Görü dönüştürücü temelli derin sahte tespiti yöntemleri

Deepfakes are synthetic media products generated by editing or swapping identities in videos or audio using artificial intelligence. Over the past few years, significant progress in deep learning has led to the development of advanced techniques to create deepfake images and videos. Although deepfake technology has legitimate uses in entertainment and education, it can be hazardous when misused. These tools, readily available to anyone, have the potential to manipulate societies, fuel hatred and hostility, violate privacy, and falsely accuse individuals or organizations of crimes. The inability of human perception to differentiate between real and fake content has spurred research into more sophisticated deepfake detection methods. Vision transformers have been increasingly used by state-of-the-art models for deepfake detection in recent studies. We propose a novel deepfake detection model, the Combinational Cross-Attention Vision Transformer (CCAViT), which integrates convolutional neural networks with Vision Transformers. CCAViT also utilizes the power of the cross-attention mechanism with combinational connections of visual transformers. This thesis aims to set a new benchmark in deepfake detection, offering robust performance against a wide range of manipulation techniques, leveraging the complementary strengths of convolutional neural networks and visual transformers. We developed an improved face detection technique for use in the data preprocessing stage. The challenges related to various stages of the deepfake detection process are explored to ensure more streamlined workflows. This thesis also provides guidance at every step in the implementation of deepfake detection methods, highlighting key focus areas to enhance detection performance.