Differentially private tabular generation

Develop a differential-privacy arm for benchmarking tabular generators, in order to determine whether methods can provide privacy guarantees without unacceptable losses in utility relative to the winning non-private baselines.

Background

The benchmark evaluates utility and privacy jointly using distance-based measures, including DCR rate and membership-inference AUC, rather than formal privacy guarantees. The results show that SMOTE achieves strong utility but has substantially higher membership-inference AUC than the deep generators, indicating a utility–privacy trade-off.

The paper identifies formal differential privacy as the sharpest unresolved issue arising from this trade-off and explicitly lists a differential-privacy arm as the first next step. Such an arm would test how differential privacy changes the comparative utility, fidelity, and privacy behavior of tabular generators in small-data settings.

References

What we would do next, in order: a differential-privacy arm, since the privacy cost of the winning baselines is the sharpest open problem here; regression and multi-class targets; and more natively small clinical datasets, because four is enough to raise the subsampling question and not enough to settle it.

— Below what training size do deep tabular generators stop beating trivial baselines? A preregistered benchmark on a size ladder of clinical and standard datasets  (2610.03500 - Shrivastava, 2 Oct 2026) in Conclusion