Cascaded classifier for improving traffic classification accuracy
Machine learning (ML) techniques have been widely applied in recent traffic classification. However, the flow-level statistics are prone to improve the accuracies for some applications; however, to reduce the accuracies for others. To address the problem, the authors propose a cascaded traffic classifier that is composed of both several binary sub-classifiers and a multiclass sub-classifier. The authors first present theorems that show how to make an optimal cascade of sub-classifiers, and then design a cascaded classification algorithm for improving the accuracy of flow-level traffic classification. In addition, to improve the classification speed, the authors propose a parallel scheme for the cascaded classifier. The authors evaluate their approaches on the traces captured from entirely different networks. Compared with the previous multiclass traffic classifiers built in one-time training process, the cascaded classifier is superior in terms of the overall accuracy and the accuracy for each application.