Imagine the scene: You’re watching a live television broadcast of your country’s leader announcing that a hostile alien force has invaded Earth and is destroying everything in their path. They urge you to flee the city immediately, and before you know it the highways are jammed with panicked people and the populous is gripped with mass hysteria. It’s not until hours later, after the looting and rioting has subdued that word reaches everyone that you’ve all been duped – the broadcast was a fake.
Welcome to the future of deepfakes.
What are deepfakes?
Simply put, they use deep learning systems to create fake content, primarily videos.
The coin deepfake was coined on Reddit in 2017 when a basic yet fully functional version of the application was introduced. The system uses machine learning processes to combine a giant library of facial images with a pre-made video to put someone’s face on someone else’s body. These images are used to teach the software how the subject’s face works – for example, how their eyes wrinkle when they smile, or how their mouth opens when pronouncing certain words – and are used in a machine learning system that compares the original video’s face with the image and tries to copy it over seamlessly.
Probably the easiest way to understand is to see it in action:
It’s Shapchat’s faceswap filter taken to a whole new level, and it’s going to change how we approach the believability of everything we see.
Technology pushed to the limit
At this stage, creating a deepfake is a complicated and computationally-heavy process. The software currently works on a limited range of CPUs and needs an expensive graphics card to do the heavy lifting. To create a convincing deepfake you need hundreds or even thousands of high-resolution images of the subject, which is why most example deepfakes star famous actors, models or politicians. Even a video of a few seconds requires days of rendering.
The results are convincing, but not 100% foolproof – yet. If the user hasn’t provided enough target images, the resulting video can have a face that glitches like a faulty hologram.
However, this is a matter of time. Just like how early computers were slow and cumbersome and required a lot of specialised knowledge to get rudimentary results out of, technology like deepfakes will eventually become fast and easy enough to create in a live environment. This is where things will get really interesting and problematic.
Samsung jumps on the deepfake AI train
Interestingly enough, Samsung have recently unveiled new technology that lets anyone generate a deepfake from one single image. In the proof of concept video, a user records their facial movements and expressions, that will be used as the basis of the target video. The system finds the face in a sample image and builds a 3D model based on it, which is then animated using the same movements and expressions. Ever wanted to see the Mona Lisa move and talk? Go to the end of this video:
These technologies pose massive ethical implications. While we’re used to seeing simulated actors in movies and television shows, the danger they pose to deceiving the public as news events cannot be understated. It’s already hard enough to know what is real on the internet, and deepfakes have taken the discourse on a strange new path.
Already we are seeing digitally altered videos of politicians and celebrities to make them appear under the influence, nervous or even saying the complete opposite of what they actually said. What we see every day is filtered through other people on social media, and if we can’t trust that content to be authentic, then we can lose trust in those who are meant to lead and guide us. This could have major damaging repurcusions on the world we live in, as bad faith actors could use this technology to change how we percieve the world around us.
Until then we'll enjoy the videos of Nicolas Cage's face being deepfaked onto every other actor imaginable.