With the rapid development of network applications, the Internet traffic classification has a profound impact on the research fields of network resource allocation, traffic scheduling and network security. The traditional flow analysis method based on machine learning has strict requirements for the feature selection and distribution of network flows, which makes it difficult to accurately and stably classify the complex and changeable flow data in practical application. In order to solve the adverse impact of the complexity of sample features on the traffic classification, a new classification method based on deep forest, which utilizes the cascade forest of decision trees and the multi-grained scanning mechanisms aiming to improve classification performance in the case of limited scale of samples and features, was proposed. The machine learning algorithms including support vector machine, random forest and deep forest were trained and tested by using Moore, which is a flow data set in public domain. The experiment results show that the classification accuracy using deep forest model reaches 96.36%, which outperforms the other machine learning models.