Using LLM-Generated Data to Create a Roman Urdu Scam Call Detector

Purpose: Scam calls are spreading at an alarming pace where it is estimated that the world will lose more than a thousand billion dollars in 2024. Current machine-learning systems to classify scam-calls are not yet generalized: most of such systems are only monolingual detectors, with the multilingual systems based on LLM models proving impractical because of their high computational costs. In addition, due to the high rate of innovation of scam-call strategies, most of the implemented models are obsolete. This paper aims to suggest and analyze a multilingual, cheap, and easily updateable architecture to detect scam-calls with the help of LLM-generated synthetic data.Design/Methodology/Approach: The paper presents a model that was trained purely on scam and non-scam conversations of the multilingual nature as generated by the LLM. Evaluation was done using a small human-written data of actual scam and non- scam call transcripts. The method focuses on scalability, linguistic flexibility, and speedy re-generation of data with the help of synthetic generation.Findings: The experimental results indicate that an experimental model that has been trained on synthetic data can transfer to actual scam-call data. The model, when tested on the human-written data, obtained an average score of more than 90 percent accuracy, and F1-score, proving the viability of synthetic multilingual training data, which can be used to detect scam-calls.Implications/Originality/Value: The study represents a solution to addressing the practical constraints of conventional scam-call detection systems which have linguistic and adaptability limitations. The suggested framework, based on the data produced by LLM, can provide multilingual coverage, help minimize computational costs, and update regularly, with minimal costs, thus being not only operationally viable but also able to adapt to changing scam-call tactics.