Email Phishing Detection Using NLP and Machine Learning
International Journal of Advanced Computer Science and Applications · 2026 · 17 (10)
DOI: 10.14569/ijacsa.2026.0171010 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Advanced Computer Science and Applications, 2026. Verfasst von Norah Aljurbua und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Phishing emails imitate legitimate correspondence and evolve to evade rule-based filters, yet many reported detection gains rest on single data splits, untested differences, and unexamined benchmark artifacts. This study compares TF-IDF classifiers (Logistic Regression, Decision Tree, SGD, and XGBoost) with recurrent networks (Simple RNN, GRU, Bi-LSTM, and Bi-LSTM with attention) on 17,085 deduplicated emails from the Phishing Email Detection dataset, using one leakage-controlled pipeline and paired significance tests. SGD achieved the highest held-out F1-score (97.33%; ROC-AUC 0.9979) and made significantly fewer errors than Logistic Regression, Bi-LSTM, and the attention model (McNemar, Holm-adjusted p ≤ 0.012). Under stratified cross-validation, however, the five leading models were statistically indistinguishable, and tuned linear models reached the highest mean F1-score (97.2%). A three-seed ablation found no gain from attention over parameter-free pooling. Provenance analysis showed that corpus-identifying tokens inflate scores: removing them lowered classical F1-scores by up to 3.5 points, and holding out Enron-like emails raised SGD’s false-positive rate on them from 0.5% to 3.8%. Under nine perturbations of 2,000 held-out emails, character-level edits altered fewer than 1.2% of SGD, GRU, and Bi-LSTM detections, whereas appended benign text evaded 10–53% of detections for every model except the Simple RNN (recall 14.9%); a heuristic context-aware boost did not mitigate this. A two-stage cascade escalating 5.6% of emails from SGD to the attention model achieved 98.52% recall at a 2.77% false-positive rate and 0.49 ms median latency, significantly outperforming the attention model alone while matching SGD alone. Well-tuned linear TF-IDF models therefore remain highly competitive.
Zitieren
APA
Aljurbua, N., Alhashmi, A. A., & Darem, A. A. (2026). Email Phishing Detection Using NLP and Machine Learning. International Journal of Advanced Computer Science and Applications, 17(10). https://doi.org/10.14569/ijacsa.2026.0171010
Vancouver
Aljurbua N, Alhashmi AA, Darem AA. Email Phishing Detection Using NLP and Machine Learning. International Journal of Advanced Computer Science and Applications. 2026;17(10). doi:10.14569/ijacsa.2026.0171010
BibTeX
@article{Aljurbua2026,
author = {Aljurbua, Norah and Alhashmi, Asma A. and Darem, Abdulbasit A.},
title = {Email Phishing Detection Using NLP and Machine Learning},
journal = {International Journal of Advanced Computer Science and Applications},
year = {2026},
volume = {17},
number = {10},
doi = {10.14569/ijacsa.2026.0171010}
}
EUVIMED