Optimizing phishing domain detection in edge networks using graph attention network algorithms

The deep application of artificial intelligence algorithms in cybersecurity has provided new approaches for phishing domain name identification. Addressing the issues of high identification latency and weak generalization capability in edge network environments, this study constructs an optimized phishing domain name identification model based on Graph Attention Networks (GAT). The model takes domain name feature graphs as input, unifies character-level and root-level feature embeddings into a unified representation. It employs a multihead attention mechanism to adaptively allocate weights between nodes and incorporates context-relational enhancement strategies to capture implicit semantic relationships. Building upon the traditional Graph Convolutional Network (GCN) architecture, a sparse attention matrix is introduced to reduce redundant propagation paths, thereby improving computational efficiency on edge devices. Experiments conducted on the PhishTank and Alexa Top 1M datasets encompassed 200,000 domain samples. Results demonstrate that the model achieves Precision, Recall, and F1-score of 95.8%, 96.5%, and 96.2% respectively on the test set, representing an average improvement of approximately 4.1 percentage points over GCN. When deployed on the Jetson Nano platform, the average inference latency is 11.6 ms, lower than GCN's 19.4 ms. The model size is only 5.2 MB, meeting lightweight and real-time requirements. The research results validate the efficiency and robustness of the graph attention mechanism in edge phishing domain name identification.