预算约束下多任务联邦学习激励机制  

Incentive Mechanism for Multi-Task Federated Learning Under Budget Constraints

作　　者：顾永跟[1] 李国笑 吴小红[1] 陶杰[1] 张艳琼 GU Yonggen;LI Guoxiao;WU Xiaohong;TAO Jie;ZHANG Yanqiong(School of Information Engineering,Huzhou University,Huzhou 313000,Zhejiang,China;School of Electronic Information,Huzhou College,Huzhou 313000,Zhejiang,China)

机构地区：[1]湖州师范学院信息工程学院,浙江湖州313000 [2]湖州学院电子信息学院,浙江湖州313000

出　　处：《计算机工程》2024年第5期149-157,共9页Computer Engineering

基　　金：浙江省现代农业资源智慧管理与应用研究重点实验室项目(2020E10017);湖州师范学院研究生科研创新项目(2023KYCX39)。

摘　　要：联邦学习是一种实现数据隐私保护的分布式机器学习范式,性能取决于数据源的质量和数据规模。客户端是理性个体,参与联邦学习将耗费计算、通信和隐私等成本,需要通过激励提高客户端的参与意愿。因此联邦学习能成功应用的关键之一是尽可能多地激励高质量数据客户端参与训练。多任务联邦学习环境下客户端拥有面向不同任务且质量不同的数据,并具有执行能力的约束。为提高多个学习任务的整体性能,在预算受限的条件下设计一种面向任务的客户选择和报酬机制。通过分析影响模型精度的重要因素,提出一种基于客户端数据样本分布特征的质量评估标准,并结合客户端成本信息,设计一种逆向拍卖的激励机制(EMD-MQMFL),实现客户端的任务指派和支付策略。从理论上分析和证明了该机制具有诚实性、个人理性以及预算可行性,并通过大量实验验证了该方法在联邦学习性能上的有效性。在MNIST、Fashion-MNIST、Cifar-10数据集上的实验结果表明,EMD-MQMFL在数据不平衡的情况下,平均模型精度比已有的机制至少提高5.6个百分点。Federated Learning(FL)is a distributed machine-learning paradigm that achieves data-privacy protection,and its performance depends on the quality and scale of the data source.The client is a rational individual,and the client's participation in FL incurs costs related to computation,communication,and privacy.Thus,the client must be encouraged to participate through incentives.One of the key factors affecting the successful application of FL is the participation of clients with high-quality data in training.In a multi-task FL environment,clients possess data that are specific to different tasks of varying quality,and their execution capabilities are limited.To improve the overall performance of multiple learning tasks,a task-oriented customer selection and a reward mechanism are designed under budget constraints in this study.By analyzing the important factors that affect the accuracy of the proposed model,a quality-evaluation standard based on the distribution characteristics of client data samples is proposed.Combining this with the client's cost information,an incentive mechanism for reverse auction(EMD-MQMFL)is designed to achieve task assignment and payment strategies for the client.This mechanism has been theoretically analyzed and proven to exhibit honesty,personal rationality,and budget feasibility.Furthermore,its effectiveness in FL performance has been verified via numerous experiments.Experimental results on the MNIST,Fashion-MNIST,and Cifar-10 datasets show that EMD-MQMFL improves the average model accuracy by at least 5.6 percentage points compared with existing mechanisms for cases involving imbalanced data.

关 键 词：联邦学习 多任务 逆向拍卖 激励机制 数据质量 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]