Karin Huangsuwan. Enhancing the performance of federated learning with diffusion models: leveraging synthetic data to address non-IID data challenges. Master's Degree(Computer Engineering). Chulalongkorn University. Office of Academic Resources. : Chulalongkorn University, 2569.
Enhancing the performance of federated learning with diffusion models: leveraging synthetic data to address non-IID data challenges
Abstract:
In the context of machine learning in healthcare, federated learning (FL) is frequently seen as an effective approach to tackling issues of data privacy and distribution. Nonetheless, many real-world datasets exhibit non-identical and independently distributed (non-IID) characteristics, meaning that data features vary across different institutions. This non-IID nature presents challenges for FL model convergence, such as client drifting, where model weights lean towards local optima rather than global optimum. To address these issues, we introduce a new framework called "FedDrip (Federated Learning with Diffusion Reinforcement at Pseudo-site)," which leverages diffusion-generated synthetic data to mitigate data-related problems in non-IID settings. Our pseudo-site concept is designed to use synthetic data without sharing actual data among participation institutions, thus preventing data leaks. This strategy can be integrated into any federated framework, including FedAvg and FedDyn. Our experiments on the NIH ChestX-ray14 dataset showed that FedDrip improves performance by 1.66%, 2.33%, and 0.64% for FedAvg, FedDyn, and FedProx, respectively, based on the AUC metric. Additionally, we conducted empirical studies to explore the impact of prompting style, prompt accuracy, and data scarcity on the inference of diffusion models, using the Fréchet inception distance (FID) metric for generative models.