Effect of Feature Selection and Machine Learning on IDS Performanc رسالة ماجستير

Abstract

With the increasing sophistication of cyber-attacks and their evolving nature alongside advancements in network infrastructures, Intrusion Detection Systems (IDS) are facing growing challenges. Machine Learning offers a promising approach to efficiently analyze the diverse datasets generated by network traffic. This research investigates the impact of feature selection on enhancing the accuracy and efficiency of IDS. By applying our feature selection model on the CICIDS2017 dataset, we identified the top 10 most relevant features that significantly improve the performance of multiple Machine Learning models. In addition to evaluating our own feature selection methodology, we also applied the feature selection models used in previous studies on our system. The results demonstrate that our approach to feature selection still outperforms these previous models in terms of both accuracy and computational efficiency. Specifically, Random Forest achieved an accuracy of 96.1%, Naive Bayes reached 82.5%, and both AdaBoost and K-Nearest Neighbors (KNN) surpassed 98% accuracy. While K-Nearest Neighbors (KNN) demonstrated excellent accuracy, it required considerably longer computational time compared to the other models. This research emphasizes the role of feature selection in optimizing IDS performance, demonstrating how our approach in selecting the most relevant 10 features enhance detection accuracy while maintaining efficient processing times. Our findings confirm that the feature selection methodology employed in this thesis provides a clear advantage over prior models, improving both detection accuracy and real time applicability of IDS.