Abstract
Most existing encrypted-traffic classification methods treat each traffic flow as an independent sample, overlooking the relationships shared by similar flows. This can amplify within-class variation and increase overlap between classes, particularly when their distinguishing characteristics are subtle and labeled data are limited. To address this problem, we propose PGCL-CG, a prototype-guided semi-supervised graph contrastive learning framework. We represent flow-level statistical features as a sparse, weighted k-nearest-neighbor graph and use GraphSAGE to aggregate information from neighboring flows and to capture their underlying structural relationships. The encoder is pretrained using self-supervised graph contrastive learning to obtain stable representations. We use a student-teacher architecture to produce more stable pseudo-labels and apply hierarchical density-based spatial clustering to filter out low-confidence assignments. We conduct extensive experiments on the two open datasets, demonstrating that PGCL-CG outperforms competitive baselines and remains robust even when only a small fraction of labels (e.g., 20% or 40%) is available.