Attention-based intrusion detection using transformer networks on UNSW-NB15

The increasing complexity and volume of network traffic in modern digital infrastructures have rendered traditional rule-based intrusion detection systems insufficient for identifying sophisticated and evolving cyber threats. This study proposes an attention-based intrusion detection framework leveraging Transformer networks to model complex dependencies within network traffic data. Using the UNSW-NB15 dataset, a widely recognized benchmark for network intrusion detection, the proposed model captures both local and global feature interactions through self-attention mechanisms, enabling effective discrimination between normal and malicious activities. The experimental setup focuses on a supervised learning paradigm, where network traffic is preprocessed, encoded, and evaluated under a clearly defined classification task. The Transformer model achieved an accuracy of 96.8%, with strong performance across precision, recall, F1-score, and AUC-ROC metrics, demonstrating its robustness in handling high-dimensional and heterogeneous cybersecurity data. To ensure methodological rigor and reproducibility, the study employs a transparent data preprocessing pipeline, strict train-test separation, and detailed reporting of model configuration and evaluation procedures. Unlike prior works that rely heavily on convolutional or recurrent architectures, this research highlights the effectiveness of attention-based models for scalable and context-aware intrusion detection. While the current work focuses on detection, it outlines a conceptual pathway for integrating decision-making mechanisms to support adaptive and automated response in future systems. The findings contribute to the advancement of intelligent cybersecurity by demonstrating that Transformer-based architectures can provide reliable and generalizable performance for real-world intrusion detection tasks.