基于秘密分享与聚合签名的安全联邦学习方案

Secure Federated Learning Based on Secret Sharing and Aggregate Signature

联邦学习技术用以解决数据孤岛问题,但在联邦学习中用户原始的数据并不能得到好的保护,且其中负责聚合梯度的聚合服务器存在篡改聚合梯度来破坏最终的模型或与部分用户勾结以获取其他诚实用户的隐私数据的可能.本文提出一个基于秘密分享与聚合签名的安全联邦学习方案,对每个用户的数据进行秘密分享以保证隐私数据安全,使用聚合签名对梯度进行签名验证以保证聚合服务器无法篡改用户的梯度数据.方案具有较高精度,相较于目前其他同类方案,效率提升约6%.

Federated learning technique can be used to solve the data isolated island problem.However,in federated learning,the user’s original data is not well protected,and the aggregation server which is used to aggregate gradients can modify aggregated gradient to generate a wrong model or collude with some users to get the privacy data from other honest users.This paper presents a secure federated learning scheme based on secret sharing and aggregated signatures.The scheme deals with the gradient of users by secret sharing algorithm to ensure its security,and uses the aggregate signature to verify the gradient to make sure that the aggregate server can not tamper with the gradients of users.The scheme has high precision,and compared with other similar schemes,the efficiency is improved by about 6%.