VeriPhish: Bridging AI Explainability and Accuracy in Phishing Detection Through XAI and LLMs

Result code in IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216224%3A14610%2F25%3A00143083" target="_blank" >RIV/00216224:14610/25:00143083 - isvavai.cz</a>

Result on the web

<a href="https://ieeexplore.ieee.org/document/11295116" target="_blank" >https://ieeexplore.ieee.org/document/11295116</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1109/ICCST63435.2025.11295116" target="_blank" >10.1109/ICCST63435.2025.11295116</a>

Result language

angličtina

Original language name

VeriPhish: Bridging AI Explainability and Accuracy in Phishing Detection Through XAI and LLMs

Original language description

Phishing attacks have evolved into a significant cybersecurity threat, becoming more frequent and increasingly difficult to defend against. While machine learning techniques show promise for detecting these threats, they typically operate as "black boxes," providing little insight into their decision-making processes. This lack of transparency undermines user confidence and reduces the effectiveness of threat response. To address this challenge, we introduce VeriPhish, a framework designed to improve phishing detection through a three-part architecture. The framework combines a machine learning classifier using domain-specific features, a dual-explanation layer that integrates LIME and SHAP for detailed feature-level analysis, and an LLM enhancement module that leverages DeepSeek v3 to convert technical explanations into understandable natural language. Our experiments demonstrate that VeriPhish achieves 98.4 % accuracy across all tested metrics, performing comparably to existing deep learning methods while offering superior explainability. The framework provides explanations with 94.2% accuracy and shows 96.8 % consistency between the LLM-generated explanations and the model's predictions. VeriPhish is available as both a fully functional GUI application and a lightweight Chrome extension, demonstrating its versatility across various deployment contexts. This work shows that high detection accuracy can be successfully paired with effective explainability in security applications, addressing the critical gap between AI-driven systems and user trust in phishing detection.

Czech name

—

Czech description

—

Type

D - Article in proceedings

CEP classification

—

OECD FORD branch

10200 - Computer and information sciences

Project

—

Continuities

I - Institucionalni podpora na dlouhodoby koncepcni rozvoj vyzkumne organizace

Publication year

2025

Confidentiality

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Article name in the collection

2025 IEEE International Carnahan Conference on Security Technology (ICCST)

ISBN

9798331523190

ISSN

—

e-ISSN

—

Number of pages

6

Pages from-to

1-6

Publisher name

IEEE

Place of publication

New York, NY

Event location

San Antonio, TX, USA

Event date

Oct 13, 2025

Type of event by nationality

WRD - Celosvětová akce

UT code for WoS article

001710575400035