ADS-Bpois: Poisoning Attacks Against Deep-Learning-Based Air Traffic ADS-B Unsupervised Anomaly Detection Models

As a core technology of the new generation air traffic management (ATM) system, automatic dependent surveillance-broadcast (ADS-B) becomes increasingly crucial and its anomaly detection is important for safeguarding flight safety and enhancing the efficiency of air traffic. Deep learning has been used for ADS-B flight trajectory prediction and anomaly detection, but it is known for its susceptibility to poisoning attacks when updated to adapt feature distributions from new ADS-B flight data. In the study, we propose time neighborhood interpolation combined back gradient descent poisoning attacks which not only make full use of the gradient information of ADS-B anomaly detection models but also fully consider temporal correlations and maneuvering characteristics of ADS-B data during the aircraft’s take-off, climb, turning, and descent phase. The experimental results show that our proposed poisoning attack method can successfully poison the four state-of-the-art deep learning-based ADS-B time series unsupervised anomaly detection models. What is more, the experimental results also show that our method is superior to the other poisoning attack methods in terms of success rate and stealthiness. To the fullest extent of our knowledge, we exhibit, for the first time, the susceptibility of ADS-B time series unsupervised anomaly detection models to poisoning attacks, which is vital in safety-critical and cost-critical air traffic industry. For ease of understanding, our proposed poisoning attack method is referred to as ADS-Bpois.