Mission-aware BeiDou spoofing defense in UAV swarms with LLM-assisted context validation

BeiDou navigation spoofing poses a significant threat to decentralized Unmanned Aerial Vehicle (UAV) swarms, as the compromise of one drone is sufficient to disrupt mission execution. OrbitGuardNet addresses this challenge by integrating transformer sequence encoders, graph neural networks (GNN), and Kalman filter residual analysis to detect spoofing signatures with high precision. A specialized large language model adds context-aware detection and enables drones to confirm threats. The control system adjusts formation, reassigns leaders, and prioritizes trusted data so the swarm can fly during attacks. Tests in simulated urban missions and real flights demonstrated a detection accuracy of up to 98.1%, with false alarms under 3%, and strong formation keeping even under threat. All validation was performed under controlled conditions using software-injected spoofing scenarios. The evaluation further included pilot real-flight tests conducted with three UAVs. Throughout these experiments, no open-air RF spoofing was employed. Consequently, the reported performance reflects controlled testing conditions rather than full-scale real-world deployment robustness. Alerts were received within 2.3 s and allowed countermeasures to be applied before navigation errors spread. Processing demands stayed within the limits of UAV hardware. Compared to other spoofing defenses, OrbitGuardNet stands out by combining accurate detection with an active, swarm-wide response against both external spoofers and compromised drones. The approach offers a practical defense for BeiDou-based UAV swarms in smart city tasks. Future work will focus on running entirely onboard, making models lighter for small UAVs, and testing larger, more complex real-world attack scenarios.