Publication

Abstract : Rare and emerging cyber attacks are hard to iden tify because of the presence of a significant imbalance in the class distribution and the absence of representative examples in the intrusion detection datasets. This paper explores the limitations of conventional classifiers, such as Support Vector Machines, Decision Trees, and Naïve Bayes, when dealing with a significantly imbalanced attack class distribution. Limitations of the existing model is overcome by combining the oversam pling and undersampling techniques with the ensemble clas sifiers life XGBoost, Gradient Boosting and Random Forest. These techniques improve the representation of the minority class by incorporating noise filtering and outlier elimination to better define the class boundaries. The proposed system is tested using standard intrusion detection datasets, NSL-KDD and CIC-IDS2017, which represent a more realistic attack class distribution. The experimental results clearly show that the proposed system achieves substantial improvements in the recall, F1-score, and overall sensitivity of the minority attack classes with a controlled false positive rate. The results validate that the proposed integrated system facilitates efficient real-time intrusion detection by effectively combining the latest imbalance learning and ensemble modeling techniques.