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

As machine learning models are increasingly deployed in sensitive domains, researchers must balance the competing demands of data privacy and algorithmic fairness. While Differential Privacy (DP) is the gold standard for protecting individual data, it often introduces statistical noise that disproportionately harms underrepresented groups. This paper investigates whether standard fairness-aware interventions—designed for original, non-private data—remain effective when applied to DP-generated synthetic tabular data. The authors conduct a systematic benchmark using the Adaptive Iterative Mechanism (AIM), a state-of-the-art marginal-based DP synthesizer, across four datasets and three intervention stages: pre-processing, in-processing, and post-processing.

The study compares four pipeline configurations: a baseline (original data), DP-only (DP synthetic data), Fair-only (original data with fairness intervention), and DP+Fair (fairness intervention applied to DP synthetic data). By testing these configurations, the authors isolate how privacy guarantees reshape the effectiveness of fairness mechanisms. They evaluate performance using standard fairness metrics, such as Statistical Parity Difference and Equalized Odds, alongside predictive utility metrics.

The results demonstrate that while DP synthetic data generation often degrades both utility and fairness, applying fairness interventions can partially recover equitable outcomes. Among the strategies tested, post-processing methods—such as Reject Option Classification and Equalized Odds Post-Processing—consistently provide the most stable fairness-utility trade-offs. These methods often reach the Pareto frontier, achieving significant fairness improvements without sacrificing excessive predictive utility. In contrast, in-processing methods, while useful, often struggle to provide the same level of disparity reduction in the context of DP synthetic data.