Cost-Efficient Edge Federated Learning Over Multiple Base Stations for the ITS Based on Connected and Autonomous Vehicles

To realize the benefits expected and ensure user privacy and data security simultaneously, a cost-efficient edge federated learning (FL) architecture over multiple base stations (BSs) is proposed for the intelligent transportation system (ITS) based on connected and autonomous vehicles (CAVs). Firstly, in the proposed FL architecture, the road side units (RSUs) are designed to train the machine learning (ML) model with the BSs equipped with edge servers collaboratively. In this way, since the autonomous vehicles do not participate in model training, the negative impact of unreliable communication caused by vehicle mobility can be eliminated. Then, considering that the limited amount of data involved within the coverage of a single base station (BS), the FL architecture over multiple BSs at network edge is proposed for better learning performance. Along this line, the joint edge aggregation and association problem is studied, and a set function optimization problem is formulated with the objective of minimizing the costs considering latency and energy consumption comprehensively. Finally, after analyzing the complexity, monotonicity, and modularity of the problem formulated, the NP-hardness optimization problem is further decomposed and transformed, and an innovative solution is proposed. The abundant simulation results demonstrate the superior performance of the cost-efficient FL architecture proposed.