Login Search
FMDB Transactions on Sustainable Computer Letters Submit Paper
    Crossref logo

    Archiving

    Portico

    Fernando Martins De Bulhão is a member of

    Crossref logo

PhishGuard: Interpretable Machine Learning for Real-Time Phishing Detection on Mobile Devices

Authors:
Farhan Tariq, Maruf Farhan

Addresses:
Department of Computer and Information Sciences, Northumbria University, London, England, United Kingdom.

Abstract:

Phishing attempts abuse human behaviour to steal login passwords, financial data, and personal information. Mobile users must quickly and accurately detect phishing emails owing to limited computational resources. Mobile phishing detection app PhishGuard uses machine learning and user-friendly explainability to deliver interpretable feedback. DistilBERT, Logistic Regression, and XGBoost were trained and tested on 57,804 safe and phishing emails. Accuracy, precision, recall, and F1 were assessed. DistilBERT surpassed Logistic Regression at 97.87% and XGBoost at 96.76% with 99.10% accuracy, 99.35% recall, and 99.10% F1-score. XGBoost and Logistic Regression performed well, while DistilBERT had the best predictive performance, mobile application balance, and phishing email detection. PhishGuard delivers a human-readable confidence score, risk indicator, and reason summary for decision-making. Abnormal words, urgency indicators, and sender dissimilarity enable Phishguard to bypass black-box constraints and boost user confidence. After execution-time investigation, DistilBERT achieved real-time mobile inference with strong recall and F1-score. A complete phishing-detection pipeline, including model performance, interpretability, and mobile viability, is provided. Ensemble learning may boost robustness; LIME and SHAP interpretability approaches need more research; continuous learning should adapt to future treatments; and practical implementation and usability user testing are ongoing. The experiment shows that machine learning with practical, interpretable outcomes can help consumers avoid phishing and make informed decisions. DistilBERT improved mobile email security, deployment, and performance.

Keywords: Email Classification; Phishing Detection; DistilBERT Model; Ensemble Learning; Logistic Regression; Confidence Score; Risk Indicator; Conventional Methods; Support Vector Machine; Random Forest.

Received on: 15/04/2025, Revised on: 02/07/2025, Accepted on: 07/08/2025, Published on: 29/06/2026

DOI: 10.69888/FTSCL.2026.000699

FMDB Transactions on Sustainable Computer Letters, 2026 Vol. 4 No. 2, Pages: 65-96

  • Views : 49
  • Downloads : 10
Download PDF