Abstract: Accurate traffic classification is the keystone of numerous network activities, so it has been a hot topic in network measurement for a long time. In recent years, Internet traffic classification using machine learning has been a new direction in this area. Nave Bayes and its improved methods have been widely used in this area, because they are simple and efficient. However, these methods depend on the probability distribution of sample space, so they have connatural instability. In order to handle this problem, a new method based on support vector machine (SVM) is proposed in this paper. This method converts the Internet traffic classification problem to a quadratic optimization problem using non-linear transformation and structural risk minimization, which performs good accuracy and stability. After the theoretical analysis, the comparison experiments with Nave Bayes and its improved methods on traffic sets are given. The experiment results validate that support vector machine method has three advantages in Internet traffic classification: Firstly, it is not necessary for flow attributes to satisfy the independence hypothesis, so feature selection is not required. Secondly, this method can work well with poor priori knowledge. Lastly, this method doesn’t use the probability distribution of sample space, so it can classify Internet traffic stably.