Generative Adversarial Networks in Synthetic Data Creation

Abstract

Synthetic data generation has become essential for training machine learning models in domains where real data is scarce, sensitive, or imbalanced. Generative adversarial networks offer a powerful framework for producing realistic synthetic samples, yet the practical question of which GAN variant to use for which data modality -- and whether GANs outperform non-adversarial alternatives -- remains inadequately addressed. This study presents a controlled comparison of six generative methods -- DCGAN, StyleGAN2, CTGAN (tabular), TimeGAN (temporal), a variational autoencoder (VAE) baseline, and diffusion models (DDPM) -- across four synthetic data applications: medical image augmentation (chest X-ray, retinal fundus), tabular data synthesis (census, credit), time series generation (financial, clinical), and class-imbalance oversampling (fraud, rare disease). Quality was assessed on three dimensions: fidelity (statistical similarity to real data via FID, MMD, and column-wise distributional tests), utility (downstream ML performance when training on synthetic data), and privacy (resistance to membership inference attacks measuring whether individual real records can be identified from synthetic output). A total of 2,880 experiments were conducted. StyleGAN2 achieved the best image fidelity (FID = 8.4 +- 1.2 on chest X-rays) and downstream utility (classifier AUC within 1.8% of real-data baseline). CTGAN produced the highest-utility tabular data (downstream AUC within 2.4% of real). Diffusion models matched or exceeded GAN fidelity on images (FID = 7.8 +- 0.9) with superior training stability but 4.2x slower generation. The privacy-utility trade-off was quantified: adding differential privacy to CTGAN (epsilon = 10) reduced membership inference advantage from 12.4% to 3.8% at a utility cost of 4.6 AUC points. A practical selection framework mapping data modality, quality requirements, and privacy constraints to recommended generative configurations is proposed