What Is a Deepfake?

In the opening session of his 2020 introductory course on deep learning, Alexander Amini, a PhD student at the Massachusetts Institute of Technology (MIT), had invited a famous guest: former US President Barack Obama.

- BEN DICKSON

“Deep learning is revolutionizing so many fields, from robotics to medicine and everything in between,” said Obama, who joined the class by video conference.

After speaking a bit more on the virtues of artificial intelligence, “Obama” made an important revelation: “In fact, this entire speech and video are not real and were created using deep learning and artificial intelligence.”

Amini’s video was, in fact, a deepfake—an AI-doctored video in which the facial movements of an actor are transferred to that of a target. Since first appearing in 2018, deepfake technology has evolved from hobbyist experimentation to an effective and dangerous tool. Deepfakes have been used against celebrities and politicians and have become a threat to the very fabric of truth.

HOW DO DEEPFAKES WORK?

Deepfake applications work in various ways. Some, like the Obama deepfake, transfer the facial movements of an actor to a target video. Others map the face of a target person onto other videos.

Like most contemporary AI-based applications, deepfakes use deep neural networks (that’s where the “deep” in deepfake comes from), a type of AI algorithm that is especially good at finding patterns and correlations in large sets of data. Neural networks have proven to be especially good at computer vision, the branch of computer science and AI that handles visual data.

Deepfakes uses a special type of neural-network structure called an “autoencoder.” Autoencoders are composed of two parts: an encoder, which compresses an image into a small amount of data; and a decoder, which decompresses the compressed data back into the original image. The mechanism is similar to those of image and video codecs such as JPEG and MPEG.