Most existing deepfake detection methods often fail to maintain their performance when confronting new test domains. To address this issue, we propose a generalizable deepfake detection system to implement style diversification by alternately learning the domain generalization (DG)-based detector and the stylized fake face synthesizer (SFFS). For the DG-based detector, we first adopt instance normalization- and batch normalization-based structures to extract the local and global image statistics as the style and content features, which are then leveraged to obtain the more diverse feature space. Subsequently, contrastive learning is used to emphasize common style features while suppressing domain-specific ones, and adversarial learning is performed to obtain the domain-invariant features. These optimized features help the DG-based detector to learn generalized classification features and also encourage the SFFS to simulate possibly unseen domain data. In return, the samples generated by the SFFS would contribute to the detector's learning of more generalized features from augmented training data. Such a joint learning and training process enhances both the detector's and the synthesizer's feature representation capability for generalizable deepfake detection. Experimental results demonstrate that our method outperforms the state-of-the-art competitors not only in intra-domain tests but particularly in cross-domain tests.

Most existing Deepfake face forgery detection algorithms suffer from the insufficient generalization performance despite that their intra-dataset detection performance is fairly good. This is because these methods mainly rely on local features that are prone to overfitting, which leads to unsatisfactory cross-dataset detection performance. In order to solve this problem, a face forgery detection method based on multi-scale spatiotemporal features and tampering probability is proposed, which helps to maintain good performance for cross-dataset testing, cross-forgery testing as well as video compression by detecting the inevitable temporal inconsistency between continuous frames in deepfake videos. The proposed detection method consists of three modules: a multi-scale spatiotemporal feature extraction module is employed to reveal the discontinuous traces of fake videos in the temporal domain, a three-dimension dual-attention module is designed to adaptively compute the correlation between multi-scale spatiotemporal features, and an auxiliary supervision module is used to predict the tampering probabilities of randomly selected pixels to form a supervision mask. Then, the proposed algorithm is compared with the baseline algorithm and the latest relevant works on large-scale public standard databases such as FF++, DFD, DFDC and CDF. Experimental results have show that the proposed algorithm has the best overall performance for cross-dataset testing and video compression, and has the above-average performance for cross-forgery testing. Meanwhile, it maintains good average performance for all intra-dataset testing. All the experiments demonstrate the effectiveness of the proposed algorithm.