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

Area of Science:

• Aviation safety
• Deep learning
• Signal processing

Background:

• Automatic Dependent Surveillance-Broadcast (ADS-B) is a key aviation surveillance system.
• Its open protocol makes it vulnerable to GPS spoofing and network hijacking.
• A secondary verification method is needed to ensure aircraft position accuracy.

Purpose of the Study:

• To propose a novel deep learning framework for secondary aircraft position verification.
• To develop a method that relies on tamper-proof electromagnetic signals, not message content.
• To enhance aviation communication security against spoofing and hijacking attacks.

Main Methods:

• Developed an End-to-End deep learning framework named TIGER V2.
• Utilized heterogeneous sensor encoders and trajectory decoders.
• Trained the model on real flight trajectories and distributed sensor signals from the OpenSky dataset.

Main Results:

• TIGER V2 achieved a Mean Distance Error (MDE) of 38.9484 km, a 10.21% improvement over the strongest baseline.
• Demonstrated enhanced longitude accuracy and point-wise error metrics for latitude.
• Ablation studies confirmed the necessity of each model component.

Conclusions:

• The proposed deep learning framework effectively reconstructs aircraft trajectories using electromagnetic signals.
• TIGER V2 offers a reliable secondary verification method, significantly improving aviation safety.
• The model addresses critical security threats posed by GPS spoofing and network hijacking.