Investigating Evasive Techniques In SMS Spam Filtering

Abstract

The prevalence of SMS spam remains a critical challenge, necessitating the development of advanced detection techniques capable of countering increasingly sophisticated evasion strategies employed by spammers. This study proposes a comprehensive SMS spam filtering framework leveraging Long Short-Term Memory (LSTM) networks. We introduce a novel SMS dataset comprising 61% legitimate messages and 39% spam, representing the largest publicly available SMS dataset to date. A longitudinal analysis of spam evolution was conducted, and semantic as well as syntactic features were extracted to enhance model performance. Comparative evaluations of multiple machine learning models, including traditional shallow classifiers and advanced deep learning architectures, were performed. Results indicate that conventional models and existing anti-spam services are highly susceptible to evasion tactics, yielding suboptimal accuracy. In contrast, the LSTM-based model achieved a superior classification performance with 98% accuracy. Despite this, certain evasion strategies continue to pose challenges, underscoring the need for further research. The findings emphasize the potential of deep learning frameworks in developing robust SMS spam detection systems.