Network Intrusion Detection Based on Self-Attention Mechanism and BIGRU

Maintaining network security is crucial, and network intrusion detection systems perform an essential utility in detecting malicious network attacks. Machine learning and deep learning techniques are broadly employed in network intrusion detection to enhance the precision of NIDS. However, the benchmark dataset applied to intrusion detection often has a significantly more significant quantity of standard traffic samples than attack traffic samples, which can negatively impact the model's accuracy. To tackle this challenge, this paper presents a proposed solution that combines the Adaptive Synthetic Sampling Method and the Tomek-Links algorithm to resample the training data and balance the dataset. Additionally, the paper suggests using a Self-Attention Mechanism and Bidirectional Gated Recurrent Unit network for intrusion detection, with different models utilized for experimental comparison. The proposed model's experimental results on the NSL-KDD benchmark dataset demonstrate a significant improvement in accuracy. The model achieved 91.4% accuracy for the two-classification task and 82.8% for the five-classification task, higher than the comparison experimental model. These results confirm that the method effectively improves network intrusion detection accuracy.