Supervised Non-Linear Dimensionality Reduction Techniques for Classification in Intrusion Detection

Intrusion detection is still a crucial issue for network security. For visualization and classification on intrusion detection, the high dimensionality should be confronted. As a nonlinear learning method, Isomap is an effective dimension reduction tool among manifold learning algorithms. However, Euclidean distance is used in Isomap which is more suitable for continuous features. Another limitation is not using the class labels of data. This paper proposes supervised nonlinear learning method S-H-Isomap which utilized class labels to measure the dissimilarity between data points and replaced Euclidean distance with HVDM distance (Heterogeneous distance function). We evaluated the new scheme with KDD CUP 1999 datasets. In the classification experiments, S-H-Isomap was compared with WeighedIso, S-Isomap, Isomap, SVM, and k-NN. Experiments results show that S-H-Isomap performs the best with higher detection rate and the lowest false positive rate.