Synthetic Data Generation Using Generative Artificial Intelligence

Abstract

The increasing demand for large-scale, high-quality datasets for machine learning has been constrained by privacy regulations, data scarcity, and high annotation costs. Generative Artificial Intelligence (GenAI) has emerged as a transformative solution through synthetic data generation—artificially created data that mimics real-world statistical properties while safeguarding sensitive information. This study investigates the effectiveness of various synthetic data generation methods across three critical dimensions: fidelity, utility, and privacy. Using a mixed-methods design combining quantitative evaluation with application-specific assessment, we examined three primary generation approaches—statistical-based methods (synthpop), deep learning methods (Generative Adversarial Networks, Variational Autoencoders), and Large Language Model-based generation—across multiple datasets and downstream tasks. Quantitative results demonstrate that classical methods such as synthpop and copula consistently achieve strong performance across both high-dimensional and low-dimensional datasets, combining low re-identification risk (ε-identifiability of 0.25-0.35) with competitive utility. Deep learning models achieve higher distributional fidelity but incur elevated privacy risks (ε-identifiability > 0.4). Notably, general utility metrics showed only weak correlation with specific task performance, emphasizing the need for multimetric evaluation. The findings reveal that data normalization techniques critically impact generation quality, with Max-Abs normalization consistently yielding more accurate and stable synthetic profiles across all models. These findings contribute to a nuanced understanding of synthetic data generation and provide practical guidelines for selecting appropriate methods based on application requirements, data characteristics, and privacy constraints.