Optimized intrusion predictions through feature selection methods

In the realm of cybersecurity, Intrusion Detection Systems are essential for protecting networks from evolving threats. This paper studies enhancing the performance of Intrusion Detection Systems in cybersecurity using Deep Neural Networks, through the integration of advanced feature selection techniques applied to the oversampled NSL-KDD dataset. The primary objective of this study is to identify relevant features crucial for improving classification accuracy. The main techniques used to identify these features are the SHAP, correlation-based feature selection, and information gain-based methods. The baseline model considered for this work takes 41 features attaining an F1 score of 98.7%. Using the top 30 features with attack-specific characteristics on SHAP explanation list, the F1 score improves to 98.8% compared to the baseline model F1 score of 98.7%. Moreover, using SHAP and correlation-based methods to identify and utilize 33 important features further enhances the F1 score to 98.9%. It is observed that information gain-based feature selection performs inferiorly to SHAP and correlation-based methods in intrusion detection systems due to its limited ability to capture feature interactions, lack of interpretability, and sensitivity to noise and redundancy. SHAP values and correlation-based methods offer more comprehensive insights into feature importance, leading to better performance and robustness in Intrusion Detection Systems. These findings underscore significant enhancements in the proficiency of the Intrusion Detection System through feature selection, thereby strengthening cybersecurity defenses against evolving threats.