In view of the characteristics that the abnormal behaviors generated in the circulation of electronic invoices tend to be diversified and complicated, this paper proposes an electronic invoice abnormal behavior detection model based on deep learning. Some sensitive information in user consumption is processed and clustered to establish a data set of abnormal behavior characteristics. A deep network learning enterprise consumption data combined with bilinear projection and attention mechanism is proposed, and an attention-enhanced bilinear layer (BL) model is introduced into the network to improve the training efficiency of the model. The validity of the proposed detection model is verified by the data of enterprise invoicing behavior provided by a financial company. The results show that the proposed model has the highest performance, and the average recognition accuracy is 93.25%; compared with ResNet, VGG and LSTM models, the accuracy is improved by 6.6%, 6.42% and 5.27%, respectively.