Abstract: The feature selection algorithm plays an important role in deep flow inspection technology; the correct recognition and classification of dataflow all need the process of flow characteristics selection, and using these flow characteristics to distinguish the different traffic types. The current information based feature selection algorithms compute the information entropy of the feature in the whole sample space, without considering the process of dynamic change in feature selection in computation, thus cannot accurately measure the correlation degree between each feature in the feature selection process, affecting the feature selection results, ultimately leading to reducing the performance of classification algorithm. This paper proposes an improved dynamic feature selection algorithms based on the information standard, considering the dynamic changes of information criteria in the feature selection process, by removing the redundant and useless information to avoid the interference of dynamic process and achieve the accuracy and efficiency for characteristics selection. Experimental data shows that, the proposed dynamic flow feature selection algorithm’s classification performance is better than that of other selection algorithms.

Key words: deep stream inspection, characteristic selection, stream indentify, dynamic information

摘要： 流特征选择算法在深度流检测技术中发挥着重要作用，数据流的正确识别和分类都需要选择流特征，通过这些流特征在业务流中的差异区分业务流类型。当前基于信息度量的特征选择算法在整个样本空间中计算特征的信息熵，没有将特征选择过程中的动态变化信息加入计算，因此不能准确地度量特征选择过程中各个特征之间的相互关系程度，冗余信息的存在影响特征选择结果，导致分类算法性能降低。提出一种改进的动态特征选择算法，该算法基于信息标准，充分考虑特征选择过程中信息标准的动态变化，通过删除由信息动态变化导致的冗余及无用信息，避免动态选择过程的干扰，达到准确并高效选择特征的目的。实验数据说明，提出的动态流特征选择算法的分类性能比当前其他选择算法较好。

关键词: 深度流检测, 特征选择, 流识别, 动态信息