in ,

Artificial Intelligence Is Being Used to Generate a New Kind of Deepfake


For the past two years, I’ve been following a woman around the internet. It sounds ominous, I know, but hear me out. Her name is Albertina Geller, and I first stumbled across her online in October 2020, on LinkedIn. She’d listed herself as a “self-employed freelancer” in Chicago. I’m also a self-employed freelancer, so we had that in common. In her bio, she said that “I learn & teach people how to be healthy, balance their gut and improve their immune system for healthy living.” I’ve had some gut and immune-system issues myself. It was a connection practically written in the stars.

But I have to admit that what first interested me about her — what led me to spend two years tracking her, at a distance — wasn’t our shared interests. It was her appearance.

Her LinkedIn photo was a straight-on headshot of a white woman, mid- to late 20s, with a pale complexion and lightly rosy cheeks. She had shoulder-length blond hair, swept neatly to one side. She gazed directly into the camera with sparkling light-green eyes and a wide smile. It was this photo that captured my interest. I had to know: Who was the person behind that smile?

Albertina — I feel I can call her Albertina — turned out to have a significant internet presence. She pursued her passion for a balanced gut on Pinterest, where she pinned images of 14-day anti-inflammatory meal plans, and smoothies to help “drink your way to gut health.” She popped up in online health forums and had her own website and blog, even if it all looked a little slapdash. Everywhere she went there was that same headshot, those same sparkling eyes, that perfect smile.

But here was the thing about the photo: If you stared long enough at it — and I did — you started to notice certain … oddities. Like the way her right ear seemed to tuck in tighter than the left. Or the way her early wrinkles seemed strangely concentrated around her right eye. If you zoomed in, you could see a faint but startling crease down one of her cheeks. Her shirt collar jutted up twice as high on one side. 

None of these dampened her appeal, in my eyes. Quite the opposite. These were her signature tells: the small, strange artifacts left behind by the algorithm that created her. Because Albertina Geller — or her face, at least — wasn’t actually human. She was the product of artificial intelligence. Specifically, she was created by a piece of software called a Generative Adversarial Network, which studies faces and then makes its own. These aren’t the familiar kind of deepfakes, which manipulate images of real people. GANs create nonexistent people. They aren’t meant to impersonate anyone, or steal an identity. They’re meant to impersonate everyone, to mimic the fundamentals of human appearance with increasing fidelity.

Since these counterfeit humans began emerging a few years ago, I’ve been quietly stalking them. Albertina was my index case, the first GAN image that drew my attention. But soon I was tracking dozens, then hundreds, then thousands of them around the web, cataloging where they surfaced, the people they purported to represent. These faces had quietly diffused out into our digital surroundings. But where had they gone? Who was using them, and to what end?

Perhaps, I thought, I could assign some humanity to Albertina. To answer for her the questions all humans have about ourselves: Who is Albertina Geller, and what is her purpose?

Tap to see the many identities of Albertina

 


They came first a few, then all at once, GAN faces like Albertina’s. As with many technological advances, it started as a purely academic exercise, a let’s-see-if-we-can-do-it sort of thing.

In 2014, a group of computer scientists at the University of Montreal proposed Generative Adversarial Networks as a new method of computer learning. Their particular interest was in training a computer to examine a set of images and then produce likenesses of its own. There are lots of good reasons to want to do this, from the medical (generating images of rare cancers to train radiologists) to the artistic (creating unique images on demand) to the commercial (generating artificial stock photos without the need for models or photo shoots).

The Montreal group’s original system was simple, but brilliant. First it would feed the software a collection of similar-looking images — of letters, vehicles, animals, or human faces. Then, in what’s called the “generator” step, the software would extract features from those images and use them to create its own. For faces, that meant extrapolating from hairlines and wrinkles, earlobes and smiles, and then drawing new, never-before-seen visages.

