EAPT: An encrypted traffic classification model via adversarial pre-trained transformers

Abstract

Encrypted traffic classification plays a critical role in network traffic management and optimization, as it helps identify and differentiate between various types of traffic, thereby enhancing the quality and efficiency of network services. However, with the continuous evolution of traffic encryption and network applications, a large and diverse volume of encrypted traffic has emerged, presenting challenges for traditional feature extraction-based methods in identifying encrypted traffic effectively. This paper introduces an encrypted traffic classification model via adversarial pre-trained transformers-EAPT. The model utilizes the SentencePiece to tokenize encrypted traffic data, effectively addressing the issue of coarse tokenization granularity, thereby ensuring that the tokenization results more accurately reflect the characteristics of the encrypted traffic. During the pre-training phase, the EAPT employs a disentangled attention mechanism and incorporates a pre-training task similar to generative adversarial networks called Replaced BURST Detection. This approach not only enhances the model’s ability to understand contextual information but also accelerates the pre-training process. Additionally, this method minimizes model parameters, thus improving the model’s generalization capability. Experimental results show that EAPT can efficiently learn traffic features from small-scale unlabeled datasets and demonstrate excellent performance across multiple datasets with a relatively small number of model parameters.