Intrusion Detection System Using Convolution Neural Network with Feature Selection for Multiclass Classification

In an era of increasing cyber threats, traditional Intrusion Detection Systems (IDS) often struggle to keep pace with evolving attack patterns and the growing volume of network traffic, leading to inefficiencies in accuracy and detection speed. This study addresses these problems by exploring the enhancement of IDS using Convolutional Neural Networks (CNNs) combined with feature selection techniques, by using Correlation Coefficient and Recursive Feature Elimination (RFE) with Random Forest (RF). Using the CIC-IDS2017 dataset, redundant features were first removed via Correlation Coefficient, and RFE-RF was then applied to retain the most informative features. Results show that using the full feature set achieved an average accuracy of 99.66%, though it required 18.87 seconds for training and 89.34 milliseconds per data point for evaluation. The Correlation Coefficient method improved accuracy to 99.77% while reducing both training and evaluation times. RFE with half of the features achieved an accuracy of 99.68%, reducing training time to 10.81 seconds and evaluation time to 83.88 milliseconds. The most efficient configuration, RFE with ten features, resulted in an accuracy of 99.46%, the shortest training time of 8.28 seconds, and an evaluation time of 77.95 milliseconds, albeit with a slight increase in evaluation loss. This research highlights the critical role of feature selection in improving both the accuracy and efficiency of IDS, demonstrating the effectiveness of Correlation Coefficient and RFE techniques and providing a foundation for future refinement of CNN-based IDS models.