A Machine Learning Approach for Intrusion Detection

Securing networks and their confidentiality from intrusions is crucial, and for this rea-son, Intrusion Detection Systems have to be employed. The main goal of this thesis is to achieve a proper detection performance of a Network Intrusion Detection System (NIDS). In this thesis, we have examined the detection efficiency of machine learning algorithms such as Neural Network, Convolutional Neural Network, Random Forestand Long Short-Term Memory. We have constructed our models so that they can detect different types of attacks utilizing the CICIDS2017 dataset. We have worked on identifying 15 various attacks present in CICIDS2017, instead of merely identifying normal-abnormal traffic. We have also discussed the reason why to use precisely this dataset, and why should one classify by attack to enhance the detection. Previous works based on benchmark datasets such as NSL-KDD and KDD99 are discussed. Also, how to address and solve these issues. The thesis also shows how the results are effected using different machine learning algorithms. As the research will demon-strate, the Neural Network, Convulotional Neural Network, Random Forest and Long Short-Term Memory are evaluated by conducting cross validation; the average score across five folds of each model is at 92.30%, 87.73%, 94.42% and 87.94%, respectively. Nevertheless, the confusion metrics was also a crucial measurement to evaluate the models, as we shall see. Keywords: Information security, NIDS, Machine Learning, Neural Network, Convolutional Neural Network, Random Forest, Long Short-Term Memory, CICIDS2017