Summary: Concerning the defects of the current clustering algorithm for its dependence only on the initial clustering center and failure in exactly distinguishing classes of non-concave shape, this paper applied the knowledge of group learning into the clustering algorithm and proposed the anomaly intrusion detection algorithm P-BFS based on graph clustering. In order to obtain more correct classification model, this algorithm introduced a measurement method of data points similarity based on the approximate function. The experimental results suggest the advantages of the application of the graph clustering algorithm in the intrusion detection system. In addition, it indicates that compared with the classical K-means clustering algorithm, P-BFS has better performance.