Digital Twin-Driven Federated Learning for Converged Computing and Networking at the Edge

Data-intensive smart applications are currently driving the emergence of edge-enabled computing power network (Edge-CPN) by orchestrating the computing powers (CPs) of edge servers, enabling the converged computing and networking at the edge. Besides, the proliferation of these applications and various smart devices (SDs) is arousing great interest in the joint training of a shared global model by massive SDs via federated learning (FL). Due to the heterogeneity and constrained resources of SDs, the FL performance in the Edge-CPN suffers from the straggler effect, reducing the efficiency of global model aggregation. To tackle this challenge and upgrade the training mode into higher degrees of efficiency and intelligence, in this article, we propose the TwinFed, a novel digital twin (DT)-driven FL framework, which fully leverages ubiquitous CPs to configure the DTs for assisting model training of stragglers. An interplay between the end and edge layers is captured into the architecture design via the hierarchical model aggregation. We develop a unified twinning pipeline to achieve the high-fidelity DTs and efficient model training. A two-stage workflow is also introduced to implement TwinFed by flexibly integrating the computing resource orchestration and training process. Finally, we conduct a case study for anomaly detection in smart factory to validate the superiority of TwinFed in testing accuracy and training loss.