Development of an Enhanced Intrusion Detection System Using Mutual Information Feature Selection and Data Classification Algorithms

Abstract

The rising complexity of cyberattacks has made intelligent Intrusion Detection Systems (IDS) vital for modern network security. This study proposes an enhanced IDS framework that integrates Mutual Information (MI) for feature selection with two classification models: Support Vector Machine (SVM) and Backpropagation Neural Network (BPNN). Using the CICIDS2017 dataset, which contains real-world attack scenarios, the study first applies the Synthetic Minority Oversampling Technique (SMOTE) to balance class distribution. MI was used to reduce over 80 original features to the top 25 most relevant, improving both detection speed and accuracy. Six model variations were tested, including baseline classifiers, MI-enhanced models, and SMOTE-MI hybrid configurations. Performance was evaluated using accuracy, precision, recall, F1-score, and ROC-AUC. Among all tested configurations, the SMOTE + MI + BPNN model achieved the best results, with 99.93% accuracy, 0.999 precision, 0.999 recall, and 0.999 F1-score. These findings demonstrate the efficacy of combining advanced feature selection with robust classifiers to improve the performance of IDS. The results indicate that combining intelligent feature selection, class balancing and deep learning can significantly enhance IDS performance by offering a scalable and effective framework. This hybrid framework can significantly enhance cybersecurity operations across various sectors, including finance, healthcare, and government systems.