Communication-Efficient Hierarchical Federated Learning for IoT Heterogeneous Systems with Imbalanced Data

Federated learning (FL) is a distributed learning methodology that allows\nmultiple nodes to cooperatively train a deep learning model, without the need\nto share their local data. It is a promising solution for telemonitoring\nsystems that demand intensive data collection, for detection, classification,\nand prediction of future events, from different locations while maintaining a\nstrict privacy constraint. Due to privacy concerns and critical communication\nbottlenecks, it can become impractical to send the FL updated models to a\ncentralized server. Thus, this paper studies the potential of hierarchical FL\nin IoT heterogeneous systems and propose an optimized solution for user\nassignment and resource allocation on multiple edge nodes. In particular, this\nwork focuses on a generic class of machine learning models that are trained\nusing gradient-descent-based schemes while considering the practical\nconstraints of non-uniformly distributed data across different users. We\nevaluate the proposed system using two real-world datasets, and we show that it\noutperforms state-of-the-art FL solutions. In particular, our numerical results\nhighlight the effectiveness of our approach and its ability to provide 4-6%\nincrease in the classification accuracy, with respect to hierarchical FL\nschemes that consider distance-based user assignment. Furthermore, the proposed\napproach could significantly accelerate FL training and reduce communication\noverhead by providing 75-85% reduction in the communication rounds between edge\nnodes and the centralized server, for the same model accuracy.\n