VTFL: A Blockchain Based Vehicular Trustworthy Federated Learning Framework

Abstract

With the rapid development of 5G and Internet of Vehicle (IoV) technology, vehicles require a mass of data-sharing to ensure the traffic safety and improve user’s driving experience. However, the traditional way of sharing the original data leads to inefficient communication and the risk of privacy leakage when data leaves the vehicle’s local-storage. The recent advances of privacy-preserving Federated Learning (FL) bring a solution to these challenges, which allows to train models locally based on user’s data and only transmit the parameters with the raw data. But the lack of identity management in the dynamic topology of vehicles makes the FL’s training process vulnerable to be attacked by malicious vehicles, which spoilages the accuracy of the global model seriously. Therefore, how to ensure the security of the FL process has become a crucial challenge. In our work, we propose a trustworthy FL framework based on blockchain, which relies on consensus mechanism to ensure safe aggregation of models, and uses digital signatures to ensure the traceability of malicious vehicles and prevent the malicious vehicles from doing harm again. Our experiment results show that the proposed scheme effectively improves the precision of global model in the malicious scene and the efficiency of model convergence.