Locally differentially private high-dimensional data synthesis  

作　　者：Xue CHEN Cheng WANG Qing YANG Teng HU Changjun JIANG 

机构地区：[1]Key Laboratory of Embedded System and Service Computing (Tongji University),Ministry of Education,Shanghai 201804,China [2]National(Province-Ministry Joint)Collaborative Innovation Center for Financial Network Security,Tongji University,Shanghai 201804,China

出　　处：《Science China(Information Sciences)》2023年第1期21-38,共18页中国科学（信息科学）（英文版）

基　　金：supported by Strategic Research and Consulting Project of the Chinese Academy of Engineering (Grant No.2022-XY-107)。

摘　　要：In local differential privacy(LDP),a challenging problem is the ability to generate highdimensional data while efficiently capturing the correlation between attributes in a dataset.Existing solutions for low-dimensional data synthesis,which partition the privacy budget among all attributes,cease to be effective in high-dimensional scenarios due to the large-scale noise and communication cost caused by the high dimension.In fact,the high-dimensional characteristics not only bring challenges but also make it possible to apply some technologies to break this bottleneck.This paper presents Sam Priv Syn for high-dimensional data synthesis under LDP,which is composed of a marginal sampling module and a data generation module.The marginal sampling module is used to sample from the original data to obtain two-way marginals.The sampling process is based on mutual information,which is updated iteratively to retain,as much as possible,the correlation between attributes.The data generation module is used to reconstruct the synthetic dataset from the sampled two-way marginals.Furthermore,this study conducted comparison experiments on the real-world datasets to demonstrate the effectiveness and efficiency of the proposed method,with results proving that Sam PrivSyn can not only protect privacy but also retain the correlation information between the attributes.

关 键 词：local differential privacy high-dimensional data synthesis sampling data privacy privacy-preserving protocols 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]