Artificial Neural Networks (ANNs), especially back-propagation (BP) neural network, can improve the performance of intrusion detection systems. However, for the current network intrusion detection methods, the detection precision, especially for low-frequent attacks, detection stability and training time are still needed to be enhanced. In this paper, a new model which based on optimized BP neural network and Dempster-Shafer theory to solve the above problems and help NIDS to achieve higher detection rate, less false positive rate and stronger stability. The general process of the authors' model is as follows: firstly dividing the main extracted feature into several different feature subsets. Then, based on different feature subsets, different ANN models are trained to build the detection engine. Finally, the D-S evidence theory is employed to integration these results, and obtain the final result. The effectiveness of this method is verified by experimental simulation utilizing KDD Cup1999 dataset.