🤖 machine learning

Where to Intervene? Benchmarking Fairness-Aware Learning on Differentially Private Synthetic Tabular Data

This paper presents the first systematic benchmark evaluating how pre-, in-, and post-processing fairness interventions perform on differentially private synthetic tabular data, revealing that post-processing methods offer the most stable trade-offs between fairness, utility, and privacy.

Original authors: Vinícius Gabriel Angelozzi, Héber H. Arcolezi

Published 2026-07-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Vinícius Gabriel Angelozzi, Héber H. Arcolezi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a city planner trying to build a fair traffic system. You have a map of the city (your data) showing where people live and where they work. You want to design traffic lights that treat everyone equally, regardless of which neighborhood they come from.

However, there's a problem: the map contains sensitive information about individual drivers. To protect their privacy, you hire a "Privacy Guard" (Differential Privacy) to blur the map. The Guard adds a little bit of "fog" to the map so no one can identify a specific driver, but the overall shape of the roads remains visible.

The Problem:
When you try to build your traffic system using this "foggy" map, you notice something strange. The fog doesn't just hide people; it accidentally makes the traffic lights work worse for people in smaller or less common neighborhoods. The system becomes unfair because of the privacy protection.

The Big Question:
The paper asks: "Where should we step in to fix the unfairness?"
Should we:

Pre-process: Try to "clean" the foggy map before we start building?
In-process: Change the rules of how we build the traffic lights while we are building them?
Post-process: Build the lights first, and then tweak the settings afterward to make them fair?

The Experiment:
The authors set up a massive test kitchen. They used four different "cities" (real-world datasets like income records and criminal justice data) and tried every possible combination:

The Baseline: Building on a clear map (no privacy, no fairness fixes).
The Privacy-Only: Building on a foggy map with no fairness fixes (this showed the problem: privacy made things unfair).
The Fairness-Only: Building on a clear map but trying to force fairness (this showed what's theoretically possible).
The Mix: Building on a foggy map and trying to fix the fairness at different stages.

The Findings (The "Recipe" for Success):

Cleaning the Map First (Pre-processing):
- Analogy: Trying to wipe the fog off the map before you start.
- Result: Some methods worked okay, but they often smudged the map so much that the traffic lights became less accurate overall. You got fairness, but you lost precision. One method (Learning Fair Representations) made the map so blurry that the lights stopped working well at all.
Changing the Rules While Building (In-processing):
- Analogy: Telling the construction crew, "Build the lights, but make sure they are fair," while they are hammering.
- Result: This was a safe, conservative approach. It kept the traffic lights accurate, but it didn't fix the unfairness very much. The "fog" was too strong for these rules to overcome.
Tweaking the Settings Afterward (Post-processing):
- Analogy: Letting the construction crew build the lights normally, and then a specialist comes in afterward to adjust the timing of the lights specifically for the neighborhoods that were getting shortchanged.
- Result: This was the winner. By adjusting the final decisions (the traffic light timing) rather than trying to fix the foggy map or change the construction rules, they achieved the best balance. They fixed the unfairness significantly without ruining the accuracy of the system.

The Takeaway:
If you are using privacy-protected data (the "foggy map") to make decisions, don't try to fix the data itself or change the training rules too much. Instead, build your model first, and then apply a "fairness filter" to the final results.

The paper specifically found that two types of "post-processing" tools (called Reject Option Classification and Equalized Odds) were the most reliable. They acted like a skilled mechanic who could fine-tune the engine after it was built, ensuring everyone got a fair ride without needing to rebuild the whole car.

In short: Privacy creates a "fog" that causes unfairness. The best way to clear that unfairness is to fix the output (the final decision), not the input (the data) or the process (the training).

Technical Summary: Benchmarking Fairness-Aware Learning on Differentially Private Synthetic Tabular Data

Problem Statement
The deployment of machine learning in high-stakes domains necessitates rigorous protection of both privacy and fairness. While Differential Privacy (DP) has become the gold standard for privacy-preserving data analysis, and fairness-aware mechanisms aim to mitigate discrimination, these objectives often conflict. Prior research indicates that DP can amplify disparities across demographic groups, a phenomenon known as "DP disparate impact." Although recent studies have observed that DP synthetic data can degrade fairness metrics, there is a lack of systematic understanding regarding whether established fairness interventions remain effective when applied to DP synthetic data. Specifically, it is unclear where in the machine learning pipeline (pre-processing, in-processing, or post-processing) one should intervene to mitigate unfairness under DP constraints without severely compromising predictive utility.

