Enhanced Intrusion Detection Through Optimized Feature Selection in Deep Neural Networks

This research examines the influence of feature selection on the effectiveness of a Deep Neural Network model for intrusion detection, utilizing the CIC-IDS 2018 dataset. To address data imbalance, we oversample the dataset and initially train a Random Forest classifier. We then evaluate the effectiveness of three feature selection methods: SHAP, Correlation-Based Feature Selection(CBFS), and Information Gain-Based Feature Selection(IGBFS). Our results show that feature selection markedly improves the DNN model's performance, with CBFS achieving the highest F1-score of 99.93% using only 57 out of 79 features. This surpasses the performance of both SHAP and IGBFS, indicating that CBFS identified a more optimal feature set for the DNN. We further analyze the specific features selected by each method, discussing potential reasons for the observed performance differences and underscoring the importance of feature selection in optimizing intrusion detection models. Our findings emphasize the effectiveness of Correlation-Based Feature Selection in enhancing model performance with a reduced number of features.