Lightweight Federated Learning Approach for Resource-Constrained Internet of Things

Federated learning is increasingly recognized as a viable solution for deploying distributed intelligence across resource-constrained platforms, including smartphones, wireless sensor networks, and smart home devices within the broader Internet of Things ecosystem. However, traditional federated learning approaches face serious challenges in resource-constrained settings due to high processing demands, substantial memory requirements, and high communication overhead, rendering them impractical for battery-powered IoT environments. These factors increase battery consumption and, consequently, decrease the operational longevity of the network. This study proposes a streamlined, single-shot federated learning approach that minimizes communication overhead, enhances energy efficiency, and thereby extends network lifetime. The proposed approach leverages the <i>k</i>-nearest neighbors (<i>k</i>-NN) algorithm for edge-level pattern recognition and utilizes majority voting at the server/base station to reach global pattern recognition consensus, thereby eliminating the need for data transmissions across multiple communication rounds to achieve classification accuracy. The results indicate that the proposed approach maintains competitive classification accuracy performance while significantly reducing the required number of communication rounds.