Towards Efficient Federated Learning via Vehicle Selection and Resource Optimization in IoV

Federated Learning (FL) is a novel distributed machine learning paradigm that holds great potential for collaborative model training in Internet of Vehicles (IoV). However, employing FL in IoV presents several challenges. First of all, the heterogeneous vehicular resources may lead to the bias of vehicle selection for model training and result in poor performance with respect to data diversity. Moreover, executing FL tasks with limited resources and energy is also a non-trivial problem. In view of this, we investigate an efficient FL framework in IoV, and then formulate the Vehicle Selection and Resource Optimization (VSRO) problem, which aims at minimizing the energy consumption and training latency without jeopardizing model accuracy. On this basis, we first propose a vehicle selection algorithm based on Lyapunov optimization method that dynamically selects vehicles to participate in each round of FL, taking into account the computational and transmission capabilities of vehicles, as well as their participation frequency in FL tasks. Furthermore, a resource scheduling algorithm is proposed for the selected vehicles, which utilizes the Lagrange dual problem and sub-gradient projection to iteratively determine the optimal training process. The simulation results demonstrate that the proposed algorithms can achieve best performance on enhancing model accuracy and simultaneously reducing system energy consumption and latency.