首页|基于最大-最小策略的纵向联邦学习隐私保护方法

基于最大-最小策略的纵向联邦学习隐私保护方法

扫码查看
纵向联邦学习(Vertical federated learning,VFL)是一种新兴的分布式机器学习技术,在保障隐私性的前提下,利用分散在各个机构的数据实现机器学习模型的联合训练。纵向联邦学习被广泛应用于工业互联网、金融借贷和医疗诊断等诸多领域中,因此保证其隐私安全性具有重要意义。首先,针对纵向联邦学习协议中由于参与方交换的嵌入表示造成的隐私泄漏风险,研究由协作者发起的通用的属性推断攻击。攻击者利用辅助数据和嵌入表示训练一个攻击模型,然后利用训练完成的攻击模型窃取参与方的隐私属性。实验结果表明,纵向联邦学习在训练推理阶段产生的嵌入表示容易泄漏数据隐私。为了应对上述隐私泄漏风险,提出一种基于最大-最小策略的纵向联邦学习隐私保护方法(Privacy preservation method for vertical federated learning based on max-min strategy,PPVFL),其引入梯度正则组件保证训练过程主任务的预测性能,同时引入重构组件掩藏参与方嵌入表示中包含的隐私属性信息。最后,在钢板缺陷诊断工业场景的实验结果表明,相比于没有任何防御方法的VFL,隐私保护方法将攻击推断准确度从95%下降到55%以下,接近于随机猜测的水平,同时主任务预测准确率仅下降2%。
Privacy Preservation Method for Vertical Federated Learning Based on Max-min Strategy
Vertical federated learning(VFL)is an emerging distributed machine learning that applies to the data distributed in various institutions to realize the joint construction of privacy preservation machine learning models.It has been widely applied to various fields such as industrial internet,financial lending,and medical diagnosis.Therefore,the privacy security research of vertical federated learning highlights its significance.Aiming at the risk of privacy leakage caused by the embedding exchanged by participants in the vertical federated learning protocol,we propose a general property inference attack initiated by the server.The adversary uses the auxiliary data and the embedding exchanged by the vertical federated learning protocol to train the attack model and steal the target privacy property of the participant.The experimental results show that the embedding representation generated by the vertical federated learning during the training and inference process can reveal the information of the personal private property.To deal with the above proposed privacy leakage risk,proposed a privacy preservation method for vertical federated learning based on max-min strategy(PPVFL),which introduces a gradient regular component to ensure the performance of the main task of the training process and adopts a construction component to hide parti-cipant's privacy property.Finally,in steel defect diagnosis industrial scenarios,compared to VFL without any de-fense method,privacy-preserving method reduces attack inference accuracy from 95%to below 55%,which is close to the level of random guessing,while the main task only dropped by 2%of the prediction accuracy.

Vertical federated learning(VFL)property inference attackprivacy preservationmax-min strategyindustrial internet

李荣昌、刘涛、郑海斌、陈晋音、刘振广、纪守领

展开 >

浙江工业大学信息工程学院 杭州 310023

浙江工业大学计算机科学与技术学院 杭州 310023

浙江工业大学网络空间安全研究院 杭州 310023

浙江大学网络空间安全学院 杭州 310007

浙江大学计算机科学与技术学院 杭州 310007

展开 >

纵向联邦学习 属性推断攻击 隐私保护 最大-最小策略 工业互联网

浙江省自然科学基金青年原创计划国家自然科学基金国家重点研发计划基金浙江省自然科学基金浙江省自然科学基金

LDQ23F020001620724062018AAA0100801LGF21F020006LGF20F020016

2024

自动化学报
中国自动化学会 中国科学院自动化研究所

自动化学报

CSTPCD北大核心
影响因子:1.762
ISSN:0254-4156
年,卷(期):2024.50(7)