The first faces produced by the generator were blobby and ill-defined. It was the next step of the GAN, the adversarial step, where the magic happened. The generator was “pitted against an adversary”: another part of the software called the “discriminator.” The discriminator examined both the original images and the ones created by the generator and tried to figure out which was which.

Think of the generator as a counterfeiter, trying to create realistic-looking fake currency, the Montreal group suggested in its first paper on GANs. The discriminator is like the police, looking through a stack of bills to detect which are counterfeit. The counterfeiter in turn sees which bills are caught as forgeries and which slip by — and then incorporates that knowledge into its next batch. “Competition in this game drives both teams to improve their methods,” the team wrote, “until the counterfeits are indistinguishable from the genuine articles.” Even in their first efforts, the GAN images were quickly on par with those made by any other, less intelligent software. A few years later, they started looking like Albertina.

Soon other researchers and companies were making GANs, feeding them all variety of image sets: cats and fruit and landscapes and ancient vases. But GANs first entered the public spotlight in 2019, when a web site called thispersondoesnotexist.com — featuring a rotating series of artificially generated faces created by software from the company Nvidia — went viral. Soon the code to create a GAN was so freely available that even a moderately competent programmer could build their own. And they did: The number of “this does not exist” sites has become a running joke in GAN circles, even spawning a site called This X Does Not Exist.

What interested me about GAN faces like Albertina’s is the way they diverge from their famous deepfake cousins. The best-known deepfakes so far have been videos: artificially generated clips like one of Barack Obama, created by BuzzFeed in 2018, mouthing lines from Jordan Peele. Or a fake Tom Cruise doing cheesy magic tricks on TikTok.

But GANs aren’t trying to hoax a specific person — they’re creating new ones. And unlike the artisanal efforts of celebrity video  deepfakes, GANs are being made at volume, churned out by the millions. This is the industrialization of fakery, restricted only by the ambition of its creators. The promise of GANs — or their menace — lies in their ubiquity, their sheer mundanity. Albertina is already living among us. You just may not notice her.

A photo of a woman's face broken up into repeating pieces

Generated Adversarial Networks aren’t manipulating images of real people — they’re creating nonexistent people.

Matthieu Bourel for Insider




Albertina Geller was born in a company called Generated Photos, sometime in early 2020. Based in the US, Generated Photos was founded only a year earlier, by a designer named Ivan Braun. With the goal of “building the next generation of media through the power of AI,” the company employed studio models to create its catalog of human faces. It then used those real images to train a GAN to spit out realistic-looking headshots.

You can find the results on the Generated Photos site, where you can filter by ethnicity, age, sex, eye color, and other attributes. Pick your favorite “joyful black child female with medium brown hair and brown eyes” or “neutral white middle-aged male with short black hair and blue eyes.” The company suggests the images are intended to replace stock photos, and its terms of service prohibit use for “any sort of illegal activity, such as defamation, impersonation, or fraud.” Individual photos can be downloaded free, and the site says “many people have found our images useful for education, training, game development, and even dating apps!” Starting at $2.99 apiece, you can also get an exclusive license to a particular face, a nonexistent person who is yours and yours alone. 

Albertina wasn’t one of those lucky few. She remained without a forever home, in an open collection of more than 2.6 million images that anyone could download and use. That’s where I found her, staring out from a grid of radiant faces. 

She stood out for more than just her looks. Generated Photos’ GANs tend to match conventional beauty standards, a product of the models that have been fed to the software. But Albertina also had a certain … humanness about her. Many GAN faces are rendered with digital artifacts that make them instantly detectable as fake. The software can produce strange and sometimes shocking deformities — comically asymmetrical eyes, faces that melt and stretch in inhuman ways. Generated Photos seems exceptional at filtering out abnormalities, and the vast majority of its images could pass for human. For a select few, like Albertina, the differences are nearly imperceptible.

