1374257 - Stealthy Cyberattack Modeling and Impact Analysis in Mixed Traffic Flow based on Deep Reinforcement Learning

Connected and autonomous vehicles (CAVs) have great potential to enhance traffic safety and mobility, but are also susceptible to cyberattacks, which can dramatically reduce these benefits. Analogous to cyber attackers targeting computer network performance, CAV cyber attackers can target traffic performance. While CAV cyberattack impacts for CAVs have been investigated, the impacts of stealthy attacks, which also commonly exist in control and wireless communication systems, have not been explored for CAVs. Stealthy attacks can potentially be more harmful due to the difficulties in detecting them. In this study, we examine the impacts of stealthy attacks on traffic flow in a mixed traffic environment containing human-driven vehicles (HDVs), normal CAVs, and compromised CAVs controlled by stealthy attackers. To understand the behavior of stealthy attackers and their impacts, a framework is proposed to investigate the interactions and behaviors of such attacks in a microscopic traffic simulation environment. Normal CAVs are designed with capabilities to smooth traffic and compromised CAVs aim at degenerating the traffic system performance. Deep reinforcement learning (DRL) agents are trained for normal and compromised CAVs and their capabilities are validated for different traffic conditions. Numerical studies indicate that traffic degeneration caused by compromised CAVs can dominate the smoothing efforts of normal CAVs. Besides, they illustrate the existence of an effective traffic-smoothing range for normal CAVs. The stealthy attacker’s behavior analysis also provides directions for detection mechanism design to defend against such advanced attacks.