摘要
联邦学习通过聚合多方本地模型成为数据共享的新模式.现有的联邦学习激励机制有效缓解了完全信息下的数据供给不足问题,但仍面临搭便车、不公平、不可信等挑战.为此,本文提出了一种基于区块链和贝叶斯博弈(Bayesian game)的不完全信息联邦学习激励机制,通过量化数据供给方的成本效用与数据需求方的支付报酬对数据交易过程建模,采用沙普利值(Shapley value)实现了数据供给方报酬分配的公平性.在交易模型中考虑到参与个体的异质性与隐私保护,将数据供给方的资源配置策略构建为不完全信息的贝叶斯博弈模型,通过优化本地模型训练策略实现对数据供给方的激励作用.本文进一步分析了激励机制的有效性与行动策略的可信性,提出一种隐私保护的贝叶斯博弈行动策略共识算法(privacy-preserving Bayesian game action strategy consensus algorithm,PPBG-AC),该算法使数据供给方在基于区块链的数据交易平台下实现了贝叶斯纳什均衡.方案对比与理论分析表明本文提出的不完全信息联邦学习激励机制保障了数据供给方利益分配的公平性与资源配置的可信性,基于实际公开数据集的仿真实验与性能评估验证了激励机制的有效性.
Federated learning(FL)has become a new form of data sharing by aggregating multi-party local models.Although the existing FL incentive system has reduced insufficient data supply under comprehensive information,it still confronts issues including free-riding,unfairness,and unreliability.Therefore,this paper proposes an incomplete information FL incentive mechanism based on blockchain and Bayesian games.The data transaction process is modeled by quantifying the cost-utility of the data providers and the payment reward of the data requesters,in which Shapley value is used to realize the fairness of reward distribution of data providers.We consider the heterogeneity and privacy protection of participating individuals.The data providers’resource allocation strategies are built as a Bayesian game model,which optimizes the local training strategy to realize the incentive effect on the data providers.Furthermore,we consider the effectiveness of the incentive mechanism,a privacy-preserving Bayesian game action strategy consensus algorithm(PPBG-AC)is proposed,which enables the data providers to realize Bayesian Nash equilibrium under a data trading platform based on blockchain.The comparison and analysis of the schemes reveal that the incentive mechanism presented in our paper assures benefit distribution fairness and resource allocation credibility.Simulation experiments and performance evaluations based on real datasets demonstrate the effectiveness of our incentive mechanism.
作者
张沁楠
朱建明
高胜
熊泽辉
丁庆洋
朴桂荣
Qinnan ZHANG;Jianming ZHU;Sheng GAO;Zehui XIONG;Qingyang DING;Guirong PI AO(School of Information,Central University of Finance and Economics,Beijing 100081,China;Pillar of Information Systems Technology and Design,Singapore University of Technology and Design,Singapore 487372,Singapore;School of Management,Beijing Union University,Beijing 100020,China)
出处
《中国科学：信息科学》
CSCD
北大核心
2022年第6期971-991,共21页
Scientia Sinica(Informationis)
基金
国家重点研发计划(批准号:2017YFB1400700)
国家自然科学基金(批准号:62072487)
北京市自然科学基金(批准号:M21036)
北京联合大学教育科学研究课题(批准号:JK202114)和北京联合大学科研专项(批准号:ZK30202101)资助项目。
关键词
联邦学习
激励机制
区块链
贝叶斯博弈
沙普利值
不完全信息
隐私保护
federated learning
incentive mechanism
blockchain
Bayesian game
Shapley value
incomplete information
privacy protection
