MetaClusterFL: Personalized Federated Learning on Non-IID data with Meta-learning and Clustering
Hui Zeng, Shiyu Xiong, Hongzhou Shi · 2024
Federated learning introduces a paradigm in machine learning, where global models are trained by aggregating parameters from clients’ locally trained models, ensuring privacy as client data does not pass through a central node. But Lots of studies have demonstrated that Non-IID data significantly impacts the accuracy of global models. To address this challenge, we propose a new personalized federated learning framework, MetaClusterFL, which combines federated meta-learning and clustering federated learning. MetaClusterFL leverages the trained meta-model to facilitate knowledge sharing during the clustering process and compute the similarity between client data distributions. It dynamically detects the optimal number of clusters for accurate client clustering without requiring additional hyperparameters. The clustered clients then train clusterlevel meta-models, which serve as the initial models for their personalized models. We conduct experiments on three benchmark FL datasets with different classification tasks to validate the performance of our approach on Non-IID data. Experimental results show that MetaClusterFL outperforms other comparative methods in the same field.