Current facial image manipulation techniques have caused public concerns while achieving impressive quality. However, these techniques are mostly bound to a single frame for synthesized videos and pay little attention to the most discriminatory temporal frequency artifacts between various frames. Detecting deepfake videos using temporal modeling still poses a challenge. To address this issue, we present a novel deepfake video detection framework in this paper that consists of two levels: temporal modeling and coherence analysis. At the first level, to fully capture temporal coherence over the entire video, we devise an efficient temporal facial pattern (TFP) mechanism that explores the color variations of forgery-sensitive facial areas by providing global and local-successive temporal views. The second level presents a temporal coherence analyzing network (TCAN), which consists of novel global temporal self-attention characteristics, high-resolution fine and low-resolution coarse feature extraction, and aggregation mechanisms, with the aims of long-range relationship modeling from a local-successive temporal perspective within a TFP and capturing the vital dynamic incoherence for robust detection. Thorough experiments on large-scale datasets, including FaceForensics++, DeepFakeDetection, DeepFake Detection Challenge, CelebDF-V2, and DeeperForensics, reveal that our paradigm surpasses current approaches and stays effective when detecting unseen sorts of deepfake videos.

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.