Scholarlink

In this paper, we explore whether incorporating prior knowledge about the data can enhance GAN performance in the context of synthetic data generation for privacy protection and identify effective methodologies for doing so. We propose three approaches for integrating auxiliary information: (1) embedding public constraints into the adversarial loss function, (2) preserving correlation structures between attributes, and (3) leveraging Bayesian networks to model attribute dependencies and encode them into Conditional GANs. Through comprehensive empirical evaluations against existing baselines, we demonstrate that Bayesian networks and public constraints significantly improve the fidelity and realism of synthetic data. Furthermore, GAN-generated synthetic data lacks inherent privacy protections, making it susceptible to privacy attacks. To address this, we incorporate DP mechanisms into the GAN framework, ensuring robust privacy guarantees while maintaining data utility. The proposed approaches are evaluated for their effectiveness in generating high-quality, privacy-preserving synthetic data, offering valuable insights for future advancements in GAN-based synthetic data generation.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Using Prior Knowledge to Improve GANs for Tabular Data Without Compromising Privacy

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sonakshi Garg,
  • Marcel Neunhoeffer,
  • Jörg Drechsler,
  • Vicenç Torra

摘要

In this paper, we explore whether incorporating prior knowledge about the data can enhance GAN performance in the context of synthetic data generation for privacy protection and identify effective methodologies for doing so. We propose three approaches for integrating auxiliary information: (1) embedding public constraints into the adversarial loss function, (2) preserving correlation structures between attributes, and (3) leveraging Bayesian networks to model attribute dependencies and encode them into Conditional GANs. Through comprehensive empirical evaluations against existing baselines, we demonstrate that Bayesian networks and public constraints significantly improve the fidelity and realism of synthetic data. Furthermore, GAN-generated synthetic data lacks inherent privacy protections, making it susceptible to privacy attacks. To address this, we incorporate DP mechanisms into the GAN framework, ensuring robust privacy guarantees while maintaining data utility. The proposed approaches are evaluated for their effectiveness in generating high-quality, privacy-preserving synthetic data, offering valuable insights for future advancements in GAN-based synthetic data generation.