Communication Dans Un Congrès Année : 2025

Fair and Privacy-preserving Synthetic Data Generation via Clustering-based Variational Autoencoder and Adversarially Debiased Wasserstein Generative Adversarial Networks with Gradient Penalty

Résumé

The increasing reliance on machine learning in sensitive domains, such as healthcare, has amplified concerns about bias and privacy in data-driven decision-making. While fairness-aware generative models aim to mitigate bias, they often depend on labeled data, limiting their applicability in unsupervised settings. Conversely, differentially private generative models ensure privacy but may still encode hidden biases. Existing methods fail to jointly optimize fairness and privacy without explicit supervision. To address this gap, we propose a hybrid generative framework that integrates clustering-based Variational Autoencoder (VAE) with Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN-GP) to generate fair and privacy-preserving synthetic data. The VAE structures latent representations under zero-Concentrated Differential Privacy (zCDP) while incorporating K-Means clustering directly in the latent space. The clustering serves as a factor to influence the generative process into producing samples that resemble real data in unsupervised settings. These structured representations along with cluster labels then guide WGAN-GP's generator toward sample generation and enhance adversarial debiasing through the Fairness Critic, which penalizes correlations between synthetic data and sensitive attributes to ensure fairness. By integrating clustering-based VAEs with WGAN-GP, our framework enforces fairness while maintaining strong privacy guarantees. Experimental results demonstrate that it outperforms existing generative models by effectively reducing bias, preserving privacy, and ensuring high data utility across multiple fairness and privacy metrics.

Fichier principal
Vignette du fichier
Clust-VAE-WGAN-GP_ecml.pdf (1.15 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-05113907 , version 1 (16-06-2025)

Licence

Identifiants

  • HAL Id : hal-05113907 , version 1

Citer

Malek Adouani, Zaineb Chelly Dagdia. Fair and Privacy-preserving Synthetic Data Generation via Clustering-based Variational Autoencoder and Adversarially Debiased Wasserstein Generative Adversarial Networks with Gradient Penalty. ECMLPKDD 2025, Sep 2025, Porto (Portugal), Portugal. ⟨hal-05113907⟩
1107 Consultations
433 Téléchargements

Partager

  • More