Abstract: The dominant methodology of flow identification and classification is based on statistical analysis, which mainly focuses on extracting efficient characteristics. However, its illogical hypothesis of characteristics independency and data independency dwarfs the classification effectiveness. Thus quantities of methods are proposed to resolve the problem of characteristics dependency, but few achievements as to data dependency. Therefore, theory of traffic fractals is introduced to identify and classify flows in consideration of data dependency, which has to be modified and adjusted to fit the practical application. Finally, theoretical evaluations indicate the validity of the revised theory, and series of experiments demonstrate the performance of this method when classifying on coarse size and classifying unknown flows.

Key words: flow, identification and classification, traffic fractals, data dependency, Hurst exponent

摘要： 关于流识别与分类，目前主流的技术是基于统计学方法，核心环节是提取有效的特征属性集。这种方法的假设条件是，特征不相关，数据不相关。正因为这种假设的不合理性，使得分类效果有限。虽然已经有很多研究在集中解决特征相关性问题，但数据相关性却难以突破。因此引入流量分形理论，该理论建立在数据相关性基础之上。通过对原有理论进行必要的修改、调整以适用于流的分类识别，并用理论证明验证其有效性，最后通过系列实验体现该方法在粗粒度分类、未知流分类等方面的实际效果。

关键词: 流, 识别与分类, 流量分形, 数据相关性, Hurst指数