Split and Federated Learning with Mobility in Vehicular Edge Computing

Abstract

Vehicular edge computing (VEC) is a promising technology to support vehicular applications that leverage machine learning (ML) technology. Due to limited resources of the vehicle, the vehicle uses Split learning (SL) to split the computation of the ML model and offload it to the VEC server (VECS). Federated learning (FL) is also used for data privacy and parallel training of the vehicles. Therefore, SplitFed learning, which combines SL and FL, enables parallel processing, which is an advantage of FL, and reduces the computational burden on the vehicle through ML model split, which is an advantage of SL. However, the SplitFed learning does not consider the mobility of device/vehicle. Therefore, we propose a SplitFed learning with mobility method to minimize the training time of the model. SplitFed learning with mobility method is a migration method of the ML model when the vehicle moves from the current serving VECS to the target VECS. Through simulations, compared with conventional SplitFed learning where the vehicle travels after 50% and 80% of training is completed, the proposed method can reduce training time by about 19-33% for LeNet and by about 22-44% for VGG16, respectively, and does not degrade accuracy of model.