A novel method for network intrusion detection based on nonlinear SNE and SVM

Abstract

In the case of network intrusion detection data, pre-processing techniques have been extensively used to enhance the accuracy of the model. An ideal intrusion detection system (IDS) is one that has appreciable detection capability overall the group of attacks. An open research problem of this area is the lower detection rate for less frequent attacks, which result from the curse of dimensionality and imbalanced class distribution of the benchmark datasets. This work attempts to minimise the effects of imbalanced class distribution by applying random under-sampling of the majority classes and SMOTE-based oversampling of minority classes. In order to alleviate the issue arising from the curse of dimensionality, this model makes use of stochastic neighbour embedding a nonlinear dimension reduction technique to embed the higher dimensional feature vectors in low dimensional embedding spaces. A nonlinear support vector machine with a radial basis function on a series of gamma values was used to build the model. The results demonstrate