Federated Learning With Dynamic Staleness Correction for Privacy Protection in Vehicular Networks

Edge intelligence combines mobile edge computing(MEC) with artificial intelligence(AI), which has great potential for improving the security and efficiency of data-driven intelligent transportation systems (ITS) and emerging Internet Of Vehicles (IOV) services. However, when big data and AI empower ITS, the information security issue in vehicular networks needs to be paid more attention. In order to protect the privacy and security of vehicular training data, we propose a privacy protection federated learning scheme with staleness asynchronous update to overcome the differences caused by vehicle heterogeneity. Moreover, unlike the traditional weighted average on the server side only by the number of samples, our solution introduces dynamic time weights according to the calculation and communication capabilities of different vehicles, to make full use of the previously trained local model. The proposed scheme is able to reduce traffic load of the network and improve learning performance while enhancing security and privacy. Experiment results demonstrate that in terms of communication cost and model accuracy, the performance of the proposed asynchronous federated learning is better than the benchmark algorithm, and it also achieves good performance under NIID data.