Using Large Language Models to Detect Smishing Messages in Telecom Networks

Scamming has emerged as a significant issue in recent years. Smishing, which is the use of text messages to deliver malicious links for phishing attacks, is one of the scamming tactics. Traditional methods for preventing smishing involve installing filtering applications on smartphones. However, due to privacy concerns, installing applications is not widely adopted. Therefore, we propose a more general solution that redirects short messages to our smishing detection system within the telecom network. Our system leverages the power of Large Language Models (LLMs) to directly detect smishing messages. Unlike traditional machine learning models that require extensive user data for training and raise significant privacy issues, using LLMs can avoid the need to collect and store sensitive user data. Our system consists of two key components. First, we parse organization names and query a domain database to obtain legitimate domains as the context of Retrieval-Augmented Generation. Second, we use LLMs to examine the content of the text messages and identify the smishing attempts and malicious links. With these two components, our system can enable widespread smishing detection and protect user privacy. To evaluate our system, we have conducted experiments on a smishing dataset. The results of our experiments demonstrate the effectiveness of our system for smishing detection.