Hybrid Relabeled Model for Network Intrusion Detection
Bhumika Patel, Zaheenabanu Somani, Samuel A. Ajila, Chung–Horng Lung · 2018
With growing web communications throughout the Internet, the need for better security protection has been intensified. Intrusion detection is important for identifying malicious activities and cyber-crimes. Recently, UNSW-NB15 dataset has been popular in the research community for network intrusion detection system as it is publicly available labelled dataset which has a hybrid of real normal and contemporary synthesized attack activities of the network traffic. The goal of this paper is to relabel an unsupervised labelled data using a hybrid approach based on supervised learning. The methodology for training model in this work includes: (i) feature selection to remove redundant and highly correlated features, (ii) clustering the training dataset to create referential labels based on the size of cluster by using selected features, (iii) creating supervised learning model using ensemble classifiers with the generated referential labels, and (iv) testing individual data-point in doubt using the generated learning model. Our results show 81.29% accuracy compared to the original labels. Further, the proposed ensemble technique using LogitBoost and Random Forest algorithms produces 90.33% accuracy with the original labels, and 99.99% accuracy with the new labels for both training and testing dataset.