Network Intrusion Detection Using Extreme Machine Learning Algorithm with Extreme Gradient Boosting for Feature Selection

Abstract

This study addresses the challenge of improving the performance of the Extreme Learning Machine model, particularly in accurately identifying minority classes in unbalanced datasets like UNSW-NB15 and NSL-KDD. The research question guiding this study is: How can we improve the ELM model's performance for better accuracy and minority class recognition in network intrusion detection? The methodology includes balancing the dataset to address the issue of poor minority class identification, using XGBoost for feature selection to reduce the curse of high data dimensionality, Particle Swarm Optimization finally used to optimize the model. The results show that the proposed approach outperformed other models when tested on the NSL-KDD dataset, achieving accuracies of 94.29% for binary classification and 89.02% for multiclass classification. However, on the UNSW-NB15 dataset, the model achieved a binary accuracy of 90.79%, which was lower than the performance of Random Forest (93.02%) and Decision Tree (92.76). In multiclass classification, the accuracy was 78.79%, underperforming compared to other state-of-the-art models. The study concludes that although the suggested approach performs well in binary classification, future studies need to focus on improving detection accuracies of datasets that are heavily unbalanced with multiple classes like UNSW-NB15 dataset.