Synthetic Data & Privacy
synthpriv
[ Open Source ][ Beta ]
Tabular data synthesis preserving differential privacy
synthpriv generates synthetic tabular data under formal differential privacy guarantees: DP generators (dp-gan with DP-SGD and RDP accountant, dp-copula with pure DP), DP marginal ECDFs, model serialization, budget/benchmark sweeps, and an HTML privacy/utility report — with CLI, Python API, and DP integrity assertions.
// Features
- Multiple generators: dp-gan (DP-SGD, AC-GAN conditioning) and dp-copula (pure DP, δ=0), plus SDV baselines (ctgan, tvae, copula-gan, gaussian-copula)
- DP marginal ECDFs with Laplace noise, ecdf_epsilon budget, and split_budget between training and marginals
- CLI: generate, sample, evaluate, sweep (epsilon ↔ utility), benchmark vs non-DP baselines, and dpcheck audit of persisted models
- Model serialization (save/load, regenerate without retraining) plus assert_dp integrity checks of the RDP budget
- Utility hardening: numeric='uniform' gaussianization, rectify_marginals with guaranteed KS, and label smoothing
- Self-contained HTML privacy/utility report with NNDR, MIA, and anonymeter attack metrics