Research map / Federated learning and privacy / Differential privacy
Learning with User-Level Privacy
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase samples per user m under the paper's i.i.d. or limited-heterogeneity assumptions | User-level private mean estimation, ERM, and SCO · User-level private learning | privacy error: O(1/√m) | Helped |
| Increase the number of users n | User-level private learning · User-level private learning | privacy error: O(1/n) | Helped |
| Use concentration-aware private mean estimation instead of noise calibrated to the full range | Mean estimation and private query answering · High-dimensional mean estimation | privacy error: error proportional to τ rather than the whole range B | Helped |
| Apply random rotation before coordinate-wise mean estimation | WinsorizedMeanHighD · ℓ₂-concentrated high-dimensional mean estimation | mean squared error: avoids a superfluous d factor | Helped |
| Use Phased ERM with successively increasing regularization and fresh, disjoint user batches | Stochastic convex optimization · Stochastic convex optimization | population excess risk: Õ(R√(G G̲)/√(mn) + R G̃√d/(n√m ε)) | Helped |
| Let each user contribute more samples when user distributions are arbitrary | User-level private learning · Arbitrary user distributions | utility | No clear effect |
| Increase the number of users rather than samples per user | User-level private learning · Regime with heterogeneity between the i.i.d. and arbitrary-distribution cases | privacy cost: O(1/n) versus O(1/√m) | Helped |
| Increase m while holding n fixed | Any considered user-level private learning task · Fixed number of users | risk: Ω(e^{-nε}) regardless of m | Hurt |
Closest papers by meaning
- Learning discrete distributions: user vs item-level privacy (NeurIPS 2020 · similarity 0.93)
- User-level Private Stochastic Convex Optimization with Optimal Rates (ICML 2023 · similarity 0.92)
- Faster Algorithms for User-Level Private Stochastic Convex Optimization (NeurIPS 2024 · similarity 0.92)
- Tight and Robust Private Mean Estimation with Few Users (ICML 2022 · similarity 0.92)
- The Limits of Differential Privacy in Online Learning (NeurIPS 2024 · similarity 0.89)
- Differentially Private Optimization with Sparse Gradients (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.