Lightweight Hybrid Feature Selection for High-Accuracy Ensemble Intrusion Detection Systems

Main Article Content

Fiona Lawrence
Rajesh kumar Nigam

Abstract

Intrusion Detection Systems (IDS) are vital for defending networks, yet model selection is complicated by dataset-dependent performance. This study benchmarks Support Vector Machines (linear, polynomial, RBF, sigmoid) against gradient-boosting ensembles—LightGBM, XGBoost, and CatBoost—on two widely used corpora, CICIDS2017 and NF-UNSW-NB15. To mitigate class imbalance and redundancy, we apply SMOTE for resampling and PCA for dimensionality reduction. Models are evaluated using accuracy, precision, sensitivity (recall), and F-measure. Across both datasets, the ensemble methods consistently deliver higher and more stable results, achieving accuracies above 98%, whereas SVM variants display pronounced variability, performing comparatively better on NFUNSW-NB15 than on CICIDS2017. These findings indicate that ensemble learners generalize more reliably under dataset shift and are therefore stronger candidates for practical IDS deployments, where robustness across heterogeneous traffic profiles is essential.

Article Details

Section
Articles