基于扩散式差分隐私的联邦学习数据保护方法  

Federated Learning Data Protection Based on Diffusive Differential Privacy

作　　者：雷靖鹏 任诚[1,2] LEI Jingpeng;REN Cheng(College of Electrical Engineering and Information,Southwest Petroleum University,Chengdu 610500,China;School of Information and Communication Engineering,University of Electronic Science and Technology of China,Chengdu 611731,China)

机构地区：[1]西南石油大学电气信息学院,成都610500 [2]电子科技大学信息与通信工程学院,成都611731

出　　处：《电讯技术》2024年第11期1765-1771,共7页Telecommunication Engineering

摘　　要：传统的联邦学习(Federated Learning,FL)差分隐私(Differential Privacy,DP)保护机制在一定程度上抵御了差分攻击,防止用户数据的泄露问题,但是引入的噪声扰动在一定程度上又影响了原本数据,导致在服务器聚合时与原本数据产生影响较大的偏差,严重影响了全局模型的准确率和收敛性。为了解决这一问题,提出了一种基于扩散式联邦学习差分隐私保护(Diffusive Differential Privacy Federated Learning,DDPFL)方法,通过在服务器端聚合之前更加精确地对噪声进行拟合,还原精度更高的数据样本,降低了对原模型的影响。在数据分布为IID和Non-IID联邦学习实际场景下验证了该方法的有效性。在数据分布为Non-IID场景下,所提方法与联邦学习原始差分隐私保护方法相比,准确率在其基础上提高了1.7%~4.6%。The traditional federated learning differential privacy protection mechanism resists differential attacks to a certain extent and prevents the problem of user data leakage,but the introduced noise perturbation affects the original data again to a certain extent,leading to the server aggregation with the original data having a large deviation,which seriously affects the accuracy and convergence of the global model.In order to solve this problem,the authors propose a diffusive differential privacy federated learning(DDPFL)based method to reduce the impact on the original model by more accurately fitting the noise before server-side aggregation and restoring data samples with higher accuracy.The effectiveness of the method is verified in real-world scenarios where the data distributions are IID and Non-IID federated learning.And under the data distribution of Non-IID scenarios,comparion between the proposed method and the original differential privacy preserving method of federated learning shows that the accuracy is improved by 1.7%to 4.6%on its basis.

关 键 词：差分隐私保护 数据泄露 联邦学习 扩散式传播 

分 类 号：TP309[自动化与计算机技术—计算机系统结构] TP181[自动化与计算机技术—计算机科学与技术]