Differential Privacy Synthetic Tabular Data Generation for Federated Learning

Abstract

Machine learning in healthcare struggles for one main reason: good data are hard to come by. Medical records are sensitive, and the rules on sharing them between hospitals are strict. A common workaround is the use of synthetic data, that is, artificial records that copy the statistics of real ones; however, on its own, such use offers no real privacy guarantee, and earlier studies show it can still leak details about the patients behind it. We introduce an approach tackling both problems at once. It builds on a previous UMAP-based generator methodology and extends it so several hospitals can work together without sharing raw records, in a federated learning form. Then, the proposed methodology adds differential privacy so every shared quantity carries a formal (εDP,δ) guarantee, holding by construction for the fully synthetic variant and under a bounded-input assumption for the partially synthetic one, tracked by a custom privacy accountant. Two versions of the methodology are tested: apartially synthetic one, where a reference hospital validates the others’ rows, and a fully synthetic one, where each centre builds its own data from aggregated cluster statistics. Both are evaluated on four publicly available healthcare tabular datasets spanning prostate cancer (PI-CAI, CIA), breast cancer (BC-MLR), and cardiovascular disease (fiv-CardioDB), to test whether the observed trends generalize beyond a single clinical domain. As a key takeaway, it is demonstrated that going federated barely affects quality, but the privacy cost depends strongly on the moment in the procedure when the noise is injected, and this pattern holds consistently across all four datasets.