面向不平衡类的联邦学习客户端智能选择算法  

An Intelligent Client Selection Algorithm of Federated Learning for Class-imbalance

作　　者：朱素霞[1,2] 王云梦 颜培森 孙广路[1,2] ZHU Suxia;WANG Yunmeng;YAN Peisen;SUN Guangu(School of Computer Science and Technology,Harbin University of Science and Technology,Harbin 150080,China;Research Center of Information Security and Intelligent Technology,Harbin University of Science and Technology,Harbin 150080,China)

机构地区：[1]哈尔滨理工大学计算机科学与技术学院,哈尔滨150080 [2]哈尔滨理工大学信息安全与智能技术研究中心,哈尔滨150080

出　　处：《哈尔滨理工大学学报》2024年第2期33-42,共10页Journal of Harbin University of Science and Technology

基　　金：黑龙江省自然科学基金(LH2021F032);黑龙江省重点研发计划项目(2022ZX01A34).

摘　　要：在联邦学习应用场景下,若客户端设备之间的数据呈现非独立同分布特征,甚至出现类不平衡的情况时,客户端本地模型的优化目标将偏离全局优化目标,从而给全局模型的性能带来巨大挑战。为解决这种数据异质性带来的挑战,通过积极选择合适的客户端子集以平衡数据分布将有助于提高模型的性能。因此,设计了一种面向不平衡类的联邦学习客户端智能选择算法—FedSIMT。该算法不借助任何辅助数据集,在保证客户端本地数据对服务器端不可见的隐私前提下,使用Tanimoto系数度量本地数据分布与目标分布之间的差异,采用强化学习领域中的组合多臂老虎机模型平衡客户端设备选择的开发和探索,在不同数据异质性类型下提高了全局模型的准确率和收敛速度。实验结果表明,该算法具有有效性。In the federated learning application scenario,if the data among the client devices shows the characteristics of nonindependence identically distribution,or even class-imbalance,the optimization goal of the client local model will deviate from the global optimization goal,which brings great challenges to the performance of the global model.In order to solve the challenges brought by this kind of data heterogeneity,it will help to improve the performance of the model by actively selecting appropriate customer terminal sets to balance the data distribution.Therefore,this paper designs a federated learning client online intelligent selection algorithm for unbalanced class FedSIMT.This algorithm does not use any auxiliary data set.Under the premise of ensuring the privacy of the client′s local data invisible to the server,it uses the Tanimoto coefficient to measure the difference between the local data distribution and the target distribution.It uses the combined Multi-armed Bandit in the field of reinforcement learning to balance the development and exploration of client devices selection,and improves the accuracy and the convergence rate of the global model under different data heterogeneity types.The experimental results show that the algorithm is effective.

关 键 词：联邦学习 类不平衡 客户端选择算法 多臂老虎机 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]