Comparative Evaluation of BERT-Based Models for Korean Deepfake Text Detection with Domain Transfer Analysis

The rapid proliferation of generative artificial intelligence (AI) has led to the widespread production of AI-generated texts that are fluent and persuasive, yet potentially prone to factual distortion and reduced information reliability. To address these concerns, this study aims to develop a Korean-based deepfake text detector and to analyze its generalization performance across domains. Specifically, we collected and curated human- and AI-generated texts from two heterogeneous domains: tourist reviews of Seongsan Ilchulbong and YouTube comments related to youth employment. Based on these datasets, four training test combinations were designed, including two within-domain settings and two cross-domain settings. The detector was implemented by fine-tuning five Korean BERT-based models KoBERT, KoELECTRA, KcELECTRA, KLUE-BERT, and KLUE-RoBERTa under identical experimental conditions. Model performance was evaluated using accuracy, precision, recall, and F1-score. The experimental results indicate that all models achieved high performance in within-domain settings. However, cross-domain evaluation resulted in a 10 15% decrease in accuracy and F1-score, highlighting the strong domain dependence of deepfake text detection. Among the models, KcELECTRA exhibited greater vulnerability to domain shifts, whereas the KLUE-BERT family demonstrated relatively stable performance across domains. These findings provide a foundation for the design of surveillance and verification systems aimed at ensuring content reliability in the era of generative AI, as well as for the development of automated detection technologies for potentially harmful or misleading text.