Machine Learning Security of Connected Autonomous Vehicles: A Systems Perspective

Machine Learning security is vital for the safe operation of Autonomous Vehicles. When Autonomous Vehicles are connected and cooperating, they form a system of systems that have shared objectives. However, adversarial environments and adversarial vehicles in the system can cause security challenges for the whole system. Current research focuses on the Machine Learning security challenges from the perspective of a single vehicle. We argue that there is a need to consider these security challenges from the perspective of multiple interconnected vehicles, as a system. In this paper, we explore these challenges from the perspective of many Connected Autonomous Vehicles as a system with respect to Machine Learning security. We include attack scenarios that demonstrate the system interactions that can lead to cascading failures, which test the resilience of the system. We also outline some of the challenges in researching this perspective, where a key challenge is identifying indicators and metrics to describe the system resilience when under attack. To observe the system, experimentation via simulation is identified as a suitable environment that can capture the complex and dynamic system interactions in this security context.