Title: An Automatic, Hybrid Mobile SMS Spam and Phishing Detection System Using Deep Learning
Authors: Michael Kinyunyu, Kenneth Brown, Yacob Norvat, Yassir Salum, Witness Mgana, Witness Suleiman, Sir koni
Volume: 10
Issue: 4
Pages: 207-214
Publication Date: 2026/04/28
Abstract:
SMS spam and phishing attacks continue to pose significant threats to mobile users worldwide, particularly in developing regions where digital literacy rates remain low and localized scam tactics emerge daily. Despite substantial advances in deep learning-based text classification achieving accuracy rates exceeding 97%, a critical implementation gap persists: no accessible, real-time mobile system exists that enables end-users to automatically verify SMS messages before acting upon them. This study addresses this gap by developing and evaluating an automatic, hybrid mobile SMS spam detection system that operates seamlessly across offline and online states. The system architecture combines a lightweight quantized on-device classifier for offline environments with a cloud-hosted CNN-GRU hybrid model achieving 98.42% accuracy for deep semantic analysis. A mobile application with background SMS interception capabilities and integrated URL safety verification was developed and tested. Evaluation results demonstrate that the proposed system achieves 98.42% accuracy, 97.89% precision, 96.74% recall, and 97.31% F1-score on the test dataset, with offline inference time averaging 0.23 seconds and online inference time averaging 1.87 seconds under standard network conditions. The system successfully processes Swahili-English code-mixed messages and detects malicious URLs with 94.5% accuracy. These findings establish that hybrid deep learning architectures can be effectively deployed as practical mobile-first solutions for real-time SMS spam detection, substantially reducing the gap between theoretical model performance and practical software engineering applications.