Split-FL: An Efficient Online Federated Learning Framework with Constrained Computation and Streaming Data

To satisfy the increasing demand for enabling large-scale machine learning for low-latency multimedia data processing and data privacy on mobile edge devices, federated learning (FL) has advanced as an important learning infrastructure. However, in many practical wireless communications applications with limited resources, the conventional FL techniques suffer from heavy communication and computation overheads arising from local training and periodic global aggregation, especially for streaming data. Moreover, data heterogeneity in the distributed system breaks the classical independent and identically distributed (IID) assumption, impacting accuracy and convergence speed. In this paper, we propose a novel FL framework of Split-FL to enhance the computational efficiency of processing data streams in edge networks, which allows model customization based on hardware constraints and enables the participation of heterogeneous devices without compromising learning performance. Specifically, Split-FL involves an online learning procedure that balances the parameters updated from training data streams. Additionally, an adaptive pruning procedure is proposed to reduce computation and communication costs during the training process. The simulation results demonstrate the superiority of Split-FL over existing methods, demonstrating its high efficiency in communication and computation.