This paper aims to develop a lightweight and computation efficient model and checking its strength against various growing video and identity manipulation types. The proposed model was trained on the video sequences obtained from the five different manipulation types of the FaceForensics++ dataset under the C23 compression setting. An XceptionNet was applied on the frames obtained from FaceForensics++ videos for feature extractionsince previous research have shown its strength to be the highest among all CNN models for deepfake detection. Then, a GRU captured temporal dependencies across frame sequences, enabling the model to detect both visual artifacts and temporal inconsistencies in deepfake videos. The findings across manipulation types reveal the strength of the proposed model and learning where the model seems to lack.
Shaheen Afroz (Sun,) studied this question.