Intrusion Detection using Local Metric Learning

Machine learning methods have been successfully applied to the intrusion detection systems. These methods learn a classifier to discretize intrusions from normal connections. Intrusion detection systems usually have high-dimensional datasets, thus the performance of these methods depends on the used distance measure. Using either Euclidean distance or cosine similarity could not represent the distance between data in the suitable manner. Additionally, some researches have shown that high dimensional intrusion detection data lies on low dimension intrinsic manifold and exploiting this structure could improve the performance of learning methods.In this paper, we propose to learn a smooth metric on manifold of data by approximating it through the use of some linear metrics. This method resolves both high dimensional and nonlinearity challenges in the intrusion detection data sets. The experimental results show that this method improves both error rate and false detection rate significantly in comparison to Euclidean distance on Kyoto2006+ with various percent of labeled data.