Phishing delivered through short text messages evolves quickly, producing scarce and imbalanced labeled data and shifting lexical patterns that challenge conventional supervised detectors. This study develops and audits a transparent detection framework on a private corpus of 15,950 Indonesian-language SMS messages, of which 13,411 unique messages (8,091 phishing, 5,320 normal) remain after removing 3,771 normalized duplicates. A class-weighted logistic regression over TF-IDF unigram–bigram features serves as the primary detector, and it is subjected to a battery of scenario evaluations: stratified five-fold cross-validation, a 2-way one/five/ten-shot episodic protocol, a controlled simulated lexical-drift stream with adaptive and static online learners, a binary class-introduction test, and an explainability pipeline built from exact linear SHAP and perturbation-based LIME. The detector attained an accuracy of 0.9806, an F1-score of 0.9838, a Matthews correlation coefficient of 0.9598, a ROC-AUC of 0.9973, and a PR-AUC of 0.9983 on a stratified hold-out. Five-fold cross-validation confirmed consistent performance with F1 scores ranging from 0.9829 to 0.9888. Episodic evaluation showed monotonic improvement from the one-shot to the ten-shot setting. Under simulated lexical drift, an adaptive online learner preserved a post-drift F1 of 0.9808 against 0.9604 for a static learner. Exact linear SHAP reproduced the decision function to within 1.8 × 10-15, LIME achieved a local fidelity of R2 = 0.921, and SHAP-based selection retaining 500 of 10,000 features (a 95% reduction) reduced F1 only modestly to 0.9741 with a mean pairwise Jaccard stability of 0.818. The findings indicate that a compact, class-weighted linear model can deliver strong, stable, and auditable SMS phishing detection, while few-shot and drift experiments delineate the boundaries within which limited-support and non-stationary performance should be interpreted.
Copyrights © 2026