Existing deepfake detectors rely on single images and superficial artifacts, limiting their ability to detect realistic forgeries and explain identity-specific differences. To address this, we propose a reference image-guided framework for explainable deepfake detection. Our approach leverages reference images of the same individual to highlight semantic differences across facial regions and determine authenticity. We also introduce XFace, a multimodal dataset containing manipulated images, reference images, and paired explanations, along with a Reference Feature Scoring module. Experimental results across diverse settings demonstrate that our method outperforms existing approaches in both accuracy and interpretability.

Heejae Jo is a Master’s student in Artificial Intelligence at Chung-Ang University, advised by Prof. Jongwon Choi. Her research focuses on multimodal deepfake detection, explainable AI, and vision-language models. She has developed novel frameworks such as reference image-guided explainable detection and constructed the XFace multimodal dataset. Her broader academic interests include trustworthy AI, adversarial robustness, and human-centered AI systems.