Towards Efficient Federated Learning Using Agile Aggregation in Internet of Vehicles

Abstract

Federated learning is an enabling technology for the services in Internet of vehicles because it can effectively alleviate privacy issues in data circulation and diversified intelligent applications. However, existing federated learning methods still confront the problem of low computational efficiency when applied to the scenario of high-dynamic vehicles. In order to address this problem, we present a multistep federated learning architecture including task release, vehicle model training, identity authentication, quality assessment, and weighted aggregation processes, where a novel agile aggregation model is established. Based on the agile aggregation model, an efficient federated learning (EFL) algorithm is proposed, which is specifically applied to the scenario of high-dynamic vehicles during the training model. We theoretically analyze its advantages over the existing traditional federated learning algorithms in terms of efficiency, training accuracy, and security. Simulations are conducted using the MNIST dataset to evaluate the performance of the proposed EFL algorithm. The results show that EFL improves the efficiency of the federated training process with effective accuracy and privacy preservation.