Personalized Federated Learning in One-Shot: A Method for Heterogeneous Data Scenarios

Federated learning is a distributed machine-learning technique that allows multiple clients to collaboratively train a model without sharing their local raw data. In existing federated learning solutions, one-shot federated learning is a promising yet challenging direction. It involves model training with just one round of communication between clients and the server, reducing communication overheads and security risks. However, most existing methods still face several challenges. First, generating additional data for training prolongs the model training time. Second, the traditional single model cannot handle data heterogeneity in real world scenarios. To address these issues, this article proposes a personalized federated learning method called FedOM under the one-communication premise. Considering the high data heterogeneity, FedOM abandons the traditional single global model architecture and generates multiple group models to overcome the generalization limitations of a single model. Additionally, based on FedOM, this study proposes FedOMF, a personalized method with a fine-tuning module. Experiments on public datasets show that in highly heterogeneous data scenarios, the proposed methods outperform baseline methods, demonstrating superior performance and great practical potential.