Synthetic Data: Generate Avatar Data on Demand

Abstract

Anonymization is crucial for the sharing of personal data in a privacy-aware manner yet it is a complex task that requires to set up a trade-off between the robustness of anonymization (i.e., the privacy level provided) and the quality of the analysis that can be expected from anonymized data (i.e., the resulting utility). Synthetic data has emerged as a promising solution to overcome the limits of classical anonymization methods while achieving similar statistical properties to the original data. Avatar-based approaches are a specific type of synthetic data generation that rely on local stochastic simulation modeling to generate an avatar for each original record. While these approaches have been used in healthcare, their attack surface is not well documented and understood. In this paper, we provide an extensive assessment of such approaches and comparing them against other data synthesis methods. We also propose an improvement based on conditional sampling in the latent space, which allows synthetic data to be generated on demand (i.e., of arbitrary size). Our empirical analysis shows that avatar-generated data are subject to the same utility and privacy trade-off as other data synthesis methods with a privacy risk more important on the edge data, which correspond to records that have the fewest alter egos in the original data.