Research Article
Intrusion Detection Based on the Combination of TMG-GAN and CNN-BILSTM-Attention
@INPROCEEDINGS{10.1007/978-3-031-96944-7_18, author={Li Wang and Ru Li and Haolong Li and Liming He}, title={Intrusion Detection Based on the Combination of TMG-GAN and CNN-BILSTM-Attention}, proceedings={Computer and Information Security. First World Conference, WCCIS 2024, Kuala Lumpur, Malaysia, September 20--22, 2024, Proceedings}, proceedings_a={WCCIS}, year={2026}, month={11}, keywords={Intrusion Detection Deep Learning Attention Mechanism Multi-Generator Generative Adversarial Network}, doi={10.1007/978-3-031-96944-7_18} }- Li Wang
Ru Li
Haolong Li
Liming He
Year: 2026
Intrusion Detection Based on the Combination of TMG-GAN and CNN-BILSTM-Attention
WCCIS
Springer
DOI: 10.1007/978-3-031-96944-7_18
Abstract
In recent years, network intrusion detection technology has achieved significant breakthroughs with the application of deep learning, playing an indispensable role in protecting network security. However, two major limitations remain: (1) an imbalance between benign and attack data in network traffic, where benign data dominates and attack data is minimal, leading to a substantial impact on model performance; and (2) the difficulty in balancing temporal and spatial features when addressing dynamic and evolving network threats, with limited attention to the spatiotemporal characteristics of network traffic. To tackle these issues, I have proposed an intrusion detection method called TGCBA which is composed of TMG-GAN, CNN, BILSTM, and ATTENTION. First, the TMG-GAN (Tabular Multi-Generator Generative Adversarial Network) method is used to generate more realistic minority class samples to augment and balance the dataset. Then, CNN (Convolutional Neural Network) is applied to extract spatial characteristics from the processed dataset, followed by BILSTM (Bidirectional Long Short-Term Memory) to capture temporal features. Finally, the Attention mechanism focuses on key temporal-spatial features, improving the model’s detection accuracy. Comparative experiments on representative datasets, CICDDoS 2019 and UNSW-NB 15, show that the proposed method outperforms existing models in terms of accuracy, recall, and false positive rate.