Continental Academic Journal of Integrative Research and Studies
e-ISSN: XXXX 1, ISSUE 1
Research Article

Evaluating Adversarial Robustness and Evasion Vulnerabilities in Machine Learning–Based Phishing URL Detectors: A Problem-Space Attack and Defense Benchmark

Received
Sep 30, 2026
Accepted
Oct 22, 2026
Published
Nov 13, 2026
Available on
Nov 13, 2026
Abstract

Phishing URL detectors are using machine learning (ML) technologies and routinely achieve F1-scores of above .95, and still new phishing campaigns are slipping through. This gap indicates that an adversary that rewrites URLs to avoid classification has a poor correlation with accuracy measured under random splits. A problem space evasion vulnerability/ layered defense benchmark is presented. We used 231,482 cleaned urls from the PhiUSIIL corpus with each split by registered domain to avoid host leakage, trained seven detectors (linear, tree-based, kernel, neural, and character-level convolutional) and attacked each with eight families of attacks which preserve the functionality of the domain and at five perturbation depths (280 conditions). Clean F1 ranged from .9464 (decision tree) to .9826 (CharCNN). All detectors appeared to be impacted when attacked, with mean attack success rate (ASR) varying from .324 for XGBoost to .509 for the decision tree, and the ordered composite attack achieving .769 against the weakest model. However, subdomain injection and benign-context wrapping, both with mean ASR of .507 and .508 respectively, were the two most harmful single families and outperformed all operators, at character level. The vulnerability was representation specific, with CharCNN being the most vulnerable to homoglyph substitution (.397) but also the most resistant to typographic substitution (.188). A four-component defense stack (SHIELD) reduced composite ASR by 76.2%–79.7% at a mean clean-F1 cost of .0059, and resulted in a mean F1 of .8886 restored from .4921. All of the degradations were significant (p < .001) and the top-10 SHAP feature overlap dropped to .28–.55 under attack. It is not so much a matter of model capacity as of manipulability of the feature space.

Keywords
Phishing URL detection adversarial machine learning evasion attacks problem-space evaluation adversarial training explainable security
INTRODUCTION

The most often reported first point of entry in cybercrime is phishing. According to the Anti-Phishing Working Group (2025), over a million unique phishing attacks were reported in a single quarter in 2025, while the Federal Bureau of Investigation (2025) lists phishing and its variants as the top reported crime category. The uniform resource locator (URL) which is the artifact upon which the victim takes action -- clicking it -- makes URL classification the default first line of automated defense across browsers, email gateways, and endpoint agents (Basit et al., 2020; Chiew et al., 2018). Blacklists have no way to generalize to zero hour domains, since the median life of a phishing site is in hours (Chiew et al., 2018). Hence, they were replaced by supervised learning over features derived from the URL and the host, including the number of digits in the URL, the ratio of digits to the length of the URL, the depth of the URL subdomain and suspicious tokens (Hannousse & Yahiouche, 2021; Sahingoz et al., 2019). Modern detectors combine character representations, transformer representations (Chiper & Ionescu, 2025; Saxe & Berlin, 2017), graph signal from redirect chains, and graph signal from hosting infrastructure (Kim et al., 2022). Nowadays, F1-scores exceeding .95 on public benchmarks are common (Prasad & Chandra, 2024; Yıldırım, 2026).

This apparent maturity hides two methodological issues. The first is exchangeability, where URLs occurring on the same registered domain are randomly split between the training and test sets, allowing models to attain high scores by memorizing the host, not necessarily by learning transferable structure (Ahamed et al., 2026; Arp et al., 2022). The second one, and it is this one that will be considered here, is the passivity, in which the test data are generated independently of the classifier. If an operator is told that long URLs are bad, they shorten them; if he's told that subdomain depth is informative, he flattens it. The adversary is an optimizer with a goal of the error of the classifier. Evasion attacks are designed to corrupt the inputs during prediction time to be misclassified while still being semantically similar to the original input (Biggio & Roli, 2018; Szegedy et al., 2014). Pierazzi et al. (2020) and Arp et al. (2022) differentiate between the feature space, which is where an attacker can modify the numeric values of a feature, and the problem space, which is where he can only apply transformations that produce a functional (real-world) object. Feature-space attacks are easy to do, but they typically don't map to any valid URL, while problem-space attacks are limited and are much more indicative of the risk that can be deployed.

Hence, we stress the detectors in this problem space by applying them to one big corpus and results are reported as a fraction of phishing URLs that survive the mutation and get to the user. The study is guided by 4 research questions. RQ1: What is the vulnerability of various detectors from different learning paradigms and which attack families cause the most damage? RQ2 – Does the effectiveness of the attack increase with increasing perturbations (depth) and can it be generalized across models? RQ3: What is the cost in clean accuracy for the robustness of the individual and layered defenses? RQ4: What is the extent to which model explanations are stable when attacked? We present a domain-disjoint, problem-space benchmark consisting of 280 model–attack–depth combinations, where explicit realizability checks are conducted; evidence that host-structure manipulation is more effective than character-level obfuscation and that vulnerability is representation-bound; and a layered-defense evaluation combined with paired significance tests and an explanation-stability analysis.

METHODOLOGY

The study used a ten-stage pipeline, where each stage maintained a versioned and checksummed artifact, enabling traceability of the outcomes back to their inputs (Fig. 1). The design was replicated across seven models, eight attack families and five perturbation depths and contrasted with a clean reference and defended configurations. The unit of analysis is the URL and the unit of inference is the model–attack–depth condition, with comparisons made on the same test instances. There are two options to highlight. This partitioning was done at the registered-domain level, and as a consequence no host is present in two partitions, eliminating leakage that contributes to artificially high accuracy based on random splits (Arp et al., 2022). Only phishing urls were attacked during evaluation, benign predictions were not affected, simulating a threat model in which attackers control the urls they attack and benign traffic is not controlled by attackers.

References
  1. Roy, P. S., Popy, L. A., Saha, A., Sifat, M. O. I., & Sunbeam, I. F. (2026). Evaluating adversarial robustness and evasion vulnerabilities in machine learning–based phishing URL detectors: A problem-space attack and defense benchmark. Contin. Acad. J. Integr. Res. Stud., 1(1), 28-39
    https://doi.org/10.5281/zenodo.24213527
How to Cite This Article
Roy, P. S., Popy, L. A., Saha, A., Sifat, M. O. I., & Sunbeam, I. F. (2026). Evaluating adversarial robustness and evasion vulnerabilities in machine learning–based phishing URL detectors: A problem-space attack and defense benchmark. Contin. Acad. J. Integr. Res. Stud., 1(1), 28-39, https://doi.org/10.5281/zenodo.24213527
Format: APA 7th Edition • DOI: https://doi.org/10.5281/zenodo.24213527