作　　者： ; ;

机构地区： 广东工业大学

出　　处： 《小型微型计算机系统》 2012年第5期1034-1040,共7页

摘　　要： 机器学习方法不依赖匹配协议端口或解析协议内容,而是利用网络流的各种统计特征识别网络应用,近年来得到了广泛关注和快速发展.本文总结了基于机器学习的网络流量分类方法自2004年来的研究进展,并且按有监督、无监督与半监督的区别进行分类、分析与比较.重点讨论了基于机器学习的网络流量分类研究的挑战与方向,即解决样本标注瓶颈、样本分布不平衡与动态变化、实时与连续分类以及分类算法可扩展性等核心问题. ML（machine learning） employs statistical network flow characteristics to assist in the IP traffic classification identification and classification,which is different with traditional methods that depend on well known application port numbers or deeply inspecting the contents of packet payloads.ML-based network traffic classification has been researched widely and developed rapidly.This survey reviews the significant works that cover the dominant period since 2004,and categorize,analyze and compare them according to their choice of ML strategies which include supervised,unsupervised and semi-supervised learning algorithms.We importantly discuss the orientations and challenges for the employment of ML-based traffic classifiers in operational IP networks.More specifically,the key issues such as sample labeling bottleneck,skewed data distribution,real-time and continuous classification and scalability of classification algorithms are discussed.

关 键 词： 机器学习 网络流 网络流量分类 统计特征

领　　域： [自动化与计算机技术] [自动化与计算机技术]