Differential privacy and algorithmic fairness don't play well together — and a new benchmark finally maps out what to do about it.
Researchers published what they describe as the first systematic evaluation of fairness interventions applied to differentially private synthetic tabular data. The study centers on AIM, the state-of-the-art marginal-based DP synthesizer. They tested pre-processing, in-processing, and post-processing mitigation strategies across four datasets and multiple group fairness metrics, sweeping a wide range of privacy budgets. The short version: DP alone reliably hurts both model accuracy and demographic parity, but layering in fairness interventions can partially restore equitable outcomes.
The finding that post-processing methods deliver the most stable fairness-utility trade-offs is practically useful. Engineers deploying DP pipelines in hiring, lending, or healthcare — the exact high-stakes domains regulators are watching — now have benchmark evidence that intervening after training is more robust than trying to bake fairness in earlier. That's not obvious, and it shifts where teams should focus effort.
All code and data are released as open-source, which matters: fairness benchmarks are only as credible as their reproducibility. The harder question the paper leaves open is whether these trade-offs hold at the privacy budgets real deployments actually use, which tend to be far tighter than the wide range surveyed here.