Semantic Aware Fusion Model for Fraudulent Uniform Resource Locator (URL) Classification in the Web Ecosystem

The authors propose a Semantic-Aware Fusion Model (SACL). Unlike conventional classification models that rely solely on lexical or statistical features, the approach incorporates semantic relationships extracted from a domain-specific knowledge graph. They conducted single-model validation experiments using 11 popular models, including the Gated Recurrent Unit (GRU) model. Subsequently, they selected three state-of-the-art machine-learning models, including ACmix, demonstrating the highest accuracy in single-model trials for pairwise combinations. They introduce the ACL method to classify fraudulent URLs based on these insights. The experiments show that the Semantic-Aware Fusion Model model surpasses the other selected baselines with a maximum accuracy of 97.6% on the comprehensive dataset. They also validated the SACL model's superior generalizability and practicality through extensive testing.