An enhancing privacy training architecture for federal vehicle networking

期刊论文
发表日期：

2025年4月23日

查看全文

Yuzhou Dai,
Qian Cheng,
Hui Li,
Dan Liao,
Ming Zhang,
Hailing Zhang,
Zhiliang Xu,
Guangxin Li

摘要

The secure sharing of massive data for federal vehicle networking has gradually become a research hotspot. Federated learning allows users to train models without sharing local raw data, which is beneficial for protecting privacy. However, attackers can infer users’ sensitive information by stealing the local model parameters uploaded by Vehicle Users (VUs). Meanwhile, due to differences in vehicle performance, lower-performance vehicles require more time for local training, which hinders the aggregation of the global model. To address these issues, this paper proposes an enhanced privacy training architecture based on federated learning, named FL-EPTA. FL-EPTA introduces Laplace noise into the objective function of local training using a functional mechanism to achieve $\epsilon $ $ϵ$ -privacy protection. Taking into account the differences in computing resources between VUs, FL-EPTA designs a VU selection algorithm on the Roadside Unit (RSU) to optimize the selection process. By formulating VU selection as a maximization problem with knapsack constraints, a greedy approach is employed to minimize training time. Theoretical analysis proves that the proposed FL-EPTA architecture satisfies $\epsilon $ $ϵ$ -differential privacy and ensures convergence. The simulation results further demonstrate that FL-EPTA achieves faster convergence, lower training loss, and shorter training time compared to existing methods.