RLGC: Reconstruction Learning Fusing Gradient and Content Features for Efficient Deepfake Detection

Kaiwen Xu, Xiyuan Hu, Xiaokang Zhou, Xiaolong Xu, Lianyong Qi, Chen Chen

Abstract: Current deepfake detection methods, which utilize noise features, localized textures, or frequency statistics, may perform well in special domains or forgery methods. But the generalization performance of these methods is often unsatisfactory because of the ignorance of mining intrinsic facial features. To address this problem, we re-evaluated the fusion of image gradient features in neural networks and delved deeper into the intrinsic structure of input images. Consequently, we propose a reconstruction-classification network that initially learns face content and gradient separately from a reconstruction perspective and then detects forged faces by fusing them together. This paper introduces three well-designed components: 1) a dual-branch feature extraction module to excite distributional inconsistencies between real and forged faces; 2) a content-gradient feature fusion module to investigate the relationship between face content and image gradient; 3) a reconstruction disparity based Bi-Directional attention module that guides the model in efficiently categorizing the fused features. Extensive experiments on large-scale benchmark datasets demonstrate that our method significantly enhances performance, especially for generalization ability, compared to state-of-the-art methods.

External IDs:dblp:journals/tce/XuHZXQC24