A Cross-Client Coordinator in Federated Learning Framework for Conquering Heterogeneity.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Federated learning (FL) protects data privacy by training models locally.
• Data heterogeneity across clients causes divergent learning targets, a key FL challenge.
• Existing FL methods struggle with non-IID (independently and identically distributed) data.

Purpose of the Study:

• To propose a federated learning framework, FedUCS, addressing divergent learning targets in heterogeneous data settings.
• To enable uniform coding rules across clients for consistent model training.
• To enhance privacy-preserving machine learning in real-world, diverse data environments.

Main Methods:

• Developed FedUCS, a federated learning framework utilizing a unified coding space.
• Introduced a cross-client coordinator for supervising uniform client coding.
• Implemented a partial memory mechanism for knowledge retention.
• Applied supervised contrastive learning to improve encoding space distinguishability.

Main Results:

• FedUCS effectively mitigates target divergence caused by heterogeneous data.
• The unified coding space ensures consistent learning across clients.
• Experimental results validate the framework's superior performance in non-IID settings.
• The partial memory and contrastive learning components enhance model robustness and accuracy.

Conclusions:

• FedUCS provides a robust solution for federated learning with heterogeneous data.
• The proposed cross-client uniform coding space is a significant advancement for FL.
• This framework enhances the practical applicability of privacy-preserving machine learning.