FeatureTransfer: Unsupervised Domain Adaptation for Cross-Domain Deepfake Detection

Figure 1

The pipeline of the proposed method FeatureTransfer. In the preprocessing stage, we obtain the face images of the video frame from the source and target domain and then feed them into CNN to extract the feature vectors. In the unsupervised domain adaptation stage, the BP-DANN consists of a feature extractor (green), a label classifier (blue), and a domain discriminator (red). The gradient reversal layer connects and to realize unsupervised domain adaptation, and it multiplies the gradient by a certain negative constant during the backpropagation-based training.