An Encrypted Traffic Classification Framework Based on Higher-Interaction-Graph Neural Network

The exponential increase in encrypted traffic presents significant challenges to conventional traffic classification methodologies. The integration of deep learning and time series analysis has emerged as a contemporary approach, but the crucial higher interaction information among encrypted traffic packets is often neglected. Based on the fact that distinct categories of encrypted traffic manifest unique spatial structures and temporal patterns, we introduce a novel deep learning framework, termed Higher-Interaction-Graph encrypted classifier (HIGEC). This framework employs graph convolutional networks (GCN) and higher interaction information among packets for the precise classification of encrypted traffic flows. Initially, our approach incorporates a newly designed module that leverages metadata to transform each encrypted traffic flow into a graph structure, preserving the temporal information and spatial structure between packets. Subsequently, we introduce an advanced graph pooling and merge layer atop the GCN architecture that dynamically derives multi-order graph representations, augmenting our adaptability to diverse network environments. Meanwhile, we propose a new encrypted traffic dataset that addresses the inherent limitations of currently available public datasets to validate the effectiveness of our model. Comprehensive testing on our proposed dataset and two more public real-world datasets demonstrates that the HIGEC model significantly enhances accuracy, outperforming existing state-of-the-art methods.