Low Resolution Deepfake Face Detection using Multi-Scale Discrete Cosine Transform and Vision Transformer

Deepfake technology uses generative adversarial networks to produce highly authentic counterfeit images and videos, which raises a significant global security concern by disseminating fake information through online platforms, especially on social media. Researchers have developed sophisticated models for deepfake detection, demonstrating high performance on high-resolution images. However, their effectiveness significantly diminishes when applied to low-resolution images. We propose a simple framework that utilizes multi-scale discrete cosine transform and vision transformer for low-resolution deepfake face detection. This framework introduces a multi-scale frequency filter fusion approach to extract subtle frequency features in the frequency domain. Our proposed network has been tested on the FaceForensics++ and Celeb-DF datasets, and it outperforms existing models, achieving superior AUC and F1 scores.