《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (7): 2209-2216.DOI: 10.11772/j.issn.1001-9081.2022060909

• 网络空间安全 • 上一篇    

边缘计算下基于区块链的隐私保护联邦学习算法

陈宛桢1, 张恩1,2(), 秦磊勇1, 洪双喜1,2   

  1. 1.河南师范大学 计算机与信息工程学院, 河南 新乡 453007
    2.智慧商务与物联网技术河南省工程实验室(河南师范大学), 河南 新乡 453007
  • 收稿日期:2022-06-23 修回日期:2022-08-27 接受日期:2022-09-05 发布日期:2023-07-20 出版日期:2023-07-10
  • 通讯作者: 张恩
  • 作者简介:陈宛桢(1998—),女,河南南阳人,硕士研究生,主要研究方向:隐私保护机器学习;
    张恩(1974—),男,河南新乡人,教授,博士,CCF会员,主要研究方向:安全多方计算;
    秦磊勇(1997—),男,河南商丘人,硕士研究生,主要研究方向:安全多方计算;
    洪双喜(1979—),男,河南新乡人,副教授,博士,主要研究方向:隐私保护机器学习。
  • 基金资助:
    国家自然科学基金资助项目(62002103);河南省科技攻关计划项目(212102210388);河南省软科学研究计划项目(212400410109)

Privacy-preserving federated learning algorithm based on blockchain in edge computing

Wanzhen CHEN1, En ZHANG1,2(), Leiyong QIN1, Shuangxi HONG1,2   

  1. 1.College of Computer and Information Engineering,Henan Normal University,Xinxiang Henan 453007,China
    2.Engineering Lab of Intelligence Business and Internet of Things of Henan Province (Henan Normal University),Xinxiang Henan 453007,China
  • Received:2022-06-23 Revised:2022-08-27 Accepted:2022-09-05 Online:2023-07-20 Published:2023-07-10
  • Contact: En ZHANG
  • About author:CHEN Wanzhen, born in 1998, M. S. candidate. Her research interests include privacy-preserving machine learning.
    ZHANG En, born in 1974, Ph. D., professor. His research interests include secure multi-party computation.
    QIN Leiyong, born in 1997, M. S. candidate. His research interests include secure multi-party computation.
    HONG Shuangxi, born in 1979, Ph. D., associate professor. His research interests include privacy-preserving machine learning.
  • Supported by:
    National Natural Science Foundation of China(62002103);Science and Technology Research Program of Henan Province(212102210388);Soft Science Research Program of Henan Province(212400410109)

摘要:

针对在边缘计算(EC)场景下进行的联邦学习(FL)过程中存在的模型参数隐私泄露、不可信服务器可能返回错误的聚合结果以及参与训练的用户可能上传错误或低质量模型参数的问题,提出一种边缘计算下基于区块链的隐私保护联邦学习算法。在训练过程中,每个用户首先使用全局模型参数在其本地数据集上进行训练,并将训练得到的模型参数以秘密共享的方式上传至附近的边缘节点,从而实现对用户本地模型参数的保护;然后由边缘节点在本地计算它们所收到的模型参数的份额之间的欧氏距离,并将结果上传至区块链;最后由区块链负责对模型参数之间的欧氏距离进行重构,进而在去除有毒的更新后,再进行全局模型参数的聚合。通过安全分析证明了所提算法的安全性:即使在部分边缘节点合谋的情况下,用户的本地模型参数信息也不会泄露。同时实验结果表明该算法具有较高的准确率:在投毒样本比例为30%时,它的模型准确率为94.2%,接近没有投毒样本时的联邦平均(FedAvg)算法的模型准确率97.8%,而在投毒样本比例为30%时FedAvg算法的模型准确率下降至68.7%。

关键词: 边缘计算, 联邦学习, 区块链, 投毒攻击, 隐私保护

Abstract:

Aiming at the problems of the leakage of model parameters, that the untrusted server may return wrong aggregation results, and the users participating in training may upload wrong or low-quality model parameters in the process of federated learning in edge computing scenarios, a privacy-preserving federated learning algorithm based on blockchain in edge computing was proposed. In the training process, firstly, the global model parameters were trained on the local dataset of each user by the users, and the model parameters obtained by training were uploaded to neighboring edge nodes through secret sharing, thereby protecting the local model parameters of the users. Secondly, the Euclidean distances between the shares of model parameters received by the edge nodes were computed, and the results of these calculations were uploaded to the blockchain. Finally, the Euclidean distances between model parameters were reconstructed by the blockchain, and then the global model parameter was aggregated after removing the poisoned updates. The security analysis proves the security of the proposed algorithm: even in the case of collusion of a part of edge nodes, the users’ local model parameter information will not be leaked. At the same time, the experimental results show the high accuracy of this algorithm: the accuracy of the proposed algorithm is 94.2% when the proportion of poisoned samples is 30%, which is close to the accuracy of the Federated Averaging (FedAvg) algorithm without poisoned samples (97.8%), and the accuracy of FedAvg algorithm is decreased to 68.7% when the proportion of poisoned samples is 30%.

Key words: Edge Computing (EC), Federated Learning (FL), blockchain, poisoning attack, privacy-preserving

中图分类号: