梯度隐藏的安全聚类与隐私保护联邦学习  

Gradient-hiding secure clustering and privacy-preserving federated learning

作　　者：李功丽[1,2] 马婧雯 范云 Li Gongli;Ma Jingwen;Fan Yun(School of Computer&Information Engineering,Henan Normal University,Xinxiang Henan 453007,China;Key Laboratory of Artificial Intelligence&Personalized Learning in Education of Henan Province,Henan Normal University,Xinxiang Henan 453007,China)

机构地区：[1]河南师范大学计算机与信息工程学院,河南新乡453007 [2]河南师范大学河南省教育人工智能与个性化学习重点实验室,河南新乡453007

出　　处：《计算机应用研究》2024年第6期1851-1861,共11页Application Research of Computers

基　　金：河南省科技攻关计划资助项目(232102211057)。

摘　　要：联邦学习是一种前沿的分布式机器学习算法,它在保障用户对数据控制权的同时实现了多方协同训练。然而,现有的联邦学习算法在处理Non-IID数据、梯度信息泄露和动态用户离线等方面存在诸多问题。为了解决这些问题,基于四元数、零共享与秘密共享等技术,提出了一种梯度隐藏的安全聚类与隐私保护联邦学习SCFL。首先,借助四元数旋转技术隐藏首轮模型梯度,并且在确保梯度特征分布不变的情况下实现安全的聚类分层,从而解决Non-IID数据导致的性能下降问题;其次,设计了一种链式零共享算法,采用单掩码策略保护用户模型梯度;然后,通过门限秘密共享来提升对用户离线情况的鲁棒性。与其他现有算法进行多维度比较表明,SCFL在Non-IID数据分布下准确度提高3.13%~16.03%,整体运行时间提高3~6倍。同时,任何阶段均能保证信息传输的安全性,满足了精确性、安全性和高效性的设计目标。Federated learning is a kind of advanced distributed machine learning algorithm,which realizes multiparty cooperative training while ensuring the user’s control over the data.However,the existing federated learning algorithms have many problems in dealing with Non-IID data,gradient information leakage and dynamic user offline.To solve these problems,this paper proposed a gradient hidden safe clustering and privacy-protecting federated learning based on quaternion,zero sharing and secret sharing techniques.Firstly,it used quaternion rotation technology to hide the first-round model gradient and achieve secure clustering stratification without altering the gradient feature distribution,so as to solve the performance degradation issue caused by Non-IID data.Secondly,this paper designed a chain zero sharing algorithm,using single strategy to protect the user model gradient mask.Then,it used the threshold secret sharing to improve the robustness against offline users.Multi-dimensional comparison with other existing algorithms shows that the accuracy of SCFL is improved by about 3.13%~16.03%under the Non-IID data distribution,and the overall running time is improved by about 3~6 times.Mean while,the security of information transmission is guaranteed at any stage,satisfying the design goals of accuracy,security and efficiency.

关 键 词：联邦学习 隐私保护 聚类 四元数 零共享 秘密共享 

分 类 号：TP390[自动化与计算机技术—计算机应用技术]