maintaining 98.6% attack detection rate on the SWaT dataset. Edge deployment on NVIDIA Jetson AGX Orin shows sub-second response times, outperforming rule-based IDS by 60% in mitigation speed.