AN ENSEMBLE LEARNING FRAMEWORK FOR NETWORK INTRUSION DETECTION AND ATTACK CLASSIFICATION

image: https://doi.org/10.65725/JCISE/2/4/011
JOURNAL OF COMPUTATIONAL INTELLIGENCE SCIENCE AND ENGINEERING (JCISE) 
ISSN: 3107-8168
Volume 2 Issue 4, Oct – Dec 2026

Abstract: 
Modern enterprise networks confront an escalating landscape of sophisticated, multi-vector cyber-attacks designed to systematically bypass conventional signature-based Intrusion Detection Systems (IDS). To address these critical security vulnerabilities, this paper proposes a high-performance, resilient Ensemble Learning Framework engineered specifically for real-time network intrusion detection and granular multi-class attack classification. The architecture implements a comprehensive data preprocessing pipeline that incorporates Z-score feature normalization, the Synthetic Minority Over-sampling Technique (SMOTE) for mitigating severe class imbalances across minority attack profiles, and Principal Component Analysis (PCA) for optimized dimensionality reduction and feature extraction. By synthesizing these preprocessing methodologies with a heterogeneous stacking ensemble comprising Random Forest (RF) for parallelized bagging, Extreme Gradient Boosting (XGBoost), and LightGBM for high-speed network telemetry processing, our framework achieves exceptional generalization across diverse and complex network topologies. Furthermore, a Level-1 meta-classifier utilizing a soft-voting aggregator aggregates prediction probabilities to render final attack decisions. Through rigorous empirical validation conducted across standard, widely recognized benchmark network datasets—namely CIC-IDS2017 and UNSW-NB15—the proposed framework demonstrates outstanding performance metrics, yielding an overall classification accuracy of 99.24%, an ultra-low false alarm rate of under 0.4%, and an aggregate F1-score of 99.12%. These empirical outcomes conclusively indicate that the proposed framework substantially outperforms baseline individual machine learning models and traditional single-algorithm architectures, making it highly viable for high-speed software-defined networking (SDN) and real-time enterprise security operations centers (SOCs).

Authors: Dr.K.Nirmala, Mrs.R.Geetha

Keywords: Intrusion Detection System (IDS); Ensemble Learning; Stacking Generalization; Random Forest; XGBoost; LightGBM; Synthetic Minority Over-sampling Technique (SMOTE); Principal Component Analysis (PCA); Network Security; CIC-IDS2017; UNSW-NB15.