Multi-Type Feature Extraction and Early Fusion Framework for SMS Spam Detection

Abstract

SMS spam is a pervasive issue that affects millions worldwide, leading to significant inconvenience, time wastage, and potential financial scams. Given the prevalence and potential harm, accurate and real-time detection of SMS spam is crucial. This paper proposes a novel approach to SMS spam detection involving five steps: preprocessing, feature extraction, feature fusion, feature selection, and classification. Our model is designed to simultaneously capture local, temporal, and global text message features using a hybrid deep learning model to enhance feature representation. We evaluated our model using the UCI dataset, comparing it with traditional and deep learning algorithms such as RF and BERT using cross-validation to ensure the robustness of our results. Our proposed method exhibited superior performance, achieving a good accuracy of 99.56%, surpassing other methods. The effectiveness of this method in SMS spam detection proved its potential for real-world implementation, where it could substantially mitigate the prevalence and impact of SMS spam.

Keywords

• CNN
• data fusion
• deep learning
• LSTM
• SMS spam detection