Explainable Reinforcement Learning Models for Adaptive Cyber Defense in Autonomous Vehicles

Abstract

[1] Cybersecurity continues to be an increasingly critical challenge for autonomous cars, and intelligent attackers can develop adversarial real-time attack strategies that are difficult to anticipate or mitigate. With conventional rule-based cybersecurity strategies and solutions, such as those used in traditional ITS (intelligent transportation system) scenarios, autonomous cars will not be able to anticipate all these possible real-time adversarial strategies to which they will be subject, thereby receiving suboptimal attack resilience. On the other hand, deep reinforcement learning (DRL)-based cyber defense strategies are able to model and optimize dynamic, multi-stage, and uncertain strategic games against adversarial black-box attacks. Unfortunately, this solution space over the state-action space of an autonomous car’s DRL policies contains adversarially optimized policies or black-box adversarial actions, and their stop effect can catastrophic.[2] There has been tremendous growth and advancement of technologies used in intelligent traffic systems (ITS), autonomous vehicles (AV), and intelligent transportation systems. This advance has made it possible to carry out various routine activities such as traffic management, dynamic route decision making, and adaptive cruise control. However, this smart mobility system employing intelligent transportation technology has the potential to be exploited by the cyber attackers. The exploitations can result in devastating outcomes such as the violation of data privacy and digital security, road traffic congestion, and disruption of local services due to malfunctioning traffic lights and automatic cars. Hence, it is vitally important to understand the inherent cyber vulnerabilities, and study the potential impacts of cyberattacks to the mixed traffic flow of various levels of connected and automated vehicles. The appropriate framework for evaluating the prevailing infrastructure and requirements of cyberattack resilient control strategies is intensely demanded.