Security Management of Interdependent Critical Infrastructures by Deep Reinforcement Learning

Critical Infrastructures (CIs) are increasingly exposed to intentional attacks that can disrupt essential services and propagate across interdependent systems. Traditional approaches to security management often rely on static gametheoretical models or optimization schemes, which have limitations in capturing the dynamic and adaptive nature of adversarial interactions and the complexity of system-of-systems interdependences. This work introduces a novel framework for the security management of interdependent multi-state CIs using Deep Reinforcement Learning (DRL). By leveraging Adversarial Reinforcement Learning (ARL), where attacker and defender are modeled as competing agents with asymmetric knowledge and budgets, the security management policy is tested against intelligent adversaries, improving its robustness and operability. The proposed approach addresses two central challenges: i ) the combinatorial explosion of possible attack scenarios, and i i ) the need for effective decision-making in high-dimensional, uncertain environments. Through continuous interaction and learning, DRL provides adaptive defense strategies that go beyond static optimization, dynamically allocating resources to mitigate risks and reduce cascading effects across CIs. The framework is demonstrated through a case study involving interdependent power and water networks. Results illustrate how the proposed DRL-based framework can support operators in designing robust, adaptive, and resource-aware security management policies.

Keywords: Critical Infrastructures, Interdependent Systems, Security Management, Adversarial Reinforcement Learning.