KoBERT와 CNN-LSTM을 활용한 보이스피싱 탐지 시스템

초록

Voice phishing has emerged as a serious social issue due to increasingly intelligent and sophisticated tactics. To effectively detect such threats, it is essential to assess both the contextual risk of conversations and the authenticity of the speaker’s voice. This paper proposes a dual-pipeline detection system that simultaneously analyzes text and audio. The text pipeline utilizes KoBERT to estimate contextual risk, while the audio pipeline employs a CNN-LSTM hybrid model to detect synthetic speech. Trained on real-world phishing cases from the Financial Supervisory Service and synthetic data generated with the so-vits-svc-fork framework, the proposed system achieved a detection accuracy of 94.9%.

키워드