利用半监督近邻传播聚类算法实现P2P流量识别

P2P traffic identification based on semi-supervised affinity propagation clustering

为了解决利用少量标记样本实现准确的P2P流量识别,提出了一种基于半监督近邻传播(AP)聚类算法的P2P流量识别方法.首先对少量样本进行标记,然后在聚类过程中为标记样本和非标记样本设置不同的参考度,使标记样本能够优先成为类代表点,进而通过样本间的消息加权更新完成聚类,最后按照相应的"标记-类别映射"规则实现对P2P流量的识别.研究了参考度与消息加权更新对识别性能的影响,实验结果显示:当标记样本的比例为5%时,对P2P流量的识别准确率高于90%,误识别率低于3%;当标记样本的比例达到15%后,识别准确率高于95%,最高可达98%,而误识别率则低于1%;识别性能随标记样本比例的提高而提高.

Abstract ： In this study a method for P2P traffic identification was proposed based on utilizing semi-supervised affin- ity propagation clustering aimed at accurately identifying P2P traffic with as few labeled samples as possible. First- ly, a small amount of samples were labeled. Secondly, the labeled as well as the unlabeled samples were con- figured with different ＂preference＂ parameters, which made it more likely for the labeled samples to become exem- plars, as opposed to the unlabeled samples. Thirdly, all samples were clustered by a weighted message passing. Finally, P2P traffic was identified by the preset ＂marks-category＂ mapping rules based on the clustering results. The influence of the ＂preference＂ parameter and weighted message passing on the identification was examined. Ex- perimental results show the true positive rate （TPrate） was above 90% and the false positive rate （FPrate） was be- low 3% when the proportion of labeled samples was 5% ; the TPrate was above 95% with a maximum of 98% while the FPrate was below 1% when the proportion of labeled samples increased to 15% and above; and increase of the proportion of labeled samples may result to performance improvement of the proposed method.