Deep Fake Video Detection Based on Enhanced Capsule Network with Golden Ratio

The fact that deepfake video content can be produced in a very realistic way causes serious social problems. Therefore, it is important to detect such fake content. In most of the studies in the literature, a certain number of random frames of the videos are selected and used in the detection processes. However, all frames are not equally important. In this study, which frames of the videos will be used are determined with the help of the golden ratio information on the face and classification is made with the capsule network. Frames selected using the golden ratio information constitute examples showing the feature differences that will enable real/fake distinction for detection tasks. The use of these examples in the training and testing stages of the used capsule network model greatly increases the detection performance. Experimental results were obtained for the Celeb-DF and DFDC-P databases. Better results were obtained in both Celeb-DF and DFDC-P databases compared to studies using the same capsule network.