Keywords: Federated learning, Regression, One-shot, Data-free
TL;DR: We consider a novel problem setting: data-free one-shot federated regression, and propose a practical method for this setting.
Abstract: We consider a novel problem setting: data-free one-shot federated regression. This setting aims to prepare a global model through a single round of communication without relying on auxiliary information, e.g., proxy datasets. To address this problem, we propose a practical framework that consists of three stages: local training, data synthesizing, and knowledge distillation, and demonstrate its efficacy with an application to bone age assessment. We conduct validation under independent and identical distribution (IID) and non-IID settings while considering both model homogeneity and heterogeneity. Validation results show that our method surpasses FedAvgOneShot by a large margin and sometimes even outperforms the proxy-data-dependent approach FedOneShot.