A Vision-Based Covert Attack and Detection System for Autonomous Vehicles Using Generative Adversarial Network

Autonomous vehicles (AVs) are advancing fast and are vulnerable to covert, intelligent cyber-attacks. In this paper, a novel covert attack and detection method is introduced. The proposed method is called VCA-HADS, which consists of a vision-based covert attack (VCA) and a hybrid adversary detection system (HADS), which can be applied to a driver assistance system (DAS) of AVs. The VCA would drive the AV out of the lane. Keeping deviation effects hidden, VCA employs a generative adversarial network to manipulate the vision sensor's output such that the AV is positioned on and aligned with the road's center line. To detect malicious behavior of AVs, a HADS is developed using a customized deep neural network, GPS data, and a road map. The DAS is also equipped with a decision-making algorithm to investigate the possibility of a VCA and to warn the driver in high-risk driving situations. To evaluate the performance of VCA-HADS in a multi-agent lane-keeping mission, various 3D driving simulations are performed. Based on the simulation results, the proposed covert attack and detection mechanism are valid and effective.