MG²FL: Multi-Granularity Grouping-Based Federated Learning in Green Edge Computing Systems

In this paper, we propose MG²FL, a novel multi-granularity grouping-based federated learning framework designed for green edge computing systems. MG²FL addresses key challenges in traditional FL—such as high communication energy consumption, model heterogeneity, and vulnerability to malicious participants—by introducing three core innovations.

First, a balanced graph partitioning mechanism is employed to group edge devices with low communication overhead and strong inter-device guidance. Second, a multi-granularity guidance model allows large-scale, fine-granularity models to effectively guide smaller models, boosting overall learning performance. Third, a credit-based model aggregation strategy ensures that only trustworthy edge devices contribute significantly to the global model, enhancing robustness against adversarial behavior.

This work is a collaboration between Tianjin University, Carleton University, and the Guangdong Laboratory of Artificial Intelligence and Digital Economy (Shenzhen).

Recommended citation: Ziming Dai, Yunfeng Zhao, Chao Qiu, Xiaofei Wang, and F. Richard Yu. "MG 2 FL: Multi-Granularity Grouping-Based Federated Learning in Green Edge Computing Systems." In GLOBECOM 2023-2023 IEEE Global Communications Conference, pp. 152-157. IEEE, 2023.
Download Paper | Download Slides