Detecting Stealthy Cyberattacks on Automated Vehicles via Generative Adversarial Networks

The emergence of vehicles with driver-assist features, including adaptive cruise control (ACC) or other automated driving capabilities, introduces the possibility of cyberattacks where a select number of automated vehicles (AVs) are compromised to drive with adversarial controls. While obvious attacks that force vehicles to crash may be easily detectable, more subtle attacks are harder to detect and could change vehicle driving behavior, resulting in a network-wide increase in congestion and fuel consumption. To address this pressing problem, we first characterize two scenarios of potential cyberattacks, namely malicious attack on individual vehicles and data injection attack on sensor measurements. Then, a generative adversarial network (GAN)-based anomaly detection model is proposed for real-time detection of such attacks. Finally, to demonstrate the effectiveness of the proposed method we conduct numerical experiments using both synthetic vehicle trajectory data, and real-world ACC trajectory data with synthetic sensor attacks injected. The results show that the proposed model is capable of detecting attacks using a short period of trajectory data.