Integrating Graph Convolutional Networks And Multi-Head Self-Attention For Real-Time Threat Perception In Encrypted Communication Environments

End-to-end encryption protocols such as QUIC increase user privacy but also create major challenges for threat detection, with payloads inaccessible to standard inspection methods. Current DL architectures such as CNNs and Transformers tend to fail in identifying the intricate temporal and structural dependencies inherent in encrypted flow patterns. Moreover, they fail to ensure high accuracy while upholding privacy in real-time detection workflows. To tackle these voids, we introduce a new hybrid model that combines Graph Convolutional Networks (GCN) and Multi-Head Self-Attention (MHSA) to conduct real-time threat detection in encrypted settings based on the UC Davis QUIC dataset. Our solution represents every flow as a node and constructs graph edges on the basis of temporal and statistical proximity so that the GCN can learn local interactions while MHSA learns global dependencies. This two-stage representation provides better feature learning and classification performance. The experiment results indicate that the model outperforms current methods with accuracy of 99.33%, precision of 99.13%, recall of 99.12%, and F1-score of 99.12%, beating benchmarks like FD-CNN (98.61%), Transformer (89%), and ConvNeXt-ST-AFF (92.16%). The model also has an AUC-ROC of more than 0.99, validating its discriminative ability. Although the model is not necessarily deployed in real time, its design and low inference latency (~30ms on 80-node graphs) qualify it for real-time or near-real-time use. This supports proactive threat monitoring in encrypted environments without the need for payload inspection. This paper greatly enhances encrypted traffic classification through a union of graph learning and attention mechanisms, providing a scalable, privacy-respecting, and high-accuracy framework for real-time threat detection in contemporary communication networks.