Poster: CMA-FL: Cognitive Multi-Agent Federated Learning for Resource-Aware Drone Communication Attack Detection

Abstract

Drone communication networks increasingly support surveillance, logistics, disaster response, and tactical coordination, where a compromised link can quickly affect mission safety. Intrusion detection systems (IDSs) in such settings are challenging because attack evidence is distributed across drones and edge nodes, while centralizing raw data adds delay, bandwidth cost, and privacy exposure. Federated learning (FL) offers a natural way to train IDS models without moving raw data; however, conventional FL still gives limited attention to mobility-sensitive link quality, resource heterogeneity, trust, and security-rule consistency when deciding which client updates should influence the global detector. This paper proposes CMA-FL, a cognitive multi-agent assisted FL framework for resource-aware intrusion detection in drone communication networks. Each selected drone-edge client trains a long short-term memory (LSTM)-based temporal detector, while task-specific client-side agents evaluate signal/data quality, security-rule consistency, resource state, local training behavior, and update utility. Server-side trust, client-selection, weighting, aggregation, and explanation agents then govern participation and reliability-weighted aggregation. On the evaluated public drone-communication benchmark, CMA-FL achieves 99.6% accuracy with 0.99 precision, recall, and F1-score and reduces response time by more than 80% relative to the evaluated FL, edge-cloud, and cloud-only alternatives. These results characterize the controlled benchmark setting, generalization to independently collected flight traces, stronger non-IID partitions, and adversarial clients remains to be established.