An Intrusion detection algorithm based on an improved ART-2 Neural Networks is proposed in this paper. Based on traditional ART-2 neural networks, a prepositive matching system and an amplitude analysis procedure are employed. The prepositive matching system is employed to hasten the pattern matching and provide stable clustering while training the ANN. It also overcomes the limitation of sensibility to noise existing in ART2. The simulation results showed that the algorithm is efficient and precise. The information of the stable clustering can be used to provide supports for decision-making of defining normal and abnormal behavior patterns.