摘要
目前联邦学习中通常采用加密技术(如同态加密、差分隐私)提高隐私保护能力,但这会导致高通信开销和低准确率.因此,针对如何兼顾隐私保护、加密效率和准确预测的问题,该文提出基于随机正交矩阵的数据屏蔽方法.在训练阶段,利用分块矩阵的思想高效生成随机正交矩阵,对参与方的原始数据进行扰动,由发起方结合扰动后的数据主导服务器训练纵向联邦逻辑回归模型,并在预测阶段应用秘密分享技术增强数据隐私保护.在4个不同规模的数据集上的实验结果表明,相较于同态加密技术,该加密方法的计算效率提高了100~800倍,模型准确率提升约0.4%,相较于差分隐私提升约1.2%,且能够有效保护原始数据.
In federated learning,privacy protection is typically enhanced through the use of encryption techniques.However,these approaches often lead to high communication overhead and reduced accuracy.To address the challenge of balancing security,efficiency and accuracy,this paper proposes a data masking method based on random orthogonal matrices.By leveraging the concept of block matrices,the method efficiently generates random orthogonal matrices to perturb original data during the training phase.The initiating party combines locally perturbed data with participants,guiding the server in training a vertically federated logistic regression model.In the prediction phase,the method use secret sharing techniques to enhance security.Experiments on four datasets of varying scales indicate a substantial improvement in computational efficiency(100~800 times)compared to homomorphic encryption.Our method improves accuracy by approximately 0.4%.In comparison to differential privacy,the improvement is around 1.2%.Additionally,our method effectively protects local data.
作者
牛琬茹
黄一珉
付海燕
王湾湾
何浩
姚明
郭艳卿
NIU Wanru;HUANG Yimin;FU Haiyan;WANG Wanwan;HE Hao;YAO Ming;GUO Yanqing(School of Information and Communication Engineering,Dalian University of Technology,Dalian 116024,China;Data Intelligence Department of INSIGHTONE Tech Co,Ltd,Beijing 100007,China)
出处
《小型微型计算机系统》
北大核心
2025年第6期1305-1311,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(62076052,62106037)资助
辽宁省应用基础研究计划项目(2022JH2/101300262)资助
中央高校基本科研业务费项目(DUT20TD110,DUT22YG205)资助
国家社科基金重大项目(19ZDA127)资助.
关键词
逻辑回归
隐私安全
联邦学习
数据失真
logistic regression
privacy security
federated learning
data distortion
