Facing airborne attacks on ADS-B data with autoencoders

Abstract

The automatic dependent surveillance-broadcast (ADS-B) represents a major change in flight tracking and it is one of the key components in building the next generation of air transportation systems. However, several concerns have been raised regarding its vulnerabilities to cyber attacks. In recent years, a new and promising approach of utilizing large-scale and publicly available flight recordings for training machine learning models that can detect anomalous flight patterns has been demonstrated as a valid countermeasure for several ADS-B attacks. The new approach differs significantly from previously proposed methods in the simplicity of its integration with the current ADS-B system. It also provides a valid countermeasure against highly sophisticated airborne attackers. However, previously proposed machine learning methods require training a different model for each flight route or geographic location to give acceptable results. This requirement limits the current solution to flights with a sufficient amount of historical data, which is unavailable in many cases such as business aviation, instructional flying, aerial work, and more. In this research, we address this limitation of previous work, by applying a differencing time-series transformation on the ADS-B data and utilizing a non-recurrent autoencoder classifier. The effectiveness of our method is compared to existing methods on several simulated trajectory modification attacks. The results of our experiments show that the proposed method achieves a ROC AUC value of 0.935-0.951, in comparison to 0.627 from existing methods when evaluated on flights that are absent from training data.