Deepfake Video Detection Based on Image Source Anomaly
Yufei Wang, Guangjun Liao · 2024
In order to detect the deepfake videos, most of the effective detection approaches need huge number of samples for training, including the real and fake samples. However, the fake samples are not easy to obtain. To find the solution, this paper proposes a deepfake video detection method based on image source anomaly, which only needs the real samples for training. The proposed method uses a neural network to extract features from the face region and its eight neighbor regions, and then uses another neural network to compare the similarity between the features from face region and each one of its neighbor regions. Finally, the average similarity score is utilized as the measure to detect deepfake video. The experimental results show that the proposed method has good detection performance, which achieved the HTER of 2.75% in the DFD dataset and 2.25% in the FF++ dataset, while it needs less samples for training. Moreover, the proposed method also has good cross-datasets performance. It had the HTER of 1.22% when training in FF++ dataset and testing in DFD dataset, which indicates that it can be used in many practical scenarios.