An Efficient Intrusion Detection Model with Traffic Image Representation and Enhanced EfficientNet Architecture

Abstract

Network intrusion detection (NID) is a fundamental technology for ensuring network security, but deploying lightweight yet highly accurate intrusion detection models (IDS) on resource-constrained devices remains challenging. To address this need, this paper proposes a lightweight intrusion detection model, IDSAtt-EffNet, contributing to existing cybersecurity frameworks. By integrating a Multi-Head Attention (MHA) mechanism, the model significantly enhances detection accuracy and feature representation capabilities while maintaining a lightweight architecture. First, network traffic data collected from real-world environments is converted into image data to facilitate analysis by convolutional neural networks (CNNs). Second, the Multi-Head Attention (MHA) mechanism is integrated into the EfficientNet-B0 model to construct the IDSAtt-EffNet architecture, enhancing its feature extraction capabilities. Final experimental results demonstrate that IDSAtt-EffNet achieves accuracies of 97.94% and 96.86% on the NSL-KDD and QAX2024 datasets, respectively, validating the effectiveness of this approach. These findings hold significant implications for the ongoing development of secure and efficient network intrusion detection solutions in resource-constrained network environments.