Split federated learning (SFL) is emerging as a practical way to train large-scale deep models on edge devices with limited resources. By partitioning the model between clients and a server, SFL reduces the computational burden on individual devices. However, system heterogeneity can slow training, prompting interest in asynchronous variants that do not wait for all clients.
The new arXiv paper introduces CoASFL, a coupled consistency optimization approach for asynchronous SFL. The authors note that asynchronous SFL alleviates the straggler bottleneck, but this benefit can come at the cost of model consistency across participants. CoASFL is designed to address that weakness.
The abstract does not yet provide experimental results or details on the optimization mechanism. It frames CoASFL as a contribution to improving reliability in asynchronous SFL, while acknowledging the existing trade-off that motivates the work.