Privacy and Utility of Private Synthetic Data for Medical Data Analyses

Abstract

The increasing availability and use of sensitive personal data raises a set of issues regarding the privacy of the individuals behind the data. These concerns become even more important when health data are processed, as are considered sensitive (according to most global regulations). Privacy Enhancing Technologies (PETs) attempt to protect the privacy of individuals whilst preserving the utility of data. One of the most popular technologies recently is Differential Privacy (DP), which was used for the 2020 U.S. Census. Another trend is to combine synthetic data generators with DP to create so-called private synthetic data generators. The objective is to preserve statistical properties as accurately as possible, while the generated data should be as different as possible compared to the original data regarding private features. While these technologies seem promising, there is a gap between academic research on DP and synthetic data and the practical application and evaluation of these techniques for real-world use cases. In this paper, we evaluate three different private synthetic data generators (MWEM, DP-CTGAN, and PATE-CTGAN) on their use-case-specific privacy and utility. For the use case, continuous heart rate measurements from different individuals are analyzed. This work shows that private synthetic data generators have tremendous advantages over traditional techniques, but also require in-depth analysis depending on the use case. Furthermore, it can be seen that each technology has different strengths, so there is no clear winner. However, DP-CTGAN often performs slightly better than the other technologies, so it can be recommended for a continuous medical data use case.