A Semi-Supervised Learning Framework for Encrypted Traffic Classification Based on Supervised Contrastive Learning and Masked Sequence Prediction Tasks

With the widespread adoption of encryption technologies, the proportion of encrypted traffic in network traffic has significantly increased, making encrypted traffic classification a key technology for enhancing network security and optimizing network performance. Although existing research on encrypted traffic classification has made certain advancements, two major challenges remain: (1) Current classification methods generally lack effective utilization of unlabeled data, limiting models to specific task scenarios and reducing their generalization ability; (2) During pre-training, existing methods often focus solely on unlabeled data, failing to fully leverage label information, which restricts further optimization of the feature space. To address these issues, this paper proposes CoMask, a semi-supervised learning framework for encrypted traffic classification that integrates supervised contrastive learning and masked sequence prediction tasks. The framework uses multi-granularity feature sequences as input and employs a cross-training strategy to collaboratively utilize both labeled and unlabeled data. Specifically, the framework utilizes unlabeled data to perform the masked sequence prediction task to enhance the generalization ability of feature representations, while supervised contrastive learning tasks are conducted using labeled data to optimize the feature space. Experimental results show that CoMask outperforms existing state-of-the-art methods on multiple public datasets, with an average F1 score improvement of 3.3%, demonstrating its effectiveness in handling encrypted traffic classification tasks in both known and unknown category scenarios.