With the widespread deployment of TLS protocols, encrypted traffic detection faces dual challenges of feature-space homogenization and dynamic attack patterns. Existing deep learning-based detection methods are constrained by flattened data representation and supervised learning paradigms, struggling to capture topological correlations in encrypted traffic. This study proposes the GSC-SAGE framework, which models traffic topology through IP-port graph construction and innovatively integrates generative subgraph contrastive learning with an enhanced graph neural network. The method employs breadth-first sampling for subgraph generation, constructs contrastive loss functions via Wasserstein Distance metrics for self-supervised graph representation learning, and synchronizes node-edge feature propagation through the E-GraphSAGE mechanism. In summary, this research provides a novel technical pathway for unsupervised graph representation learning in cybersecurity applications.