Generation and Evaluation of Privacy-Preserving Tabular Synthetic Data Using Self-Attention Generative Adversarial Networks

Abstract

The rapid advancements in Artificial Intelligence (AI) and Machine Learning (ML) have led to the generation of synthetic data using Generative Adversarial Networks (GANs). The generation of high-fidelity synthetic data has become increasingly important in various fields, including healthcare, finance, and social sciences, where data privacy and scarcity pose significant challenges for training ML models. In healthcare and finance, where data privacy is paramount, synthetic data provides a valuable alternative to real data, facilitating research and model development. This paper presents a Self-Attention GAN. This GAN significantly improves the performance of tabular data by incorporating a Self-Attention mechanism in the Discriminator compared to other existing GANs. We evaluated our model using various tabular data, comparing its performance with existing GAN-based approaches. Our results demonstrate that the inclusion of Self-Attention improves the ability of the model to capture complex dependencies within the input data, leading to synthetic datasets that closely mirror the statistical properties of the original data. The improved synthetic data ensures privacy while retaining high utility for machine learning tasks.