Drop-in differential privacy for sklearn devotees

What it does Diffprivlib is IBM’s general-purpose differential privacy library for Python. It offers mechanisms, models, data-analysis tools, and a privacy budget accountant for building differentially private machine learning workflows. The models module copies the scikit-learn API—GaussianNB and others use familiar fit and predict calls—so you can experiment with privacy-preserving training without learning a new dialect.

The interesting bit The library treats privacy loss as a composable resource: the BudgetAccountant tracks cumulative spend across operations using advanced composition techniques, which is the kind of bookkeeping that is easy to describe in a paper but tedious to get right in code. It also exposes raw mechanisms for experts who want to build custom algorithms from scratch.

Key highlights

• Models for clustering, classification, regression, dimensionality reduction, and preprocessing that mirror sklearn estimators.
• A BudgetAccountant that tallies total privacy loss across multiple queries or model fits.
• Differentially private histograms that reuse NumPy’s histogram interface.
• Low-level privacy mechanisms exposed for researchers implementing bespoke algorithms.
• Explicitly labeled for research and education; production deployments require IBM engagement.

Caveats

• The README states the public release is intended for research and education only, so shipping it to production involves talking to IBM first.
• Some estimators, such as GaussianNB, emit warnings unless you supply a bounds parameter, so the zero-config experience is limited.

Verdict Worth a look if you are a researcher or student who wants to prototype differentially private machine learning inside the scikit-learn ecosystem. Look elsewhere if you need a library that is cleared for immediate production use without vendor coordination.