Research map / Federated learning and privacy / Differential privacy
Algorithms for bounding contribution for histogram estimation under user-level privacy
Yuhan Liu, Ananda Theertha Suresh, Wennan Zhu, Peter Kairouz, Marco Gruteser
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 |
|---|---|---|---|
| Select the contribution threshold privately using the dataset-specific objective instead of the median threshold | Unbounded-domain histogram estimation · All reported synthetic and Sentiment140 settings | relative loss: i.i.d., d=50: 0.0015 vs 0.0397; non-i.i.d., d=50: 0.0124 vs 0.6957; Sentiment140, d=1000: 0.1483 vs 0.5912 | Improved |
| Select the contribution threshold privately using the dataset-specific objective instead of the 90% quantile threshold | Unbounded-domain histogram estimation · All reported synthetic and Sentiment140 settings | relative loss: i.i.d., d=50: 0.0015 vs 0.0049; non-i.i.d., d=50: 0.0124 vs 0.1645; Sentiment140, d=1000: 0.1483 vs 0.2699 | Improved |
| Use private threshold selection rather than the best threshold chosen non-privately | Unbounded-domain histogram estimation · i.i.d. synthetic, d=50, d=100, and d=200 | relative loss: 0.001 -> 0.0015; 0.0026 -> 0.0026; 0.0048 -> 0.0048 | No clear effect |
| Use private threshold selection rather than the best threshold chosen non-privately | Unbounded-domain histogram estimation · non-i.i.d. synthetic, d=50, d=100, and d=200 | relative loss: 0.0094 -> 0.0124; 0.0175 -> 0.0224; 0.0461 -> 0.0462 | No clear effect |
| Use private threshold selection rather than the best threshold chosen non-privately | Unbounded-domain histogram estimation · Sentiment140, d=100 and d=1000 | relative loss: 0.0076 -> 0.0285; 0.1250 -> 0.1483 | Mixed |
| Estimate the bounded-domain Gaussian clipping threshold with DP-SGD rather than the DP-M-quantile heuristic | Bounded-domain histogram estimation · Sentiment140 and SNAP | estimation error | Improved |
| Use DP-M-quantile to set the clipping threshold | Bounded-domain histogram estimation · Sentiment140 and SNAP | estimation error | Hurt |
| Apply Poisson-based post-processing to the clipped count estimate to reduce clipping bias | Count estimation under non-i.i.d. Poisson user counts · Count estimation | squared error of the mean estimate | Improved |
Closest papers by meaning
- Bounding User Contributions: A Bias-Variance Trade-off in Differential Privacy (ICML 2019 · similarity 0.91)
- Differentially Private Anonymized Histograms (NeurIPS 2019 · similarity 0.90)
- Anonymized Histograms in Intermediate Privacy Models (NeurIPS 2022 · similarity 0.90)
- Learning discrete distributions: user vs item-level privacy (NeurIPS 2020 · similarity 0.88)
- It's My Data Too: Private ML for Datasets with Multi-User Training Examples (ICML 2025 · similarity 0.87)
- Private and Personalized Frequency Estimation in a Federated Setting (NeurIPS 2024 · 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.
