VQ-FedDiff: Federated Learning Algorithm of Diffusion Models With Client-Specific Vector-Quantized Conditioning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Generative models like DDPMs create realistic images but require sensitive data.
• Federated learning (FL) trains models on decentralized data while preserving privacy.
• Existing FL methods for generative models often focus on GANs, not DDPMs.

Purpose of the Study:

• To propose a new algorithm for training denoising diffusion probabilistic models (DDPMs) within federated learning (FL) frameworks.
• To enable the generation of high-quality synthetic images from sensitive, decentralized datasets while ensuring data privacy.
• To develop a personalized approach for training diffusion models that maintains data security.

Main Methods:

• Introduced VQ-FedDiff, a novel algorithm specifically designed for training DDPMs under FL settings.
• Focused on personalized model training to enhance image quality (FID) and maintain data privacy.
• Evaluated performance on both independent and identically distributed (IID) and non-IID data settings.

Main Results:

• VQ-FedDiff demonstrated state-of-the-art performance in federated learning of diffusion models.
• Achieved high-quality image generation, competitive with locally trained models, across photorealistic and medical datasets.
• Effectively preserved data privacy in decentralized learning scenarios.

Conclusions:

• Diffusion models can be efficiently trained on decentralized, sensitive data using the proposed VQ-FedDiff algorithm.
• VQ-FedDiff offers a privacy-preserving solution for generating high-quality synthetic images in FL settings.
• The method shows significant potential for applications in sensitive domains like healthcare and finance.