In 2017, the ‘Synthesizing Obama’ program modified footage of the former President’s speeches to a different audio track. The techniques used to adapt video of a person’s face is considered to be one of the earliest ‘deepfakes’, though the techniques used weren’t all AI-based.
Though this was done for academic research, deepfake videos are often used for fake news or for bogus celebrity sex tapes. High-quality deepfakes deploy AI to train on hours of genuine videos of the target. This has been made easier with the rise of open-source tools like Faceswap. Most of these use an ‘autoencoder’. This reduces images to a lower dimensional latent space, then derives information like facial features, body type, and posture. These can be coupled with generative adversarial networks (GANs), which use algorithms to see if a frame is generated, then modify it. This can make deepfake videos look extremely realistic.