非独立同分布数据环境下的联邦学习激励机制设计  

Design of Federal Learning Incentive Mechanism in Non-IID Data Environment

作　　者：李秋贤 周全兴[1] LI Qiuxian;ZHOU Quanxing(Kaili University,Kaili 556011,China)

机构地区：[1]凯里学院,贵州凯里556011

出　　处：《现代信息科技》2024年第22期30-35,共6页Modern Information Technology

基　　金：黔东南州科技计划项目(黔东南科合J字〔2023〕106号);2022年度凯里学院规划课题(2022YB08);扶持市(州)高校质量提升工程项目(院办发〔2022〕10号-32);贵州省科技计划项目(黔科合基础-ZK〔2023〕一般440)。

摘　　要：在联邦学习环境中,非独立同分布(Non-IID)数据的存在对模型性能和用户参与度提出了严峻挑战。为了应对这些挑战,文章提出了一种基于博弈论和深度强化学习的新型激励机制,以提升非IID数据环境下的联邦学习效果。通过设计中央服务器和用户的收益函数,综合考虑通信成本、计算成本和本地模型精度,公平衡量用户贡献,并利用博弈论模型和深度强化学习算法优化用户参与策略。实验结果表明,所提出的激励机制显著提升了模型的精度和用户的参与度,有效地缓解了非IID数据分布对联邦学习性能的负面影响,从而增强了整个系统的性能和稳定性。In the Federated Learning environment,the existence of Non-Independent Identically Distributed(Non-IID)data poses a serious challenge to model performance and user engagement.To address these challenges,this paper proposes a new incentive mechanism based on game theory and Deep Reinforcement Learning,to improve the Federated Learning effect in Non-IID data environment.Byd esigning the payoff function of the central server and the user,considering the communication cost,computing cost and local model accuracy,the user contribution is measured fairly,and the user participation strategy is optimized by using the game theory model and the Deep Reinforcement Learning algorithm.The experimental results show that the proposed incentive mechanism significantly improves the accuracy of the model and the participation of users,and effectively alleviates the negative impact of Non-IID data distribution on Federated Learning performance,so as to enhance the performance and stability of the whole system.

关 键 词：联邦学习 博弈论 非独立同分布 激励机制 深度强化学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]