Advanced Hierarchical Deep Learning Approach for Intrusion Detection: Hyperparameter Tuning and Performance Evaluation on the CICIDS Dataset

Abstract

Intrusion Detection Systems (IDS) are essential in the communication security and integrity of contemporary network architectures. This study introduces a novel advanced hierarchical deep learning model for network intrusion detection using the CICIDS dataset-2019. To improve the detection accuracy, the proposed model architecture builds several deep learning layers embedded with hyperparameters' optimization. Additional preprocessing steps, including over-sampling and min-max scaling, as well as feature selection via Random Forest, were applied to enhance the models' ability to generalize good performance. The chosen performance indices, precision, recall, F1-score, and accuracy, clearly reflect the model stability: Accuracy = 0.98 on 40k test samples. Operations of the confusion matrix also provided a high level of support to the model precision and recall for benign and attack classes. In the same token, ROC and precision-recall curves further provided validation of the model for the differentiation between normal and anomalous behaviours. The evaluation of feature importance, based on the model, included decisions about which elements of network intrusion were most relevant and called for improvement.