A Data Privacy Risk Assessment Model for Federated Learning in the Internet of Vehicles

To address the issue of privacy leakage in the communication process of nodes performing federated learning (FL) in the Internet of Vehicles (IoV), and the challenge of defining and quantifying privacy when dealing with non-plaintext data interactions, we designed and implemented a privacy risk assessment framework for vehicular node data based on membership inference attacks. Firstly, we analyze the advantages of using membership inference attacks to assess the privacy of vehicular nodes. We then construct black-box and white-box simulation attacks, considering the different background knowledge that attackers may possess during the FL process. The results of these multi-source simulation attacks are then converted into privacy risk scores for the nodes. Finally, we validate the accuracy of the proposed privacy metrics by comparing the metric values before and after applying privacy protection measures. Experimental results show that the privacy risk assessment method for FL communication based on simulation attacks proposed in this paper exhibits strong generalizability across multiple privacy datasets, and effectively captures the privacy risk of node data during the FL process in IoVs.