Keywords
CPS, Meta-learning, Security

Technologies
CyberBattleSim, Python, DQL

Traditional Cyber-Physical System (CPS) security relies on static defense mechanisms that fail to adapt to sophisticated cyber-attacks. We investigate the application of adaptive learning algorithms to model the interaction between attackers and defenders. We also simulate a network environment where an attacker learns to exploit vulnerabilities of the system while a defender applies self-learning techniques to protect the network and prevent from future attack by patching the system.

Preliminary findings show that a defender capable of analyzing past attacks and learn their patterns can significantly increase the "Mean Time to Compromise" (MTTC). This research provides a robust framework for transitioning from reactive response to predictive cyber-defense. Specifically, the implementation of a Self-Learning Defender (SLD) utilizing meta-learning and Dijkstra path-prediction algorithms resulted in a 93% reduction in attacker success rates.

By identifying and pre-emptively blocking preferred lateral movement paths, the framework delayed initial system breaches by over seven-fold, effectively neutralizing the co-evolutionary advantages of AI-driven threats.