twinify
github.com
- API
- Self-hosted
What twinify is
twinify is an open-source software package for generating synthetic counterparts of sensitive tabular datasets while applying differential privacy. It learns probabilistic models with ε and δ settings, letting users choose privacy levels; smaller values indicate stronger privacy. The package provides NAPSU-MQ and differentially private variational inference (DPVI). DPVI can create models automatically or use user-defined NumPyro models, and it handles categorical, continuous, and mixed data. NAPSU-MQ is intended for fully categorical datasets. Automatic modeling can represent missing values by estimating their probability under an assumption of independent missingness across features. You can use twinify as a Python library or as a command-line tool that reads CSV files. NumPyro, JAX, and d3p support its modeling and private training implementation. Stable installation is available with pip from PyPI, while a development version can be installed from a cloned repository. The code is licensed under Apache License 2.0 and is classified for scientific research audiences.
Who it suits
twinify suits scientific research users working with sensitive tabular data who need synthetic datasets and configurable differential privacy. It is also aimed at Python users or command-line workflows involving CSV files.
Good to know, and what to watch
- Differential privacy uses configurable ε and δ parameters
- DPVI handles categorical, continuous, and mixed data
- Available as Python library and CSV command-line tool
- Automatic modeling can account for missing values
- Licensed under Apache License 2.0
- NAPSU-MQ supports only fully categorical datasets
- NAPSU-MQ may run long with many feature dimensions
- DPVI produces an approximate posterior
- DPVI omits extra differential-privacy uncertainty
The verdict
twinify combines two privacy-preserving inference methods with both library and command-line use. Check the data-type and inference limitations before selecting a method for a project.
What the maker publishes
- Data types
- DPVI can handle categorical, continuous, or mixed data; NAPSU-MQ is currently suitable only for fully categorical data. source
- DPVI limitation
- DPVI supports categorical, continuous, and mixed data, but produces an approximate posterior and does not explicitly capture the additional uncertainty due to differential privacy. source
- Inference methods
- It implements NAPSU-MQ and differentially private variational inference (DPVI). source
- Installation
- The README says a stable version can be installed from PyPI with pip, or installed from a cloned repository for the development version. source
- Installation
- The stable version is installed with pip from PyPI, and the development version can be installed from the cloned GitHub repository. source
- Integrations
- The implementation relies on NumPyro for modeling and inference, JAX for CPU and GPU kernels, and d3p for differentially private training routines. source
- Integrations
- The implementation relies on NumPyro for modeling and inference, JAX for CPU and GPU kernels, and d3p for differentially private training routines. source
- Intended audience
- The package metadata classifies twinify for scientific research audiences. source
- License
- The code base is licensed under the Apache License 2.0. source
- License
- The codebase is licensed under Apache License 2.0. source
- Limitations
- NAPSU-MQ may run for a long time on datasets with many feature dimensions, while DPVI approximates the true posterior and does not explicitly capture additional uncertainty due to differential privacy. source
- Maintainer organization
- The DPBayes GitHub organization describes itself as providing differential privacy software from the Finnish Center for Artificial Intelligence FCAI. source
- Maker
- The DPBayes GitHub organization describes itself as differential privacy software from the Finnish Center for Artificial Intelligence (FCAI). source
- Maker details
- The opened maker page does not state a headquarters or founding year. source
- Methods
- It implements NAPSU-MQ and differentially private variational inference (DPVI). source
- Missing values
- Automatic modeling handles missing values by modeling their probability, assuming missingness is independent across features. source
- Missing values
- Automatic modeling handles missing values by assuming they are missing at random, independently of whether other feature values are missing. source
- Modeling
- DPVI supports automatic modeling and user-defined models written with NumPyro. source
- Modeling
- DPVI automatic modeling builds mixture models from user-specified feature distributions; users can also supply NumPyro models. source
- NAPSU-MQ limitation
- NAPSU-MQ currently supports only fully categorical data and may run for a long time on datasets with many feature dimensions. source
- Privacy controls
- Users can set ε and δ privacy parameters, with smaller values indicating stronger privacy. source
- Privacy method
- It learns probabilistic models under differential privacy, with ε and δ parameters for setting the privacy level. source
- Purpose
- twinify is a software package for privacy-preserving generation of synthetic twins of sensitive tabular datasets. source
- Purpose
- twinify generates privacy-preserving synthetic twins of sensitive tabular datasets. source
Checked against the maker's pages on 5 Oct 2026.