Network security is one of the most important issues in the field of computer science. The network intrusion may bring disaster to the network users. It is therefore critical to monitor the network intrusion to prevent the computers from attacking. The intrusion pattern identification is the key point in the intrusion detection. The use of the support vector machine (SVM) can provide intelligent intrusion detection even using a small amount of training sample data. However, the intrusion detection efficiency is still influenced by the input features of the ANN. This is because the original feature space always contains a certain number of redundant data. To solve this problem, a new network intrusion detection method based on nonlinear dimensionality reduction and least square support vector machines (LS-SVM) is proposed in this work. The Isometric Mapping (Isomap) was employed to reduce the dimensionality of the original intrusion feature vector. Then the LS-SVM detection model with proper input features was applied to the intrusion pattern recognition. The efficiency of the proposed method was evaluated with the real intrusion data. The analysis results show that the proposed approach has good intrusion detection rate, and is superior to the traditional LSSVM method with a 5.8 % increase of the detection precision.