Network Intrusion Detection using Supervised & Un-Supervised Machine Learning Algorithms

In recent years, with the enormous volumes of data produced every day because of the rapid increase in computer/mobile usage and digital computer network, the defense of the computer system becomes more and more crucial. This brings a growing concern in information security and analysis of data, as network hackers discover new network intrusion attacks day by day. Network Intrusion detection system prevents network attacks by monitoring the computer network and analyzing the data in the computer system or the computer network. There have been various researches conducted to find an efficient solution to prevent network intrusion to ensure privacy and security in the computer network. But with the emergence of high volume, variety, and velocity of data in the computer network, the traditional techniques for network intrusion detection face difficulty in the data analysis process to detect attacks. Machine learning techniques prove to be a very effective tool for analysis of abnormal behavior as it provides supervised and unsupervised methods for clustering classification and regression. These techniques can identify any anomalous behavior in computer network traffic. The principal aim of this paper is to develop a machine learning model which can work on a dataset with no output label i.e., no dependent variable and should be able to classify any abnormal behavior in the network traffic, to satisfy this aim a combination of the unsupervised and supervised machine learning algorithms is used to develop a hybrid machine learning model which will cluster and classify any abnormal behavior in network traffic. The CICIDS2017 dataset is used to train and test the proposed model. The algorithms used to develop the hybrid machine learning model are Principal Component Analysis/PCA (unsupervised machine learning) used for dimensionality reduction, Kohonen Self Organized Maps (unsupervised machine learning) used for visualization and clustering and Artificial Neural Networks (supervised Machine learning) used for regression and classification. The accuracy calculated for the evaluation of the developed hybrid model is 82%, with a False Positive of 9.40% and False Negative of 6.64%.