A Reliable Encrypted Traffic Classification Method Based on Attention Mechanisms

In online encrypted traffic classification, existing machine learning methods face the challenge of manually extracting flow-level statistical features. Deep learning methods encounter complex preprocessing issues, requiring models to balance complexity and accuracy. Therefore, this paper proposes a reliable method for encrypted traffic classification based on attention mechanism. The method designs a neural network model based on the Transformer architecture. The model employs embedding operations to extract byte features and utilizes position encoding to capture spatial features, which enables efficient packet-level classification of encrypted traffic. By using only the first 40 bytes of the packet header as input, the model avoids analyzing the payload, thereby reducing the risk of privacy breaches. Moreover, the simplified input features reduce the time overhead for feature extraction, enhancing training speed and making the model suitable for deployment in high real-time scenarios. This paper validates the proposed method experimentally using two public datasets: ISCX VPN-nonVPN and ISCX Tor-nonTor. The experimental results indicate that the proposed method outperforms existing approaches, achieving an improvement of at least 1.5% in classification accuracy and 1.6% in F1 score. Additionally, with a memory usage of 533 MB, it meets the performance requirements for low-resource scenarios.