Novel deep learning-based IoT network attack detection using magnet loss optimization

Abstract

The increasing prevalence of Internet of Things (IoT) devices across various industries has raised critical security concerns due to their inherent vulnerabilities and high interconnectivity. While traditional security mechanisms have shown limitations in effectively securing large IoT networks, machine learning (ML) and deep learning (DL) methods have been explored to tackle the attack detection problem in this domain. However, existing approaches still lack optimal regularization and have limited comprehensiveness in validation across different IoT-centric datasets. To address these challenges, this research proposes the extension of the Deep Magnet Autoencoder (DMAE) and introduces a novel approach, the Cascade Deep Magnet Autoencoder (CDMAE), leveraging the Magnet Loss optimization as regularization for better class distinction through local separation in latent space. This enhanced class clustering strengthens attack detection by maximizing inter-class separation while compactly grouping data points of the same class, leading to more precise identification of benign and malicious traffic. Extensive experiments conducted on three contemporary IoT datasets, CIC-BoT–IoT, CIC-ToN–IoT, and CICIoT2023, demonstrate that our proposed models are able to produce meaningful latent representations with powerful discrimination between benign and malicious IoT network data. Empirical insights for fine-tuning the model are also provided through supplementary experiments. Comprehensive results show that the proposed methods significantly boost classification across different IoT datasets with high metric scores, outperforming other approaches.