Network Intrusion Detection Method Based on CNN-BiLSTM-Attention Model

Abstract

Network intrusion detection plays a crucial role in cybersecurity, requiring robust and intelligent methods to identify and mitigate malicious activities. This study explores an advanced deep learning-based approach utilizing NSL-KDD, UNSW-NB15, and CIC DDoS2019 datasets to enhance intrusion detection accuracy. A comprehensive set of machine learning and deep learning techniques is employed, incorporating convolutional and recurrent neural network architectures to capture spatial and sequential dependencies in network traffic data. The proposed method integrates deep feature extraction and sequential pattern recognition with an attention mechanism to enhance detection performance by focusing on critical attack characteristics. Experimental evaluations demonstrate that the Voting Classifier (BaggingDT + BagRF) achieves the highest performance, attaining 98.1% accuracy on the CIC-DDoS2019 dataset, 99.9% accuracy on the NSL-KDD dataset, and 92.9% accuracy on the UNSW-NB15 dataset. These results highlight the effectiveness of ensemble learning in improving detection precision, recall, and generalization across diverse intrusion patterns. The findings indicate that integrating multiple classifiers and optimizing feature representation significantly enhances network intrusion detection, providing a scalable and efficient solution for real-world cybersecurity applications.