Rozpoznatelnost deepfake videí

The development of generative artificial intelligence has brought not only new possibilities in audiovisual content synthesis but also increased risks associated with the manipulation of digital reality. This bachelor's thesis focuses on the detection of deepfake videos, highly realistic audiovisual recordings generated using artificial intelligence to mimic real people or events. The primary goal is to compare the effectiveness of two deep learning models – MesoNet and XceptionNet – in detecting deepfake content, emphasizing the evaluation of accuracy, sensitivity, and robustness against video compression. The experimental results indicate that the XceptionNet model shows higher sensitivity in detecting manipulated content, while the MesoNet model achieves more stable accuracy when recognizing real videos. This thesis provides an overview of current approaches to deepfake detection, identifies key challenges, and suggests potential directions for further development of detection methods.