ОПТИМИЗАЦИЯ КОММУНИКАЦИОННЫХ ЗАТРАТ В ДВУХУРОВНЕВОМ ФЕДЕРАТИВНОМ ОБУЧЕНИИ ДЛЯ ОБНАРУЖЕНИЯ КИБЕРАТАК В IOT-СЕТЯХ

Abstract

This paper addresses the challenge of reducing communication overhead in federated learning for distributed IoT networks whilst maintaining high-quality cyberattack detection. The CICIoT2023 dataset, containing data on network traffic and attacks in an Internet of Things environment, was used as the basis for the study. A one-dimensional convolutional neural network (1D-CNN) was used for classification, and a comparison was made between the classical single-layer FL architecture and the proposed two-layer hierarchical scheme with intermediate aggregation gateways. Experiments were conducted on 20 simulated devices over 21 training rounds. The results showed that the two-level architecture achieves an accuracy of 88.82%, which is practically equivalent to the result of the single-level model (88.87%), whilst the difference in Macro F1 and ROC-AUC also remains negligible. The main advantage of the proposed approach lies in a significant reduction in the load on the global server: the number of global parameter transfers is reduced from 420 to 15, i.e. by a factor of 28, which corresponds to a 96.4% reduction in traffic. Thus, the proposed architecture demonstrates an effective balance between classification quality and communication efficiency and can be regarded as a promising solution for scalable cybersecurity systems in the IoT.