Wondering where she might have surfaced on the internet, I plugged her photo into the search-by-image offerings on Google and Bing. This produced Albertina’s LinkedIn and Pinterest profiles, among a dozen other webpages. Eventually I added more specialized image searchers, like Yandex and TinEye. (Try it yourself: There is a handy Google Chrome extension called “Search by Image” that allows you to explore all of them at once.) From there I began following her around the internet, trying to piece together her intentions.

She was, first and foremost, a prolific answerer of questions on health-related message boards. It wasn’t just gut health, either. On an obsessive-compulsive-disorder forum, Albertina offered feedback as to whether constant “sexual thoughts” constituted OCD. On a site called Anxiety Central she offered heartfelt responses to questions about everything from coronavirus tests to antibiotics. On the Q&A forum Quora she answered hundreds of queries, on topics as varied as constipation, rowing machines, and “some creative ideas for decorating cookies with sprinkles.” She edited Wikipedia pages, commented on cake recipes, and made precisely one post on a site called singaporemotherhood.com. 

From her answers, she didn’t appear to be a bot. There was someone, or some group, behind her face — creating her profiles, pinning her recipes, writing her blog posts. That someone had gone through considerable efforts to transform Albertina’s blond headshot, Pinocchio-like, into a real human.

A gif showing pieces of a woman's face being revealed and then hidden

There was someone, or some group, behind Albertina’s face — creating her profiles, pinning her recipes, writing her blog posts.

Matthieu Bourel for Insider




For most of its history, of course, photography has aimed in the opposite direction: to take something real, something human, and fix it in place. In its early years, the photograph was hailed not just as a technology to preserve the world but as a new way to examine it. It’s the reason Frederick Douglass was captivated by its possibilities. Photos, he believed, offered a way to render Black Americans with a dignity that could counter their dehumanization by American society. He sat for hundreds of portraits, famously becoming the most photographed man of the 19th century. Photographs have the power, he said, to “make our subjective nature objective, giving it form.”

Over the next century, we came to experience the myriad ways that seeming objectivity could be manipulated and propagandized. Photographs could be deceptively staged, captioned with lies, or selectively cropped. Indeed, they were selectively cropped by their very nature, presenting a slice of reality rather than reality itself. “Photographs attract false beliefs the way flypaper attracts flies,” Errol Morris wrote in “Believing Is Seeing,” his 2011 investigation into photographic truth. “Because vision is privileged in our society and our sensorium. We trust it; we place our confidence in it. Photography allows us to uncritically think. We imagine that photographs provide a magic path to the truth.” 

With the twin developments of digital photography and the internet, issues once debated in highbrow photography critiques began to metastasize into our everyday life. Driven by the ubiquity of powerful cameras and the ease of Photoshop, photographs increasingly slipped their anchor to reality. There were canaries in the coal mine, like the viral photo of the “tourist guy,” standing on the observation deck of the World Trade Center, an approaching plane edited in below him. In one decade, we went from magazine-cover touch-ups of movie stars to powerful photo filters carried in our pockets, allowing us to instantly manipulate our own images before sharing them with the world.

At the same time, photographs became simple to steal and repurpose, in the service of catfishing or other mischief. They became trivial to propagandize, in the service of ideology. And when the trick was exposed — as it was with Donald Trump’s inaugural photos, amateurishly manipulated to show a larger crowd — it only multiplied our cynicism about what we were seeing.

Even with all that, though, our trust in photos has remained stubborn. However tenuous or suspect the connection, my impulse is still to believe that what I’m seeing in a photograph is somehow real, that at some level, it exists. GANs are here to sever that connection, at scale. Not with a manipulated reality but with a wholly manufactured one.  

