Attention Mechanism for Attacks and Intrusion Detection

In the rapidly evolving field of cybersecurity, effectively detecting and preventing network-based attacks is critical to safeguarding vital infrastructures. Traditional Network Intrusion Detection Systems (NIDS) often struggle to address the complexity and sophistication of modern cyber threats. To overcome these limitations, this thesis introduces a novel deep learning-based NIDS framework that integrates TabNet with an Attention Mechanism to improve both detection accuracy and interpretability. Leveraging the CIC UNSW-NB15 Augmented Dataset, which includes nine diverse attack categories alongside benign traffic, the proposed system employs advanced preprocessing techniques, such as SMOTE, to address class imbalance and data variability. Experimental results indicate that the model achieves an overall accuracy of 74%, excelling in the detection of benign traffic and reconnaissance attacks but encountering challenges with rare attack types, including worms and shellcode. These findings underscore the promise of attentionbased deep learning models for enhancing NIDS performance and emphasize the need for future research to refine detection capabilities for rare and complex attacks.