Learning to Detect Deepfakes via Adaptive Attention and Constrained Difference

Abstract

Since facial forgery techniques have made remarkable progress, the area of forgery detection attracts a significant amount of attention due to security concerns. Existing methods attempt to utilize convolutional neural networks (CNNs) to mine discriminative clues for forgery detection. However, most of these coarse-grained and vanilla methods struggle to extract subtle and multiscale clues in forgery detection. To address such problems, we propose a well-designed deep learning framework, named SCA-Net, to exploit subtle, multiscale and multiview clues. Specifically, our framework consists of a skipped channel attention module (SCM), a constrained difference module (CDM) and an adaptive attention module (AAM). First, the skipped channel attention module is used as the backbone to extract sufficient different information, including low-level and high-level features. Then, the constrained difference module captures manipulation clues from the input image based on constrained characteristics. Finally, the adaptive attention module captures multiscale features represented by facial forgery. Moreover, we introduce a combined loss to address the learning difficulty of our framework. The experimental results demonstrate that the proposed model has great detection performance compared with other face forgery detection methods in most cases.