面向联邦学习激励优化的演化博弈模型  

Evolutionary Game Model for Federated Learning Incentive Optimization

作　　者：孙跃杰 赵国生[1] 廖祎玮[1] SUN Yuejie;ZHAO Guosheng;LIAO Yiwei(School of Computer Science and Information Engineering,Harbin Normal University,Harbin 150025,China)

机构地区：[1]哈尔滨师范大学计算机科学与信息工程学院,哈尔滨150025

出　　处：《小型微型计算机系统》2024年第3期718-725,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61202458,61403109)资助;黑龙江省自然科学基金项目(LH2020F034)资助;哈尔滨市科技创新研究基金项目(2016RAQXJ036)资助.

摘　　要：针对联邦学习中参与者虚报训练成本导致激励不匹配的现象,提出了面向联邦学习激励优化的演化博弈模型.首先在联邦学习系统中建立了联邦参与者-联邦组织者演化博弈模型,设计模型质量评估算法对参与者提交的模型进行质量评估,去除低质量模型的同时量化参与者训练成本.然后结合信誉度指标提出优化的激励分配方法,通过求解演化博弈的稳定策略得到不同初始状态下的最优收益策略.最后仿真实验表明参与者激励收益方面,与平均分配法和个体收益分享法相比诚实参与者的收益提升了70%和57.4%,虚报参与者收益降低了65%和69.5%,策略选择方面,所提模型能合理选择收益策略.In response to the phenomenon of incentive mismatch caused by participants′misreporting of training costs in federal learning,an evolutionary game model for incentive optimization of federal learning is proposed.Firstly,the federated participant-federal organizer evolutionary game model is established in the federated learning system,and the model quality evaluation algorithm is designed to evaluate the quality of the models submitted by the participants and quantify the training cost of the participants while removing the low-quality models.Then the optimal incentive allocation method is proposed by combining the credibility index,and the optimal payoff strategy under different initial states is obtained by solving the stable strategy of the evolutionary game.The final simulation experiments show that in terms of participant incentive gains,the gains of honest participants are improved by 70%and 57.4%compared with the average allocation method and the individual gain sharing method,and the gains of misrepresented participants are reduced by 65%and 69.5%.In terms of strategy selection,the proposed model can reasonably select the gain strategy.

关 键 词：联邦学习 演化博弈 激励机制 复制动态方程 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]