AI Against Smishing in Kenya: Culturally Adapted SMS Scam Detection for Digital Trust.

Abstract

Kenya’s mobile-financial ecosystem, driven by MPesa and a mobile penetration rate above 133%, has advancedcommerce and financial inclusion but also exposed users to SMSbased scams exploiting linguistic diversity, cultural trust, andpsychological manipulation. Existing fraud-reporting mechanisms such as keyword filtering and manual reporting, remain reactiveand inadequate against these evolving threats. To fill this gap, thisstudy develops a culturally adapted, machine learning–drivenapproach to SMS scam detection tailored to Kenya’s multilingualenvironment. Using a crowdsourced dataset of 738 SMS messages (427 scams, 311 legitimate), XGBoost achieved 83.8% accuracywith strong precision and F1-scores, while Logistic Regression offered superior recall (91.8%) in detecting fraudulent content. SHAP analysis revealed urgency cues, code-switching features,and social proof language as the most influential predictors ofscams. The proposed model, designed for integration via APIs intoUSSD platforms, enables real-time detection and fosters user confidence through explainable alerts. This study provides ascalable, context-aware framework for fraud prevention inmobile-first economies and actionable insights for telecom operators, fintech platforms, and policymakers seeking to strengthen digital trust