Content-based Vehicle Selection and Resource Allocation for Federated Learning in IoV
Siyu Wang, Fangfang Liu, Hailun Xia
Abstract
In order to use datasets collected from multiple vehicles to train a machine learning model while ensuring vehicle user privacy, federal learning framework was introduced into the Internet of Vehicles. Federated learning is a distributed learning framework. Under the federated learning framework, the packet error rate and wireless bandwidth have a great influence on the global model training process because the vehicle exchanges model parameters with the central server through the wireless channel. With limited bandwidth, the central server needs to select a more appropriate subset of vehicles candidates to participate in federated learning. In this paper, image classification is taken as a typical application in the Internet of vehicles. The dataset contents of different vehicles are different, and the selection of different subsets of vehicles will affect the accuracy and convergence rate of the global model. Therefore, an algorithm of vehicle selection and wireless resource allocation based on dataset content is proposed. Vehicles selection and wireless resource allocation are designed as an optimization problem by joint considerations of vehicle computing resource, datasets, and wireless resources with the goal of maximizing loss function decay of the global model. Finally, simulation with the CIFAR-10 dataset verifies that the vehicle selection and resource allocation algorithm based on the dataset content is superior to the baseline methods in terms of model accuracy and convergence rate.