Lexical-Feature Machine Learning versus Naive Keyword Matching for Phishing URL Detection Under Adversarial Evasion Attacks

Phishing remains one of the most widespread vectors for credential theft, and machine learning (ML) has become a common tool for detecting malicious URLs automatically. However, many practical detectors, especially in beginner and small-scale deployments, still rely on naive string/keyword matching rather than ML. This paper presents a controlled comparative study of a naive keyword-matching filter against a logistic-regression detector trained on lexical/structural URL features (length, character entropy, digit ratio, subdomain count, suspicious top-level domains). Both detectors are evaluated on a labeled dataset of legitimate and phishing-style URLs before and after four evasion techniques are applied: typosquatting, subdomain injection, Unicode homograph substitution, and a combined attack. Results show that the naive filter’s detection rate collapses to 0% under typosquatting and combined attacks and to 6.7% under homograph substitution, while in this controlled synthetic evaluation the lexical-feature ML detector retains 93.3–100% recall across all four fixed transformations. We analyze why the selected structural features are more robust than exact substring matching under the transformations evaluated, note the one condition (subdomain injection) that does not degrade either detector, and discuss the limitations of a synthetic, model-agnostic evaluation setting together with directions for future work.