Deepfake detection with quantum transfer learning and class attention architecture

Today's widespread use of the internet, along with the increase in the production of digital content, has brought with it major issues of information security and manipulation. Deepfake detection is a process where various deep learning and artificial intelligence techniques are used to identify fake elements in artificially generated videos. In this thesis, an innovative model for fake video detection is developed using a combination of Quantum Transfer Learning (QTL) and Vision Transformer (ViT) architectures. In the study, training was performed on the DFDC (Deepfake Detection Challenge) dataset and a system that can detect spatial and temporal inconsistencies was created by combining QTL and ViT technologies. In addition to the existing preprocessing studies in the literature on the dataset, a unique data processing method suitable for the requirements of deep learning models has been developed. The quantum computing advantages of QTL and the global feature extraction capabilities of ViT in visual data are combined. The results show that the proposed method achieves remarkable success in deepfake detection with an accuracy of 90% and an ROC AUC of 0.94. The success of the model is compared with other studies on the DFDC dataset and it is found that it offers a prominent solution in terms of resource efficiency. The findings reveal that the proposed QTL-ViT-based system provides a strong foundation for deepfake detection and makes a significant contribution to the literature. For future work, it is recommended to test the model on real quantum devices and apply it on larger datasets.