Short Message Service (SMS) spam continues to pose a significant threat to user privacy, security, and trust in mobile communication systems. Traditional rule-based and classical machine learning approaches rely heavily on surface- level lexical features, which makes them vulnerable to obfuscation and evolving spam strategies. In this paper, we propose a robust SMS spam detection framework based on a fine-tuned RoBERTa transformer model that captures deep contextual semantics of short text messages.

To validate the effectiveness of the proposed approach, we con- duct a comparative evaluation against classical baseline models, including Naive Bayes, Support Vector Machines, and Logistic Regression, using the publicly available SMS Spam Collection dataset. Experimental results demonstrate that the transformer- based model achieves near-perfect performance on a balanced evaluation subset, significantly outperforming traditional classi- fiers.

Furthermore, we present a real-time web-based deployment of the proposed system using a Streamlit interface, enabling interactive and user-friendly spam detection. All components of the framework, including the source code, trained model, and live demonstration, are publicly released to support reproducibility. This work bridges the gap between state-of-the-art natural lan- guage processing techniques and practical spam filtering systems