Research on gradient leakage defense in federated learning for vehicular networks based on reputation mechanism
Jiaheng Li, Qinmu Wu, Yang Liu · 2025
With the rapid advancement of vehicular networks and intelligent transportation systems, the demand for secure and privacy-preserving data sharing in collaborative environments has grown significantly. Federated learning, which is increasingly applied in vehicle-to-everything (V2X) communications, allows vehicles to collaboratively train models by sharing gradients instead of raw data, preserving data privacy. However, this approach is still vulnerable to privacy attacks, such as deep leakage from gradients (DLG), where adversaries can reconstruct local data from gradients. To address this challenge, we propose a defense mechanism tailored for vehicular federated learning, combining a reputation mechanism with gradient sparsification. This method assesses and monitors node behavior, excluding malicious participants, while reducing data leakage risks.