Machine Learning-based Approach for Scam Message Detection in Kiswahili Using Natural Language

Scam messages transmitted via SMS and digital platforms present significant cybersecurity threats, particularly in regions where mobile money services are widely used. In Tanzania, where Kiswahili is the dominant language of communication, existing scam detection systems—primarily designed for English—fail to capture the linguistic and contextual nuances of Kiswahili messages. This study presents a machine learning-based approach for detecting scam messages written in Kiswahili using Natural Language Processing (NLP) techniques. A labelled dataset of 5,000 real-world messages was compiled through public sources, user contributions, and collaboration with law enforcement. Four models—Support Vector Machine (SVM), Random Forest (RF), Bidirectional Long Short-Term Memory (Bi-LSTM), and Swahili-BERT—were developed and evaluated using accuracy, precision, recall, and F1-score. Swahili-BERT outperformed all other models, achieving a detection accuracy of 96.3%, demonstrating the efficacy of transformer-based models in low-resource language settings. The system was deployed through a Flask-based API for real-time use on mobile platforms. This research advances Kiswahili NLP, strengthens digital security infrastructure in Tanzania, and offers a scalable solution to combat SMS-based fraud in mobile money ecosystems