dpmm (Differentially Private Marginal Models)

dpmm (Differentially Private Marginal Models) is an open-source Python library for generating synthetic tabular data with differential privacy guarantees. Synthetic data consists of artificial records that preserve the statistical patterns of an original dataset. It can be shared or used to develop and test AI models where real personal data could not be used. However, synthetic data can still leak information about real individuals unless it is generated with formal privacy protection. dpmm addresses this by applying differential privacy, which limits how much any single person's data can influence the result. It was developed by SAS and presented at the 2025 Theory and Practice of Differential Privacy (TPDP) workshop.

The library includes three widely used marginal models: PrivBayes, MST and AIM. These models measure simple statistics about the original data, such as how often combinations of values in a few columns appear together, and add calibrated noise to them. They then generate synthetic records that match these noisy statistics. The authors report that their implementations produce more useful synthetic data and offer more functionality than alternative implementations.

dpmm follows best practices to provide end-to-end privacy guarantees. This means that privacy is protected across the whole process, including preparing the data, and not only when the model is trained. The library also addresses well-known vulnerabilities in differential privacy software. It is designed to be easy to install and highly customisable, so it can serve a wide range of users.

Auto-discovered on 2026-08-05 by OECD Catalogue Automation