Synthetic Data for Scam Detection: Leveraging LLMs to Train Deep Learning Models

This paper presents a novel approach to training scam detection models using synthetic data generated by Large Language Models (LLMs). We propose single-agent and multiagent methods for data generation and train six deep learning architectures—LSTM, BiLSTM, GRU, BiGRU, CNN, and BERT—to classify conversations as scam or non-scam. Our experiments demonstrate that models trained on synthetic data achieve high accuracy on both generated test sets and real-world scam conversations. The models perform well even with limited conversation turns and when analyzing only the suspect’s messages, indicating potential for early scam detection and privacypreserving applications. Our findings highlight the efficacy of synthetic data in overcoming real-world dataset limitations for scam detection. We make the dataset and trained models publicly available to facilitate further research and development in this critical area of fraud prevention.