Robust Attack Detection Framework Using Pretrained CNN Model for the Edge Industrial IoT Networks

Abstract

The rapid expansion of edge industrial Internet of things (Edge-IIoT) has transformed industrial operations while introducing critical security challenges at the network edge. The growing sophistication of cyber attacks targeting Edge-IIoT networks, particularly in resource-constrained industrial environments, necessitates advanced detection mechanisms capable of identifying and classifying diverse attack patterns at the edge. This article presents a comprehensive edge-centric attack detection framework leveraging pretrained deep learning models for securing Edge-IIoT networks. Our methodology uses five state-of-the-art pretrained models, GoogleNet, AlexNet, EfficientNetB0, ResNet50, and MobileNet, evaluated on the Edge-IIoTset dataset comprising 2,219,201 network flow samples across 15 distinct attack classes. The framework efficiently processes many input features extracted from edge network traffic, including basic network characteristics, protocol headers, and industrial application-level attributes specific to Edge-IIoT environments. The experimental results demonstrate that GoogleNet achieves the highest accuracy of 97% and lowest performance degradation compared to other pretrained models with AlexNet at 96.85%, EfficientNetB0 at 96.81%, ResNet50 at 96.7%, and MobileNet at 96.42% in edge environments. Furthermore, our proposed approach significantly outperforms existing Edge-IIoT security studies using the same dataset by up to 4.2%.