Development of deep learning-based hybrid architectures for deepfake detection

Deep learning technologies have become one of the applications that are becoming increasingly popular. One of these is applications called deepfake. Basically, deepfake is a new technology that allows images, sounds and videos to be artificially created or modified to create misleading content that looks like real. Although it has positive effects in areas such as education, entertainment and research, it is also misused such as fake news, propaganda, reputation damage and blackmail. For this reason, various deep learning-based techniques have been proposed to detect deepfakes in the fight against forgery. In this study, which was carried out to contribute to this fight, 3 different deepfake algorithms using the Xception and EfficientNet methods were used. In addition, a new deepfake detection method was presented by combining it with Choquet Fuzzy Integral. Our proposed method takes 3 different algorithms that are good in their own fields and gathers the accuracy values and fuzzy membership values that each algorithm can work on its own under a single roof using Choquet Fuzzy Integral, and thus, it has significantly increased the accuracy rate of deepfake detection by signing a study that has not been done before in the field of deepfake. Experimental results obtained using FaceForensics ++, DFDC, Celeb-DF-v2 and DeepFake-TIMIT-HQ dataset show that the proposed approach based on Choquet fuzzy integral technique outperforms single classifiers for deepfake classification and reaches the highest AUC of 99.8%. In this method, a more effective result can be obtained by using other effective models. In the second proposed method, Image Transformer (ViT), which is the latest technology used in image processing technologies and is up-to-date, is used. In the method, together with ViT, the MesoNet method, which is a pioneer in deepfake detection methods in the literature and is very lightweight in terms of architecture, is used. Although the proposed method was trained on a small amount of dataset, it showed an acceptable performance with an accuracy rate of 92.6%. In addition, a real-time deep detection application is designed for users to analyze their images. We believe that the proposed method will inspire researchers and will be further developed.