摘要
针对基于中央服务器的联邦学习存在的梯度泄露、投毒攻击和单点故障问题,设计一种基于联盟链的可信联邦学习方法。在客户端上传梯度前对梯度进行稀疏化,有效降低梯度泄露风险,缓解通信压力;根据参数质量筛选优质梯度进行全局聚合,降低投毒攻击对模型的影响;提出一种基于双因子贡献的区块链共识算法,在有效激励联邦学习客户端参与训练的同时,提高了系统的稳定性。实验结果表明,基于联盟链的可信联邦学习方法抗投毒攻击能力提高了30%,能够有效降低梯度泄露和系统中心化程度,提高了联邦学习的安全性。
Aiming at the problems of gradient leakage,poisoning attacks and single points of failure in the federated learning training model based on the central server,a trustworthy federated learning method based on consortium blockchain was designed.The gradients were sparsed before the clients upload the gradients,which effectively reduced the risk of gradient lea-kage and relieved the communication pressure.According to the parameter quality,the high-quality gradients were screened according to the parameter quality for global aggregation to reduce the impact of poisoning attacks on model training.A blockchain consensus algorithm based on two factor contribution was proposed.The algorithm improves the stability of the system while effectively incentivizing the federated learning client to participate in the training.Experimental results show that the anti-poisoning attack capability of the proposed method is improved by 30%,which can effectively reduce gradient leakage and system centralization,and improve the security of federated learning.
作者
王金鹏
戴欢
李凯程
唐毅
付保川
WANG Jin-peng;DAI Huan;LI Kai-cheng;TANG Yi;FU Bao-chuan(School of Electronics and Information Engineering,Suzhou University of Science and Technology,Suzhou 215009,China;Suzhou Heshu Blockchain Application Research Institute Limited Company,Suzhou 215000,China)
出处
《计算机工程与设计》
北大核心
2023年第12期3545-3553,共9页
Computer Engineering and Design
基金
国家自然科学基金项目(61702354、61876121)
苏州科技大学科研基金项目(XKZ2017004)
苏州科技大学教改基金项目(SKJG18_05)
江苏省物联网移动互联技术工程重点实验室开放课题基金项目(JSWLW2017004)
研究生科研创新计划基金项目(SKSJ18_012、SJCX19_0963)。
关键词
区块链
联邦学习
机器学习
梯度压缩
共识算法
激励机制
隐私保护
blockchain
federated learning
machine learning
gradient compression
consensus algorithm
incentive mechanism
privacy protection