Year: 2025
Type: preprint
Abstract: Synthetic data arise to solve problems of data scarcity, privacy regulation, and bias in machine learning. However, trade-offs between fidelity, privacy, computational costs, and scalability of generative models have not yet been explored. The work presents an extensive benchmarking and critical evaluation of major generative frameworks for synthetic data generation-GANs, VAEs, diffusion models, and recent hybrid architectures. Through standardized metrics across image and tabular datasets, we show that such hybrid architectures, namely Diffusion-GANs, can drastically improve fidelity (32% reduction in Fréchet Inception Distance) while reducing computation cost by 40% compared to regular GANs. Our technical evaluation is followed by a systematic study of the real-world applications in healthcare, finance, and autonomous systems, covering the ethical, environmental, and governance issues regarding synthetic data. Our findings provide a balanced roadmap that intertwines technical advancement with responsible deployment strategies, making synthetic data generation a foundation for scalable and privacypreserving AI innovations.
Author: Anil Kumar Shukla
Institution: Infosys (India)
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4411254014