End-to-end deep learning for flight trajectory reconstruction from multi-station ADS-B measurements

In the field of aviation safety, ADS-B is widely adopted as an active broadcast-based aviation surveillance system, enabling aircraft to broadcast their real-time position information via onboard devices. However, the open communication protocol relied upon by ADS-B has led to increasingly frequent GPS spoofing and network hijacking attacks, posing significant threats to communication security. A reliable secondary verification method is urgently needed to revalidate the position information broadcast by aircraft. To address this issue, this paper proposes a deep learning framework that does not rely on specific message content but directly calculates aircraft trajectories through tamper-proof electromagnetic signals. Based on real flight trajectories and distributed sensor signals collected from the OpenSky dataset, we trained an End-to-End neural network, innovatively introducing heterogeneous sensor encoders and trajectory decoders, and demonstrated the effectiveness of the proposed model through empirical experiments. In comparison experiments, TIGER V2 achieved the lowest MDE of 38.9484 km, corresponding to a 10.21% reduction relative to the strongest non-TIGER sliding-window baseline, while also improving longitude accuracy and most latitude point-wise error metrics. Finally, through ablation experiments and visualization analysis, we demonstrate the necessity of each component of the model and the overall effectiveness of trajectory reconstruction.