A Novel Feed-Forward Neural Network Model for Intrusion Detection Systems

Abstract

This paper presents a novel Feed-Forward Neural Network (FFNN) model for Intrusion Detection Systems (IDS), demonstrating superior performance in detecting network intrusions compared to traditional machine learning models. Using benchmark datasets, our proposed FFNN achieved the highest accuracy (99.33%), precision (99.33%), recall (99.33%), and F1 score (99.29%), underscoring its exceptional ability to accurately and reliably identify intrusions. While the K-Nearest Neighbour (KNN) and RandomForestClassifier models exhibited high ROC AUC scores (99.82% and 99.41%, respectively), their lower Cohen Kappa scores (61.53% and 56.24%) indicated less consistency in predictions. Conversely, the Support Vector Machine (SVM) recorded the highest Cohen Kappa score (94.56%), signifying strong agreement between predicted and actual classifications, but it fell short in overall accuracy (94.23%) and ROC AUC (91.35%). The Naïve Bayes (NB) model performed robustly with a Cohen Kappa score of 86.44% and high precision (98.23%) and recall (98%), yet did not surpass the FFNN's overall performance. This comprehensive evaluation illustrates that the FFNN model provides a balanced and reliable approach to intrusion detection, combining high accuracy, precision, recall, and consistency, making it an optimal solution for effectively mitigating cyber threats in network systems.