About the journal
Browse
Collections
Multimedia collections
Authors & reviewers
ARE-GNN: Attribute Representation Enhanced Graph Neural Network for Robust Encrypted Malicious Traffic Classification
Mingyuan XIA , Yunchun LI , Wei LI , Minyuan XIAO , Chen CHEN , Tong GUAN
With the widespread adoption of encryption in network communications, classifying encrypted malicious traffic has become an important challenge in network security. Modeling network traffic as graphs and applying graph neural networks has emerged as an effective approach. Recent work models flows as graph nodes and connects them through edges based on shared fields such as protocol fingerprints and server identifiers, where each shared field defines a relation type. However, the fields used for edge construction are not specific to any single traffic class. Flows of different classes may carry the same field values due to shared network environments and protocol implementations, and can therefore become connected. Attackers may further amplify the misleading relations through protocol mimicry. We define this as inter-flow relation confusion. In the presence of inter-flow relation confusion, graph neural networks classify each flow by aggregating information from neighbors under different relation types, and features from other classes enter the target flow representation during this process, reducing class discriminability. Existing graph-based methods for encrypted malicious traffic classification have sought to correct misleading relations, but attention-based weighting remains constrained by the neighborhood structure. We propose ARE-GNN to tackle this problem. The model pre-trains a node attribute representation solely from flow features and incorporates it into attention-based aggregation to reduce the influence of misleading neighbors. A joint training objective combines classification loss with supervised contrastive regularization to improve intra-class compactness and inter-class separability of the embeddings. Experiments on several public encrypted traffic datasets show that ARE-GNN outperforms all baselines. On the dataset with the most severe inter-flow relation confusion, it improves F1 by over 4 percentage points.
Encrypted Traffic Analysis / Malicious Traffic Classification / Graph Neural Networks / Contrastive Learning
Higher Education Press 2026
/
| 〈 |
|
〉 |