Deep fake video detection by fusion of optimized capsule networks based on golden ratio

Deepfake videos are fake digital content produced using advanced artificial intelligence techniques and have become a major threat in recent years. Detecting these videos is critical for protecting public safety and personal privacy. In the deepfake video detection process, frame selection is performed using the golden ratio information on the face. The golden ratio is calculated using ratios that ensure the face appears aesthetically pleasing and symmetrical, and the most suitable frames are selected according to this criterion. This method improves the overall performance of the detection process and optimizes processing time. In this study, the stages of using feature extractors, capsule network-based classification networks, and their fusion were applied for deepfake video detection. First, deep network models such as VGG19, Efficient-Net B0, and Efficient-Net B4 were used to extract meaningful features from video frames. After the feature extraction process, these features were input into capsule network-based classifier networks. Capsule network architectures such as Capsule Forensics and ArCapsNet were used. Experimental results show that the proposed method provides higher accuracy and reliability compared to existing deepfake video detection methods. The study offers a new and effective approach to deepfake video detection, making a significant contribution to the literature in this field.