Abstract: This paper dedicates to propose a supervised local decision Hierachical Support Vector Machine (HSVM) learning model for anomaly intrusion detection in high dimensional feature space. The binary-tree structure of HSVM presents a divide-and-conquer algorithm for complex anomaly intrusion detection problem, i.e., the training signal for supervising local decision at each internal node is constructed according to information gain criterion. The embedded SVMs at internal node are trained on local optimized feature subsets standing on the sensitivity degrees of a margin to features. The experimental results suggest that the proposed anomaly intrusion detection method can gain learning model with better stability under the local decision supervisal of training signals. Further, it also achieves competitive detection accuracy and higher detection efficiency with condensed feature information.