As it happened, even Albertina was a victim of a kind of multiple reality. Her face on Generated Photos was a popular one, and my searches turned up other personas who claimed it as their own. Her image appeared with another LinkedIn account, for Zoya Scoot, a “marketing specialist” in Cleveland. Scoot posted jargon-heavy sales talk to a corresponding Twitter account, and wrote about the future of the metaverse. On Amazon, the image was deployed by J.R. Wily, a self-published  author of grammatically-challenged  children’s books. On a Russian camera-repair site, she was a satisfied customer named Leonova Margarita, while on a Serbian gift-shop site she took the form of a customer-service specialist. On YouTube she became Mary Smith, auteur of a porridge-recipe video, or Maria Ward, who’d violated the terms of service and gotten herself suspended.

Each search I conducted produced not just new uses of Albertina’s photo but other blond-haired GANs, ones the search engines deemed look similar enough to hers. What did GANs most look like, after all, besides one another? The whole thing had a fractal quality to it, an unending feedback loop of smiling, blond-haired faces.

Rows of smiling blonde haired faces

The whole thing had a fractal quality to it, an unending feedback loop of smiling, blond-haired faces.

Matthieu Bourel for Insider




The profligacy of Albertina-like images made me wonder whether there might be some way to automate my efforts, to capture the GAN phenomenon more broadly. So we decided to reverse engineer it. I turned to Angela Wang, a data reporter at Insider, who created a bit of code that would examine thousands of images from Generated Photos, run image searches on each, and rank which were used most often, and where. The program generated a spreadsheet with thousands of sites that had used the most popular GANs.

As I began working my way through them, I noticed they fell into several general categories. Most numerous were the “testimonial” group: GAN faces deployed next to a bit of positive customer feedback. These seemed only mildly deceptive. I’m not sure anyone would be surprised to find fake images accompanying testimonials on an Estonian bitcoin exchange or an online CBD seller. 

Then there were the “placeholders,” GAN images left behind on what seemed to be half-finished websites. Most of them seemed abandoned, ghost companies floating along in the Pacific garbage patch of the internet, their artificial faces peering out from the portholes. Who knew what ambitions were once poured into purrdiva.com, a cat site that had planned to give away cat toys on its launch.

Then there was what I came to think of as the “vaguely nefarious”: GANs usually found on Russian and Chinese sites of dubious intent, or on LinkedIn, attempting to impersonate journalists or other professionals. These appeared to have been deployed on behalf of scams or influence campaigns, and often they came and went as fast as I could find them.

Most intriguing to me, though, were what I came to think of as the “about us” category: companies whose entire staffs were represented by GAN images. Companies like Platinum Systems, “a team of passionate designers, developers, and strategists who create mobile experiences that improve lives,” with a GAN team of 12. Or biggerstars.com, a Hollywood news site whose entire editorial team featured fake images. There you could find reporting by the likes of the “editorial journalist” Daniel T. Hammock — whose GAN featured a bizarrely puffed-out shirt — with non-scoops like “George Clooney and Julia Roberts: On Set In Queensland!”

Whenever I dug into one of these companies, there was usually a kernel of reality behind them. Take Informa Systems, a company in Texas that sells law-enforcement training materials. The company’s site listed the Austin Police Department, the state of Nebraska, and the Los Angeles Police Department among its clients. According to Texas public records, the company really exists. And the police in Austin, Texas, really did contract for its services. But photos on its “Meet Our Team” page were almost all GANs, from CEO “Greg Scully” to the chief marketing officer “Roger Tendul.” (Tendul, a swarthy man with a beard and thick eyebrows, I’d seen before. His photo turned up on 30 other sites.) The irony was almost too perfect: law-enforcement officers being trained to spot criminals on a system created by a company full of fake employees.

Tap to see more fake “About Us” pages

 

The only real human at Informa appeared to be the chief implementation officer Mark Connolly, whose photo looked genuinely imperfect, and whose name appeared on company documents. When I called to ask Connolly about it all, I got an answering service.

Not surprisingly, most of the “about us” sites ignored my inquiries. But I finally reached one, an Austrian test-prep company called takeIELTS. The site’s About Us page featured a staff of 14 smiling employees, each with a robust biography. A closer look revealed textbook anomalies: Lena, a supervisor, wearing only one earring. Felix, the lead designer, with one side of his face shaved closer than the other. Emilia, a practice-test examiner, wearing glasses that lacked the frames to hold their lenses.

