Business Analytics for Cyber Fraud Resilience Using Explainable AI to Detect Adaptive Synthetic Identities in Digital Account Opening A locked, U.S.-bank-aligned digital account-opening study
Journal of Intelligent Decision Making and Information Science · 17.09.2026 · 3 (12s) · 2898-2916
DOI: 10.17762/jidmis.v3.3714 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Intelligent Decision Making and Information Science, online erschienen am 17.09.2026. Verfasst von Aftab Hossain Sazu und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Synthetic identity fraud combines genuine and fabricated identity elements, while generative AI makes convincing digital artifacts easier to create. However, limited evidence shows whether plausible, detector-independent applications can evade account-opening models under fixed investigation capacity or whether routine temporal updating improves resilience. Using the one-million-row Bank Account Fraud benchmark, we developed models on Months 0–4, selected the static model in Month 5, selected and calibrated a drift-aware model in Month 6, and conducted a locked Month 7 test. We compared a static CatBoost model with a two-month recency-weighted version. A detector-independent Gaussian-mixture copula, used as a statistical proxy for scalable AI-assisted composition, generated 1,500 conventional attacks, each paired with an adaptive version using benign-generated identity, financial, or device attributes. Each scenario included 96,843 natural applications and 1,500 attacks; the 984 highest-risk cases were reviewed. Uncertainty was estimated using 10,000 paired bootstrap resamples of 30 generation-seed clusters. The static model detected 294 conventional attacks (19.6%) but only 105 adaptive attacks (7.0%), a difference of −12.60 percentage points (95% CI −14.80 to −10.53), supporting the vulnerability hypothesis. The drift-aware model detected 78 adaptive attacks (5.2%), 1.80 points below the static model (95% CI −2.53 to −1.07); therefore, the mitigation hypothesis was not supported. Vulnerability persisted at 0.5%, 2%, and 5% review capacities. Plausible composite adaptation materially reduced detection at binding review capacity, and recency weighting did not restore resilience. These findings establish a reproducible benchmark vulnerability, not real-bank effectiveness or control bypass. Banks should treat calibration, temporal updating, attack robustness, and identity verification as distinct controls and validate attack-aware challengers alongside layered verification.
Zitieren
APA
Sazu, A. H., Alam, M. M., & Sholomer, A. (2026). Business Analytics for Cyber Fraud Resilience Using Explainable AI to Detect Adaptive Synthetic Identities in Digital Account Opening A locked, U.S.-bank-aligned digital account-opening study. Journal of Intelligent Decision Making and Information Science, 3(12s), 2898-2916. https://doi.org/10.17762/jidmis.v3.3714
Vancouver
Sazu AH, Alam MM, Sholomer A. Business Analytics for Cyber Fraud Resilience Using Explainable AI to Detect Adaptive Synthetic Identities in Digital Account Opening A locked, U.S.-bank-aligned digital account-opening study. Journal of Intelligent Decision Making and Information Science. 2026;3(12s):2898-2916. doi:10.17762/jidmis.v3.3714
BibTeX
@article{Sazu2026,
author = {Sazu, Aftab Hossain and Alam, Md Mahbubul and Sholomer, Aitymon},
title = {Business Analytics for Cyber Fraud Resilience Using Explainable AI to Detect Adaptive Synthetic Identities in Digital Account Opening A locked, U.S.-bank-aligned digital account-opening study},
journal = {Journal of Intelligent Decision Making and Information Science},
year = {2026},
volume = {3},
number = {12s},
pages = {2898--2916},
doi = {10.17762/jidmis.v3.3714}
}
EUVIMED