Research map / Federated learning and privacy / Differential privacy
A Simple and Practical Method for Reducing the Disparate Impact of Differential Privacy
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.
Shows that simple stratification (run DP separately per group, then recombine with public group sizes) reduces disparate impact of DP while costing little global accuracy, supported by theory for DP mean estimation and experiments on DP synthetic data.
Problem
Differentially private mechanisms add noise that degrades utility unevenly across demographic subgroups, producing disparate impact; the paper asks whether a simple baseline can reduce subgroup disparities without hurting population-level statistics.
Approach
Stratify: compute independent DP estimates or synthesizers for each group (disjoint, so parallel composition gives no extra privacy cost), then recombine them using public group-size weights to get global statistics; theory via Dirichlet-distributed group sizes shows the aggregation cost is small.
Key results
- Stratified GEM reduces parity error by at least 200% relative to vanilla GEM on Folktables employment prediction.
- For epsilon >= 1, stratified synthesizers differ from non-stratified classifier accuracy by at most 2%.
- For a New York ACS example with alpha about 0.13, expected stratification error is claimed to be about 80% better than the worst-case bound.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Stratify Coinpress and aggregate group estimates | Coinpress · Synthetic Gaussian mixture | global mean estimation error | No clear effect |
| Stratify Coinpress by protected groups | Coinpress · Folktables employment prediction, SCHL, WKHP, and JWMNP | mean estimation error | Helped |
| Stratify Coinpress by protected groups | Coinpress · Folktables employment prediction, intersectional groups | parity error | Helped |
| Stratify private synthesizers | MST, GEM, and AIM · Folktables employment prediction | classifier accuracy: ≤ 2% | Mixed |
| Stratify GEM by groups | GEM · Folktables employment prediction | parity error: ≥ 200% | Helped |
| Stratify GEM | GEM · Folktables data synthesis, ε ≤ 1.0 | parity error | Helped |
| Stratify AIM | AIM · Folktables data synthesis, ε < 1.0 | parity error | Helped |
| Stratify AIM | AIM · All 3-Way Marginal workload | parity error | Hurt |
| Increase privacy budget | Stratified synthesizers · Folktables employment prediction | classifier accuracy | Helped |
Surprising findings
- Naive stratification can improve global mean estimation error over non-stratified CoinPress for meaningful subgroup structures such as SCHL and WKHP.
- Stratification only seemed to harm overall synthetic data utility in low privacy regimes (epsilon < 0.1).
Limitations
- Relies on access to accurate public group-size estimates for aggregation, which may not be available in some settings.
- No theoretical guarantees for stratified private data synthesizers; they are only empirically evaluated.
- Targets parity error and may trade off worse performance on other metrics.
Experimental setup
Datasets: Folktables ACS New York 2018 synthetic Gaussian mixture
Benchmarks: Folktables employment prediction Folktables mobility Folktables ACS New York
Metrics: normalized mean estimation error parity error classifier accuracy on synthetic data 3-way marginal workload error demographic parity false negative rate difference
Baselines: CoinPress MST GEM AIM
Builds on: CoinPress MST GEM AIM MWEM
Closest papers by meaning
- Fairness Issues and Mitigations in (Differentially Private) Socio-Demographic Data Processes (AAAI 2025 · similarity 0.90)
- Optimal Survey Design for Private Mean Estimation (ICML 2025 · similarity 0.88)
- Improving the Variance of Differentially Private Randomized Experiments through Clustering (ICML 2025 · similarity 0.87)
- Privacy Profiles for Private Selection (ICML 2024 · similarity 0.87)
- Mean Estimation in the Add-Remove Model of Differential Privacy (ICML 2024 · similarity 0.87)
- A Generalized Shuffle Framework for Privacy Amplification: Strengthening Privacy Guarantees and Enhancing Utility (AAAI 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.