Graph neural network-based computer network anomaly detection and algorithm efficiency evaluation

Abstract

Computer network security is a critical concern in the era of increasing cyber threats. Traditional intrusion detection systems rely on handcrafted rules and statistical features that fail to capture the complex topological relationships inherent in network traffic. This paper proposes a graph neural network (GNN)-based framework for computer network anomaly detection that models network traffic as dynamic graphs and applies Graph Attention Networks (GAT) combined with Gated Recurrent Units (GRU) for spatial-temporal anomaly detection. We further provide a comprehensive algorithm efficiency evaluation covering inference time, memory usage, throughput, and floating-point operations. Experiments on the CIC-IDS2017 and UNSW-NB15 datasets demonstrate that the proposed GAT-GRU model achieves 95.23% F1-score and 0.98 AUC, outperforming baseline methods including MLP, GCN, and GraphSAGE, while maintaining acceptable computational overhead for real-time deployment.