It's funny how a very simple DL architecture any beginner could prepare ended up in a prohibition act - essentially training two convolutional autoencoders, one on the face one wants to replace (extracting dataset by e.g. tracking face with OpenCV in target video), the second on many pictures of the face one wants to replace (e.g. scraping from Internet or known videos), then essentially gluing together encoder from one and decoder from the other, assuming latent variables match, and suddenly a fake video is done.
IMO if you really want to control deep fakes, you should use a blockchain and track all processing steps, starting from image/video acquisition. I don't want to give anyone ideas, but they will do it anyway.
IMO if you really want to control deep fakes, you should use a blockchain and track all processing steps, starting from image/video acquisition. I don't want to give anyone ideas, but they will do it anyway.