一种基于带宽分配的联邦学习激励机制  

An incentive mechanism with bandwidth allocation for federated learning

作　　者：郭英芸 高博 张志飞[1,2] 张煜[3] 熊轲[1,2] GUO Yingyun;GAO Bo;ZHANG Zhifei;ZHANG Yu;XIONG Ke(School of Computer and Information Technology,Beijing Jiaotong University,Beijing100044,China;Engineering Research Center of High Speed Railway Network Management,Ministry of Education,Beijing Jiaotong University,Beijing 100044,China;State Grid Energy Research Institute Co.,Ltd.,Beijing 102209,China)

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044 [2]北京交通大学高速铁路网络管理教育部工程研究中心,北京100044 [3]国网能源研究院有限公司,北京102209

出　　处：《物联网学报》2022年第4期82-92,共11页Chinese Journal on Internet of Things

基　　金：国家自然科学基金资助项目(No.61872028);中央高校基本科研业务费专项资金(No.2021JBM008,No.2022JBXT001)。

摘　　要：联邦学习(FL,federated learning)是一种新兴的机器学习范式,它可以充分利用移动众包资源进行去中心化数据训练。然而,在无线网络中部署FL面临网络带宽有限、移动用户自私等挑战。为了应对这些挑战,提出了一种基于带宽分配的激励机制(IMBA,incentive mechanism with bandwidth allocation)。IMBA考虑用户数据质量和计算能力的不同设计支付方案,以激励高数据质量用户贡献其计算资源,进而提升模型训练精度。通过最小化训练时间和支付成本权重和确定最佳支付与带宽分配方案,通过优化带宽分配降低训练时延。实验表明,IMBA能够有效提高训练精度,降低训练时间,并帮助服务器灵活权衡训练时间和支付报酬。Federated learning(FL)is an emerging machine learning paradigm that can make full use of crowd sourced mobile resources for training on decentralized data.However,it is challenging to deploy FL over a wireless network because of the limited bandwidth and clients’selfishness.To address these challenges,an incentive mechanism with bandwidth allocation(IMBA)was proposed.Considering the difference between clients'data quality and computing power,IMBA designs a payment scheme to incentivize high-quality clients to contribute their computing resources,thus improving the training accuracy of the model.By minimizing the weight sum of training time and payment cost,the optimal payment and bandwidth allocation scheme was determined,and the training delay was reduced by optimizing bandwidth allocation.Experiments show that IMBA effectively improves training accuracy,reduces the training delay and helps the server flexibly balance training delay and hiring payment.

关 键 词：联邦学习 激励机制 带宽分配 STACKELBERG博弈 训练质量 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]