AI-Driven Intrusion Detection System for Edge Computing Environments Using Lightweight Deep Learning Models

Abstract

The increasing deployment of Internet of Things (IoT) devices and edge computing services has introduced significant cybersecurity challenges due to the distributed nature and resource constraints of edge environments. Conventional intrusion detection systems (IDSs), particularly computationally intensive deep learning approaches, may introduce excessive latency and memory requirements, limiting their suitability for real-time edge deployment. This study proposes a Lightweight Attention-Based Convolutional Neural Network (LA-CNN) for efficient intrusion detection in edge computing environments. The proposed framework combines data preprocessing, feature optimization using correlation analysis, Mutual Information (MI), and Recursive Feature Elimination (RFE), depthwise separable convolution, and channel-attention-based feature learning to improve detection performance while reducing computational requirements. The proposed model is evaluated using the CICIoT2023 dataset under a multi-class intrusion detection setting and compared with conventional machine learning and deep learning approaches. Experimental results demonstrate that LA-CNN achieves 98.34% accuracy, 98.12% precision, 98.41% recall, and 98.26% F1-score, with a false alarm rate of 1.59%. Furthermore, the proposed model requires only 8.7 ms inference time and 91 MB memory, representing approximately 52.7% lower inference latency and 37.2% lower memory consumption than the conventional CNN baseline. These results demonstrate that integrating feature optimization, lightweight convolution, and attention-based feature representation can provide an effective balance between intrusion detection performance and computational efficiency. The proposed framework therefore offers a promising solution for real-time cybersecurity monitoring in resource-constrained edge computing environments.