Command-line interface¶

Entry point: synthpriv (module docstring in synthpriv.cli).

synthpriv generate¶

Train a generator and produce synthetic data.

synthpriv generate --data real.csv --epsilon 8 --rows 5000 \
  --save demo_model.sz -o synthetic.csv

Options:

--data/-d (required) — CSV with the real data.
--method/-m (default: ``ctgan``) — generator registry key (dp-gan, dp-copula, gaussian-copula, ctgan, tvae, copula-gan).
--rows/-n (default: 5000) — number of synthetic rows.
--epochs (default: 300) — training epochs (dp-gan only).
--epsilon — differential privacy budget: trains dp-gan (DP-SGD) or dp-copula (pure DP) with this ε.
--ecdf-epsilon — extra DP budget for the dp-gan marginal ECDFs (requires --epsilon; dp-gan only).
--output/-o (default: ``synthetic.csv``) — output CSV.
--save — persist the trained synthesizer (for synthpriv sample).

synthpriv sample¶

Generate rows from a persisted synthesizer without retraining: training with DP costs once; reloading regenerates the dataset with the same privacy guarantee (accounted epsilon) in seconds.

synthpriv sample --model demo_model.sz --rows 5000 -o resample.csv

Options:

--model/-m (required) — synthesizer saved with generate --save.
--rows/-n (default: 1000) — number of synthetic rows.
--output/-o (default: ``synthetic.csv``) — output CSV.

synthpriv evaluate¶

Evaluate utility and privacy of some synthetic data and render the HTML report.

synthpriv evaluate --real real.csv --synthetic synthetic.csv \
  --epsilon 8 -o report.html

Options:

--real/-r (required) — CSV with the real data.
--synthetic/-s (required) — CSV with the synthetic data.
--epsilon — DP budget used when generating (labels the report).
--delta (default: 1e-5) — associated DP delta.
--output/-o (default: ``report.html``) — output HTML report.

synthpriv sweep¶

Train a DP generator with several epsilons and plot the privacy/utility curve.

synthpriv sweep --data real.csv --generator dp-copula \
  --epsilons "0.1,0.5,1,2,5,50" -o sweep_report.html

Options:

--data/-r (required) — CSV with the real data.
--generator/-g (default: ``dp-gan``) — DP-capable generator: dp-gan or dp-copula.
--epsilons/-e (default: ``0.1,0.5,1,2,5,50``) — comma-separated budgets (50 ≈ almost no DP).
--delta (default: 1e-5) — DP delta.
--epochs (default: 50) — epochs per point (dp-gan only).
--rows/-n (default: 2000) — synthetic rows per point.
--output/-o (default: ``sweep_report.html``) — output HTML report.

synthpriv benchmark¶

Compare a DP generator (several epsilons) vs non-DP SDV generators.

synthpriv benchmark --data real.csv --generator dp-gan \
  --epsilons "1,5,50" --baselines gaussian-copula -o bench.html

Options:

--data/-d (required) — CSV with the real data.
--generator/-g (default: ``dp-gan``) — dp-gan or dp-copula.
--epsilons/-e (default: ``1,2,5,10,50``) — comma-separated budgets.
--delta (default: 1e-5) — DP delta.
--baselines/-b (default: ``gaussian-copula``) — comma-separated non-DP SDV generators, or all.
--epochs (default: 100) — dp-gan epochs (ignored by dp-copula).
--numeric (default: ``mode``) — dp-gan numeric encoding: mode (GMM) or uniform (gaussianized).
--rectify-marginals (flag) — rectify marginals at sampling (requires --numeric uniform; dp-gan only).
--ecdf-epsilon — DP-ECDF marginals budget (requires --numeric uniform; dp-gan only); the point’s reported total is training + this value.
--baseline-epochs (default: 0) — epochs of the SDV baselines (0 = each generator’s default).
--rows/-n — synthetic rows per point (default: same as real).
--output/-o (default: ``benchmark_report.html``) — output HTML report.

synthpriv dpcheck¶

Validate that a persisted model’s DP guarantee is not exceeded (see assert_dp()).

synthpriv dpcheck --model demo_model.sz --tolerance 0.05

Options:

--model/-m (required) — synthesizer saved with generate --save.
--tolerance (default: 0.05) — relative margin allowed over the declared budget.