Heterogeneity and Privacy in Modern Learning
summary
Survey of concrete models and provable separations. Proves DP meta‑learning ≡ DP multitask but shows joint differential privacy is strictly stronger than the billboard/public‑representation model by giving a mean‑estimation separation where error scales with dimension d under billboard but only with number of released tasks t under JDP. Gives algorithms that weight noisy local outputs by epsilon and match lower bounds for 1D and d‑dimensional mean and distribution learning. Defines Bayesian coordinate DP and a clustered LDP layering scheme with provable error bounds.
AI-written summary · about
chapters (7)
- 0:00Privacy in heterogeneous learning
- 1:47Private personalization in multitask learning
- 9:04Differential privacy for multitask learning
- 14:51Privacy budgets per user
- 20:58Feature-level privacy with Bayesian DP
- 29:42Algorithm structure and clustering
- 32:37Q&A
related talks
36mLearning without Interaction Requires Separation
37mUser-level and federated local differential privacy
17mData privacy and reliable learning
58mLarge-Scale Private Learning, Part III
01h 14mTutorial: Privacy, Part II
33mPrivacy of Decentralized Machine Learning
26mAaron Roth, Associate Professor, University of Pennsylvania, NYC 2017
45mProving Differential Privacy via Relational Types