Formal (ε, δ)-Differential Privacy • Rényi RDP Composition • Copula Fidelity

Provable Differential Privacy Dataset Synthesis

Generate high-fidelity synthetic tabular datasets with mathematically bounded privacy leakage. Formally immune to database reconstruction, linkability, and membership inference attacks.

quickstart.sh
$ git clone https://github.com/KELLERBABG/Differential-Privacy-Synthesizer.git
$ cd Differential-Privacy-Synthesizer && pip install -e .
$ python -m dpsynth.cli synthesize --input patients.csv --epsilon 1.0 --delta 1e-6 --output synthetic.parquet

Rényi Differential Privacy (RDP)

Standard composition theorems overestimate privacy decay exponentially under multi-query workloads. DPSynth implements Rényi divergence accounting:

D_α(M(D) || M(D')) ≤ ε(α)
ε_{total} = min_α { ∑ ε_i(α) + log(1/δ) / (α - 1) }

This allows 5x-10x more synthesis queries on the same target dataset before exhausting the corporate privacy budget.

Multi-Attribute Copula Preservation

Independent marginal noise destroys cross-column feature interactions required by machine learning pipelines:

  • Empirical Copula Modeling: Captures dependency rank structures across high-dimensional joint distributions.
  • Wasserstein Distance Guarantees: Continuous verification ensures statistical divergence between original and synthetic columns remains strictly minimized.
  • Outlier Sanitization: Private quantile trimming prevents anomalous singleton data points from skewing synthetic distribution tails.