Classifying imbalanced Internet traffic based PCDD

Internet traffic classification is important for network traffic engineering and management. Previous studies have shown that machine learning-based traffic classification methods can obtain high classification accuracy in a static classification context. On the other hand, Internet traffic flows are dynamic, and the classification model needs to be updated at certain intervals. This study first examined the concept drift situation in multi class Internet traffic classification. The classifier obtains high classification accuracy for the majority class over a long period, but looses the classification accuracy for the minority class within a short time. This suggests that concept drift occurs easily in the minority class. To adapt the dynamic classification context, this paper proposes a classification framework based on the concept drift detection method. The experiment results on real traffic datasets showed that the proposed approach can promptly detect concept drift for each class and improve the recall of the minority class while maintaining high overall accuracy.