Data-Driven DLT3 Federated Deep Reinforcement Learning for Secure and Efficient Autonomous Driving

As autonomous vehicles (AVs) become increasingly widespread, the intelligent driving control and safety concerns have emerged. Recent advents in Internet of Things (IoTs) and 6G technologies have vastly boosted AVs’ sensing and communication capabilities, thereby enabling the high-speed and real-time data exchange. However, efficiently utilizing data while safeguarding the privacy remains challenges. Thus, we propose a federated learning (FL) framework with the privacy protection to facilitate the deep reinforcement learning (DRL) for longitudinal vehicle safety tracking control. First, the proposed framework enables the efficient information sharing and policy updates via encrypted policy gradient transmission. Then, the twin delayed deep deterministic policy gradient (TD3) method makes the velocity decision in the local model, while the dual long short-term memory (LSTM) network captures the temporal trajectory information. Finally, stacked denoising autoencoders (SDAE) is used to extract high-level environmental features, and the dual-LSTM-TD3 (DLT3)-SDAE model is introduced to enhance the policy decision efficiency and accuracy. Experimental results demonstrate that within the FL framework supported by homomorphic encryption, the DLT3-SDAE-based AV agent surpasses other DRL methods in terms of collision avoidance and average rewards, in the LAF environment. The safety, efficiency and comfort of the DLT3-SDAE AV agent are also evaluated and demonstrated.