ASFL: An Adaptive Model Splitting and Resource Allocation Framework for Split Federated Learning

A team of researchers from academic institutions has published a breakthrough paper on arXiv titled "ASFL: An Adaptive Model Splitting and Resource Allocation Framework for Split Federated Learning." The work addresses a critical bottleneck in federated learning (FL) where edge devices with limited computational power struggle with high latency and energy consumption during AI model training. The proposed ASFL framework introduces adaptive model splitting, allowing parts of the neural network to be trained on a central server while other parts remain on client devices, creating a hybrid approach that leverages both distributed and centralized resources.

Technically, ASFL formulates a joint optimization problem balancing learning performance (convergence rate) with system efficiency (delay and energy). The researchers developed an Online Optimization Enhanced Block Coordinate Descent (OOE-BCD) algorithm to solve this complex problem iteratively, handling long-term constraints and the coupling between model splitting decisions and resource allocation. Experimental results demonstrate dramatic improvements: ASFL reduces total delay by up to 75% and energy consumption by up to 80% compared to five baseline federated learning schemes, while also achieving faster convergence. This represents a significant advancement for deploying AI on wireless networks with heterogeneous devices.