一种基于合同理论的可激励联邦学习模型  

An Incentivized Federated Learning Model Based on Contract Theory

作　　者：王鑫[1,4,5] 李美庆[2] 王黎明 余芸 杨漾 孙凌云 WANG Xin;LI Meiqing;WANG Liming;YU Yun;YANG Yang;SUN Lingyun(College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China;College of Computer Science and Technology,Xidian University,Xi’an 710071,China;Digital Grid Research Institute Co.Ltd,China Southern Power Grid,Guangzhou 510663,China;Zhejiang University-China Southern Power Grid Joint Research Centre on AI,Hangzhou 310058,China;College of Computer Science and Technology,Zhejiang University,Hangzhou 310058,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310023 [2]西安电子科技大学计算机科学与技术学院,西安710071 [3]南方电网数字电网研究院有限公司,广州510663 [4]浙江大学南方电网人工智能创新联合研究中心,杭州310058 [5]浙江大学计算机科学与技术学院,杭州310058

出　　处：《电子与信息学报》2023年第3期874-883,共10页Journal of Electronics & Information Technology

基　　金：国家重点研发计划(2020YFB0906000,2020YFB0906004)。

摘　　要：针对目前较少研究去中心化联邦学习中的激励机制设计,且已有联邦学习激励机制较少以全局模型效果为出发点的现状,该文为去中心化联邦学习加入了基于合同理论的联邦学习激励机制,提出一种新的可激励的联邦学习模型。使用区块链与星际文件系统(IPFS)取代传统联邦学习的中央服务器,用于模型参数存储与分发,在此基础上使用一个合同发布者来负责合同的制定和发布,各个联邦学习参与方结合本地数据质量选择签订合同。每轮本地训练结束后合同发布者将对各个本地训练模型进行评估,若满足签订合同时约定的奖励发放条件则发放相应的奖励,同时全局模型的聚合也基于奖励结果进行模型参数的聚合。通过在MNIST数据集以及行业用电量数据集上进行实验验证,相比于传统联邦学习,加入激励机制后的联邦学习训练得到的全局模型效果更优,同时去中心化的结构也提高了联邦学习的鲁棒性。In view of the fact that there is rare research on the incentive mechanism design in decentralized federated learning,and the existing incentive mechanisms for federated learning are seldom based on the global model effect,an incentive mechanism based on contract theory,is added into decentralized federated learning and a new incentivized federated learning model is proposed.A blockchain and an InterPlanetary File System(IPFS)are used to replace the central server of traditional federated learning for model parameter storage and distribution,based on which a contract publisher is responsible for contract formulation and distribution,and each federated learning participant chooses to sign a contract based on its local data quality.The contract publisher evaluates each local training model after each round of local training and issues a reward based on the agreed-upon conditions in the contract.The global model aggregation also aggregates model parameters based on the reward results.Experimental validation on the MNIST dataset and industry electricity consumption dataset show that the proposed incentivized federated learning model outperforms traditional federated learning and its decentralized structure improves its robustness.

关 键 词：联邦学习 激励机制 合同理论 去中心化 电力大数据 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]