What Is A Deepfake?


    Most people have heard of, or at least seen (perhaps unwittingly), a deepfake. Some may not know the term “deepfake” but are aware of “those CG face videos.” Deepfakes are an AI-based technology that has been around for a few years now and have only been growing. Despite their relative popularity, though, I don’t think a lot of people know exactly what they are, much less how they work. So, allow me to explain to the best of my understanding (I’m no expert).

    First off, there’s a variety of software available to create deepfakes—there’s no single “DeepFake” AI or program. However, all the programs available perform the same task in essentially the same manner, which I will explain shortly. Not only are there a variety of software available to create deepfakes, but there are also a variety of deepfakes. What do I mean by this? Well, everyone familiar with deepfakes knows that they are used to replace a person’s face. But did you know that there are also voice deepfakes? That’s right, we’re at the point where we can clone a person’s voice, and in some cases it’s pretty convincing.

    Deepfakes, especially of the face variety, have become well known in recent years, though it’s not just because they allow you to stick anyone into any given video. It’s because they make the technology to do it so accessible.

    Let’s look at Hollywood for a moment. They’ve been doing face replacements since 1993’s Jurassic Park. (Check out the second clip of this video and see if you can notice the face replacement!). Of course, the method at the time was highly time consuming, but it’s something that has been improved upon and nearly mastered by visual effects artists who are doing it all manually. Deepfakes now give this ability to anybody with access to a computer, and they largely automate the process. You, the user, will have to feed hundreds of images of the face you want into the software. The AI will do the rest, scanning each image and figuring out where all the facial features are and how to reconstruct them in 3D space. Once that’s done, you can apply the face to any video of any person you want, though it will likely need some manual tweaks from a professional to look truly convincing.

    So how does this all work? I’m going to get a bit technical, so bear with me. Deepfakes use a type of neural network called an autoencoder. An autoencoder is made up of an encoder, which is what reduces the images of the face into something readable by the computer; and a decoder, which reconstructs the image from the computer’s “language” to something that we humans can see and understand on the screen. To add onto this, the best deepfakes also rely on something called a generative adversarial network (GAN), which is built in as part of the decoder. The GAN consists of the decoder and what’s called a discriminator. The discriminator’s job is to look at the images that the decoder outputs and try to determine if they’re real images or deepfakes. If the discriminator decides they’re fake, then the images are generated again and again until they are flawless in the eyes of the discriminator. This is what makes a good deepfake good. They’re nearly impossible to tell from the real thing because they’ve already been through intense scrutinization during their creation process.

    If you’d like to learn some tricks on how to spot a deepfake, check out this video by the CG artists at Corridor: https://www.youtube.com/watch?v=5PbqGAmUnVU.


Don't be fooled!
Davis Allred

Comments

Post a Comment

Popular Posts

Questions And Lessons From "I, Robot" (2004)

Image
It's the year 2035, and the society now has the help of robots. These robots follow three laws integrated into their system; 1) They cannot harm a human being or, through inaction, allow a human being to come to harm. 2)They must do whatever they're told by a human being as long as such orders don't conflict with law one. 3) They have to defend themselves as long as such defense doesn't conflict with laws one or two. One day, the writer of the three laws, Alfred Lanning, apparently jumps out of a window at U.S. Robotics. The majority of the Chicago Police Department believe he committed suicide, but Detective Del Spooner, who hates robots, thinks he was murdered, and the number one suspect is a robot who calls himself Sonny. However, if it was Sonny, then it means he would've had to have broken the three laws. With the help of Dr. Susan Calvin, Spooner must now discover the truth before it's too late. — Synopsis by  Ridley Lavine Pick something from I, Robot th...
Read more

[Essay] The Real Problems of Deepfakes

Image
The Real Problems of Deepfakes             Deepfakes—defined by machine-learning experts Yisroel Mirsky and Wenke Lee as “content generated by an artificial intelligence, that is authentic in the eyes of a human being,” particularly “the generation and manipulation of human imagery”—have become increasingly popular in recent years. They rely on the rapidly developing technology of “deep learning.” Deep learning is, essentially, the process of a computer teaching itself new information through trial and error. In the case of deepfakes, the computer is teaching itself how human faces look and how they can be manipulated. Deepfakes have “lower[ed] the technological barriers required to create high-quality manipulations” (Lutz & Bassett). Their ease of access lends well to any individual or group seeking to create a time- and cost-efficient video for what is often a malicious purpose. Because of these malicious uses, it would be s...
Read more