A Study of Feature Reduction Techniques and Classification for Network Anomaly Detection

摘要

Due to the launch of new applications the behavior of Internet traffic is changing. Hackers are always looking for sophisticated tools to launch attacks and damage the services. Researchers have been working on intrusion detection techniques involving machine learning algorithms for supervised and unsupervised detection of these attacks. However, with newly found attacks these techniques need to be refined. Handling data with large number of attributes adds to the problem. Therefore, dimensionality based feature reduction of the data is required. In this work three reduction techniques, namely, Principal Component Analysis (PCA), Artificial Neural Network (ANN), and Nonlinear Principal Component Analysis (NLPCA) have been studied and analyzed. Secondly, performance of four classifiers, namely, Decision Tree (DT), Support Vector Machine (SVM), K Nearest Neighbor (KNN) and Naive Bayes (NB) has been studied for the actual and reduced datasets. In addition, novel performance measurement metrics, Classification Difference Measure (CDM), Specificity Difference Measure (S_PDM), Sensitivity Difference Measure (S_NDM), and F1 Difference Measure (F1DM) have been defined and used to compare the outcomes on actual and reduced data-sets. Comparisons have been done using new Coburg Intrusion Detection Data Set (CIDDS-2017) dataset as well widely referred NSL-KDD dataset. Successful results were achieved for Decision Tree with 99.0 percent and 99.8 percent accuracy on CIDDS and NSL-KDD datasets respectively.