非独立同分布数据下联邦学习的隐私保护算法  

Privacy-preserving algorithms for federated learning under non-independent and non-identically distributed data

作　　者：张愈杰 龙士工[1,2] 张珺铭 刘光源 ZHANG Yu-jie;LONG Shi-gong;ZHANG Jun-ming;LIU Guang-yuan(State Key Laboratory of Public Big Data,Guizhou University,Guiyang 550025,China;School of Computer Science and Technology,Guizhou University,Guiyang 550025,China;School of Computer Science and Technology,Guizhou Polytechnic of Construction,Guiyang 551400,China)

机构地区：[1]贵州大学公共大数据国家重点实验室,贵州贵阳550025 [2]贵州大学计算机科学与技术学院,贵州贵阳550025 [3]贵州建设职业技术学院计算机科学与技术学院,贵州贵阳551400

出　　处：《计算机工程与设计》2025年第4期1047-1055,共9页Computer Engineering and Design

基　　金：国家自然科学基金项目(62062020)。

摘　　要：针对联邦学习中客户端数据统计多样性以及隐私保护问题,提出一种基于差分隐私的联邦学习聚类模型循环交换算法(DPFed-CMRE)。通过聚类将数据分布类似的客户端分为一类,通过各类模型循环交换使全局模型适应客户端数据的不同分布,同时应用差分隐私技术保护客户端数据隐私,通过自适应梯度裁剪优化噪声分配,降低模型的性能损失。为验证算法的有效性,在3个标准数据集上进行大量实验,实验结果表明,提出算法提升了联邦学习在客户端数据高度非独立同分布(Non-IID)以及高隐私保证情况下的性能。In response to the issues of client data diversity and privacy protection in federated learning,a federated learning clustering model with a differential privacy-based cyclic model rotation algorithm(DPFed-CMRE)was proposed.Clients with similar data distributions were clustered into groups and models among these groups were cyclically exchanged to adapt the global model to the different data distributions of clients.Differential privacy techniques were applied to protect client data privacy.Adaptive gradient clipping was used to optimize noise allocation,reducing the performance loss of the model.To validate the effectiveness of the algorithm,extensive experiments were conducted on three standard datasets.Result of experiments show that the proposed method enhances the performance of federated learning in situations where client data is highly non-independent and iden-tically distributed(Non-IID)and under high privacy guarantees.

关 键 词：联邦学习 差分隐私 自适应 非独立同分布 聚类 梯度裁剪 隐私保护 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]