FedGraph-ID: A Federated Graph Learning Framework for Intrusion Detection in UAV Networks Under Adversarial Settings

Intrusion detection in UAV networks requires decentralized learning mechanisms that preserve data privacy while remaining robust to Byzantine failures caused by compromised drones. Existing federated learning approaches suffer from two fundamental limitations: (1) they rely on flat traffic feature representations that ignore the intrinsic graph structure of network communications, and (2) they implicitly assume benign client behavior, resulting in severe performance degradation even under moderate adversarial participation. This paper presents FedGraph-ID, a federated graph learning framework for robust intrusion detection in adversarial UAV networks. FedGraph-ID introduces a sliding-window graph construction mechanism that transforms raw network flows into temporal graph representations, enabling Graph Convolutional Networks (GCNs) to jointly capture spatial communication patterns and temporal attack dynamics. To tolerate Byzantine behavior, we design HYDRA, a hybrid robust aggregation strategy that combines statistical anomaly filtering with trust-aware weighted model fusion, effectively suppressing malicious updates while preserving convergence. Extensive experiments on the CICIDS-2017 benchmark and a newly developed UAVIDS-2025 dataset show that FedGraph-ID achieves 96.29% detection precision in benign settings and maintains 92.21% precision with up to 50% Byzantine clients, a regime in which existing methods fail. Under extreme adversarial conditions, FedGraph-ID outperforms state-of-the-art defenses by up to 79 F1-score points, demonstrating strong robustness and scalability for federated intrusion detection in UAV networks.