A Layered Reference Model for Penetration Testing with Reinforcement Learning and Attack Graphs

This paper considers key challenges to using reinforcement learning (RL) with attack graphs to automate penetration testing in real-world applications from a systems perspective. RL approaches to automated penetration testing are actively being developed, but there is no consensus view on the representation of computer networks with which RL should be interacting. Moreover, there are significant open challenges to how those representations can be grounded to the real networks where RL solution methods are applied. This paper elaborates on representation and grounding using topic challenges of interacting with real networks in real-time, emulating realistic adversary behavior, and handling unstable, evolving networks. These challenges are both practical and mathematical, and they directly concern the reliability and dependability of penetration testing systems. This paper proposes a layered reference model to help organize related research and engineering efforts. The presented layered reference model contrasts traditional models of attack graph workflows because it is not scoped to a sequential, feed-forward generation and analysis process, but to broader aspects of lifecycle and continuous deployment. Researchers and practitioners can use the presented layered reference model as a first-principles outline to help orient the systems engineering of their penetration testing systems.