Federated learning in IoT networks

Abstract

The development of the Internet of Things (IoT) creates specific requirements for machine learning related to the distributed nature of data, device resource constraints, and information privacy. Federated Learning (FL) is a distributed machine learning paradigm that enables model training on edge devices without centralizing raw data. The article provides a systematic review of federated learning methodologies for IoT networks. Adaptive optimization algorithms (EAFO, FedEAFO), communication compression techniques, privacy preservation methods (FedHDPrivacy), and architectural solutions for heterogeneous IoT environments are analyzed. Specific IoT-FL challenges are considered: statistical heterogeneity (non-IID data), system heterogeneity of devices, energy constraints, and network connectivity dynamics. A multidimensional classification of FL methods by architecture, optimization, and privacy protection criteria is presented. Experimental results demonstrate achieving communication compression up to 100x with minimal accuracy loss. Open problems of scalability to millions of devices and adaptation to dynamic topologies are identified. Promising directions are defined: continual learning, integration with 5G/6G networks, and neuromorphic computing for ultra-low-power IoT devices.