AMFL: Asynchronous Multi-level Federated Learning with Client Selection

Synchronous Federated Learning (FL) may suffer from increased training time and costs. To address this issue, Asynchronous Federated Learning (AFL) has been proposed. Furthermore, traditional single-level FL with just one cloud server and multiple clients may incur long communication delays, due to the absence of intermediate nodes. As a solution, Hierarchical FL (HFL) and Multi-level FL have been proposed to overcome this limitation. In this work, we integrate Asynchronous FL and Multi-level FL, by employing a creative Client Selection method to avoid the accumulation of outdated updates during multi-level aggregation, called Asynchronous Multi-level Federated Learning with Client Selection (AMFL) method. We evaluate AMFL's performance with that of Asynchronous FL, Hierarchical FL, Multi-level FL, and other stateof-the-art baselines. The results indicate that AMFL converges faster than these methods.