Transformer-Based Intrusion Detection on the NSL-KDD Benchmark: A Performance Assessment Against Conventional Classifiers
Network intrusion detection plays a crucial role in cybersecurity, aiding in the detection of malicious traffic and the ability to distinguish between attacks and normal network traffic. This paper compares a Transformer-based model for network intrusion detection on the NSL-KDD benchmark dataset with a binary classification approach. The architecture proposed uses a 41-input-feature projection to 64 dimensions, followed by two Transformer encoder blocks with 4 attention heads processing the 64-dimensional features. The confusion matrix for the resulting model on the evaluation set of 30,000 records is 98.72% accuracy, 98.56% precision, 98.78% recall, and 98.67% F1-score. The false-positive and false-negative rate are 1.34% and 1.22%, respectively. When compared with conventional classifiers, the reported accuracy of the Transformer is lower than the accuracy of XGBoost (99.44%) and Random Forest (99.31%) while being superior to the reported accuracy of the multilayer perceptron (98.56%) and close to the reported accuracy of Gradient Boosting (98.80%). These findings demonstrate that the evaluated Transformer model can achieve a high classification performance on the selected NSL-KDD records, although it does not outperform all conventional classifiers. The study provides a benchmark-based assessment of the model and highlights the importance of consistent evaluation procedures when comparing intrusion detection approaches. Further research may investigate feature-tokenized Transformer architectures and evaluate their performance on datasets representing industrial control system networks.
Authors
- Awatif Alqahtani (ORCID: https://orcid.org/0000-0002-9794-6390)
- Hanan Alwadi
Institutions
- King Saud University (SA)
Publication Details
- Journal
- Electronics
- Published
- 2026-11-16
- DOI
- https://doi.org/10.3390/electronics15225316
- Primary Topic
- Network Security and Intrusion Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- King Saud University