Trust-Aware Benchmarking of GAN, VAE, and Diffusion Models for Synthetic Data in Image and Tabular Domains

Synthetic data generation using generative AI models offers a promising solution to challenges in data availability, privacy, and fairness. However, comparative insights across model families and data modalities remain limited, especially when trust-related dimensions such as fairness, privacy, and efficiency are considered alongside fidelity.This paper presents a trust-aware benchmarking study across both visual (CIFAR-10, MNIST) and tabular (Adult Income) domains, evaluating representative baselines—vanilla GAN/WGAN-GP, standard and β-VAE, DDPM and classifier-guided DDPM, and hybrids such as VAE-GAN and Latent Diffusion. Advanced models including StyleGAN2/3, BigGAN, CTGAN, Diffusion Transformers, and GigaGAN are acknowledged to situate the findings within the evolving landscape of foundation-scale generators.Performance is assessed using a multi-objective framework that integrates fidelity (FID, precision/recall), fairness (demographic parity), privacy leakage resistance, and computational efficiency. Results show that hybrid latent diffusion models achieve near-diffusion fidelity (FID 10.2 vs. 8.5 on CIFAR-10; 7.8 vs. 6.2 on MNIST) while reducing sampling time by over 70%. On tabular data, hybrids balance accuracy (84.7%) and fairness (0.93), whereas classical GANs and VAEs exhibit trade-offs between fidelity, efficiency, and fairness.To the best of our knowledge, this is the first study to benchmark GANs, VAEs, diffusion, and hybrid models across both image and tabular data using a unified, trust-aware evaluation framework. By providing reproducible, cross-domain comparisons, this work offers practical guidance for selecting and deploying generative models in trust-sensitive applications such as healthcare, finance, and autonomy.