基于GAN的拆分纵向联邦学习重建攻击  

GAN-based split vertical federated learning for reconstruction attack

作　　者：唐琳 冯秀芳[1] 陈永乐[2] TANG Lin;FENG Xiu-fang;CHEN Yong-le(School of Software,Taiyuan University of Technology,Jinzhong 030600,China;College of Computer Science and Technology(College of Data Science),Taiyuan University of Technology,Jinzhong 030600,China)

机构地区：[1]太原理工大学软件学院,山西晋中030600 [2]太原理工大学计算机科学与技术学院(大数据学院),山西晋中030600

出　　处：《计算机工程与设计》2024年第12期3531-3537,共7页Computer Engineering and Design

基　　金：山西省重点研发计划基金项目(202102020101007)。

摘　　要：针对拆分纵向联邦学习的参与者在训练过程中输出的中间结果容易泄露大量隐私的问题,提出一种重建攻击Re_GAN。利用生成式对抗网络学习图像的先验知识,优化生成式对抗网络的输入,使重建图像和真实图像的中间结果逼近来重建参与者的隐私图像。在衡量中间结果时,使用分片沃瑟斯坦距离捕捉图像的特征。实验结果表明,Re_GAN在MNIST数据集、Fashion-MNIST数据集和CIFAR-10数据集上均能重建参与者图像,表明了拆分纵向联邦学习隐私存在泄露的风险。A reconstruction attack Re_GAN was proposed to address the issue of participants in split vertical federated learning,where the intermediate results output during the training process are prone to leakage of a large amount of privacy.The generative adversarial network was used to learn the prior knowledge of the images.The input of the generative adversarial networks was optimized to approximate the intermediate result of the reconstructed image and the real image to reconstruct the participant’s private image.The intermediate result was measured using the Sliced Wasserstein distance to capture the features of the image.Experimental results indicate that Re_GAN is able to reconstruct participant images on the MNIST dataset,Fashion-MNIST dataset,and CIFAR-10 dataset,indicating the risk of privacy leakage in split vertical federated learning.

关 键 词：纵向联邦学习 拆分学习 重建攻击 生成式对抗网络 隐私泄露 机器学习 分布式系统 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]