Assisted driving system based on federated reinforcement learning

Abstract

For the visually impaired persons, the partial loss of vision brings some challenges to vehicle driving. How to assist the visually impaired to safely control the vehicle is an important issue. Reinforcement learning is used to train the driving control models on the cloud server based on the data collected from several vehicles. However, the images gathered by the camera on a vehicle imply the privacy of the drivers and passengers. In order to balance the model accuracy and the privacy protection, we design an assisted driving system based on federated reinforcement learning, called DFRL, which supports the secure model aggregation among several vehicles. Each vehicle trains the local model through the deep deterministic policy gradient algorithm, and transmits the encrypted weights to the server. Then, the server takes step-based federated averaging to obtain the global model, and distributes it to each participant. Finally, each vehicle decrypts the model and updates its local model accordingly. In this way, only the encrypted parameters rather than original data are shared between the vehicles and the cloud server, which stops the attackers to recover the data with privacy. The experiments on TORCS show that, DFRL improves the accuracy of driving control, while providing privacy protection for the users.