SR-GNN: Spectral Residuals with Graph Neural Networks for Anomaly Detection in ADS-B Data

The Automatic Dependent Surveillance–Broadcast (ADS-B) system is an essential component of the air traffic system. However, its protocol lacks data encryption and authentication, making it highly vulnerable to various deceptive attacks. ADS-B data represents typical multivariate time series, and existing solutions have limitations as they fail to capture relationships between features. In this paper, we propose a new self-supervised framework called SR-GNN. It leverages spectral residual (SR) to identify and remove point anomalies and utilizes two graph neural networks to capture relationships between time series. Finally, the model is jointly optimized using prediction and reconstruction. To address this issue, we create an evaluation dataset using commonly used trajectory modification methods. Through extensive experimentation, we demonstrate that the proposed method significantly outperforms other models.