Dual-Branch Graph Attention Network for Cyber-Physical Intrusion Detection System for Unmanned Aerial Vehicles

Unmanned aerial vehicles (UAVs) are increasingly involved in critical missions which has dramatically raised their vulnerability to cyber-physical attacks. Existing intrusion detection systems (IDSs) operate with separate processing of cyber and physical telemetry through synchronized or handcrafted features, which limits their robustness against complex attacks. In this paper, we propose a novel Dual-Branch Graph Attention Network (Dual-GAT) architecture that models cyber and physical UAV data streams as asynchronous graphs. Our method independently learns modality-specific representations through parallel graph attention mechanisms and fuses them at the decision level for improved attack classification. To evaluate the effectiveness of this approach, a publicly available UAV cyber-physical dataset is used. The dataset includes records from benign operations, Denial-of-Service (DoS) and replay attacks Our Dual-GAT model demonstrates superior performance compared to cyber-only and physical-only systems and early-fusion approaches by detecting replay attacks that need domain context.