基于流统计特性的网络流量分类算法

针对传统基于单个流统计特性的网络流量分类算法识别率低、分类算法复杂的问题，在分析各类应用协议的基础上，发现了一组易于获取、可有效区分不同业务的网络流量特征。将这一组特征应用于网络流量分类，可以有效解决以往对等网络（P2P）业务识别率低下的问题；同时利用该组特征仅需采用多项逻辑斯谛回归算法即可实现网络流量的分类，较传统流量分类算法有较低的复杂度。实验结果表明，该组特征用于分类还具有较好的泛化特性，只需较少量训练样本即可在在较长时间内保持较高的识别率。

A Network Traffic Classification Algorithm Based on Flow Statistical Characteristics

Based on analysis of application protocols, a group of multi-flow characteristics with low complexity, high quality is proposed to mitigate the problem of low recognition rate and high implementation complexity associated with the traditional flow classification algorithms using single flow statistics. These characteristics can effectively identify peer-to-peer (P2P) traffic in network flow classification, and improve the recognition rate of the traditional algorithms. They also enable the use of multinomial logistic regression algorithm to classify the network flow, and reduce the complexity of the traditional algorithms. Experiment results show that the proposed characteristics can achieve good generalization, and only need a small number of training samples to get a model that can maintain good performance for a long time.