Dimension Reduction in Intrusion Detection Using Manifold Learning

Manifold learning is an emerging and promising approach in nonlinear dimension reduction. Representative methods include locally linear embedding (LLE) and Isomap. However, both methods fail to guarantee connectedness of the constructed neighborhood graphs. We propose k variable method called kv-LLE and kv-Isomap to build connected neighborhood graphs so as to enhance the robustness. The applicability of above two modified dimension reduction methods is examined by combining with the classifier of one class SVM. We evaluate new schemes with the KDD dataset and UNM dataset. Experiment results demonstrate higher detection rate and significant lower false positive rate on kv-Isomap method. The kv-LLE method is more suitable for kinds of text classification tasks. And the algorithm kv-Isomap performs much better than kv-LLE. We also have an unexpected gain in marking singular points.