Intrusion detection can monitor network transmis-sion in real-time. It is an active security protection technology, which plays a great role in network security. In this paper, a method based on a hybrid autoencoder and decision tree is proposed to conduct intrusion detection. The autoencoder is trained through positive sample data to make its parameters fit the normal flow. The gap between normal samples and abnormal samples is distinguished by calculating the loss value, and the gap is standardized as a newly generated feature. This method can not only avoid the information loss caused by dimensionality reduction of high-dimensional data but also ensure speed and accuracy. The intrusion detection algorithm with hybrid auto encoder and decision tree obtained by the method proposed in this paper is stronger than using decision tree alone and many common machine learning methods. For example, compare the decision tree method 1.74 % better in accuracy, 2.16% better in precision, 1.47% better in recall, 1.81 % better in fscore.