Integrating domain constraints in tabular data generation process

Salijona Dyrmishi, Mihaela C. Stoian, Eleonora Giunchiglia, Maxime Cordy

Keywords: Deep Generative Models, Tabular Data, Adversarial Attack

Abstract: Deep Generative Models (DGMs) excel at generating synthetic tabular data but struggle to enforce domain-specific constraints essential in applications like ML robustness testing. To overcome this, we introduce Constrained Deep Generative Models (C-DGMs), which use a Constraint Layer to ensure that generated data adhere to predefined rules while staying true to the original distributions. We extend this approach to create Constrained Adversarial DGMs (C-AdvDGMs), which generate adversarial examples that both satisfy domain constraints and effectively assess the robustness of machine learning models.

Submission Number: 3