密码学报（中英文）2025，Vol.12Issue(3)：545-564,20.DOI:10.13868/j.cnki.jcr.000780
车联网环境下面向异构数据的隐私保护联邦学习
PPHSFL:Privacy-Preserving Federated Learning Towards Heterogeneous System in IoV
摘要
Abstract
With the rapid development of technologies in the fields of Internet of Vehicles(IoV)and intelligent manufacturing,the integration of IoV and artificial intelligence has become a new direc-tion for the advancement of smart transportation.This study proposes a privacy-preserving federated learning framework,for information sharing in IoV environments,aiming to address the issues of incon-sistent client training conditions,data drift,and model data privacy leakage in data sharing for IoV.The mobility and distributed nature of smart vehicles pose special challenges to federated learning.The framework incorporates performance-enhanced heterogeneous data federated learning algorithm,which modifies the training optimization steps and propose a generalized update approach consider-ing additional factors during the aggregation process.Essentially,it employs a normalized averaging approach,ensuring faster loss convergence in heterogeneous scenarios.To prevent inference of privacy information by honest yet curious servers and external adversaries from transmitted parameters,exist-ing solutions employ differential privacy mechanisms,which add noise to local parameters,to protect them from leakage.Nevertheless,the added noise can disrupt the learning process and degrade the effectiveness of the trained models.A performance-boosting differential privacy algorithm is proposed,which introduces regularization terms in local optimization objective function,to improve the robust-ness of training models to injected noise,reduce the impact of noise on IoV devices during training,and provide total privacy budget statistics.The framework's performance is evaluated on synthetic and real-world heterogeneous datasets.The first algorithm of framework outperforms existing hetero-geneous data optimization algorithms,while the another algorithm demonstrates better classification performance at the same level of privacy protection compared to existing differential privacy-based federated learning algorithms,making it more suitable for IoV applications.关键词
车联网/联邦学习/非独立同分布数据/差分隐私/效用隐私平衡Key words
Internet of Vehicles/federated learning/non-IID/differential privacy/utility-privacy tradeoff分类
信息技术与安全科学引用本文复制引用
张宇,咸鹤群..车联网环境下面向异构数据的隐私保护联邦学习[J].密码学报（中英文）,2025,12(3):545-564,20.基金项目
国家自然科学基金(62102212) （62102212）
山东省自然科学基金(ZR2021QF030)National Natural Science Foundation of China(62102212) （ZR2021QF030）
Natural Science Foundation of Shandong Province(ZR2021QF030) （ZR2021QF030）
