Exclusive online! The American president declares war on China in a live video link!
You see the screaming headline and, naturally, you click play at once. On the screen, in a dramatic tone, the leader of the world's greatest power really does say that World War III has begun. Unreal! What now?! Within no time everyone online is writing about what they've seen, and the shock turns into an instinct for self-preservation…
Deepfake. Forgive me, but we still don't have a Bulgarian word for the technology. Either way, what matters more is knowing what it can do and why it can be dangerous. It is now entirely possible to create a video in which it looks utterly convincing that someone is saying things chosen by the author of the clip.
Here's an example:
That video appeared three years ago, and the technology has improved quite seriously since then.
Today you can easily “put words in the mouth” of anyone you like, whatever words you choose. And if the job is done to standard, the forgery becomes practically indistinguishable from the truth.
In Deep Water
There's no doubt about it: our lives already depend on artificial intelligence. And on all the technologies hiding behind that almost mystical and, at the same time, worn-out term.
Today's world is a powder keg, only a digital one. Any word can become a fuse, especially if it's spoken by someone important enough. Even the owners of the global social networks can't stop this with their algorithms, and sometimes, it's only logical to suspect, they might provoke it themselves.
“Of course he said it, I saw him say it!”
This is the next, more dangerous version of “It says so on the internet.” Never mind that the internet doesn't “say” anything; what matters is what kind of site hosts a given claim, what kind of author wrote it and what sources they used.
Yes, all sorts of fake news keeps getting shared online, absurd quotes obviously never uttered by the person they're attributed to. But honestly, if you saw the president of a great power announce the start of the next world war on video, wouldn't that be a good deal more convincing?
By Order of the Queen
By tradition, on December 15 every year the British queen addresses her subjects, much like the New Year's speeches Bulgarian presidents customarily give. At the end of 2020, however, for the first time there were not one but two speeches… by the same queen. Sort of. Because Channel 4 produced a themed alternative address with entirely made-up words and a deepfake Elizabeth II in the lead role:
“So much of our world today comes to us through these phone screens. Which brings me back to the question of trust. Is what we see and hear always what it seems?”
So asked the virtual queen. The idea, once again, was to draw attention to the danger that comes with the advance of this technology. Rather impressive, actually… Garnished at the end with a healthy dose of good old British humor:
Did that go well? Can I have a cup of tea now?
Fauxtography
We all remember well how, slowly but irreversibly, image-editing technologies, lumped together as “Photoshop,” changed our attitude toward pictures forever. All that's left for us now is to make not-so-kind jokes online about some pop-folk singer who overdid the retouching so badly that she ended up with fewer fingers, say, or without a belly button. That's why, for a long time now, whenever we see an image, we keep in the back of our minds that it doesn't always show the truth. So much so that we can't even recognize some “over-Photoshopped” celebrity when we see them in person.
With video, though, things were different. Yes, many film effects are created entirely artificially, but it never seemed possible for the actor to be recreated with technology. There are already games that offer a more realistic view from an airplane window than the real thing, but human characters still seemed quite hard to reproduce realistically.
Or at least they did, before deepfake arrived.
God Is High, the Fake Is Deep
Perhaps a rough translation of the technology's name would be something like “a deep, in-depth forgery.” People first started talking about it in 2018. Although we always associate it with video, it's really a general principle that uses artificial intelligence to process video, audio or photos so that they resemble the real person.
In fact, it was deepfake that made digital humans possible. Its main purpose is to create visual products that cannot be told apart from a human being. It's especially successful with video, because resolution there is generally lower, details are coarser and small imperfections are easier to “smudge over.” With earlier technologies that wasn't much of an advantage, because they failed at the bigger challenge—convincingly recreating human features, facial expressions and fluid movement.
And so you can make a clip in which a person very realistically does whatever you want them to. Try to guess which field, according to 2019 data, accounted for 96 percent of the videos made with this technology?
That's right. It's about po…
Going Deeper
No, not politics. Porn. Yes, this is the next level after silicone dolls: now, with the help of technology, you can see and hear your favorite star say how much she wants you. But that's another, no less interesting topic, one that deserves separate attention, especially its psychological aspects.
Still, while we're on the subject… What if this technology turns out to be so good that a compromising clip can be made of any politician, celebrity or businessperson?
For now, deepfake does well specifically with faces and facial expressions. And that is built into the technological principles it works on.
