A Very Brief Introduction to DeepFakes
The internet is once again hit with a very cool yet scary thing, heavily related to the buzzwords “Deep Learning” and “Neural Networks”. Deepfakes is a technology not long ago developed using those two buzzwords I just mentioned, to swap an object in a video with whatever you like, but it is often used to swap faces since it would work a lot better if you’re swapping two similar things, such as swapping a face with another face. There are two applications that the public can use, which are FakeApp and faceswap. People have been using this technology for fun, and maybe to spread some kind of fake news. Everyone can go to Youtube and just search for one, and there’s a ton of them already, ranging from Mark Zuckerberg to Donald Trump as a deepfake subject. This is such a big deal, because it contributes a lot to the flow of information on the media.
But first of all, how does it work? Why is it related to AI buzzwords? Well, it turns out that it implements what is called a neural network. A neural network is basically a computing method that loosely resembles how the brain works. It’s got a node and an edge (you can treat it like some kind of a string connecting each node), and a lot of math. Deepfakes works using a type of neural network called an autoencoder. Simply put, it receives an image as an input, and then the autoencoder compresses it, then learns how to reconstruct the input image using the features that it has. It consists of two parts, which is the encoder and the decoder. The encoder is the one doing the compressing and learning the features, whereas the decoder transforms the input image by replacing the original image with the decoded input.
We input a set of images of two persons, say person 1 and 2, to the encoder, which will learn the features of both of the persons. Then, we separate the decoder into decoder 1 and 2. Decoder 1 works to redraw the input image with the features of person 1, whereas decoder 2 works with the features of person 2. If we input an image of the first person, then decode using decoder 2, we will get an image of person 1 with the facial features of person 2, while still maintaining the expressions of the original input.
Now that we have a very simplified idea of deepfakes, what is it for then? Is it only for trolling people, for fun? It turns out, not really. Some people does it only for fun, to see their favorite actors in the films they never played, to make fun of people, and other harmless stuff (probably). But it also has the potential to be dangerous. The potential can be seen on Youtube, where a deepfake video of a character’s face from TV series getting swapped with Donald Trump’s face, and I’d say if the title doesn’t say deepfakes and I’m a few years younger, I’d totally believe it. Deepfakes are getting really realistic and it’s getting harder for people to recognize it, so it is pretty believable and will be even more believable.
But is deepfakes making believable news more scarce than it already is? You don’t really need deepfakes to convince fake news to people. And this hype is pretty similar to that Photoshop one from back then, how Photoshop can transform an image to something else. What really matters is how we handle them. We just have to be skeptical on the internet. If we don’t really trust a news source, then it simply is not worth to share.
