Deepfakes have become a social problem due to their various negative uses, including political manipulation, privacy violations, and economic scams. To
address these issues, it is essential to develop accurate deepfake detection models. However, if th...
Deepfakes have become a social problem due to their various negative uses, including political manipulation, privacy violations, and economic scams. To
address these issues, it is essential to develop accurate deepfake detection models. However, if the dataset used to validate a deepfake detection model
contains low-quality data, the performance of the detection model may be overstated. In this study, we use FaceForensics++ and Celeb-DF datasets, which
are commonly used in deep fake research, to analyze the quality of various images using quality evaluation metrics and generation model evaluation
metrics. Based on this, we investigate the correlation between dataset quality and the performance of Face X-ray, UCF, F3Net, and Meso4Net deep fake detection models.
Our results show that the quality of the dataset measured by PSNR, MS-SSIM, BRISQUE, and Pose Error among various image quality evaluation techniques correlates with the performance of the detection model. Instead of the existing random data selection, the evaluation data selection that considers the quality
of the dataset shows up to 4.21% lower confidence in video frame selection and up to 26.73% lower confidence in frame-by-frame thrshold verification than
random selection, confirming that strict deep fake detection model verification is possible. These findings are expected to contribute to the correct validation and performance improvement of deep fake detection technology.