基于相似性迁移的网络流量分类方法

Network traffic classfication method based on similarity transfer

针对传统网络流量分类方法要求训练和测试数据分布一致、训练数据充足的假设在实际中难以满足的问题,引入迁移学习理论对其研究,提出从特征属性和样本域内相似性两个角度对网络流量数据优化,改进TrAdaBoost(boosting for transfer learning)算法的权重更新机制使其适应多分类任务,添加抑制因子解决源域权重转移现象。在数据集Moore上对该方法检验,结果表明,当目标样本量不足,该模型能根据相似性迁移,有效避免负迁移问题,相比其它流量分类模型具有更高的分类准确率。

Aiming at the problem that the assumption that the traditional network traffic classification methods require uniform distribution of training and test data with sufficient training data is difficult to be met in practice, the transfer learning theory was introduced. The network traffic data were optimized from the perspectives of feature and similarity. The multi classification task was adapted by adjusting the weight update mechanism of the TrAdaBoost(boosting for transfer learning) algorithm. The source domain weight transfer phenomenon was solved by the inhibitory factors. The method was tested on the Moore. The results show that when the target sample is insufficient, the model can transfer according to similarity, and the negative transfer can be effectively avoided. Compared with other traffic classification models, the model has higher classification accuracy.