LiteRT-IDSNet: A Lightweight Hybrid Deep Learning Framework for Real-Time Intrusion Detection in Industrial IoT Using the RT-IoT 2022 Dataset
Roseline Oluwaseun Ogundokun, Pius Adewale Owolawi, Etienne Van Wyk · 2025
The rapid expansion of Industrial Internet of Things (IIoT) infrastructures has exposed mission-critical systems to sophisticated cyber threats, thereby demanding highly accurate, low-latency, and computationally efficient intrusion detection systems (IDS). In response, this paper introduces LiteRT-IDSNet (Lightweight Real-Time Intrusion Detection Network), a novel hybrid deep learning framework tailored for real-time IIoT security. The proposed architecture integrates compact feature encoding with parallel ReLU- and Sigmoid-activated branches to enable strong multi-scale feature learning with low inference cost and simple architecture. Its effectiveness was verified by training and evaluating the LiteRT-IDSNet on the RT-IoT 2022 dataset with 123,117 instances distributed across 12 attack classes. Its performance was compared with that of a baseline fully connected neural network (BaseIDS) under the same training specifications for 50 epochs. Experimental results show that LiteRT-IDSNet achieved training accuracy of 99.61% and validation accuracy of 99.60%, outperforming the baseline model with 99.47% training and 99.59% validation accuracy. Moreover, LiteRT-IDSNet generalized better with lesser validation loss and improved detection rates for minority attack classes, which was reflected by confusion matrix heatmaps and classification reports. These results validate LiteRT-IDSNet as a suitable real-time IIoT cybersecurity system offering an optimal trade-off between detection efficiency and computational expense. The model's lightweight design and high classification fidelity make it well-suited for deployment in edge-based IIoT systems where resources are limited and security demands are high.