Synthetic Data Generators: The Deep Dive into Synthetic Realities

Speaker: Abdul Majeed, PhD., University of Warwick, England Synthetic data is a cornerstone for AI development as it resolves data scarcity and privacy challenges. Synthetic data is artificially generated information that reproduces important patterns and relationships of real data without directly copying its records. Fundamental topics include generation methods-such as statistical sampling, simulation/agent-based methods, data augmentation, and rule-based approaches-along with data types, preprocessing, train-test separation, and evaluation of realism, utility, diversity, and privacy. Advanced methods include generative adversarial networks (GANs), variational autoencoders (VAEs), diffusion models, large language models, and domain-specific simulators for producing tabular, image, text, time-series, and graph data. Important advanced concepts include conditional generation, rare-event synthesis, fairness, distribution shift, privacy attacks, differential privacy, federated generation, and model-based validation. This lecture introduces synthetic data approaches (rule-based, statistical-based, hand-engineered, etc.), generator types (GANs, GAN variants, VAEs), adversarial learning frameworks, and their powerful ability to generate realistic data. Applications in vision and synthesis are explored across disciplines (e.g., medical, transport, ambient assisted living, LLMs, etc.). Advanced data generators such as vision models, federated models, and LLMs will also be discussed in relation to conventional approaches. Lastly, the challenges related to bias, fairness, privacy, and attack vectors will also be discussed in detail. Some discussion around metrics will also be provided to highlight synthetic use across domains and scenarios. Some discussion around both attacks (e.g., launching a membership attack) and defense vectors (defense against attribute/membership leakage) of the synthetic data will also be provided. Lastly, some uses of SD in agentic AI and LLMs will also be discussed to provide information on the latest trends in this line of work. Finally, challenges will be presented through detailed synthesis of state-of-the-art literature to demonstrate where this technology is heading, which will allow participating researchers to pursue the latest topics/challenges.