在线多任务学习的骨干网网络流量分类研究

Backbone Network Traffic Classification Based on Online Multi-task Learning

网络流量分类是机器学习与网络安全领域中的一个研究热点.针对高速骨干网上网络流量的高速性与演化特性,基于在线稀疏学习算法FTPRL,提出一种在线多任务特征选择学习算法-MT-FTPRL.使用了Per-Coordinate学习率,对每个特征的学习率分别考虑,与全局学习率相比更具优势;提出一个在线多任务学习的网络流量分类框架,通过多个网络流之间的信息共享,提取一组拥有良好判别能力的共同特征子集;在实验部分构造了一个基于真实的骨干网网络流量的MAWI数据集,并通过对比实验对提出的算法及分类框架进行验证.实验表明,算法有着满意的分类准确性和检测效率,且能在多个网络流中提取一组共同的特征子集,提高分类系统的鲁棒性,更适应网络流量动态演化的特点.

Network traffic classification is a hot topic in machine learning and network security. Aiming at the features of high speed and evolutionary of network traffic on high speed backbone network,an online multi-task feature selection algorithm-MT-FTPRL is proposed,based on online sparse learning algorithm-FTPRL. Using the Per-Coordinate learning rate,this learning rate for each feature is considered separately,which is more advantageous than the global learning rate. Proposed an online multi-task learning based network traffic classification framework,extracting a set of common features with good discriminative ability,by sharing information between multiple tasks; A MAWI data set based on real backbone network traffic is constructed,and the proposed algorithm and classification framework are verified by comparison experiment. The experiment shows that,the proposed algorithm has excellent classification accuracy and detection efficiency,also can extract a set of common feature subsets in multiple network streams to improve the robustness of the classification system,and adapt to the features of dynamic evolution of network traffic.