FedEGG: Federated Learning with Explicit Global Guidance
Kun Zhai , Yifeng Gao , Yunhao Feng , Wei Gao , Xingjun Ma , Yu-Gang Jiang
Federated learning (FL) is a promising paradigm for privacy-preserving collaborative learning, but its convergence is often hindered by non-IID data distributions, limiting its effectiveness in real-world deployments. Existing approaches alleviate this challenge through client-side optimization constraints, adaptive client selection, or the use of pre-trained models and synthetic data. We reinterpret these methods from a unified perspective and argue that they all introduce an implicit guiding task to regularize and steer local training. Motivated by this insight, we propose to explicitly incorporate an explicit global guiding task into FL to improve convergence and performance. To this end, we present FedEGG, a new FL algorithm that constructs a well-defined and easy-to-optimize global guiding task from a public dataset with the assistance of large language models (LLMs). By coupling the original FL objective with this guiding task, FedEGG effectively combines the strengths of federated and centralized learning. We further provide a comparative analysis of FedEGG and FedAvg under the stated assumptions, characterizing how cross-task heterogeneity and guiding strength affect the correction introduced by the guiding task and deriving a sufficient condition for beneficial guidance. Extensive experiments show that FedEGG consistently outperforms state-of-the-art FL methods under both IID and non-IID settings, and further improves their performance when used as a plug-in guidance mechanism.
Federated learning / explicit global guidance / convergence
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(9530KB)
