A Novel Method with Transformers for Fine-grained Encrypted Traffic Classification.

With the rise of cyber threats and data breaches, encrypted traffic has become increasingly important in ensuring online privacy and security. However, it also brings significant security challenges. Existing studies have shown that encrypted traffic from applications or web services can be identified using various methods. Nevertheless, there is limited generalization ability across multiple scenarios and difficulties in distinguishing similar traffic traces between web pages of the same application. To overcome the limitations of existing methods, we propose ANT-ET, a novel learning method tailored for fine-grained webpage analysis. ANT-ET extracts the contextual semantics from raw traffic and represents it as a semantic vector. These vector representations are subsequently utilized for classification through a classification layer and discriminator. The experimental results show that ANT-ET can effectively understand the semantics of encrypted traffic across various scenarios, achieving F1-score greater than 95% in four public encrypted traffic tasks, outperforming existing methods. In addition, we collected a real-world traffic dataset to verify the effectiveness of ANT-ET for fine-grained webpage detection.