An XAI-Integrated Transformer Architecture for Proactive Email Spam and Abuse Detection

Abstract

This paper introduces an XAI-Integrated Transformer Architecture in proactive email spam and abuse detection. The
semantic patterns of email text get captured using a lightweight DistilBERT encoder and the metadata features consisting of the sender-domain indicators, header anomalies, URL statistics, and structural cues are processed by a CatBoost classifier. These two representations are combined in order to enhance resistance to changing spam strategies and content obfuscation. Imbalance is tackled by class-weighted training and SMOTE-ENN resampling but SHAP-based explainability can give global and token-level explanation of model decisions. Examples indicate that the hybrid design is more precise, recalls more information and its interpretability is higher than classical machine-learning designs, deep recurrent networks, and standalone transformer baselines. The solution provides a transparent and scalable solution appropriate in an enterpriselevel email security.