Reconstruction-Based Anomaly Detection for ADS-B Surveillance Data : A Temporal Attention LSTM Autoencoder Prototype for DigitalAir Traffic Services

Sammanfattning

Modern Air Traffic Management (ATM) heavily relies on Automatic Dependent Surveillance-Broadcast (ADS-B) for real-time aircraft tracking. However, the inherent lack of encryption and authentication in the ADS-B protocol exposes digital air traffic services to significant cybersecurity risks, including data tampering, spoofing, and injection attacks. While traditional cryptography-based security solutions introduce significant deployment overhead and latency, machine learning offers a non-intrusive alternative for verifying data integrity through trajectory consistency checking. This thesis proposes a reconstruction-based anomaly detection framework tailored for multivariate ADS-B surveillance data streams. We design and implement a Temporal Attention Long Short-Term Memory (LSTM) Autoencoder prototype capable of modeling the complex temporal dynamics of aircraft trajectories. By utilizing an unsupervised learning paradigm, the model learns the normal underlying patterns of legitimate flights and detects anomalies based on reconstruction errors. To bridge the gap between deep learning complexity and operational trust in safety-critical aviation domains, we integrate Explainable Artificial Intelligence (XAI) using SHapley Additive exPlanations (SHAP) to provide feature-level interpretability for detected anomalies. The framework is evaluated using a curated dataset of real-world trajectories combined with simulated attack vectors, including trajectory deviation and spoofing scenarios. The experimental results demonstrate that the Temporal Attention LSTM Autoencoder achieves high detection accuracy and robust performance against various anomaly profiles. Furthermore, the SHAP-based explanations successfully isolate the specific data features contributing to the anomaly scores, providing air traffic controllers with actionable insights. This research underscores the potential of combining deep sequence modeling with explainability to enhance the security and trustworthiness of next-generation digital air traffic services.