Research map / Federated learning and privacy / Differential privacy
Differentially Private Bootstrap: New Privacy Analysis and Inference Strategies
Zhanyu Wang, Guang Cheng, Jordan Awan
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 deconvolution on DP bootstrap estimates instead of the asymptotic CI or the earlier conservative method (ablation) | DP bootstrap CI for population mean · Population mean, clamped normal data, n=10000, 90% CI | coverage and width: deconvolution CIs similar to non-private bootstrap; asymptotic method always over-covers with larger width | Improved |
| set B=2000 mu^2 (constant SNR) for deconvolution CIs (ablation) | DP bootstrap with deconvolution · Population mean, mu=1, 0.5, 0.3, 0.1 | coverage: B=2000\mu^2 | Improved |
| increase B among values giving satisfactory coverage (B=20,180,500,2000 at mu=1) (ablation) | DP bootstrap with deconvolution · Population mean, mu=1 | CI width: the largest B gives the shortest CI | Improved |
| use the strongest privacy guarantee mu=0.1 (ablation) | DP bootstrap with deconvolution · Population mean, 90% CI | coverage and width: mu=0.1 gives over-coverage and larger width | Hurt |
| compare NoisyVar parametric method to DP bootstrap deconvolution | NoisyVar · Population mean, clamped normal data | coverage: NoisyVar has under-coverage and less width for all mu | Hurt |
| use a larger sample size n for DP bootstrap CIs in logistic regression (ablation) | DP bootstrap (deconvolution and asymptotic) · 2016 Canada Census logistic regression, 1-GDP, B=100 | CI width and P(CI covers 0): as wide as non-private bootstrap CI at n=100000; P(CI covers 0) approx 0 when n>=30000 | Improved |
| DP bootstrap versus DP-CI-ERM at small n | DP bootstrap · 2016 Canada Census logistic regression, 1-GDP | CI width: wider than DP-CI-ERM when n is small | Hurt |
| use deconvolution CIs with larger n in quantile regression (ablation) | DP bootstrap deconvolution · 2016 Canada Census quantile regression tau=0.5, 1-GDP | CI width relative to non-private bootstrap: similar to non-private bootstrap when n>=30000; P(CI covers 0) approx 0 when n>=1000 | Improved |
| use Theorem 3 mixture bound on bootstrap privacy instead of assuming bootstrap is free | DP bootstrap with 1-GDP Gaussian mechanism · Privacy analysis | privacy guarantee: bootstrap no longer satisfies 1-GDP | Hurt |
Closest papers by meaning
- On Inherent Privacy of Posterior Sampling: A Unified R\'enyi-Divergence Framework (NeurIPS 2026 · similarity 0.89)
- A Simple and Practical Method for Reducing the Disparate Impact of Differential Privacy (AAAI 2024 · similarity 0.89)
- Analysis of Differentially Private Synthetic Data: A Measurement Error Approach (AAAI 2024 · similarity 0.88)
- Private Statistical Estimation via Truncation (NeurIPS 2025 · similarity 0.88)
- Locally Optimal Private Sampling: Beyond the Global Minimax (NeurIPS 2025 · similarity 0.87)
- Privacy Guarantees in Posterior Sampling under Contamination (NeurIPS 2026 · similarity 0.87)
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.
