Adaptive deep learning method for network traffic identification in open environments

This letter proposes an open set recognition (OSR) method in the unknown network traffic domain based on spatial feature learning. The recognition risk of unknown traffic classes is first analyzed, and some interesting insights into this problem is then provided. Based on this, an open set recognition method is proposed with the combination of spatial feature representation and recognition, where a central loss function is implemented for the representor to establish separable clusters for the N known traffic classes. Then, by training a compact closed boundary for each cluster using One-Class Support Vector Machine (OCSVM), which can effectively reduce the risk of open space errors. Probability modeling is then performed based on Extreme Value Theory (EVT), with decision scores from features to closed boundaries used for probabilistic modeling to achieve the final recognition process. Experimental results on the USTC-TFC2016 and CTU-13 datasets shows that the proposed method outperforms some current state-of-the-art in terms of both balanced accuracy and openness.