Federated hierarchical MARL for zero-shot cyber defense

Cyber defense systems face increasingly sophisticated threats that rapidly evolve and exploit vulnerabilities in complex environments. Traditional approaches which often rely on centralized monitoring and static rule-based detection, struggle to adapt to new, crafted, and novel attack patterns. This paper presents the Adaptive Zero-Shot Hierarchical Multi-Agent Reinforcement Learning (AZH-MARL) framework, a novel approach that integrates hierarchical reinforcement learning, zero-shot learning capabilities, and federated knowledge sharing to build resilient cyber defense systems. The hierarchical structure decomposes complex defense tasks into specialized sub-tasks managed by agents, reducing the learning problem's complexity and enabling more efficient coordination. The zero-shot learning component allows the framework to recognize and response to previously unseen attack patterns through semantic mapping. Furthermore, the federated learning learning component facilitates for knowledge sharing across network domains while preserving data privacy, enabling collaborative defense without exposing sensitive information. The detailed evaluation demonstrates that our approach significantly outperforms existing methods across a range of scenarios. It achieves a high detection rate of 94.2% for known attacks and 82.7% for zero-day exploits, while maintaining a low false positive rate of 3.8%. This robust performance extends to the most sophisticated threats, achieving an 87.3% containment rate against Advanced Persistent Threats (APTs). The framework's zero-shot capability is underpinned by a semantic mapping accuracy of 89.3%, which enables rapid adaptation to novel threats. Consequently, the mean response time is reduced by 35% for known attacks and 42% for zero-day exploits compared to the best-performing baseline. Finally, the federated learning architecture proves highly efficient, reducing communication overhead by 45% while preserving privacy. These results collectively demonstrate our framework's potential to set a new standard for resilient and adaptive cyber defense in complex, distributed environments.