Research map / Federated learning and privacy / Differential privacy
Privacy-Preserving Data Release Leveraging Optimal Transport and Particle Gradient Descent
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PrivPGD, a differentially private tabular data generator that matches privatized 2-way marginals by minimizing a sliced Wasserstein distance over particles via GPU-parallel particle gradient descent, with pluggable DP domain-specific penalty terms.
Problem
Generating differentially private synthetic tabular data from noisy marginals; existing generation steps (e.g., PGM) are memory/runtime heavy with many marginals, ignore data geometry, and are hard to extend with domain constraints.
Approach
Replace the PGM graphical-model generation step with particle gradient descent that minimizes a sliced-Wasserstein (SW_2^2) divergence between particle marginals and the projected privatized signed marginals, using an ordered embedding of the domain so geometry is preserved, and adding DP differentiable penalty terms for constraints.
Key results
- At epsilon=2.5, PrivPGD achieves the lowest covariance error on most datasets (e.g., Inc. 0.001 vs 0.014 for PGM+AIM) and the best downstream error on several datasets, ranking best or second-best in most metrics.
- PrivPGD consistently outperforms PGM+MST and most query-based methods (GEM, RAP) across datasets, and is competitive with PGM+AIM, but is outperformed by PGM-based methods on TV distance.
- PrivPGD's advantage diminishes at smaller epsilon: at epsilon=0.2 it loses ground to PGM+AIM and Private GSD on several metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PrivPGD instead of PGM+MST for marginal-based synthesis | DP tabular data synthesis · Across datasets | downstream error | Helped |
| Use PrivPGD instead of PGM+MST for marginal-based synthesis | DP tabular data synthesis · Across datasets, except Diabetes thresholding queries | covariance and query errors | Helped |
| Use PrivPGD instead of PGM+AIM | DP tabular data synthesis · Across datasets, except ACS Mobility thresholding queries | covariance and query errors | Helped |
| Use PrivPGD instead of query-based methods | DP tabular data synthesis · Most evaluated scenarios | overall performance | Helped |
| Use PrivPGD instead of Private GSD | DP tabular data synthesis · ACS Public Coverage and some other scenarios | overall performance | Mixed |
| Minimize sliced Wasserstein distance rather than total variation distance | PrivPGD · Across 9 tabular datasets | average SW1 distance | Helped |
| Increase domain-specific regularization strength | PrivPGD · ACS Income and ACS Employment | domain-specific threshold query error | Helped |
Surprising findings
- PrivPGD minimizes SW1 distance rather than TV, and PGM-based methods (which target TV) achieve lower TV error on several datasets such as Inc. and Taxi, so the TV advantage does not hold uniformly.
- At low privacy budgets (epsilon=0.2), PGM-based methods and Private GSD often beat PrivPGD, which the authors attribute to PGM's maximum-entropy inductive bias under high noise.
Limitations
- Performance edge over PGM-based methods shrinks at small privacy budgets (epsilon=0.2 and 1.0), where PGM's maximum-entropy bias helps.
- Geometry-preserving embedding is ordinal; features like race need other embeddings, which are left to future work.
- Only Gaussian mechanism privatization and fixed hyperparameters are evaluated; the domain-specific constraint experiments are limited to a single thresholding-query example.
Experimental setup
Datasets: ACS Income ACS Employment ACS Mobility ACS Traveltime ACS Public Coverage Medical charges Black Friday NYC Taxi Green December 2016 Diabetes 130-US
Benchmarks: ACS Income ACS Employment ACS Mobility ACS Traveltime ACS Public Coverage Medical charges Black Friday NYC Taxi Green December 2016 Diabetes 130-US
Metrics: downstream error covariance error counting query error thresholding query error average SW1 distance average TV distance
Baselines: PGM+AIM PGM+MST Private GSD RAP GEM
Builds on: PGM MWEM Private GSD sliced Wasserstein distance
Closest papers by meaning
- Do You Really Need Public Data? Surrogate Public Data for Differential Privacy on Tabular Data (NeurIPS 2025 · similarity 0.87)
- Scalable DP-SGD: Shuffling vs. Poisson Subsampling (NeurIPS 2024 · similarity 0.87)
- Instance-Specific Asymmetric Sensitivity in Differential Privacy (NeurIPS 2024 · similarity 0.87)
- Private Evolution Converges (NeurIPS 2025 · similarity 0.87)
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.86)
- Perturb-and-Project: Differentially Private Similarities and Marginals (ICML 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.