Research map / Federated learning and privacy / Differential privacy
Differentially Private Post-Processing for Fair Regression
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 a differentially private post-processing algorithm that makes a given regressor statistically fair (approximate KS-based statistical parity) by privately estimating group output distributions via histogram density estimation with Laplace noise, then computing a Wasserstein barycenter and optimal transports.
Problem
Learning fair regressors under statistical parity while preserving differential privacy of training data, since exact group fairness conflicts with DP.
Approach
Post-process any pre-trained regressor: privately estimate group-conditional output PMFs via histogram density estimation with Laplace noise and isotonic-regression renormalization, then compute the (KS-relaxed) Wasserstein barycenter and optimal transport maps; analysis reveals a bias-variance trade-off in the number of bins, with fewer bins favoring fairness.
Key results
- On Law School with epsilon=0.1, exact statistical parity is achieved with k=1 bin but at MSE 0.6772, showing the bias-variance trade-off in bins.
- Results are insensitive to privacy budget epsilon unless very strong privacy or very large k is used, since estimation error dominates DP noise when n is large relative to k/(w_a eps^2).
- Theory gives risk bound O(sqrt(k|A|/n) + k|A|/(n eps)) plus 8/k discretization bias, and fairness bound with only variance terms.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use k=1 histogram bin instead of a larger bin count (ablation) | Private fair post-processing · Law School | KS distance (ΔSP) | Helped |
| Use k=1 histogram bin instead of a larger bin count (ablation) | Private fair post-processing · Law School | MSE: 0.6772 MSE | Hurt |
| Use fewer histogram bins (ablation) | Private fair post-processing · Law School | KS distance (ΔSP) | Helped |
| Use fewer histogram bins (ablation) | Private fair post-processing · Law School | MSE | Hurt |
| Increase histogram bins from k=36 to k=60 (ablation) | Private fair post-processing · Law School, ε=0.1 | error-fairness trade-off | Hurt |
| Demand stricter differential privacy (ablation) | Private fair post-processing · Law School | error-fairness trade-off | Hurt |
| Relax fairness tolerance α from exact statistical parity (ablation) | Private fair post-processing · Law School and Communities & Crime | MSE | Helped |
| Increase privacy budget ε (ablation) | Private fair post-processing · Law School | error-fairness trade-off | Helped |
Surprising findings
- Using fewer bins always favors fairness at the expense of error, so the fairness guarantee contains no bias term.
- The paper reports that the cost of discretization is usually insignificant compared to fairness gains, unless the model is already fair.
Limitations
- Only the attribute-aware setting is studied; attribute-blind extension left to future work
- Experiments use ground-truth responses with identity regressor (X=Y, f=Id), isolating post-processing only
- Pure epsilon-DP requires an interval [s,t] known from domain knowledge; hyperparameter tuning (k, alpha) itself may leak privacy
Experimental setup
Datasets: Communities & Crime Law School
Benchmarks: Communities & Crime Law School
Metrics: mean squared error KS distance statistical parity violation excess risk
Builds on: Wasserstein barycenter fair post-processing (Chzhen et al.) Laplace mechanism differential privacy post-processing
Closest papers by meaning
- Differentially Private Fair Learning (ICML 2019 · similarity 0.90)
- Improving Fairness and Privacy in Selection Problems (AAAI 2021 · similarity 0.88)
- Geometry of Relaxed Fair Regression: A Unified Framework for Aware and Unaware Settings (NeurIPS 2026 · similarity 0.86)
- Differentially Private Model Merging (NeurIPS 2026 · similarity 0.86)
- Instance-Specific Asymmetric Sensitivity in Differential Privacy (NeurIPS 2024 · similarity 0.86)
- Provable Privacy with Non-Private Pre-Processing (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.