Advancing Security and Trust in WSNs: A Federated Multi-Agent Deep Reinforcement Learning Approach

Wireless Sensor Networks (WSNs) show significant potential through their ability to collect and analyze real-time data, notably enhancing various sectors. The new emerging security threats present a severe risk to the security and reliability of WSNs. Data-driven Artificial Intelligence (AI) leverages WSNs data to deal with new emerging threats like zero-day attacks. However, AI-based models suffer from poor adoption due to the lack of realistic/up-to-date attack data. Recently, Multi-Agent Deep Reinforcement Learning (MARL) has gained significant attention for enhancing Intrusion Detection Systems (IDS) capabilities. MARL offers improved flexibility, efficiency, and robustness. However, this requires data sharing, leading to network bandwidth consumption and slower training. Additionally, the curse of dimensionality hampers its benefits, given the exponential expansion of the state-action space. Privacy-aware collaborative methods such as Federated Learning (FL) emerge as a new approach, enabling decentralized model training across a network of devices while preserving the privacy of each participant. In this context, we introduce a novel framework (MAF-DRL) that leverages FL and MARL to efficiently detect WSN-based attacks. MAF-DRL enables distributed learning across multiple agents with adaptive, flexible, and robust attack detection. We also introduce a trust-based scheduling mechanism that dynamically allocates resources based on agent reliability. This trust-aware approach allows FL systems to adapt to changing network conditions and device behaviors. By prioritizing reliable devices, our method improves the energy efficiency of WSNs and enhances the resilience and effectiveness of the distributed FL paradigm. Finally, we assess the robustness of our framework by testing it against real-world WSN attacks. This evaluation demonstrates its efficiency for secure and communication-efficient federated edge learning across various agents.