Tri-stream multi-model architecture for real-time detection of BeiDou signal manipulation in UAV swarms.

This study presents a novel framework for detecting and mitigating BeiDou Global Navigation Satellite Systems (GNSS) spoofing in decentralized Unmanned Aerial Vehicle (UAV) swarm systems and addressed a critical vulnerability in autonomous aerial operations. The research tackles the problem of coordinated spoofing attacks that compromise swarm navigation by injecting counterfeit satellite signals, which disrupt formation stability and mission execution. To solve this, the proposed methodology combines Kalman filter-based residual tracking, a Tri-Stream Deep Fusion model incorporating CNN, LSTM, and GNN, and a transformer-based Large Language Model trained on UAV telemetry for contextual anomaly validation. The evaluation was conducted using a high-fidelity software-in-the-loop simulation platform that replicates various spoofing scenarios under realistic signal conditions. Simulated assessment outcomes showed that the system achieved a detection accuracy of 97% ± 0.6%, maintained swarm cohesion with trajectory deviations under 5 m, and sustained a 97% mission completion rate even during multi-source adversarial interference. These findings demonstrate that integrating statistical, temporal, and relational learning with decentralized consensus can enable real-time swarm resilience. This research contributes a scalable, sensor-compatible approach to GNSS security in UAV networks and offers a foundation for future work in edge-deployable LLM optimization, spoofing-jamming co-detection, and field validation in urban GNSS-denied environments.