Rethinking Perturbation-Based Training-Free Method for Deepfake Face Detection

Training-free deepfake image detection aims to discern whether inputs are authentic or synthetic by directly conduct evaluation on testing samples. Prior approaches predominantly measure the similarity between original images and perturbation-generated versions. Leveraging pre-trained foundation models like DINOv2, these methods typically deliver remarkable detection performance on natural images. However, their effectiveness diminishes significantly when applied to deepfake face detection, particularly for faces of varying resolutions. Additionally, the computational demands are high due to the complexity of these foundational models, hindering the application in real-world scenarios. To overcome these challenges, we elaborate a simple yet effective Upsampling-Perturbation-Downsampling method for training-free deepfake face detection. This approach enhances both the robustness against diverse input resolutions and the efficiency of detection process. Extensive experiments on our augmented DeepFakeFaceForensics dataset demonstrate that our approach significantly outperforms state-of-the-art methods.