一种基于异步联邦学习的安全聚合机制  

A secure aggregation mechanism based on asynchronous federated learning

作　　者：秦宝东 杨国栋 马宇涵 QIN Baodong;YANG Guodong;MA Yuhan(School of Cyberspace Security,Xi an University of Posts and Telecommunications,Xi an 710121,China;School of Information Engineering,Chang an University,Xi an 710064,China)

机构地区：[1]西安邮电大学网络空间安全学院,陕西西安710121 [2]长安大学信息工程学院,陕西西安710064

出　　处：《西安邮电大学学报》2023年第1期50-61,共12页Journal of Xi’an University of Posts and Telecommunications

基　　金：青海省基础研究计划项目(2020-ZJ-701)。

摘　　要：针对异步联邦学习的客户端数据隐私保护难度高、存在推理攻击等安全问题,提出一种基于异步联邦学习的安全聚合机制。根据客户端异步学习的特征,利用秘密分享与Paillier同态加密等技术在客户端选择自己的秘密份额用于掩盖客户端模型参数,服务器则利用拉格朗日插值法恢复总秘密用于获取聚合的全局参数,在保留异步模型低开销与高精度的优势下,具备抵御推理攻击的能力,使模型更加可靠实用。实验结果表明,所提安全聚合机制在每轮迭代中,客户端加密的平均耗时为0.226 s,服务器安全聚合的平均耗时为0.363 s。与模型训练相比,安全聚合产生的时间开销极小,且提高了模型的安全性。A secure aggregation mechanism for asynchronous federated learning is propose to resolve the security issues such as high difficulty in protecting client data privacy and inference attacks of asynchronous federated learning.Based on the characteristics of client-side asynchronous learning,secret sharing and the Paillier homomorphic encryption are adopt by the client to pick its share of secrets for masking client-side model parameters,and the Lagrangian interpolation is use by the server to recover the total secrets for obtaining the aggregated global parameters.While retaining the advantages of low cost and high precision of asynchronous models,it also has the ability to resist inference attacks,making the model more reliable and practical.Experiments show that in each iteration of the proposed security aggregation mechanism,the average time for client encryption is 0.226 seconds,and the average time for server security aggregation is 0.363 seconds.Compared with model training,the time overhead of the security aggregation is extremely small,and the security of the model is improved.

关 键 词：联邦学习 异步更新 推理攻击 安全聚合 秘密分享 Paillier加密 

分 类 号：TP309.7[自动化与计算机技术—计算机系统结构]