I managed to arrange a phone call with the company’s “chief growth officer,” Lukas (hobbies: “mountain biking, swimming and spending time with my family in the suburbs”). After a few minutes of chitchat, I gingerly asked whether the people at the company were, well, real.

He paused. “Some of them are, yes,” he said. “But some of them not.” They had a lot of part-time employees, he said. People came and went, and they’d chosen GANs just to keep it all looking uniform.

“Is there any way to tell which ones are real and which ones are not?” I said.

“I can’t tell you right now,” he said. “But I can look into it.” 

Then came an awkward question. “I don’t mean this at any judgmental way,” I said. “But are you real?”

“Yeah,” he said. “I’m real.”

Well, sort of. His first name was really Lukas, he said, and he was actually the CEO of the company. The CEO on the site was a fiction, as was the information on Lukas’ LinkedIn, including his last name and employment history.

What baffled me, as I told Lukas, was that the company seemed to be quite successful. I’d found hundreds of online reviews from genuine-seeming clients, extolling how the service had helped them. So why did he feel the need to fake his employees?

“It conveys the right message that it’s a big company working with professionals,” he said.

That, I guessed, was most likely the rationale behind a lot of the “about us” sites. That purveyors of small companies might find customers or investors easier to persuade if they displayed a few extra employees on staff. Or that, with the increased attention to corporate inclusion (at least in some parts of the world), they might want to project a level of diversity they hadn’t actually obtained. That these would seem, to them, like small fudges. Who cares what the employees at a test-training site look like in the first place, they might say.

But the problem with a little bit of fakery is that when it’s uncovered, it’s hard to escape the feeling that it’s hiding something deeper, more nefarious. When I checked back in on the takeIELTS site a few months later, Lukas had removed all the GANs from the About Us page and changed the company’s name.


Rows of different GAN created faces

First-wave GANs like Albertina are already finding themselves surpassed by AI images showing off a wider range of expressions, angles, and backgrounds.

Generated Photos



I soon learned I wasn’t the only person trying to distinguish GANs from real humans. As the images have proliferated, so too have efforts to suss them out. And it turns out that the same algorithms used for generating GANs can also be employed to detect them, by spotting the digital artifacts they carry. To train them, you simply feed the algorithms GAN images, instead of real photos. This, in turn, has created its own type of GAN-on-GAN arms race, reflected in academic papers like “Making GAN-Generated Images Difficult To Spot: A New Attack Against Synthetic Image Detectors.” As the police get better at detecting forgeries, the counterfeiters up their game. As a result, first-wave GANs like Albertina are already finding themselves surpassed by AI images showing off a wider range of expressions, angles, and backgrounds. 

In a sense, GANs like Albertina are a testament to the way fakery already enjoys a place of pride on the internet. Anonymity is built into its very fabric, and there’s tremendous power in choosing whatever visual image we wish to represent ourselves online. But that same anonymity, in the hands of industrial-scale creators of artificial photographs, has a way of scrambling the brain. The human race is now so routinely turning to the Albertina Gellers of the world for answers to our questions, trivial and profound, that perhaps we no longer care who or what is behind the image. We are already listening to people whom we not only don’t know but have no evidence are who they claim to be — or are even human at all. Are they real people deploying fake images, or fake people deploying fake images? Does it matter?

As far back as 2013, a team of engineers at YouTube hit upon a phenomenon it called “the inversion”: the point at which the fake content we encounter on the internet outstrips the real. The engineers were developing algorithms to distinguish between authentic human views and manufactured web traffic — bought-and-paid-for views from bots or “click farms.” Like the discriminator in a GAN, the team’s algorithms studied the traffic data and tried to understand the difference between normal visitors and bogus ones.

