Research map / Federated learning and privacy / Differential privacy
Bayesian Estimation of Differential 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 |
|---|---|---|---|
| Joint posterior of (FNR,FPR) with credible intervals instead of separate Clopper-Pearson intervals | Bayesian approach for epsilon estimation · SST2 no DP, CIFAR10 (avg and worst case), N=1000 | credible interval width vs Clopper-Pearson: -34% to -40% | Helped |
| Joint posterior of (FNR,FPR) with credible intervals instead of separate Clopper-Pearson intervals | Bayesian approach for epsilon estimation · CIFAR10 worst case eps=10 and SST2 eps=4 | meaningful interval vs trivial (0, inf) | Helped |
| Jeffreys intervals instead of Clopper-Pearson for FNR/FPR | epsilon confidence intervals · CIFAR10 and SST2 (N=1000) | interval width vs Clopper-Pearson: -0% to -19% | Helped |
| Heuristic: draw m samples per model (m=10,100,1000) instead of one model per sample (ablation) | Bayesian approach on SST2 no DP · SST2 no DP, N=1000 | credible interval width: 1.6 (m=1) vs 1.4 (m=10, 100, 1000) | No clear effect |
| Heuristic with worst-case challenge points: m=10 and m=100 draws per model (ablation) | CIFAR10 DP eps=10 Bayesian estimate · CIFAR10 worst case | CDF absolute error vs m=1 baseline: 0.05 | No clear effect |
| Heuristic with worst-case challenge points: single model with m=1000 samples (ablation) | CIFAR10 DP eps=10 Bayesian estimate · CIFAR10 worst case | CDF error vs m=1 baseline | Hurt |
| Heuristic with average-case challenge points: m=10, 100, 1000 samples per model (ablation) | CIFAR10 DP eps=10 Bayesian estimate · CIFAR10 average case | lower bound density error vs m=1 baseline: 0 (m=10), 0.001 (m=100, 1000) | No clear effect |
| Heuristic with m samples per model on SST2 DP eps=4 (ablation) | Bayesian estimate on RoBERTa SST2 · SST2 eps=4 | CDF max error vs m=1 baseline: max 0.07 | No clear effect |
| Heuristic with m samples per model on SST2 trained without DP (ablation) | Bayesian estimate on non-DP RoBERTa SST2 · SST2 no DP | CDF error vs m=1 baseline: 0.12 (m=10) | Hurt |
Closest papers by meaning
- Attack-Aware Noise Calibration for Differential Privacy (NeurIPS 2024 · similarity 0.90)
- One-shot Empirical Privacy Estimation for Federated Learning (ICLR 2024 · similarity 0.89)
- Privacy Guarantees in Posterior Sampling under Contamination (NeurIPS 2026 · similarity 0.89)
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.89)
- Instance-Specific Asymmetric Sensitivity in Differential Privacy (NeurIPS 2024 · similarity 0.88)
- On Inherent Privacy of Posterior Sampling: A Unified R\'enyi-Divergence Framework (NeurIPS 2026 · similarity 0.88)
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.