Enhancing IoT Security Through Multilayer Unsupervised Learning and Hybrid Models
Atharva Vetal, J Lekha, Sweety Nain · 2024
This research addresses the challenge of limited unsupervised learning in current IoT security research, which heavily relies on labelled datasets, hindering the detection of unknown threats. To overcome this constraint, the study proposes a sophisticated methodology integrating K-means clustering, autoencoders, and a hybrid model (combining both). The aim is to enhance detection capabilities without being reliant on prior labelled data. Emphasizing the need to go beyond traditional models, the research underscores the significance of incorporating a diverse range of smart home IoT devices to gain comprehensive insights. Tests conducted on the N-BaIoT dataset, which incorporates authentic traffic data from nine commercial IoT devices afflicted with Mirai and BASH-LITE infections, demonstrate the effectiveness of the suggested models. K-means clustering demonstrates excellence in precision, recall, and F1-scores, particularly in Doorbell and Thermostat categories. The Hybrid model consistently achieves high precision and recall metrics across various device categories by leveraging the strengths of both Kmeans and autoencoder techniques. Notably, the Autoencoder model stands out for its exceptional ability to achieve a perfect 100% detection rate for anomalies across all devices. This study highlights the robust performance of the proposed unsupervised learning models, emphasizing their strengths and potential areas for refinement in enhancing IoT network security.