Authors - Mohammed Rasol Al Saidat, Khaled Shaalan, Suleiman Y. Yerima Abstract - This paper presents a technical bibliometric review of artificial intelligence for bilingual Arabic–English smishing detection. It maps the evolution of machine learning, deep learning, and NLP for SMS phishing detection, identifies Arabic–English linguistic and security challenges, and grounds the review in an empirical audit of a public Arabic SMS spam corpus. The methodology pairs a PRISMA-style protocol and bibliometric mapping with the parsing of 1,494 SMS records (747 ham, 747 spam) from a public GitHub dataset. The field has shifted from rules and classical feature engineering to CNN, LSTM, Bi-LSTM, transformer, and large language model approaches. The dataset audit reveals strong class-conditional cues: spam messages are far longer and far richer in digits, phone numbers, short codes, URLs, currency markers, and Latin residues. A reproducible bilingual character TF-IDF baseline with a Linear SVM reached 0.9779 mean accuracy and 0.9775 F1 under stratified three-fold validation, matching the referenced CNN-Bi-LSTM model (0.9699 accuracy, 0.9707 F1). Robust bilingual smishing detection therefore requires hybrid text–security features, Arabic morphology-aware normalization, Unicode safety screening, external validation, and explainable deployment.
Open Zoom