Mitigating False-Positive Cascades in URL-Based Phishing Detection: A Structure-Factorized Character Encoder with Identity-Preserving Consistency Regularization

While URL-based phishing detectors often achieve almost 100% accuracy on the PhiUSIIL benchmark, a recent robustness test revealed that almost all of these detectors label legitimate sites as phishing if a benign-looking subdomain is added in front of a URL. The study examine why this is the case, and suggest a remedy. A fast audit reveals that all the benign urls that are in PhiUSIIL are bare https://www. homepages. Removing the scheme and the www. prefix lowers the in-distribution accuracy of a character-level CNN from 0.999 to 0.901, but raises its AUROC on an independent benchmark of URL-Phish phishing and Tranco benign URLs from 0.54 to 0.97, indicating that much of the reported accuracy reflects dataset artifacts. Then propose SF-CharNet, a structure-factorized encoder that processes the registrable domain, the subdomain and the path in separate branches fused by a learned gate, trained with identity-preserving consistency regularization: a label-free divergence penalty between predictions on a URL and on a variant that keeps its registrable domain. Under registrable-domain-disjoint evaluation, consistency regularization matches adversarial training on perturbation families seen during training (benign false-positive rate of about 3%), and no method generalizes to a held-out subdomain-injection family. Under cross-source shift, however, the combined model reduces the false-positive rate on perturbed external benign URLs from 0.90 with adversarial training to 0.18, at 1.9 ms per URL on a CPU. These results support cross-source, perturbation-aware evaluation as the default for URL-based phishing detection.