Exposing Deepfake Face Forgeries with Guided Residuals

For Deepfake detection, residual-based features can preserve tampering traces and suppress irrelevant image content. However, inappropriate residual prediction brings side effects on detection accuracy. Meanwhile, residual-domain are easily affected by some operations such as lossy compression. Most existing works exploit either spatial-domain or features, which fed into the backbone network for feature learning. Actually, both types of mutually correlated. In this work, we propose an adaptive fusion based guided residuals (AdapGRnet), fuses in a reinforcing way, detection. Specifically, present fine-grained manipulation trace extractor (MTE), is key module AdapGRnet. Compared with prediction-based residuals, MTE avoid potential bias caused prediction. Moreover, attention mechanism (AFM) designed to selectively emphasize channel maps adaptively allocate weights two streams. Experimental results show that AdapGRnet achieves better accuracies than state-of-the-art four public fake face datasets including HFF, FaceForensics++, DFDC CelebDF. Especially, accuracy up 96.52% HFF-JP60 dataset, improves about 5.50%. That is, robustness works.