Methodology
This work presents the first systematic benchmark evaluating fairness interventions on differentially private synthetic tabular data. The study centers on the Adaptive Iterative Mechanism (AIM), identified as the state-of-the-art marginal-based DP synthesizer for tabular data, with supplementary results using the Maximum Spanning Tree (MST) synthesizer.

The benchmark evaluates four pipeline configurations across four datasets (Adult, COMPAS, ACSIncome, and a synthetic BiasOnDemand generator) and three classifiers (XGBoost, Logistic Regression, Random Forest):

Baseline: Training on original, non-private data without fairness interventions.
DP-only: Training on DP synthetic data without fairness interventions.
Fair-only: Applying fairness mechanisms to original data without DP.
DP+Fair: Combining fairness mechanisms with DP synthetic data.

The study assesses three categories of fairness interventions:

Pre-processing: Reweighing (RW), Disparate Impact Remover (DIR), and Learning Fair Representations (LFR).
In-processing: Exponentiated Gradient Reduction (EGR) and Grid Search Reduction (GSR).
Post-processing: Reject Option Classification (ROC), Equalized Odds Post-Processing (EqOdds), and Calibrated Equalized Odds Post-processing (CEOP).

Performance is measured across a wide range of privacy budgets ( $\epsilon \in \{0.05, \dots, 20\}$ ) using utility metrics (Accuracy, F1-score) and group fairness metrics (Model Accuracy Difference, Equalized Opportunity Difference, Statistical Parity Difference). The experimental design ensures that fairness post-processing is calibrated on real held-out data, adhering to the DP post-processing property, with ablation studies confirming that DP-compliant calibration yields consistent qualitative results.

Key Results
The empirical analysis reveals distinct trade-offs depending on the intervention stage:

Pre-processing: Interventions like RW and LFR can reduce group disparities introduced by DP, but they consistently incur a significant cost in predictive utility. LFR enforces strong parity (driving metrics near zero) but often leads to unstable training outcomes and substantial utility degradation. DIR shows limited corrective power under the uniform preprocessing pipeline, performing similarly to the DP-only baseline.
In-processing: Methods like EGR offer a conservative trade-off, preserving utility close to the DP-only baseline while achieving moderate, bounded reductions in disparity. However, they generally fail to reach the fairness levels achievable by the "Fair-only" setting. GSR provides negligible systematic benefits over DP-only.
Post-processing: This stage emerges as the most effective and robust. ROC and EqOdds consistently achieve the strongest reductions in disparity (particularly for Equalized Opportunity and Statistical Parity) while maintaining utility within a bounded degradation relative to the DP-only baseline. They frequently appear on or near the Pareto frontier of the fairness-utility trade-off. CEOP exhibits more heterogeneous behavior and is less reliable as a default choice.

Statistical validation using paired Wilcoxon signed-rank tests confirms that DP+Fair significantly recovers fairness degradation (specifically for EOD and SPD) caused by DP-only, and that post-processing strategies significantly outperform pre- and in-processing strategies in terms of fairness-utility trade-offs.

Significance and Claims
The authors claim to provide the first systematic evaluation of fairness interventions applied to classifiers trained on DP synthetic tabular data. The primary contribution is identifying that post-processing is the most reliable intervention stage for mitigating bias in this context. The paper argues that while DP alone degrades both utility and fairness, applying fairness interventions—specifically post-processing methods like ROC and EqOdds—can partially restore equitable outcomes while preserving competitive utility.

The study highlights that the effectiveness of interventions is not uniform; it depends heavily on the choice of DP synthesizer (AIM vs. MST) and the specific fairness metric. The authors emphasize that their findings are specific to marginal-based DP synthesizers, which are currently state-of-the-art for tabular data, and that the separation of DP generation and fairness intervention allows for the characterization of how established mechanisms behave under fixed privacy budgets. All code, data, and experimental artifacts are released to ensure reproducibility.

Technical Summary: Benchmarking Fairness-Aware Learning on Differentially Private Synthetic Tabular Data

More like this