Security-Focused Training Model of Reinforcement Learning in Autonomous Vehicles

Reinforcement Learning (RL) in autonomous ve-hicles (AV s) is expected to enhance the safe maneuvering of AV s to improve road safety. However, existing literature on AV s focuses on the impacts of image perturbations as adversarial examples (AEs) during the testing phase. Limited attention has been given to more intrusive types of AEs, such as vehicles with adversarial intent to induce accidents on the road proactively. Without addressing this type of AEs, the learned policy remains vulnerable to different types of AEs, making the RL unusable in AV networks, given the nature of cyber-physical systems (CPS), for which negative consequences include accidents, property losses, and injuries. We focus on the training phase to address this gap and fortify the learned policy using our expanded AE definitions. This paper presents our approach to realizing this training model to build a more robust policy against adversaries.