基于RRCM框架的联邦学习激励机制  

Federated Learning Incentive Mechanism Based on RRCM Framework

作　　者：王文鑫 赵奕涵 张健毅 WANG Wenxin;ZHAO Yihan;ZHANG Jianyi(Beijing Electronic Science and Technology Institute,Beijing 100070,P.R.China)

机构地区：[1]北京电子科技学院,北京市100070

出　　处：《北京电子科技学院学报》2022年第4期54-62,共9页Journal of Beijing Electronic Science And Technology Institute

基　　金：国家重点研发计划项目(项目编号:2018YFB1004100);中国科学院网络测评技术重点实验室(中国科学院信息工程研究所)项目(项目编号:KFKT2019-004)

摘　　要：随着社会各界对于数据隐私的不断重视,借助模型传输的联邦学习技术近年来被广泛研究。尽管该技术不断成熟,但激励机制的研究相对较少,成为了制约技术落地的短板。传统联邦学习框架中,中心服务器向参与方分配相同激励会对高贡献者不公,这将导致高贡献者不再提供任务需要的模型。部分激励框架在一定程度上解决了无区别对待的问题,然而其方法缺乏保护措施,所以系统聚合时存在被敌手恶意攻击的风险。在联邦商业化模式中,参与方申请加入系统需要提供成本,然而现有分薪方式没有很好解决激励的分配。针对以上问题,本文提出了RRCM框架,通过设置声誉系统、奖惩措施和成本利息机制实现联邦系统协作公平性。在基准数据集上进行的实验表明,与同类方法相比,本文设计的RRCM框架能实现较高公平性。在使得联邦系统设计合理的同时,RRCM框架又能吸引更多优质参与方加入到联邦系统。With the increasing emphasis on data privacy from all walks of life,federated learning technology using the model transmission is widely studied.Despite continuous improvement,researches on the incentive mechanism still relatively lack,being the weakness restricting the full implementation of the federated learning technology.In the traditional federated learning framework,central server assigning same incentive to the participants brings unfairness to the participants with high contribution,leading to the condition that the participants with high contribution no longer provide models required by the tasks.The partial incentive framework addresses the issue of equal treatment,but it lacks protection measures,causing the risk of malicious attack when the system is aggregated.In the federated commercialization model,participants provide the cost when applying to join the system.However,existing salary sharing methods do not provide a commendable solution to the incentive distribution.To address the above-mentioned problems,in this paper a RRCM framework is proposed,where reputation system,incentives and disincentives measures,and cost-interest mechanism are established to achieve the cooperation fairness of the federated system.Experiments on benchmark dataset show that the proposed RRCM framework could achieve higher fairness compared with similar methods.With the RRCM framework,federated system will be designed more reasonably,and more quality participants will be attracted to participating in the federated system.

关 键 词：联邦学习 激励机制 RRCM框架 公平性 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]