Cyber Attack-Defense Game Strategy Solving Based on Reinforcement Learning and Self-play
Jie Zhang, Yunfeng Luo · 2024
Cybersecurity is fundamentally a strategic game between two opposing entities. Strategy analysis from a game-theoretic perspective enables participants to make more informed and rational decisions. In contrast to traditional game theory problems, the dynamics of cyber attack-defense are in constant flux, and conventional game analysis methods often prove inadequate due to their failure to model the extensive state space accurately. Consequently, this study proposes the adoption of reinforcement learning and self-play techniques, utilizing the representational power of deep learning to comprehend the state space, and exploiting the decision-making prowess of reinforcement learning. The introduction of self-play allows agents to iteratively refine their strategies throughout the confrontation, thereby facilitating the strategic evolution of the involved agents. When strategies that exclude self-play are compared under identical initial conditions, the proposed method has improved the attacker's score to at least 0.985, while simultaneously preventing a deduction of 0.015 points for the defender. This attests to the method's effectiveness and its practical applicability.