합성 데이터와 XAI를 활용한 온디바이스 텍스트 기반 보이스피싱 탐지 시스템

초록

Voice phishing scams are becoming more sophisticated, while data sparsity and black-box models hinder user trust in automated detection. we propose an on-device, text-based detection system that combines LLM-based synthetic data augmentation and explainable AI, where an LLM generates synthetic examples of new scam patterns, an LSTM-Attention detector highlights salient tokens, and a lightweight LLM explainer produces rationales, and in experiments LSTM-Attention and KoBERT achieved F1-scores above 99.5% on calls longer than 30 seconds while the Gemma 2 2B explainer reached BLEU-4 28.69 and ROUGE-1/2/L 62.3/44.41/54.76, highlighting trade-offs between accuracy, latency, and model size for on-device deployment.

키워드