Metadata-Driven Business Email Compromise Detection Under Adversarial Evasion: Robustness Analysis and Adaptive Recovery

Abstract

Business Email Compromise (BEC) exploits organisational trust, impersonation, and metadata manipulation to evade detection systems. Although machine learning models can achieve high accuracy in email security tasks, their robustness against adaptive adversaries remains underexplored. This paper presents a unified robustness evaluation framework for metadata-driven BEC detection in privacy-constrained environments. Using the Enron email corpus, controlled BEC instances are generated and evaluated through a multi-level adversarial evasion benchmark (L0–L4). Logistic Regression, Random Forest, and XGBoost are assessed under baseline, adversarial, and adversarially retrained conditions. Results show that despite strong baseline performance, detection capability degrades sharply under structured evasion and collapses under the strongest evasion level. Adversarial retraining restores performance, with ensemble models recovering most of their baseline detection capability. These findings demonstrate the importance of staged adversarial evaluation, robustness-aware metrics, and adaptive retraining for trustworthy AI-driven BEC detection systems.