Exploring Utility-Fairness-Privacy Trade-offs in Synthetic Sensitive Tabular Data
Year: 2026
Type: conference-paper
Abstract: Sharing sensitive tabular healthcare data requires generative models that balance the fundamentally conflicting objectives of statistical utility, demographic fairness, and individual privacy. Conventional configuration relies on scalarized search, which reduces this multidimensional tension to a single score and returns a single operating point, leaving the broader trade-off landscape unexplored. We introduce Evo-Clust-VAE-WGAN-GP, a framework treating generative model configuration as a tri-objective optimization problem over structural and training hyperparameters using NSGA-II to approximate the Pareto front. Rather than a head-to-head optimizer contest, we conduct a landscape characterization study, using Simulated Annealing and Bayesian Optimization as scalarized reference probes to expose what single-point strategies structurally cannot access. Evaluating across five healthcare benchmarks (HIV, Stress, Obesity, Heart Failure, Pediatric), results show that scalarized probes consistently collapse onto dataset-insensitive regions. In contrast, evolutionary search uncovers well-distributed, dataset-dependent Pareto fronts ranging from continuous to fragmented. These findings demonstrate that a single configuration compromise cannot adequately represent generative model behavior, establishing population-based search as a principled framework for exposing the full trade-off landscape in privacy-sensitive data generation.
Source: HAL
Authors: Adouani, Malek, Chelly Dagdia, Zaineb
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214134674