A Multi-Scenario Traffic Classification Method Based on Pretrained Encoder and Text Convolutional Neural Network

The current network traffic application scenarios are complex and diverse. The traditional traffic classification methods are obviously no longer suitable for the complex traffic of today. Most of these methods are usually only designed for traffic classification in single scenarios, and they yield mixed performance results when applied to multiple traffic datasets. With the advancement of deep learning, the current predominant methods for complex traffic classification rely on pre-trained models. These models effectively address the issue of weak semantics in encrypted traffic and can achieve high performance in classifying such traffic. We propose a multi-scenario traffic classification approach based on an existing benchmark model. This approach treats the traffic as textual sequence information and processes it as text. It involves extracting the 12-layer [CLS] vectors of the pre-trained encoder and combining them with textual convolutional neural networks. This approach yields features for key information as well as a combination of low-level and high-level semantics. The F1 score for the encrypted traffic ISCX-VPN dataset reaches 99.66%, while for the malicious intrusion traffic ISCX-IDS2012 dataset, it reaches 99%. These scores are higher than those of most benchmark models. Additionally, performance on the USTC-TFC dataset is also on par with the best current malware traffic classification models, with an F1 score of 99.50%. Experiments have proven that the model presented in this paper can effectively distinguish network traffic in various scenarios.