Detecting Healthcare Phishing Through Official Domain Grounding and Redirect Chain Analysis
Year: 2026
Type: preprint
Abstract: Healthcare phishing impersonates a named institution. It often arrives as a genuine Bitly, SendGrid or Microsoft link and reaches a lookalike lab-result or hospital-portal page only after one or more redirects. We ask whether grounding detection in a registry of official hospital domains helps, build a detector to test the question, and measure the answer under a protocol designed to let the claim fail. The detector uses a registry of 67 healthcare brands and their 79 official hostnames, computes features over the resolved redirect chain, not the first hop, and returns a brand-constrained counterfactual evidence card. We find that a simple official-domain rule, not a gradient-boosted model trained on 20\,000 URLs, carries the impersonation signal: on 457 lookalikes the rule recalls 0.978 against the model's 0.569, and on combination-TLD lookalikes, which place an official second-level domain under a different top-level domain, the model recalls 0.000 while the rule recalls at least 0.952, a class of evasion to which a learned classifier is blind. Official-domain false positives, measured out-of-fold by leave-one-brand-out, are 0 of 79 (Wilson 95\% confidence interval 0.000 to 0.046); removing the same brand from the registry raises them to 56 of 79, so registry coverage, not model capacity, is the binding constraint. Redirect-chain features detect wrapper-then-unknown chains: of 200 such chains, 198 carry no brand claim and the mismatch feature alone recalls 0.010. Correcting a reserved-TLD confound in our own synthetic slice lowers lookalike recall from 1.000 to 0.569. We release the registry, slices and harness.
Source: Research Square
Language: English
Topic: Spam and Phishing Detection
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Keyword: phishing detection
Open Access status: green
OpenAlex ID: W7234864509
