Abstract: Network intrusion detection systems (NIDS) provide a better solution to network security than other traditional network defense technologies, such as firewall systems. This paper proposes an intrusion detection model that combines deep belief network (DBN) and local preserving projection (LPP). The DBN was used for feature learning of the original data, and the LPP was used to fuse the deep features to further remove redundant and irrelevant features. Softmax classifier was used for classification. In addition, the accuracy, detection rate, false alarm rate and other classification indicators of this method on the NSL-KDD data set and UNSW-NB15 data set were studied and compared with the conventional machine learning classification method and the latest model method in other literature. The experimental results show that the DBN-LPP model improves the comprehensive performance of intrusion detection system, and its performance is better than traditional machine learning classification methods and other methods. This paper provides a new research method for intrusion detection.