BDFL: A Byzantine-Fault-Tolerance Decentralized Federated Learning Method for Autonomous Vehicle

Autonomous Vehicles ( $AV$ s) take advantage of Machine Learning (ML) for yielding improved experiences of self-driving. However, large-scale collection of $AV$ s' data for training will inevitably result in a privacy leakage problem. Federated Learning (FL) is proposed to solve privacy leakage problems, but it is exposed to security threats such as model inversion, membership inference. Therefore, the vulnerability of the FL should be brought to the forefront when applying to $AV$ s. We propose a novel Byzantine-Fault-Tolerant (BFT) decentralized FL method with privacy-preservation for $AV$ s called BDFL. In this paper, a Peer-to-Peer (P2P) FL with BFT is built by extending the HydRand protocol. In order to protect their model, each $AV$ uses the Publicly Verifiable Secret Sharing(PVSS) scheme, which allows anyone to verify the correctness of encrypted shares. The evaluation results on the MNIST dataset have shown that introducing decentralized FL into $ AV$ area is feasible, and the proposed BDFL is superior to other BFT-based FL method. Furthermore, the experimental results on KITTI dataset indicate the practicality of BDFL on improving performances of multi-object recognition in $ AV$ areas. Finally, the proposed PVSS-based data privacy preservation scheme is also justified its characteristic of no side-effect on models' parameters by the experiments on the MNIST and KITTI datasets.