Dual-Layer FL and Blockchain Empowered High Accurate Edge Training Framework

With the popularization of the smart city and the explosion of vehicles, the emergence of large amounts of data has boosted research. Distributed machine learning (DML) is an effective solution to improve the accuracy and timeliness. However, schemes based on traditional DML have the problems of high network delay, vehicle privacy leakage, and falling to provide customized models for multiple users. To solve the above problems, this paper constructs a dual-layer federated learning and blockchain-empowered high-accuracy edge training (DFLB-ET) framework for assisted driving. First of all, a dual-layer semi-asynchronous federated learning (FL) mechanism based on blockchain and local Directed Acyclic Graph are proposed to improve the efficiency and accuracy of FL and achieve high-accuracy model sharing among edge servers. Secondly, a regional node selection algorithm is proposed based on the mobility and model accuracy of vehicles, so as to better utilize the computational resources of Road-Side Units. Simulation results show that the proposed DFLB-ET framework outperforms both the local training and synchronous/asynchronous FL schemes in terms of the training delay and model accuracy.