基于深度学习的联邦学习中数据隐私保护方法  

A Deep Learning-based Data Privacy Protection Method for Federated Learning

作　　者：田根源 TIAN Genyuan(School of Information Engineering,Zhumadian Vocational and Technical College,Zhumadian 463000,China)

机构地区：[1]驻马店职业技术学院信息工程学院,河南驻马店463000

出　　处：《火力与指挥控制》2025年第1期189-194,共6页Fire Control & Command Control

基　　金：河南省科技攻关计划资助项目(212102210515)。

摘　　要：拆分学习可在不共享原始数据的情况下,由客户端同服务端协作训练深度学习模型,成为隐私保护领域的研究热点。然而,拆分学习仍面临数据重构攻击,其威胁着参与者的敏感信息。提出基于二进制拆分学习的数据隐私保护算法。BSLP算法将客户端所训练的本地模型进行二值化,降低由拆分层输出值导致的数据泄露损失。BSLP算法引用差分隐私机制,在数据中添加噪声,进而泛化数据。以4个典型的数据集进行实验,分析BSLP算法的分类准确率和隐私保护性能。分析结果表明,提出的BSLP算法在MNIST数据集上的分类准确率达到97%,而KL散度为3.68,验证了BSLP算法具有较强的隐私保护性能的事实。Split learning(SL)enables privacy protection field become a research hotspot by allowing clients to collaboratively train a deep learning model with the server without sharing raw data.However,split learning still faces data reconstruction attacks that threaten participants'sensitive information.Therefore,binary split learning-based data privacy protection(BSLP)algorithm is proposed.In BLDP algorithm,the binarization of the local model trained by the client is conducted and the data leakage loss caused by the output value of split layer is reduced.At the same time,the BSLP algorithm quotes the differential privacy mechanism to add noise to the data and then generalize the data.Four typical datasets are experimented and the classification accuracy rate and privacy protection performance of BSLP algorithm are analyzed.The analysis results show that the classification accuracy rate of the proposed BSLP algorithm on the MNIST dataset is 97%,and the KL divergence is 3.68,which verifies the fact that the BSLP algorithm has stronger privacy protection performance.

关 键 词：联邦学习 隐私保护 拆分学习 二值化 差分隐私 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]