Graph neural networks for dynamic network traffic classification

Abstract

Network traffic classification remains a challenging task essential for managing network resources and security. This study evaluates Graph Neural Networks (GNNs) for this purpose, leveraging their capability to handle graph-structured data effectively. We conducted comprehensive experiments across multiple studies: optimizing GNN architectures for different granularities, assessing model robustness against adversarial data, and proposing a novel graph representation inspired by BLINC. Our findings demonstrate that GNNs outperform traditional methods in accuracy and resilience, particularly in complex classification tasks. Both, the BLINC representation and GATConv layers, showed huge potential for traffic identification, but need further study. We also introduce insights into model interpretability using GNNExplainers. This research contributes to advancing GNN applications in network traffic analysis, emphasizing their potential to enhance network management practices and security measures.