Abstract: This paper proposes a characteristic selection algorithm based on information metric, which includes coarse grain selection and fine grain selection. The coarse grain selection calculates the cross-entropy between different characteristics and different business categories, and chooses the most representative characteristics using in flow classification. The fine grain selection calculates the consistency between characteristics to eliminate redundant characteristics. Experimental result shows that, when the characteristics selected in the proposed algorithm are used in data flow classification, the precision rate and recall rate are higher than the other similar algorithm, and this algorithm has lower complexity.

Key words: Deep Flow Inspection(DFI), characteristic selection, information metric, flow classification, mutual information, gain ratio

摘要： 提出一种基于信息度量的流特征选择算法。该算法可分为粗粒度选择和细粒度选择2个选择步骤。粗粒度的选择通过计算特征集合中各个特征与不同业务类别的互信息，选择在流分类中最具代表性的特征。对于选取的这些特征进行细粒度的选择处理，通过计算已选特征集合中特征之间的一致性，排除多余的特征。实验结果表明，该算法遴选出的特征在用于数据流分类时，准确率和召回率都较同类算法高，且时间复杂度较低。

关键词: 深度流检测, 特征选择, 信息度量, 流分类, 互信息, 增益比

CLC Number: 

• TP393