RingSFL: An Adaptive Federated Learning System for Heterogeneous Clients

Jinglong Shen, Nan Cheng, Zhisheng Yin, Yuchuan Fu, Wenchao Xu, Song Guo

Abstract: Federated learning (FL) has the ability to protect user privacy during cooperative training. However, vanilla FL lacks adaptation to heterogeneous clients and suffers from privacy leakage. This proposal demonstrates the prototype system of RingSFL, which integrates FL with a model split mechanism for client heterogeneity and data privacy preservation. RingSFL forms a ring topology among clients. Instead of local training, the model is split and trained across clients in a pre-defined direction. By properly setting the propagation lengths, the training efficiency is significantly enhanced. Additionally, since the local models are blended, the privacy risks is mitigated.