A Federated Learning-Based Hybrid Architecture for Multi-Class Attack Detection in UAVs Communications

초록

Unmanned Aerial Vehicle (UAV) networks face cybersecurity challenges due to their dynamic topology, distributed nature, and limited resources. Conventional centralized intrusion detection systems are impractical for UAV environments because of privacy concerns, bandwidth constraints, and intermittent connectivity. This paper proposes FL-GCN, a privacy-preserving federated learning(FL) framework that integrates Graph Convolutional Networks (GCN) with distributed model training for multi-class attack detection in UAV communications. The proposed system models UAV interactions as graph-structured data, where the nodes represent UAVs and the edges represent communication flows. The hybrid detection mechanism combines GCN-based node embeddings with edge-level multilayer perceptron (MLP) classifiers to effectively distinguish between attack types. The framework employs FL to enable collaborative model training across distributed UAV clients without sharing raw traffic data, thereby preserving privacy. We investigated three aggregation strategies (FedAvg, FedProx, and FedNova), where FedNova demonstrated superior convergence stability and detection performance. Experimental evaluation shows that FL-GCN achieves 96% accuracy with robust performance across multiple attack classes, including Blackhole, Flooding, Sybil, and Wormhole attacks. The results demonstrate lower false-positive and false-negative rates than those of federated CNN, RNN, and MLP baselines, establishing FL-GCN as an effective, scalable, and privacy-conscious intrusion detection solution for UAV networks.

키워드