Differential Privacy Made Practical
summary
Differential privacy is shown with concrete choices: noise from a Laplace distribution for provable epsilon bounds, Gaussian only under ε,δ relaxations, and practical demos in an IPython notebook plus Python repos for scale and deep learning. Real deployments named are Android keyboard, Apple quicktype and US Census. Epsilon values matter: epsilon of one is treated as a practical target, epsilon ≥3 is much weaker, and deep‑learning DP yields huge utility loss (examples report near‑random accuracy even by ε≈100).
AI-written summary · about
chapters (13)
- 0:00Differential privacy intro
- 1:07Why DP matters today
- 3:32Hiding individual contributions
- 7:43Noise for privacy and utility
- 10:30DP mathematical definition
- 12:17Epsilon parameter explained
- 14:41Epsilon practical implications
- 17:32Balancing privacy and utility
- 18:59Repeated queries and budget
- 20:26DP in deep learning
- 23:58Contact tracing with DP
- 26:04Summary and open questions
- 26:51Q&A
related talks
01h 14mTutorial: Privacy, Part II
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
27mYou've got trust issues, we've got solutions: Differential Privacy
26mAaron Roth, Associate Professor, University of Pennsylvania, NYC 2017
40mDifferential Privacy with Elliptical Stochastic Processes
31mWhen Existing Techniques in Linear Regression Preserve Differential Privacy
31mGaussian Differential Privacy (GDP)
27mPrivacy-preserving statistical analyses