ANALISIS SELEKSI FITUR PADA NETWORK INTRUSION DETECTION DENGAN NATURAL LANGUAGE PROCESSING

Abstract

This study proposes a Natural Language Processing (NLP)-based feature selection approach for Network Intrusion Detection Systems (NIDS) using the CICIDS2017 dataset. To address high computational complexity caused by high dimensionality, network flows are normalized using Min-Max Scaling and transformed into textual representations via a consistent sentence template. The text data is then weighted using Term Frequency–Inverse Document Frequency (TF-IDF), followed by Chi-Square (Chi²) feature selection to isolate the most discriminative features. This process yields a dataset comprising 691,406 text documents with 302 textual features. Performance is evaluated using Random Forest, Linear Support Vector Machine (Linear SVM), and Naïve Bayes classifiers. Experimental results show that Random Forest achieves the highest accuracy at 97.3%, followed by Linear SVM at 96.4%, and Naïve Bayes at 83.2%. The confusion matrix further confirms that the vast majority of network traffic instances are correctly classified. These findings demonstrate that transforming tabular network data into textual formats using NLP, combined with TF-IDF and Chi-Square feature selection, successfully produces representative features, enhances intrusion detection performance, and effectively reduces data complexity in NIDS.