Reliability Aware Fusion and Mask Specific Decision Calibration for Phishing Detection with MAG-Phish X
Computer and Decision Making: An International Journal · 02.11.2026 · 3 (6S) · 2672-2684
DOI: 10.1922/comdem.v3i6s.19008 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computer and Decision Making: An International Journal, online erschienen am 02.11.2026. Verfasst von Le Thi Phuong Anh und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Phishing detectors must combine incomplete evidence without transferring decision thresholds indiscriminately between acquisition conditions. This study develops a score-level instantiation of MAG-PhishX and examines whether entropy-aware fusion, probability calibration, and evidence-mask-specific thresholds improve the reliability of decisions. A reproducible pilot uses PhiUSIIL with 235,370 distinct URLs after exact deduplication. Registrable domains are partitioned into separate classifier-training, fusion-training, calibration, policy-selection, and test subsets. Eighteen URL features and 26 webpage-derived features support histogram gradient-boosting classifiers, static and adaptive fusion baselines, and controlled missing-evidence tests. On the primary test set of 22,599 URLs, entropy-aware fusion achieves a Brier score of 0.000547, reduced to 0.000348 by temperature scaling. However, feature concatenation achieves a lower Brier score of 0.000040, and several methods approach perfect ranking; the pilot therefore does not establish a general superiority of adaptive fusion. More critically, reusing the complete-evidence calibrated threshold when all webpage features are unavailable raises the false-positive rate to 100%. Selecting a separate threshold on held-out URL-only evidence reduces this rate to 0.825%, with 99.486% phishing recall. Three domain-partition seeds support the calibration analysis. These findings distinguish probability estimation from operational threshold validity and motivate evaluation under acquisition-dependent missingness. The paper specifies subsequent raw-HTML, DOM, screenshot, and temporal-graph experiments using complementary datasets. The current evidence supports a decision-calibration pilot, rather than an empirical validation of the complete deep multimodal architecture.
Zitieren
APA
Anh, L. T. P., Bao, T. M., Nguyen, V., & Le, D. (2026). Reliability Aware Fusion and Mask Specific Decision Calibration for Phishing Detection with MAG-Phish X. Computer and Decision Making: An International Journal, 3(6S), 2672-2684. https://doi.org/10.1922/comdem.v3i6s.19008
Vancouver
Anh LTP, Bao TM, Nguyen V, Le D. Reliability Aware Fusion and Mask Specific Decision Calibration for Phishing Detection with MAG-Phish X. Computer and Decision Making: An International Journal. 2026;3(6S):2672-2684. doi:10.1922/comdem.v3i6s.19008
BibTeX
@article{Anh2026,
author = {Anh, Le Thi Phuong and Bao, Tran Minh and Nguyen, Van-Nui and Le, Dac-Nhuong},
title = {Reliability Aware Fusion and Mask Specific Decision Calibration for Phishing Detection with MAG-Phish X},
journal = {Computer and Decision Making: An International Journal},
year = {2026},
volume = {3},
number = {6S},
pages = {2672--2684},
doi = {10.1922/comdem.v3i6s.19008}
}
EUVIMED