HYBRID FEATURE-OPTIMIZED NETWORK INTRUSION DETECTION SYSTEM

Abstract

The rise of sophisticated cyberattacks demands intelligent threat detection beyond traditional signature-based Network Intrusion Detection Systems (NIDS), which fail against unknown attacks. This study presents a hybrid machine learning-based NIDS using the UNSW- NB15 dataset to classify ten attack types and normal traffic. A triple-hybrid feature selection approach (Chi-Square, Mutual Information, and XGBoost importance) is employed to retain highly discriminative attributes, while SMOTE–Tomek balancing addresses class imbalance. XGBoost serves as the final classifier, achieving 84.88 accuracy and 83.71 macro-F1 using stratified 5-fold cross-validation, surpassing baseline models. SHAP integration improves in- terpretability by explaining feature contributions to predictions. The model is deployed in a Streamlit interface for real-time intrusion monitoring, demonstrating its readiness for practical cybersecurity defense.