Generally speaking, most systems of network traffic identification are based on features. The features may be port numbers, static signatures, statistic characteristics, and so on. The difficulty of the traffic identification is to find the features in the flow data. The process is very time-consuming. Also, these approaches are invalid to unknown protocol. To solve these problems, we propose a method that is based on neural network and deep learning a hotspot of research in machine learning. The results show that our approach works very well on the applications of feature learning, protocol identification, and anomalous protocol detection.