Federated learning for generating synthetic data: a scoping review.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Federated Learning (FL) trains models across decentralized clients, enhancing privacy by keeping data local.
• Synthetic data generation creates artificial datasets that mimic original data properties, minimizing disclosure risks.
• Federated synthesis combines FL and synthetic data generation to create global datasets without raw data sharing.

Purpose of the Study:

• To review current research and practices of using FL for synthetic data generation.
• To identify methods, evaluation practices, and research gaps in federated synthesis.

Main Methods:

• A scoping review systematically mapped and described the literature on FL for synthetic data generation.
• 69 articles published between 2018 and 2023 were included.
• Extracted information covered data types, model architectures, and evaluation of utility and privacy.

Main Results:

• 30% of studies focused on synthetic data generation, with 6 producing tabular data.
• 59% of studies focused on data augmentation.
• All 21 studies performing federated synthesis utilized deep learning, primarily Generative Adversarial Networks.

Conclusions:

• Federated synthesis is an emerging method for creating global synthetic datasets while preserving client data privacy.
• The field requires further exploration of privacy risks associated with different methods.
• Standardized methods for measuring privacy risks in federated synthesis are needed.