Automated penetration testing, powered by reinforcement learning (RL), has gained prominence for reducing human effort and increasing reliability. However, dealing with the rapidly expanding scale of modern network infrastructure, the current RL-based methods have shown some limitations. They often struggle with the huge action spaces in large networks, resulting in less effective and monotonous penetration testing strategies. To solve this challenges, a research team led by Dr. Xin Liu published their new research on 15 Mar 2025 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

Their research introduces an innovative RL agent, CLAP, aiming to change how we do network security assessments. CLAP deals with the challenge by proposing the coverage mechanism, a novel neural network, to deal with large action spaces in large networks. It employs a Chebyshev decomposition critic to identify diverse adversary strategies.

Experimental results demonstrate CLAP's superiority, reducing attack operations by almost 35% compared to other methods. Notably, CLAP enhances training efficiency and stability, enabling effective penetration testing over large-scale networks with up to 500 hosts. The agent also excels in discovering pareto-dominant strategies that are both diverse and effective in achieving multiple objectives. CLAP's unique approach and impressive results mark a significant leap forward in cybersecurity research, addressing the challenges posed by the expanding scale of modern network environments.

DOI: 10.1007/s11704-024-3380-1