Enhanced Intrusion Detection System Using a Two‐Staged Feature Selection Method

ABSTRACT Intrusion detection (ID) systems are essential tools for safeguarding networks against cyber‐attacks. With the increasing sophistication and frequency of these attacks, developing ID systems that are both accurate and efficient is crucial. However, high‐dimensional datasets can hinder their efficiency and increase computational costs. This paper proposes a novel two‐stage feature selection method (GIGA) to optimize and enhance ID systems by reducing dimensionality while also improving detection accuracy. The first stage employs Gini impurity (GI) to filter out features with less importance, followed by a Genetic Algorithm (GA) with a decision‐tree‐based fitness function to identify the most relevant subset of features. Experiments on the CIC‐IDS2017, CSE‐CIC‐IDS2018, and CIC‐DDoS2019 datasets demonstrate notable improvements: test accuracy increases from 99.31% to 99.52%, 96.01% to 97.19%, and 97.95% to 99.98%, respectively, while the False Positive Rate (FPR) decreases from 0.71% to 0.53%, 3.88% to 1.04%, and 0.03% to 0.01%. The number of features is significantly reduced from 71 to 8, 70 to 4, and 69 to 8 for the datasets, respectively. The proposed method improves detection accuracy across machine learning models like Random Forest and Decision Tree while also reducing false positives and negatives. By addressing the key challenges in dimensionality and performance, GIGA offers a scalable and robust solution for enhancing ID systems.