FedDE: A Data Augmentation Strategy to Combat the Non-IID Issue in Federated Learning

Federated Learning (FL) is a distributed machine learning framework that protects data privacy by allowing multiple end devices to train machine learning models collaboratively without sharing raw data. However, data on various end devices are typically Non-Independent and Identically Distributed (Non-IID), leading to difficulties in model convergence, client model drift, and reduced generalization ability. Moreover, highly heterogeneous Non-IID data require more communication rounds, resulting in high communication costs and significantly reduced model accuracy. We propose a FL framework(FedDE) based on Denoising Diffusion Probabilistic Models (DDPM) and Knowl-edge Distillation from the data and model fusion perspective to tackle these challenges. In the FedDE framework, the DDPM is used as a class balancer to alleviate the problem of uneven label distribution. Additionally, the server uses labeled data for ensemble distillation, flexibly aggregating local clients and reducing the number of communication rounds. We conducted extensive experiments on multiple public benchmark datasets such as MNIST, FashionMNIST, and CIFAR10. The test results demonstrate that our proposed method outperforms state-of-the-art methods in accuracy and significantly reduces the number of communication rounds.