As deepfake technology advances, distinguishing deepfakes from real content has become increasingly challenging, leading to the necessity for effective deepfake detection methods. Traditional image-based detection can identify artifacts from trained d...
As deepfake technology advances, distinguishing deepfakes from real content has become increasingly challenging, leading to the necessity for effective deepfake detection methods. Traditional image-based detection can identify artifacts from trained deepfakes, but often fail to generalize to unseen deepfakes. To enhance generalization, this study analyzes the impact of stylized images on deepfake detection. We generate stylized images with two distinct styles utilizing CLIPstyler and train the Xception model on these images. Comparative experiments demonstrate that using stylized images, generated by style transfer, for training can improve the generalization capability of image-based detector. This work presents new directions in the field of deepfake detection.*