Face Forgery Detection with Long-Range Noise Features and Multilevel Frequency-Aware Clues

Figure 7

The visualization includes ground-truth, Grad-CAM visualizations of feature maps learned by the baseline models (EfficientNetB4 and Swin-B), and our framework, corresponding to different columns. We separately display the Grad-CAM visualizations of real and four different forgery methods in each row for scenarios c23 and c40. Note that the models are trained on different methods under two compression rates in FF++.