Blake Livingston, an engineer who led the team at the time, told me the algorithm was working from a key assumption: “that the majority of traffic was normal.” But sometime in 2013, the YouTube engineers realized bot traffic was growing so rapidly that it could soon surpass the human views. When it did, the team’s algorithm might flip and start identifying the bot traffic as real and the human traffic as fake.

“It wouldn’t necessarily happen as one big cataclysmic event,” recalled Livingston, who now works in education technology. “It would probably happen little by little, where one signal might flip over and start penalizing good traffic.” In the end, the engineers simply tweaked the algorithm to rely on more than just which type of traffic was greater. To be more discerning, in other words, and not trust its first instinct.

It’s the kind of thing we’ve all started doing, in a world in which photographs can represent an alternate reality as easily as our own. In 2018, the writer Max Read speculated that the internet had already hit the inversion — that it had crossed the tipping point from mostly real to mostly fake content. It’s impossible to measure, but certainly plausible. “Even if we aren’t near the point where there’s more fake content and activity on the internet than real, people’s level of trust and credulity is probably already adjusting to that,” Livingston said. “There’s an erosion of belief in a consensus reality.” 

At the time Read was writing, GANs were still in their infancy. If we hadn’t already reached the inversion online by then, we are now marching toward it at double time, led by armies of Albertinas. And once we begin defaulting to suspicion instead of trust in our online lives, it’s not hard to see how we might start “penalizing the good traffic” in the real world as well, to use Livingston’s formulation. The result is a sort of any-reality-goes world in which, say, a president could create a universe of lies, and then summon a mob to the Capitol to try to make it real. When our powers of discrimination are so degraded that the truth of a photo is down to a coin flip, it’s not surprising that we start to see deepfakes everywhere.

A GIF of a woman's face becoming distorted over time

Some days, I half expected to walk into a café and see Albertina sitting there, staring back at me.

Matthieu Bourel for Insider




In my search for Albertina, I’d experienced my own kind of personal inversion. I spent so much time looking for GANs, I started to see them everywhere. Certain kinds of headshots began to look GANish, even if I couldn’t quite specify why. When I fed them into my search protocols, I was often correct. But not always. Once in a while the results surprised me, turning up a real human behind an artificial-looking face. Some people, it seemed, just had that GAN look: a perfection in their posture and a universality in their smile. Some days, I half expected to walk into a café and see Albertina sitting there, staring back at me.

On the internet, meanwhile, I slowly started to piece together who she was, or at least who her creators wanted her to be. She seemed to have been brought to life for a kind of artisanal marketing, promoting companies one by one across the web. Arranging her posts into a timeline, it became clear that she’d been systematic about it: promoting a maker of probiotics, a seller of cake sprinkles, a virtual OCD treatment site, an exercise-machine company

When I tried her at the various emails that appeared online, no one ever replied. But along the way she’d let slip clues to the puppeteer behind her. She’d often answer questions about what to me seemed her true passions: where to get the best Rajasthani handicrafts in India, or the best sites to buy kurtis, a kind of traditional Indian dress. She’d weighed in on her favorite Indian news outlets. On one comment site, she’d revealed an IP address in Maharashtra. On another, someone using the handle “Albertina Geller” has been flagged as sending marketing spam from an Indian company.

Not long ago, after a few months away from my search, I checked in on Albertina. She’s stayed busy as always, posting glowingly about a new company called Truein. The company’s software, called Staff Attendance, tracks when employees come and go at the office. “Truein have become quite the sensation overnight, if we may say so,” she wrote recently. 

And the thing that made Albertina so excited about the company? The technological edge it offers that makes its software superior to its competitors?

Facial recognition.


Evan Ratliff is the host and writer of the podcast Persona: The French Deception, and the author of The Mastermind: A True Story of Murder, Empire, and a New Kind of Crime Lord.

 





Source link

What do you think?

How artificial intelligence is changing the face of cybersecurity

TikTok profits from livestreams of families begging