Abstract: Network traffic classification is widely used in research fields such as network resource allocation, traffic scheduling and intrusion detection systems. With the popularization of encryption protocols and the rapid development of network traffic, the traffic classifier based on deep learning has gradually attracted the attention of researchers due to its feature of automatically extracting features and high classification accuracy. But there has been no research on the credibility of network traffic classification. This article proposes a trustworthy deep learning model to classify encrypted network traffic. The proposed algorithm is based on the idea of RBF network and uses a new loss function and centroid update scheme for training. By using gradient penalty to force detection of input changes, it can effectively detect out-of-distribution data. On the two public ISCX VPN-nonVPN and USTC-TFC2016 traffic data sets, the proposed algorithm achieves 98.55% of the AUROC index, and has achieved the best out-of-distribution detection effect compared with similar algorithms. Extensive experimental results show that the proposed algorithm has high classification performance and can effectively detect out-of-distribution traffic data, which improves the credibility of the traffic classification model.

Key words: traffic classification, deep learning, uncertainty estimate, gradient penalty, out-of-distribution detection