Abstract: In this paper, a network flow protocol recognition method based on convolution neural network depth learning is proposed. Taking advantage of the similarity between network flow data and image, the network flow data is directly used as convolution neural network input, without the work of selecting and extracting the feature of network flow. Using supervise learning to train network flow protocol identification model, network flow protocol identification is realized. In this paper, a convolution neural network with three convolution layers is constructed. The network protocol recognition model is generated experimentally. The recognition rate of the underlying network protocol is 93.33%. The method has the ability to learn and extend. It can be scalable to the identification of traffic containing malicious code and identification of application traffic.