A Categorically Reweighted Feature Extraction Method for Anomaly Detection

Anomaly detection is an important research problem in diverse computer security areas. Reliability of normal and abnormal samples is generally different in practice. We propose a scheme with a categorical re-weighting parameter to utilize this categorical reliability difference for feature extraction in anomaly detection. It is shown that this re-weighting parameter provides a method to tune the decision hyperplane between normal and abnormal classes. Based on two existing feature extraction algorithms, two new feature extraction algorithms using the proposed scheme are designed, which generalize the existing methods. Experiments show that the proposed methods outperform the previous state-of-arts in terms of both classification accuracy and robustness on synthetic and real-world data sets.