To solve low detection rate problem of intrusion for a small number of samples in mass unbalanced datasets, an anomaly detection based on Synthetic Minority Oversampling Technique (SMOTE) and Deep Belief Network (DBN), called SMOTE-DBN method, was proposed. Firstly, SMOTE technology was used to increase the number of samples in minority categories. Secondly, on the preprocessed balanced data set, the dimensionality of the preprocessed high-dimensional data was reduced by unsupervised Restricted Boltzmann Machine (RBM). Thirdly, the model parameters were finely tuned by Back Propagation (BP) algorithm to obtain the lower low-dimensional representation of the preprocessed data. Finally, softmax classifier was used to classify the optimal low-dimensional data. The simulation experiment results on KDD1999 show that, compared with DBN method and Support Vector Machine (SVM) method, the detection rate of SMOTE-DBN method is increased by 3.31 and 7.34 percentage points respectively, and the false alarm rate is decreased by 1.11 and 2.67 percentage points respectively.