A Transformer-Based Smishing Detection and Risk Alert System for Protecting Digitally Vulnerable Populations

Smishing, a social engineering attack via text messages, has seen a surge in frequency and damage, particularly among digitally vulnerable groups like older adults. Existing technology-centric detection systems often fail to prevent real-world harm for these populations. This study proposes a Transformer-based smishing detection and risk alert system tailored for their protection. By integrating and refining datasets from SmishTank and Kaggle, we evaluated five pre-trained models: DistilBERT, BERT, RoBERTa, ELECTRA, and DeBERTa-v3-small. To mitigate class imbalance, weighted sampling and Focal Loss were employed, alongside a hold-out strategy for generalization assessment. The RoBERTa-based model demonstrated superior performance, achieving an accuracy of 0.9542, an F1-score of 0.8194, and a ROC-AUC of 0.9824 on an unseen dataset. A Brier score of 0.0357 further validated its high predictive reliability. The system goes beyond binary classification by providing probability-based risk scores and real-time alerts to both users and guardians. This approach links detection directly to actionable responses, significantly enhancing practical prevention for digitally vulnerable users.