In the bachelor's thesis we present techniques and procedures for generating deepfake videos. These are videos that were subjected to manipulations with deeplearning techniques. In our thesis we specialized on the topic of deepfakes, where the facial area was manipulated. They are usually made with the help of special generative adversarial networks - GAN.
Such videos represent a major problem in the spread of fake news, political propaganda, destruction of individual's public image, production of pornographic content, extortion, etc.
In the thesis we describe in detail different types of fake videos from deepfake database FaceForensics++. We also present our own method for potential creation of a subset of the mentioned database using the latest generative diffusion models. These models progressively generate images (or videos) from latent noise. We use a specific open source diffusion model called stable diffusion, that is used for generating images from text and image prompts.
We describe multiple techniques, that we used and tested in our experiment and analyze their quality and success. We also comment on the utility and danger posed by fake videos generated by diffusion models.
|
