Interpretable DP

A suite of tools for statistics and machine learning with interpretable risk-based privacy guarantees for easy auditability and communication in regulated environments.

Technical Layer

Audience: Researchers, differential privacy experts

riskcal

Tools for computing f-DP trade-off curves for differentially private algorithms, and calibrating their noise scale to operational privacy risk measures (attack advantage, or attack TPR and FPR).

pip install riskcal

gdpnum

Correct numerical accounting in terms of Gaussian differential privacy for more interpretable and auditable guarantees in privacy-preserving machine learning.

pip install gdpnum

Application Layer

Audience: Researchers, data scientists

Coming soon: Synthetic data with interpretable privacy guarantees.

About

Publications

This software suite has emerged as a result of several independent scientific collaborations.

Lausanne University Hospital EPFL Harvard University Hasso Plattner Institute University of Helsinki

People

Bogdan Kulynych Bogdan Kulynych Principal Investigator
Juan Felipe Gomez Juan Felipe Gomez Development

Funding

The effort is partially funded by the Swiss National Science Foundation (SNF):
Bridging Regulatory Data Protection Standards and Model Sharing in Healthcare