Practical Considerations for Differential Privacy
summary
Seven years of building DP systems at Google produce concrete lessons: production libraries used by hundreds of teams, a recurring inbounding sensitivity problem where clipping a receipt bound trades bias for noise and forces choices about storing user IDs, and a fragile privacy budgeting reality where epsilons add (epsilon 0.1 ≈ 11% utility change) and one non‑DP action counts as infinite loss. Practical deployments therefore require DP-aware tooling for bound selection, accounting for cumulative epsilon, and empirical risk tests.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:13What is differential privacy
- 4:42DP implementation challenges
- 5:05Sensitivity bounding
- 8:30Multiple user contributions
- 9:13Privacy loss accounting
- 12:23Philosophical limits of DP
- 15:55Role and limits of epsilon
- 17:19Risk-based privacy metrics
related talks
01h 14mTutorial: Privacy, Part II
01h 05mMarkets and Incentives for Transacting on (Personal) Information, Part III
20mDesigning for User Privacy: Integrating Differential Privacy into Ad Measurement...
26mAaron Roth, Associate Professor, University of Pennsylvania, NYC 2017
01h 00mPANEL: Statistical Theory, Privacy and Data Analysis
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
13mEvaluating the Usability of Differential Privacy Tools with Data Practitioners
01h 01mFrom Theory to Practice: Crafting Differential Privacy Systems with Haskell