面向全局不平衡问题的基于贡献度的联邦学习方法  

Contribution-based Federated Learning Approach for Global Imbalanced Problem

作　　者：吴飞[1] 宋一波 季一木[2] 胥熙 王木森 荆晓远[3] WU Fei;SONG Yibo;JI Yimu;XU Xi;WANG Musen;JING Xiaoyuan(College of Automation,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;School of Computer Science,Wuhan University,Wuhan 430072,China)

机构地区：[1]南京邮电大学自动化学院,南京210003 [2]南京邮电大学计算机学院,南京210003 [3]武汉大学计算机学院,武汉430072

出　　处：《计算机科学》2023年第12期343-348,共6页Computer Science

基　　金：国家自然科学基金(62076139);之江实验室开放课题(2021KF0AB05);未来网络科研基金项目(FNSRFP-2021-YB-15);南京邮电大学1311人才计划。

摘　　要：联邦学习在保护各方数据隐私的前提下,协同多方共同训练,提高了全局模型的精度。数据的类不平衡问题是联邦学习范式中具有挑战的问题,联邦学习中的数据不平衡问题可分为局部数据不平衡和全局数据不平衡,目前针对全局数据不平衡问题的相关研究较少。文中提出了一种面向全局不平衡问题的基于贡献度的联邦学习方法(CGIFL)。首先,设计了一种基于贡献度的全局判别损失函数,用于调整本地训练过程中的模型优化方向,使模型在训练中给予全局少数类更多的关注,以提高模型的泛化能力;然后,在全局模型更新阶段,设计了一种基于贡献度的动态联邦汇聚策略,优化了各节点的参与权重,更好地平衡了全局模型的更新方向。在MNIST,CIFAR10和CIFAR100这3个数据集上进行实验,实验结果表明了CGIFL在解决全局数据不平衡问题上的有效性。Under the premise of protecting the data privacy,federated learning unites multiple parties to train together to improve the accuracy of the global model.Class imbalance of data is a challenging problem in the federated learning paradigm.Data imba-lance in federated learning can be divided into local data imbalance and global data imbalance.At present,there are few researches on global data imbalance.This paper proposes a contribution-based federated learning approach for global imbalance problem(CGIFL).First,a contribution-based global discriminant loss is designed to adjust the model optimization direction in the local training process and make models give more attention to the global minority classes in training to improve the generalization ability of models.And a contribution-based dynamic federated aggregation algorithm is designed to optimize the participation weight of each node and better balance the updating direction of the global model.Experimental results on MNIST,CIFAR10 and CIFAR100 datasets demonstrate the effectiveness of CGIFL in solving the problem of global data imbalance.

关 键 词：联邦学习 数据不平衡 多方协同 图像分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]