PHISHREV+: An ASP-Based Post-Hoc Defeasible Machine Learning Framework for Phishing Detection

Mainak Sen, Kumar Sanka Ray, Amlan ChakrabartiarXiv 2026 · arXiv

TL;DR

PHISHREV+ is a hybrid framework that enhances machine learning phishing detectors by adding a logic-based reasoning layer. By treating ML predictions as "defeasible beliefs" that can be overturned by new evidence, it reduces false positives without the need for expensive model retraining, all while providing clear, graph-based explanations for every decision.

What Problem Does This Paper Address?

Phishing remains a massive cybersecurity threat, with over a million attacks recorded in just one quarter of 2026. While modern Machine Learning (ML) classifiers are highly accurate, they face three major hurdles:

  1. Fixed Decision Boundaries: Once trained, a model's logic is frozen. Incorporating new attacker tactics requires "wholesale retraining," which is computationally expensive and operationally disruptive.
  2. Alert Fatigue: Even a low false-positive (FP) rate generates a flood of wrong alerts, exhausting security analysts.
  3. The "Black Box" Problem: Standard models can't easily explain why they changed their mind or how specific contextual evidence overrides a statistical prediction.
Anatomy of a phishing attack.

Methods

The authors propose a dual-stage framework. In the first phase, four standard classifiers (SVM, KNN, Decision Tree, and Random Forest) generate an initial prediction. In the second phase, this prediction is treated as a defeasible belief—a conclusion that is accepted for now but can be retracted if contradictory evidence appears.

This retraction happens in a Non-Monotonic Reasoning (NMR) layer powered by Answer Set Programming (ASP). The researchers used Exploratory Data Analysis (EDA) to derive four hierarchical rules:

  • Rule 1 (The Trigger): If a website contains standard HTML meta tags (description, keywords, author), it is likely legitimate, and a "phishing" prediction should be revised.
  • Rules 2–4 (The Defeaters): These act as exceptions to Rule 1. If the URL is excessively long (>63 chars), lacks a 'www' prefix, or counter-intuitively embeds the domain name in the page title, the revision is "defeated," and the original phishing alert stands.

Crucially, this logic is implemented using the Clingo ASP solver, allowing for declarative, order-independent rules that can be updated in linear time ($O(n)$) without touching the underlying ML models.

Datasets

The study utilized a widely adopted benchmark dataset [27] containing 11,430 URLs (split 50/50 between phishing and legitimate). The dataset includes 87 features covering lexical, content, domain, and external service data. The authors also introduced a novel binary feature, meta_info_present, specifically to trigger their reasoning layer.

Key Insights and Ideas

  • Post-Hoc Refinement: You don't have to fix the model if you can fix the output. By wrapping "black box" ML in a "white box" logic layer, the system gains the benefits of both statistical pattern recognition and symbolic reasoning.
  • Counter-Intuitive Signals: The EDA revealed that 82.6% of phishing URLs in their subset embedded the domain in the page title, compared to only 56.8% of legitimate ones. Attackers do this to appear credible, making it a powerful "defeater" signal for the reasoning layer.
  • Explainability via XASP: The framework uses eXplainable ASP to generate Directed Acyclic Graphs (DAGs). Unlike SHAP or LIME, which show feature importance, XASP shows the actual logical derivation—exactly which rules fired and why a belief was retracted.

Key Results

The framework achieved consistent performance gains across all tested classifiers:

  • False Positive Reduction: SVM (41 $\rightarrow$ 39), KNN (47 $\rightarrow$ 46), DT (62 $\rightarrow$ 61), and RF (52 $\rightarrow$ 51).
  • Zero False Negative Increase: The system maintained complete parity with the base classifiers' false negative rates, ensuring no new phishing threats were missed due to the reasoning layer.
  • Efficiency: The entire NMR layer processed over 9,000 decisions in just 0.313 seconds. Adding a new rule is an $O(1)$ edit that avoids the multi-second (or minute) cost of retraining models like Random Forests or SVMs.

The paper highlights that existing approaches lack this defeasibility mechanism, as shown in their summary of the state-of-the-art:

Summary of Machine Learning Based Phishing Detection Approaches

Limitations and Future Work

The authors acknowledge that their framework was evaluated on a single benchmark dataset. The effectiveness of the meta_info_present trigger might degrade if phishers adapt by routinely including meta tags in their fraudulent sites. Furthermore, the conservative nature of the 4-rule hierarchy resulted in a low revision rate (0.05%), potentially missing other legitimate revision opportunities.

Future work aims to extend PHISHREV+ to multiple datasets to analyze "feature drift" over time and to incorporate more complex contextual signals like domain age and SSL certificate characteristics. The authors also see potential for this "defeasible ML" paradigm in other high-stakes fields like medical diagnosis and intrusion detection.

Summary updated

Abstract

Phishing website detection remains a critical cybersecurity challenge. Although machine learning classifiers achieve high accuracy on benchmark datasets, they suffer from a fundamental limitation: once trained, their decision boundaries are fixed and cannot incorporate new contextual evidence without retraining the model. This work presents \textbf{PHISHREV+}, a hybrid framework that instantiates the notion of \textit{post-hoc defeasible machine learning}. This framework does not replace the classifier. It treats the initial classifier predictions as defeasible beliefs and revises them through an Answer Set Programming based non-monotonic reasoning(NMR) layer. The framework operates in two phases. In the machine learning phase, four classifiers (SVM, KNN, DT and RF) are trained on the 87 feature benchmark dataset. In the post-hoc NMR phase, four defeasible rules derived from exploratory data analysis of the training set, and validated through Chi-square and Mann-Whitney U tests, formally revise initial predictions. The proposed framework achieves consistent false positive reduction across all our classifiers SVM (41$\rightarrow$39), KNN (47$\rightarrow$46), DT (62$\rightarrow$61), and RF (52$\rightarrow$51) while maintaining complete false negative parity with zero increase overhead. A comprehensive ablation studies across seven rule configurations confirms that the hierarchical combination of all four defeasible rules is necessary to achieve this optimal balance. Instance level explainability is provided through eXplainable ASP (XASP), which generates formal derivation graphs showing which defeasible rules fired and why the initial belief was retracted, a capability absent from all existing phishing detection approaches.
BibTeX
@misc{sen2026phishrev,
  author = {Mainak Sen and Kumar Sanka Ray and Amlan Chakrabarti},
  title = {{PHISHREV+: An ASP-Based Post-Hoc Defeasible Machine Learning Framework for Phishing Detection}},
  year = {2026},
  eprint = {2611.09273},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.09273},
}
Figure