Multi-Granularity Federated Learning by Graph-Partitioning

This work introduces GP-MGFL, a federated learning framework that leverages balanced graph partitioning to group clients based on capability similarity and applies cross-granularity guidance to improve local and inter-group model learning. Additionally, a blockchain-based credit scoring system is integrated to handle trust management and leader selection during the aggregation process.

GP-MGFL is designed to address the challenges of heterogeneity, trust, and communication efficiency in federated learning, especially in edge-cloud environments with diverse devices and unreliable participants.

The authors are affiliated with the College of Intelligence and Computing, Tianjin University, and Nanyang Technological University, Singapore.

Recommended citation: Ziming Dai, Yunfeng Zhao, Chao Qiu, Xiaofei Wang, Haipeng Yao, and Dusit Niyato. "Multi-Granularity Federated Learning by Graph-Partitioning." IEEE Transactions on Cloud Computing (2024).
Download Paper