Leveraging Pixel Difference Feature for Deepfake Detection

Abstract

The rise of Deepfake technology poses a formidable threat to the credibility of both judicial evidence and intellectual property safeguards. Current methods lack the ability to integrate the texture information of facial features into CNNs, despite the fact that fake contents are subtle and pixel-level. Due to the fixed grid kernel structure, CNNs are limited in their ability to describe detailed fine-grained information, making it challenging to achieve accurate image detection through pixel-level fine-grained features. To mitigate this problem, we propose a Pixel Difference Convolution (PDC) to capture local intrinsic detailed patterns via aggregating both intensity and gradient information. To avoid the redundant feature computations generated by PDC and explicitly enhance the representational power of a standard convolutional kernel, we separate PDC into vertical/horizontal and diagonal parts. Furthermore, we propose an Ensemble Dilated Convolution (EDC) to explore long-range contextual dependencies and further boost performance. We introduce a novel network, Pixel Difference Convolutional Network (PDCNet), which is built with PDC and EDC to expose Deepfake by capturing faint traces of tampering hidden in portrait images. By leveraging PDC and EDC in the information propagation process, PDCNet seamlessly incorporates both local and global pixel differences. Comprehensive experiments are performed on three databases, FF++, Celeb-DF, and DFDC to confirm that our PDCNet outperforms existing approaches. Our approach achieves accuracies of 0.9634, 0.9614, and 0.8819 in FF++, Celeb-DF, and DFDC, respectively.