DPSynth
Differentially private (DP) synthetic data generation for tabular data. Provides DP Gaussian mixture models, DP Gaussian copulas, DP histogram marginals and a Private Aggregation of Teacher Ensembles (PATE) synthesizer for mixed-type data, together with a standardized utility evaluation framework (univariate fidelity, propensity score MSE, multivariate dependence, downstream task performance), empirical disclosure risk auditing (membership inference, attribute disclosure, record linkage) and privacy budget accounting (basic, advanced and Renyi DP composition). The implementation follows Dwork and Roth (2014) <doi:10.1561/0400000042> and Dwork et al. (2006) <doi:10.1007/11681878_14> for the DP mechanisms, Papernot et al. (2017) <doi:10.1145/3133956.3133982> for the PATE synthesizer, and Woo et al. (2009) <doi:10.2202/1557-4679.1203> for the disclosure risk evaluation framework.
- An introduction to causal inference
- Calibrating Noise to Sensitivity in Private Data Analysis
- Practical Secure Aggregation for Privacy-Preserving Machine Learning
- The algorithmic foundations of differential privacy
This page was built for software: DPSynth
