Abstract: An algorithm of cluster with outliers(CO) is proposed and its insensitivity to outliers in real datasets is analyzed.Anomalous intrusion data often do appear far from the normal network connections,essentially,they are outliers.A CO-based unsupervised anomalous detection method with a new distance definition of heterogeneous dataset is presented.By training data without label,the parameters in CO algorithm are regarded as a classification model to predict which cluster the current data belong to.Its validity is also discussed.Experimental results on the dataset KDDCUP99 comparing with other methods demonstrate that the proposed method has promising performance.