FeatureTransfer: Unsupervised Domain Adaptation for Cross-Domain Deepfake Detection

Recently, various Deepfake detection methods have been proposed, and most of them are based on convolutional neural networks (CNNs). These suffer from overfitting the source dataset do not perform well cross-domain datasets which different distributions dataset. To address these limitations, a new method named FeatureTransfer is proposed in this paper, two-stage combining with transfer learning. Firstly, The CNN model pretrained third-party large-scale can be used to extract more transferable feature vectors videos target domains. Secondly, fed into domain-adversarial network backpropagation (BP-DANN) for unsupervised domain adaptive training, where real or fake labels, while unlabelled. experimental results indicate that effectively solve problem greatly improve performance cross-dataset evaluation.