Multi-Granularity Federated Learning by Graph-Partitioning

✨ Proposed Method

This paper proposes a Graph-Partitioning Multi-Granularity Federated Learning (GP-MGFL) method built on a consortium blockchain. To reduce overall communication overhead, the framework uses a balanced graph partitioning algorithm to group edge clients, which minimizes high-cost communications while ensuring effective intra-group guidance. Furthermore, the system introduces a cross-granularity guidance mechanism where fine-granularity models guide coarse-granularity models to fully leverage data heterogeneity and enhance accuracy. To maintain security, a dynamic credit model is implemented to adjust clients’ contributions to the global model and automatically select group leaders for model aggregation.

📊 Experimental Results

• Accuracy Enhancement: The GP-MGFL algorithm achieves an accuracy that is 5.6% higher than that of ordinary blockchain-based federated learning (BFL) algorithms.
• Superior Grouping: Compared to other grouping methods, such as greedy grouping, the proposed GP-MGFL approach improves accuracy by approximately 1.5%.
• Robust Security: In scenarios involving malicious clients, the method demonstrates strong robustness, achieving a maximum accuracy improvement of 11.1% over baseline models.

🤝 Collaborating Institutions

Tianjin University; Guangming Laboratory of Artificial Intelligence and Digital Economy (SZ); Beijing University of Posts and Telecommunications; Nanyang Technological University

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 13, no. 1 (2024): 18-33.
Download Paper