四川大学学报（自然科学版）2025，Vol.62Issue(3)：761-768,8.DOI:10.19907/j.0490-6756.240298
基于矩阵高斯机制的经验风险最小化隐私保护
Matrix Gaussian mechanism for private protection in empirical risk minimization
摘要
Abstract
Privacy-preserving computing aims to protect individuals'sensitive information from disclosure or misuse during the process of data analysis or computation.As an important privacy-preserving computing ap-proach,differential privacy adds noise to sensitive data for privacy preservation,where a balance between the security and utility of data is critical.Excessive noise heavily reduces the utility of the data,while,con-versely,it fails to guarantee the security of the data.To strike a balance,the variance of the added noise should be as small as possible while still meeting the privacy protection requirements.Focusing on the privacy preservation in empirical risk minimization,this work adopts the matrix Gaussian mechanism(MGM)to pro-tect the input data.In order to address the issue of excessive noise in matrix noise addition mechanisms caused by the deflation method,this work uses the hypothesis testing theory to derive the noise variance under the framework of f-differential privacy(f-DP).Experiments on real dataset show that the proposed method re-duces the variance of the MGM significantly and achieves higher utility in practical applications.关键词
隐私保护计算/差分隐私/矩阵高斯机制/经验风险最小化Key words
Privacy-preserving computing/Differential privacy/Matrix Gaussian mechanism/Empirical risk minimization分类
数理科学引用本文复制引用
文胜兰,王雅轩,周杰..基于矩阵高斯机制的经验风险最小化隐私保护[J].四川大学学报（自然科学版）,2025,62(3):761-768,8.基金项目
数据安全防护与智能治理教育部重点实验室(四川大学)创新引导项目(SCUSACXYD202303) （四川大学）
四川省自然科学基金(2024NSFSC0444) （2024NSFSC0444）
