A Round-Based Network Attack Detection Model Using Auto-encoder In IoT-Edge Computing
Hooman Hamidpour, Omid Bushehrian · 2023
As the applications of IoT continue to expand, the need for practical security measures becomes increasingly vital. In previous studies, supervised machine learning models have been proposed to classify attack and benign network traffic. However, they relied on a complete dataset consisting of enough labeled samples. In this study we have proposed a round-based network intrusion detection system using deep auto-encoder neural networks that is trained with a few benign traffic samples initially and subsequently retrained in the next rounds with the self-classified traffic samples, allowing the model to evolve. This solution could address the challenge of insufficient datasets generated in an IoT-Edge network. We also implemented the Federated Learning (FL) method to establish a global detection model at the cloud layer across the edge devices. Our experiments on the N-BaIoT dataset showed that a round-based approach significantly outperforms one-time learning, with an average Fl-Score improvement of 21.39% due to the presence of overlapping patterns of benign and attack samples observed in the N-BaIoT dataset.