Multi-attentional Deepfake Detection

Detection of Deepfake Videos Using Long Distance Attention

“…As aforementioned, there are not only one forgery property space, in fact, we adopt 12 such attention module to produce different attention maps of different latent space, and linearly combined into m final attention maps for a robust and efficient reason [22], more discussion is given in VI-B.…”

“…And to recalibrate the importance between regions, the attention maps generated by a single frame are adopted to the feature maps of the backbone network. As textural features exist in shallow features [22], we make the attention works with the first several layers of the backbone. More specifically, the input image I which is used for the backbone and the spatial attention model is reshaped to the resolution 398 × 398 and 224 × 224 respectively.…”

“…All of the methods mentioned above take the deepfake detection as a vanilla binary classification problem. However, as the counterfeits become more and more realistic, the differences between real and fake ones will become more and more subtle and local which making such global feature-based vanilla solutions work not well [22].…”

“…Fine-grained classification aims to classify very similar categories, such as species of the bird, models of the car, and types of the aircraft [23]. Since the deepfake detection and fine-grained classification share the same spirit, that learning subtle and discriminative features, in [22], the deepfake detection is reformulated as a fine-grained classification task. And a convolutional attention module with 1 × 1 is adopted to make a network focus on the subtle but critical regions.…”

“…But it is not directly for the convolutional attention mechanism, especially when the kernel is small. The global pooling may be a choice to assembling global information, however, the weak forgery clues will be averaged by this operation, and resulting in a loss of distinguishability [22].…”