Total variation distance privacy: Accurately measuring inference attacks and improving utility

Jingyu Jia, Chang Tan, Zhewei Liu, Xinhao Li, Zheli Liu, Siyi Lv, Changyu Dong

Abstract: Differential privacy (DP) is a general approach to defend against inference attacks, but hard to balance the privacy-utility trade-off for some complex data analysis tasks. To improve the utility of data analysis, a weaker privacy definition with a more accurate estimate of privacy risk may be a straightforward and effective solution. Total variation distance (TVD) privacy is an appropriate tool for this issue, but it has not been adequately studied. In this paper, we systematically study TVD privacy and explore its applications. We provide a comprehensive theoretical analysis of TVD privacy and demonstrate its advantage in measuring privacy risks with the example of membership inference attacks. Our work indicates that TVD privacy is a helpful tool in estimating privacy risks and has the potential to be widely used as a general privacy definition.