基于联邦学习的边缘计算隐私保护方法

Edge Computing Privacy Protection Method Based on Federated Learning

针对边缘计算下联邦学习模型参数传递的安全性问题,提出一种基于联邦学习的边缘计算隐私保护方法(EC2PM)。该方法首先利用本地差分隐私(LDP),对参与联邦学习(FL)的边缘设备训练的模型参数添加数据扰动;然后通过调整隐私参数ε控制隐私损失的大小;最后将边缘计算与联邦学习进行结合,模型训练全程无需上传本地数据,实现了在保障边缘用户安全共享数据时,边缘设备的数据本地化训练和模型聚合,解决了边缘用户数据的隐私安全问题。对比实验结果表明,该方法的准确率为86.87%,不仅能够确保聚合模型的准确率而且达到保护模型参数的效果,同时能够满足安全性要求较高的边缘计算场景。

An edge computing privacy protection method based on federated learning(EC2PM)was proposed to address the security problem of federated learning model parameter delivery under edge computing.Firstly,this method added data perturbation to the model parameters trained by the edge devices participating in federated learning(FL)by using local differential privacy(LDP).Secondly,to control the size of privacy loss by adjusting the privacy parameterε.Finally,combining the edge computing with federated learning,the model training was conducted without uploading local data throughout,which achieves data localization of edge devices while guaranteeing secure data sharing among edge users training and model aggregation so as to solve the privacy security problem of edge user data.The comparison experimental results show that the accuracy of the method can reach 86.87%,which not only ensures the accuracy of the aggregated model but also achieves the effect of protecting the model parameters,and can meet the edge computing scenarios with high security requirements.