Abstract

As urban infrastructures become increasingly autonomous incorporating connected vehicles, smart-traffic systems, and digitally orchestrated emergency services they face escalating risks from coordinated cyberattacks that exploit their interdependencies. This manuscript proposes a framework leveraging deep reinforcement learning (DRL) to secure autonomous urban systems against such advanced threats. The proposed approach models the urban infrastructure as a networked multi-agent system in which defender agents learn policies to detect, contain and recover from attacks in real-time; adversarial behaviour is simulated via virtual attacker agents to improve robustness. Through this learning paradigm, the defender adapts dynamically to previously unseen threat vectors, reducing reliance on static rule-based safeguards. Empirical evaluations in a simulated urban operations environment show that DRL-based defenders reduce system downtime and service disruption under coordinated multi-vector attacks compared to baseline reactive strategies. The results highlight not only improved resilience but also reduced false-positive intrusion responses, enabling smoother continuity of service. The contributions include (i) modelling of the coordinated attack-defence scenario within autonomous urban systems; (ii) a DRL architecture tailored for real-time decision-making in complex multi-agent domains; and (iii) experimental validation demonstrating measurable gains in resilience and adaptability. The findings suggest that integrating DRL into urban defence architectures offers a promising pathway for future smart-city cybersecurity.