Implementation of a Machine Learning-Based Spam Filter for Nigerian Mobile Phone SMS Platforms

SMS spam and fraudulent messages are prevalent on Nigerian mobile networks and cause significant consumer harm, ranging from financial loss to data theft, yet device-level spam filtering remains inadequate for most Nigerian mobile phone users. This project implemented a machine learning-based SMS spam classifier designed for deployment as an Android application on Nigerian mobile handsets. A Naïve Bayes classifier and a Support Vector Machine classifier were trained on a bilingual dataset of 8,200 labelled SMS messages, comprising 5,000 English-language messages from the UCI SMS Spam Collection and 3,200 Pidgin, Hausa, and Yoruba-inflected Nigerian SMS examples collected and labelled by a research team. Feature extraction used TF-IDF weighted bag-of-words with 1,500 features. The models were compared against a baseline keyword blacklist filter. The SVM classifier outperformed Naïve Bayes on the test set, achieving 96.8 percent precision and 94.3 percent recall for spam detection. The keyword blacklist achieved 81.2 percent precision and 78.4 percent recall. The Android application integrated the trained SVM model as a TensorFlow Lite inference module, classifying incoming messages within 280 milliseconds. False positive rate for legitimate messages was 2.1 percent. The application ran without noticeable battery impact on Android 8.0 and above devices. The study recommends extending training data collection to include OTP spoofing and bank fraud message patterns common on Nigerian mobile networks to improve detection of emerging Nigerian-specific spam typologies.

Keywords: SMS spam filter, machine learning, SVM, Nigerian mobile networks, text classification