摘要
Abstract
With the popularization of automotive connectivity technology,federated learning has become an important means of addressing data privacy and security issues.However,accompanying member inference attacks and communication costs remain to be improved.In this paper,a federated differential privacy method is proposed to defend against member inference attacks,and the tripartite gradient technology and model compression are introduced to further reduce communication costs.The differential privacy experiment is performed.The variances of different Gaussian noise distributions are compared.It is found that the accuracy rate of federated differential privacy is closer to that of the scheme without privacy protection(especially when C>1)in comparison with the traditional differential privacy.It is observed in the tripartite gradient experiment that on the datasets MNIST,Cifar10,Cifar100 and SVHN,the percentage reduction of the training gradient reaches 93.33%,93.56%,93.60%and 93.74%,respectively,which indicates that the tripartite gradients can reduce communication costs more effectively.In the layer sensitivity experiment,it is found that the accuracy rate at rate=85%,rate=90%and rate=95%is almost the same as that when uncompressed(rate=100%).It is proved that the proposed method is effective in defending against member inference attacks and reducing communication costs.关键词
三元梯度/三元联邦学习/车联网/通信效率/联邦差分隐私/模型压缩Key words
tripartite gradient/tripartite federated learning/Internet of Vehicles/communication efficiency/federated differential privacy/model compression分类
电子信息工程
