Research map / Federated learning and privacy / Differential privacy
Near-Optimal Private Linear Regression via Iterative Hessian Mixing
Omri Lev, Moshe Shenfeld, Vishwak Srinivasan, Katrina Ligett, Ashia C. Wilson
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use Iterative Hessian Mixing instead of AdaSSP and LinMix for DP-OLS | IHM (T=3, k=6*max{d,log(4T/rho)}, C=C_Y) · 33 UCI linear regression datasets, eps in [0.1,10] | normalized excess empirical risk | Improved |
| reduce IHM iterations from T=3 to T=2 on datasets with L(theta*) close to ||Y||^2 (ablation) | IHM · Solar, Pendulum, Fertility, Forest, Kin40k, TamiElectric, 3droad, Pumadyn32nm; eps <= 1 | excess empirical risk | Improved |
| increase IHM iterations and sketch size (T=6, k=12*max{d,log(4T/rho)} vs T=3) (ablation) | IHM · Synthetic data with lambda_min(X^TX) about n/d, high eps | excess empirical risk | Improved |
| omit the -eta^2 theta_t term from the noisy gradient (ablation) | IHM · empirical evaluation | excess empirical risk | Improved |
| apply sketch only to X (Hessian mixing) rather than to concatenated (X,Y) | IHM vs LinMix · DP-OLS, well-conditioned X | noise level / excess risk | Improved |
| compare IHM with DP-GD baseline (T=3, same clipping, lr 0.25) | IHM · UCI datasets, Figures 8-9 | excess empirical risk | Improved |
| IHM under frequent clipping (violating theorem assumptions) | IHM · Synthetic low-rank covariates, more than 40% of iterations clipped | excess empirical risk | Improved |
| use LinMix instead of AdaSSP | LinMix · UCI datasets | excess empirical risk | Mixed |
Closest papers by meaning
- Private Gradient Descent for Linear Regression: Tighter Error Bounds and Instance-Specific Uncertainty Estimation (ICML 2024 · similarity 0.90)
- Public-data Assisted Private Stochastic Optimization: Power and Limitations (NeurIPS 2024 · similarity 0.87)
- Better Locally Private Sparse Estimation Given Multiple Samples Per User (ICML 2024 · similarity 0.86)
- On Inherent Privacy of Posterior Sampling: A Unified R\'enyi-Divergence Framework (NeurIPS 2026 · similarity 0.86)
- Differentially Private High-dimensional Variable Selection via Integer Programming (NeurIPS 2025 · similarity 0.86)
- On the Computational Complexity of Private High-dimensional Model Selection (NeurIPS 2024 · similarity 0.86)
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.
