Enhancing Generalization of Korean Deepfake Text Detection through Multi-Feature Fusion across Domain and Model Variations

Existing BERT-based deepfake text detection methods perform well when the training and test data domains are similar; however, their performance degrades significantly under domain shift. To address this limitation, this study proposes a fusion-feature-based deepfake text detection model designed to improve generalization performance in unseen environments.<br/> The proposed model integrates three complementary feature types: semantic features derived from BERT embeddings, token-level predictability features based on generation probability and rank statistics from language models, and surface-level stylometric features capturing structural characteristics of text. These heterogeneous features are combined into a unified fusion feature vector to enhance robustness against variations in both domain and text generation mechanisms.<br/> To evaluate generalization performance, two experimental settings are designed. First, a cross-domain evaluation is conducted by training and testing the model on different domains, specifically tourism review texts and youth employment related YouTube comments. Second, to examine the impact of generative model evolution, detection performance is compared using deepfake texts generated by GPT-2 and GPT-5.<br/> The experimental results show that the proposed fusion feature based detector maintains relatively stable performance under domain mismatch conditions, outperforming single-feature-based detectors. Furthermore, the fusion approach consistently demonstrates superior performance across different generative model settings, indicating its robustness to advancements in text generation models. These findings suggest that the proposed approach can effectively enhance the reliability of deepfake text detection in real-world applications, particularly in environments where domain variability and evolving generation models pose significant challenges.