Privacy-Preserved Federated Reinforcement Learning for Autonomy in Signalized Intersections

Abstract: Control of autonomous vehicles (AV) in complex signalized intersection scenarios presents a significant challenge. Successful navigation in such environments necessitates a substantial amount of data for training autonomous driving models. However, relying solely on a single AV for this data limits the scope of training scenarios, and the sensory and computational capabilities may fall short in complex intersection settings. The emergence of Connected and Autonomous Vehicles (CAVs) introduces a transformative concept by facilitating the real-time exchange of traffic information among vehicle networks. With this new paradigm shift, a vehicle's driving behavior will no longer be solely based on the driver's limited and incomplete observation. Leveraging shared information allows CAVs to exhibit significantly improved driving behaviors that are more accurate, reliable, and efficient. Nevertheless, the imperative of preserving user privacy and ensuring data security poses a huge challenge in the context of AI-enabled CAVs. Addressing this challenge while still enabling information sharing is pivotal. To achieve this, this paper proposes a federated reinforcement learning framework designed to establish a privacy-preserving knowledge-sharing strategy. This strategy aims to enhance the precision and adaptability of autonomous driving at signalized intersections. The proposed framework efficiently aggregates and reuses the knowledge learned by diverse CAVs operating in varying intersection environments. This collective knowledge empowers more reliable decision-making for autonomous driving, making intersections safer and more efficient.