In the context of the Internet of Things (IoT), malicious traffic detection and network flow classification are critical tasks for ensuring cybersecurity. With the rapid proliferation of edge devices and the widespread adoption of data encryption technologies, traditional traffic classification methods are becoming increasingly ineffective. Existing machine learning and deep learning approaches typically rely on handcrafted feature representations and struggle to effectively model the complex structural information characterized by heterogeneity and temporal dynamics inherent in IoT traffic. This limitation hampers their ability to fully capture potential semantic patterns. To address these challenges, this paper proposes a novel model architecture based on a parallel multi-channel fusion graph neural network (PMCF-GNN). The proposed method eliminates the need for manually engineered features and enables end-to-end learning on encrypted IoT traffic. Moreover, it efficiently extracts both packet-level structural features and flow-level temporal features in resource-constrained device environments. To accommodate diverse data substructures, a three-channel parallel training strategy is designed, significantly improving classification accuracy in IoT scenarios. Extensive experiments conducted on multiple real-world traffic datasets demonstrate that the proposed approach outperforms mainstream models in encrypted traffic identification tasks, exhibiting superior adaptability and robustness.