Talk by Audra McMillan (Apple Inc)
summary
Defines local pan-privacy as protection against a continual on‑device intruder who sees every internal state and the server output, proves an information‑theoretic lower bound giving additive error ~sqrt(N T)/ε for counting users with at least one event, and shows that switching to computational DP plus re-randomizable public-key encryption recovers the same utility and message size as the local model by keeping encrypted state, re‑randomizing when no event, replacing with an encryption of one on event, and applying randomized response to ciphertexts; rerandomizability is shown necessary and tightening the T dependence is open.
AI-written summary · about
chapters (7)
- 0:00Local pan-privacy intro
- 2:29Pan-privacy model defined
- 8:44Information-theoretic lower bounds
- 15:10Computational differential privacy
- 19:51Re-randomizable encryption algorithm
- 23:24Histogram estimation with pan-privacy
- 26:39Summary and open problems
related talks
58mLarge-Scale Private Learning, Part III
37mUser-level and federated local differential privacy
29mPrivacy-preserving methods: Building secure projects
26mAaron Roth, Associate Professor, University of Pennsylvania, NYC 2017
44mLocally Private Statistical Inference: Communication, Samples, and Shared Randomness
27mOn Distributed Differential Privacy and Counting Distinct Elements
42mDifferential Privacy: From Theory to Deployment
26mLocally Private Histograms in All Privacy Regimes