Scalable and Generalizable RL Agents for Attack Path Discovery via Continuous Invariant Spaces
Date:
In this talk, my second Ph.D. paper was presented at RAID 2025. The paper addresses identifying critical attack paths in a network—sequences of vulnerabilities an attacker can exploit—and the limitations of current RL approaches, which require retraining for each network due to discrete input/output representations. We propose continuous and invariant input/output spaces for RL agents, enabling transferable policies across diverse networks and vulnerability sets. The contributions also included an extended scenario generation pipeline and an enhanced outcome approximation module.