Network traffic classification serves as a cornerstone for network management and security monitoring. Real-world traffic data naturally exhibits multimodal characteristics, encompassing metadata and payload fields. However, existing methods either rely on a single modality or adopt heuristic strategies for multimodal fusion, lacking the capability to model deep cross-modal interactions. As a result, models struggle to fully exploit the multimodal semantics of data, which constrains classification performance.

To address these challenges, our research center proposes the HetraNet framework. Built upon the pretraining and fine-tuning paradigm, this framework introduces a novel multimodal network architecture. In addition, three carefully designed pretraining tasks are incorporated to enhance cross-modal understanding and improve data representation quality. Extensive experiments on six public benchmark datasets demonstrate that HetraNet achieves highly competitive performance. Even under resource-constrained conditions with a significant reduction in labeled training samples, it maintains high performance and classification stability.

This work has been accepted by ICIC 2026, a recommended Class C conference by the China Computer Federation (CCF). The first author of this paper is Jinxia Wei, a Senior Engineer at our center, and the corresponding author is Chun Long, a Professorate Senior Engineer. This research is supported by the National Key Research and Development Program of China (Grant No. 2023YFC3304704).

Architecture Diagram of the HetraNet Framework

Related Publication：Jinxia Wei, Zilong Huang, Wei Wan, Yuhai Lu, Mu Zhu and Chun Long*, “HetraNet: A Hierarchical Encoder Pre-training Framework for Network Traffic Representation and Classification”‌International Conference on Intelligent Computing (ICIC), Toronto, Canada, 2026.