Utility Analysis for Differentially-Private Pipelines
summary
An open‑source Python module inside PipelineDP models how contribution bounding and Laplace noise trade off to affect accuracy, estimating error per hyperparameter configuration by replacing sampling and noise steps with statistical models. L1 sensitivity is the product of max_partitions_contributed and max_contribution_per_partition, so raising one reduces sampling bias but increases Laplace variance; a demo finds max_partitions_contributed=4 as optimal for the synthetic restaurant data. Outputs are analytical estimates (not differentially private) and run in Colab and GitHub.
AI-written summary · about
chapters (8)
- 0:00Differential privacy pipelines
- 1:51Hyperparameters and utility challenges
- 3:01Laplace mechanism example
- 4:48Contribution bounding parameters
- 7:44Utility analysis framework
- 9:54Parameter tuning challenges
- 10:58Utility analysis demo
- 11:40Open-source and future steps
related talks
09mExpanding Differentially Private Solutions: A Python Case Study
44mDifferentially Private Inference for Regression Coefficients
43mPrivacy Amplification by Sampling in Practice
13mEvaluating the Usability of Differential Privacy Tools with Data Practitioners
44mAccuracy Disparities and Social Choices in the Design of Privacy Mechanisms
01h 01mFrom Theory to Practice: Crafting Differential Privacy Systems with Haskell
10mCompiling Python Programs into Differentially Private Ones
27mYou've got trust issues, we've got solutions: Differential Privacy