Explainable and Efficient SMS Spam Detection Using Lightweight Attention-Based Sequential Models

SMS spam detection has become a critical challenge as fraudulent messages continue to evolve in sophistication, targeting mobile users through phishing links, fake prize notifications, and credential-harvesting schemes. Existing detection systems often lack interpretable decision-making, are evaluated only on a single dataset, and are too computationally heavy for deployment in resource-constrained environments. This paper proposes a lightweight Attention-LSTM architecture that addresses these limitations within a unified framework. A soft-attention mechanism is applied over LSTM hidden states to generate interpretable context vectors, and SHAP-based word-level attribution is used to provide explainable predictions. The model is trained on the UCI SMS Spam Collection dataset and evaluated on the Enron-Spam corpus and a recent SMS-2024 dataset to test cross-domain and temporal robustness. Experimental results show that the proposed model achieves high accuracy and F1-score while maintaining low parameter count and fast inference time, making it suitable for real-time deployment in mobile and telecom environments. The results demonstrate strong generalization across datasets and robustness against evolving spam patterns. Note: This paper is currently under review in the journal Applied Soft Computing.