一种基于TCM主动学习的P2P流识别技术  

Peer-to-Peer traffic identification using TCM based active learning

作　　者：戴磊[1,2] 云晓春[2] 张永铮[1,2] 吴志刚[1,2] 

机构地区：[1]中国科学院计算技术研究所,北京100190 [2]中国科学院研究生院,北京100039

出　　处：《高技术通讯》2010年第7期674-679,共6页Chinese High Technology Letters

基　　金：863计划(2007AA01Z444;2007AA01Z474;2007AA010501;2007AA01Z467);国家自然科学基金(60703021;60573134)资助项目

摘　　要：针对目前基于机器学习的流识别仍然存在着建立分类模型需要大量适用的训练数据,训练数据的标记需要依赖领域专家,因而导致工作量及难度过大和实用性不强的问题,采用主动学习技术提取少量高质量的训练样本进行建模,并结合支持向量机(SVM)分类算法提出了一种基于直推信任机(TCM)的样本筛选方法。实验结果表明,相对于已有的流识别方法,这种方法能够在仅依赖少量高质量训练样本的前提下,保证较高的召回率及较低的误报率,更适用于现实网络环境。To solve the problem that present identifications of Peer-to-Peer （P2P） traffic based on machine learning are still too difficult and time-consuming and are still unpratical due to the need for obtaining adequate qualified training data for the supervised classifiers to model traffic patterns and the great dependence on the domain experts, in marking the training data, the authors introduce the active learning method to select the most qualified data for training and propose a transductive confidence machine （TCM） based instance selection method for support vector machines （SVM）. The experimental results demonstrate that the proposed method is able to guarantee a high recall rate and low false positives by using a small quantity of high qualified data. Therefore, it is more suitable for the real network applications than the traditional ones.

关 键 词：支持向量机(SVM) 主动学习 直推信任机(TCM) 机器学习 不确定性采样 

分 类 号：TP393.02[自动化与计算机技术—计算机应用技术]