Adversarially Resilient Detection of Business Email Compromise Using Hybrid Machine Learning
Journal of High-Frequency Communication Technologies · 2026
DOI: 10.58399/ecfa6495 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of High-Frequency Communication Technologies, 2026. Verfasst von Aliyu Yahaya Zakari und 5 weiteren Autor:innen. Die Arbeit nennt 21 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Tertiary Education Trust Fund.
Abstract (Originalsprache)
Business Email Compromise (BEC) has emerged as one of the most financially devastating and strategically sophisticated forms of cyber-enabled fraud, leveraging advanced social engineering techniques to circumvent conventional email security infrastructures. Existing detection mechanisms, predominantly rule-based or static in nature, exhibit limited adaptability to the dynamic, context-aware, and linguistically nuanced strategies employed by modern attackers. This study proposes an adversarially resilient hybrid detection framework that synergistically integrates Natural Language Processing (NLP), classical machine learning models (Support Vector Machines and Random Forest), and deep learning architectures, including Long Short-Term Memory (LSTM) networks and Bidirectional Encoder Representations from Transformers (BERT). To address the critical challenge of limited labeled BEC datasets, a controlled synthetic data augmentation strategy was implemented using a fine-tuned Generative Pre-Trained Transformer (GPT), enabling the generation of high-fidelity adversarial email samples. A comprehensive hybrid feature engineering approach was adopted to capture the multifaceted characteristics of BEC emails, encompassing linguistic, structural, metadata, stylometric, and contextual attributes. Model training and evaluation were conducted using stratified cross-validation, with performance assessed through accuracy, precision, recall, F1-score, and Area Under the Receiver Operating Characteristic Curve (AUC-ROC). Model interpretability was enhanced through SHapley Additive exPlanations (SHAP), providing transparent insights into feature contributions. Empirical results demonstrate that the LSTM model achieved superior performance, attaining an accuracy of 98.5%, significantly outperforming Random Forest (95.3%), Support Vector Machines (94.8%), and baseline rule-based approaches (85.4%). The proposed framework demonstrates strong potential for real-world deployment within enterprise email security ecosystems. Future work will focus on multilingual detection, real-time system integration, and large-scale validation within operational Security Operations Center (SOC) environments.
Förderung
- Tertiary Education Trust Fund
Zitieren
APA
Zakari, A. Y., Abiodun, O. I., Omolara, A. E., Owolabi, O., Adeyinka, O. P., & Oluwadare, O. S. (2026). Adversarially Resilient Detection of Business Email Compromise Using Hybrid Machine Learning. Journal of High-Frequency Communication Technologies. https://doi.org/10.58399/ecfa6495
Vancouver
Zakari AY, Abiodun OI, Omolara AE, Owolabi O, Adeyinka OP, Oluwadare OS. Adversarially Resilient Detection of Business Email Compromise Using Hybrid Machine Learning. Journal of High-Frequency Communication Technologies. 2026. doi:10.58399/ecfa6495
BibTeX
@article{Zakari2026,
author = {Zakari, Aliyu Yahaya and Abiodun, Oludare Isaac and Omolara, Abiodun Esther and Owolabi, Olumide and Adeyinka, Oyadiran Phillip and Oluwadare, Olusayo Samson},
title = {Adversarially Resilient Detection of Business Email Compromise Using Hybrid Machine Learning},
journal = {Journal of High-Frequency Communication Technologies},
year = {2026},
doi = {10.58399/ecfa6495}
}
EUVIMED