Deeply Wrong
The technology owes its name to the field of deep learning. It uses generative adversarial networks (GANs, as they're affectionately known). These are machine-learning algorithms that receive a set of information and process it so as to create entirely new content while following the principles they've “picked up.” They're at the heart of many of the projects with which artificial intelligence threatens to put journalists out of work, and they're also relied upon in attempts to get computers to create art.
In a GAN, two segments of the software process the information in parallel. The first tries to create content as similar as possible to the target, while the second—a kind of evil editor—checks whether it can spot the differences from the real thing and, in most cases, “sends” the product back to its “colleague” for revision. So only processed short excerpts that have passed through a rather strict filter ever reach the user.
Obviously it still has imperfections, but these can be smoothed out. All the facial expressions are generated by the neural network, so they're often unnatural. The generated voice is monotonous and lacks emotion. Computers have another advantage, though—unlike actors, they don't get tired; they can generate new suggestions endlessly, improve themselves and give programmers even more hints on how to help them along. In other words, with a bit more work it's already entirely possible to create truly indistinguishable deepfake videos.
The Plot Deepens
The GAN was created in 2014 by Ian Goodfellow, then a student at Stanford University. Over the following years he would switch jobs several times between Google and Elon Musk's projects in artificial intelligence.
It wasn't him, however, who came up with deepfake as the specific application of this type of technology. In that same year, 2018, a user on the content-sharing network Reddit managed to gain enormous popularity by creating porn clips in which he very realistically replaced the faces of the actresses with those of various celebrities.
Unfortunately, his account was quickly shut down, but his nickname survived. Yes, that's right, he had decided to call himself “Deepfake.”
After that the technology became quite popular for jokes, witty clips and other things (…by now it's probably clear what).
There are quite a few apps that use deepfake, and some of them are free. The breakthrough in this direction came, again in 2018, with FakeApp. FakeApp and Adobe After Effects were exactly what was used for the notorious Barack Obama clip we started with.
And then… deepfake somehow faded as a topic; other gags were more popular during the past year, which caught us all with our pants down. The British queen example seems to have been the only notable event on the subject lately.
The Deep State
Today the leader is DeepFaceLab, which (according to its creators) is used to make more than 95 percent of the deepfake videos in the world.
Like any other technology, deepfake can be used to cut or to get cut. It's still at the start of its development curve. It isn't that well known yet, and it never became a trend, because at first glance it doesn't change lives. But at second glance… its uses go far beyond taking work away from the coaches on Kato dve kapki voda, the Bulgarian version of the celebrity impersonation show Your Face Sounds Familiar.
Would you like to see Clark Gable again, but this time starring in Titanic? Or Apostol Karamitev, one of the great Bulgarian film actors of the twentieth century? Art really could be different once GANs stumble onto perfection in imitation. Or maybe you're fed up with real actors, and you can “assemble” a more perfect one yourself, exactly as you imagine him…
And then… we have to come back to the biggest dangers (no, not that Clark Gable couldn't possibly sink so spectacularly with a wet mustache).
In the electronic powder keg that is today's world politics, it really takes far too little to start a fire that sweeps everything away. Tension—between states, between ethnic groups and on every other front—can quickly escalate from one small spark. Until now, our prime suspect for such a spark would usually be good old fake news. And that has already happened.
But what will happen when sufficiently perfected deepfake technology is added to the arsenal of manipulation? The example we started with is just one, but anyone can add more with a little imagination. Then come the opportunities for blackmail and for manipulating people. That's why it's important to be prepared.
Seeing Is Believing
And to know what the neural networks hiding behind artificial intelligence are already capable of. During the American Senate elections last year there was even talk of laws specifically regulating deepfake.
The interesting thing is that, because of all of the above, the very technologies that create forgeries can help us in the fight against them.
Google, for example, is developing a project in which, once again, neural networks are used to catch the differences. “I gave you life, and I will take it,” says good old Taras Bulba. Because of the principle on which GANs work, however, this risks turning into an endless race. Since the second part of their neural network is built to “catch” imperfections before they reach the viewer, it is in direct competition with the software that would spot them. A battle takes shape in which the winner will change easily.
Amazon, for its part, organized a competition with an impressive prize fund for technologies that “catch” deepfakes. One of the Pentagon's research units, meanwhile, published an algorithm that recognizes tampering in video thanks to… unnatural blinking.
This digital story could bring plenty of analog problems. In our time it's very hard to tell a forgery apart, whatever it may be; the abstract and the virtual blur the boundaries. But reality is still here to hold them in place.
