A Review on Detection of Advance Fee Fraud Using Natural Language Processing

Advance Fee Fraud (AFF) remains one of the most persistent and financially damaging forms of cyber-enabled crime, with global losses running into hundreds of millions of dollars annually and a demonstrated capacity to adapt its linguistic strategies to evade automated filters. This review synthesizes the current state of research on the automatic detection of AFF messages using Natural Language Processing (NLP), with particular emphasis on the Bag-of-Words (BoW) model and the supervised machine learning classifiers commonly paired with it. Drawing on theoretical perspectives from information theory, social engineering, linguistics, and machine learning, the review traces how AFF messages have evolved from repetitive, template-driven text to highly personalized, emotionally calibrated narratives, and how detection research has responded with richer preprocessing, feature engineering, and classification strategies. It also situates BoW-based approaches against more recent transformer-based, multimodal, and adversarially robust methods, arguing that BoW retains practical relevance where interpretability, computational efficiency, and small-data performance are priorities. Three persistent gaps are identified: the limited integration of psychologically and linguistically informed features into BoW pipelines, the scarcity of cross-lingual evaluation, and the absence of systematic adversarial robustness testing for AFF-specific systems. The review concludes by outlining a theoretically grounded, lightweight BoW-based framework as a practical direction for future detection systems, particularly in resource-constrained deployment settings.