Joint Optimization of Model Partition and Resource Allocation for Split Federated Learning Over Vehicular Edge Networks

Split federated learning (SFL) has been regarded as an efficient paradigm to enable federated learning and reduce the computation burdens at the devices by allowing them to train parts of the model. However, deploying SFL over resource-constrained vehicular edge networks is challenging, and a cost-effective scheme is necessitated to minimize the total time and energy consumption of vehicular devices. To this end, we use an improved reinforcement learning method to present a joint optimization scheme that can efficiently determine the optimal model partition point for each vehicular device and the optimal allocations of the computing resource and bandwidth resource among all vehicular devices. Experimental results validate the effectiveness and performance advantages of our proposed scheme.

I. Introduction

Federated learning (FL) is a promising decentralized learning paradigm that has great potential to be deployed in vehicular edge networks [1]. FL allows each vehicular device to keep its own data and transmit the model parameters to the edge server for aggregation, alleviating the communication burden and privacy leakage [2]. However, with the development of artificial intelligence, the parameter size of deep learning models has increased rapidly, resulting in significant difficulty in efficiently performing the complete model calculation on resource-constrained vehicular devices [3]. Moreover, due to the mobility of vehicles, some vehicular devices are unable to complete the model training and parameter uploading within the communication coverage of a base station co-located with an edge server, making it challenging to maintain FL stability in vehicular edge networks [4].