IBM/differential-privacy-library

What it solves

It provides a way to perform machine learning and data analysis while protecting the privacy of individuals in the training dataset. It allows developers to experiment with differential privacy (DP) and understand how adding privacy constraints impacts the accuracy of their models.

How it works

The library implements differential privacy by adding controlled randomness to calculations. It provides a set of "mechanisms" (the core building blocks of DP) that are used within its machine learning models. It integrates closely with the scikit-learn API, allowing users to swap standard classifiers for differentially private versions (like GaussianNB) by specifying a privacy budget parameter called epsilon.

Who it’s for

It is designed for scientists and engineers who are learning about differential privacy, researchers prototyping new DP algorithms, and developers exploring the trade-off between data privacy and model performance.

Highlights

• Scikit-learn Compatibility: Models run like standard sklearn classifiers, making them easy to integrate into existing workflows.
• Comprehensive Model Suite: Includes DP versions of tools for classification, regression, clustering, dimensionality reduction, and pre-processing.
• Privacy Budget Tracking: Includes a BudgetAccountant to track total privacy loss using advanced composition techniques.
• Data Analysis Tools: Provides generic tools for private data analysis, such as differentially private histograms.