Enhanced Voice Phishing Detection Using an LLM-Based Framework for Data Augmentation and Classification

Existing voice phishing detection models based on call transcripts often suffer from limited generalizability due to insufficient scenario diversity and the absence of ambiguous samples in data. To address these challenges, we propose an integrated framework that leverages GPT-4o, a large language model (LLM), to generate realistic call transcripts from actual fraud cases and to build an expert-guided phishing detection model. Using case reports from the Financial Supervisory Service (FSS) and Korean call transcripts from the KorCCVi dataset, we generate Korean phishing call transcripts that capture underrepresented fraud tactics. In addition, we generate non-phishing call transcripts that retain phishing-like linguistic patterns by removing or attenuating core fraudulent cues, thereby enabling training on ambiguous cases. The generated data are quantitatively evaluated for linguistic naturalness, scenario diversity, and detection difficulty. To assess the sample-level detection difficulty in semantic space, we introduce the metric Class Centroid Distance Variability (CCDV). We also propose the Domain Expert LLM, implemented using GPT-4o, a prompt-engineered detection model that incorporates six analytical criteria validated by a domain expert. The model not only improves detection performance but also produces structured analytical reports that enhance interpretability. In experiments, the Domain Expert LLM achieves an F1 score of 0.9686 on previously unseen and ambiguous transcripts, substantially outperforming conventional models such as RandomForest and KoBERT, which yield an average F1 score of approximately 0.70.