Reinforcement Learning for Autonomous Resilient Cyber Defense

This Black Hat conference talk explores UK Defence research on applying Reinforcement Learning (RL) to automated cyber defense decision-making systems that can respond at machine speed to sophisticated cyber attacks. Learn about the development of advanced simulators, tools, and adversarial models designed to improve defender robustness. The presentation covers promising approaches including Multi Agent RL (MARL) and deep RL combined with heterogeneous Graph Neural Networks (GNNs). Discover practical demonstrations across various domains including Cyber First Aid, industrial control systems, and autonomous vehicles, showing that autonomous cyber defense is feasible on representative networks. The talk features insights from a large collaborative team of scientists and engineers from organizations including DSTL, QinetiQ, Frazer-Nash Consultancy, Cambridge Consultants, BAE Systems, BT, and BMT, who share their progress and plans to expand high-fidelity projects in this critical defense area.