Enhanced Intrusion Detection in IIoT Networks: A Lightweight Approach with Autoencoder-Based Feature Learning

TL;DR

The paper tackles intrusion detection in IIoT under severe class imbalance and multiclass scenarios by proposing a cost-sensitive autoencoder that performs dimensionality reduction and discriminative feature learning for edge deployment. The method compresses 24-dimensional inputs to a bottleneck of $6$ features and uses class weights to emphasize minority attack types, optimizing a reconstruction loss $L = \frac{1}{N} \sum_{i=1}^{N} w_{y_i} (x_i - \hat{x}_i)^2$. Evaluations on Edge-IIoTset show a peak accuracy and F1 of 99.94%, with ultra-fast Jetson Nano inferences (~0.185 ms per instance), validating strong performance under imbalance and multiclass conditions while remaining suitable for resource-constrained edge devices. The work demonstrates practical impact for real-time IIoT security by combining robust detection with lightweight, edge-oriented deployment, outperforming several baselines in both accuracy and efficiency.

Abstract

The rapid expansion of the Industrial Internet of Things (IIoT) has significantly advanced digital technologies and interconnected industrial systems, creating substantial opportunities for growth. However, this growth has also heightened the risk of cyberattacks, necessitating robust security measures to protect IIoT networks. Intrusion Detection Systems (IDS) are essential for identifying and preventing abnormal network behaviors and malicious activities. Despite the potential of Machine Learning (ML)--based IDS solutions, existing models often face challenges with class imbalance and multiclass IIoT datasets, resulting in reduced detection accuracy. This research directly addresses these challenges by implementing six innovative approaches to enhance IDS performance, including leveraging an autoencoder for dimensional reduction, which improves feature learning and overall detection accuracy. Our proposed Decision Tree model achieved an exceptional F1 score and accuracy of 99.94% on the Edge-IIoTset dataset. Furthermore, we prioritized lightweight model design, ensuring deployability on resource-constrained edge devices. Notably, we are the first to deploy our model on a Jetson Nano, achieving inference times of 0.185 ms for binary classification and 0.187 ms for multiclass classification. These results highlight the novelty and robustness of our approach, offering a practical and efficient solution to the challenges posed by imbalanced and multiclass IIoT datasets, thereby enhancing the detection and prevention of network intrusions.

Figures (4)

• Figure 1: Architecture of the proposed cost-sensitive autoencoder.
• Figure 2: Autoencoder Loss Curve
• Figure 3: Proposed AutoEncoder based IDS Module
• Figure 4: Training and Validation Loss for Tabnet and BiLSTM for Binary and Multiclass Classification