Differential Privacy: From Theory to Deployment
summary
Differential privacy is deployed on iOS 10 at ~300 million devices to discover new keyboard words in over 10 languages while protecting users via local differential privacy. The system combines randomized response with a count-min sketch and a Hadamard basis rotation to compress a one-hot vector into one bit communication per client, achieving heavy-hitter recovery down to about √(n log S) frequency with server time ~O(n log S) and practical accuracy; patents and press follow the deployment.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:45What is Differential Privacy
- 1:23Learning from Private Data
- 4:51Other Applications
- 10:44Local Differential Privacy
- 12:31Privacy Knob
- 14:43Premise
- 15:34Private Heavyhitters
- 17:35Communication Constraints
- 19:11Client Side
- 22:22Problem
- 24:07Solution
- 26:17Computation
- 27:08Clientside Computation
- 30:48Differentially Private
- 31:34Sampling
- 31:59Bread and Jam
- 38:12Engineering Challenges
- 38:37Theory Bounds
- 40:05Challenges
- 41:02Summary
related talks
01h 00mDifferential Privacy: Fundamentals to Forefront I
09mReplication: The Effect of Differential Privacy Communication on German Users'...
58mLarge-Scale Private Learning, Part III
19mPractical Considerations for Differential Privacy
10mCommunicating Differential Privacy Guarantees to Data Subjects
14mDifferentially-Private Software Frequency Profiling under Linear Constraints
42mDifferential Privacy
31mIs your model private?