How Protective Are Synthetic Data?

摘要

This short paper provides a synthesis of the statistical disclosure limitation and computer science data privacy approaches to measuring the confidentiality protections provided by fully synthetic data. Since all elements of the data records in the release file derived from fully synthetic data are sampled from an appropriate probability distribution, they do not represent "real data," but there is still a disclosure risk. In SDL this risk is summarized by the inferential disclosure probability. In privacy-protected database queries, this risk is measured by the differential privacy ratio. The two are closely related. This result (not new) is demonstrated and examples are provided from recent work.