摘要
联邦学习中用户的数据数量不同得到的聚合权重不同,数据质量不同也会对参与联邦训练的用户聚合权重产生影响。针对传统联邦学习中因单一因素确定聚合权重导致的贡献不公平问题,并且基于用户的数据数量和用户的数据质量提出一种基于数据质量评估的公平联邦学习方案。首先,结合相对熵定义了评估公平标准。然后,运用熵权法定义用户数据质量计算方法,根据用户数据质量得分和数据数量得分计算用户的综合得分,并用综合得分作为用户的贡献。最后,根据用户的综合得分定义用户的聚合权重设计公平的隐私保护联邦学习方案。实验分析表明,所提出的方案比传统联邦学习方案更加具备公平性。
Different amounts of data for users in federated learning give different aggregate weights,and different data quality will also have an impact on the aggregate weights of users participating in federated training.Aiming at the problem of unfair contribution caused by the determination of aggregate weights by a single factor in traditional federated learning,this paper proposes a fair federated learning scheme based on the amount of data of users and the data quality of users.First,the criteria for assessing fairness are defined in conjunction with relative entropy.Then,the entropy method is used to define the user data quality calculation method,and the user’s comprehensive score is calculated according to the user data quality score and the data quantity score,and the comprehensive score is used as the user’s contribution.Finally,according to the user’s comprehensive score,the user’s aggregate weight is defined to design a fair privacy protection federated learning scheme,and the experiment shows that the scheme proposed in this paper is fair than the traditional federated learning scheme.
作者
杨秀清
彭长根
刘海
丁红发
汤寒林
YANG Xiuqing;PENG Changgen;LIU Hai;DING Hongfa;TANG Hanlin(College of Computer Science and Technology,Guizhou University,Guiyang 550025;State Key Laboratory of Public Big Data,Guizhou University,Guiyang 550025;Guizhou Big Data Academy,Guizhou University,Guiyang 550025;College of Information,Guizhou University of Finance and Economics,Guiyang 550025;Guizhou Shujubao Network Technology Co.,Ltd.,Guiyang 550025)
出处
《计算机与数字工程》
2022年第6期1278-1285,共8页
Computer & Digital Engineering
基金
国家自然科学基金项目“数据共享应用的块数据融合分析理论与安全管控模型研究”(编号:U1836205)
国家自然科学基金项目“基于机器学习的图数据自适应差分隐私保护模型与算法研究”(编号:62002081)
国家自然科学基金项目“图数据的结构信息熵模型及理性隐私保护方法研究”(编号:62002080)
贵州省科技计划基金项目“贵州省公共大数据安全与隐私保护科技创新人才团队”(编号:黔科合平台人才[2020]5017)
贵州省教育厅自然科学项目“高维图结构数据隐私量化及分析模型研究”(编号:黔教合KY字[2021]140)资助。
关键词
联邦学习
机器学习
公平性
数据质量评估
熵权法
federated learning
machine learning
fairness
data quality evaluation
entropy method
