A/D/O by MINI | Deepfakes: the art of deception

Deepfakes: the art of deception

Part design trend, part tech-driven dystopian tool, this practice is radically transforming everything from politics to pornography.

Even if you haven’t heard of deepfakes, chances are they have already had an impact on your life. Coined in 2017, deepfakes is the colloquial term for a visual technique that utilizes a generative adversarial network (GAN) – combining an image or voice with artificial intelligence, superimposed over existing images and videos – to make an entirely new, and entirely false image. The result is a “real looking" photo or video that is largely computer-generated, and realistic to an uncanny degree.

While scientists hope this technology will someday help with mapping capabilities for space exploration and driverless cars, create digital voices for the voiceless, transform visual effects for movies and video games, and provide invaluable information to the medical field, detractors fear they may have opened a pandora’s box of identity theft, false information and digital confusion, and that deepfakes may even pose a threat to democracy. But what is the immediate impact of deepfakes on society, and how might it evolve?

These forgeries are not only hard to spot, they are everywhere. Recently it was revealed that AI technology had been used to superimpose celebrity faces onto pornographic stills to create racy clips of actresses Scarlett Johansson and Gal Gadot. In June, the app Deepnudes was shut down when it was discovered that users could feed it photos of ordinary women and transform them into naked models.

Deepfakes of political figures like Barack Obama could be used to create fake news

But this pales in comparison to its potential uses. Recently, actor and director Jordan Peele, in a Youtube PSA clip for Buzzfeed, demonstrated how easy it is to create a realistic "digital Barack Obama" and convincingly mimic his language patterns. There’s no question that false news and doctored social media posts affected the 2016 election, but there is strong evidence that this new variety of digital forgery will be used to generate misinformation in the upcoming presidential election cycle.

And deepfakes have the potential to wreak havoc on the financial sector as well. Recently Symantec, a global cybersecurity company, has reported several attacks on private financial organizations using AI: in each case a company’s “CEO” called a trusted senior financial officer ordering an urgent money transfer, imitating the CEOs’ speech patterns with an AI program that had most likely been trained on audio data culled from publicly-available clips, many originating from social media posts. In these cases, millions of dollars were stolen from each company.

“During my time at Google, I noticed a shift from one [recognizable] female voice to an array of voices indicating different genders, backgrounds, locales," said an anonymous former AI contractor with Google, who hypothesizes that in the future, rather than just a Siri-like bot, we may soon have AI that can mimic regional inflections and subtle vocal intonations. "I'm sure some major music label isn't far away from developing a version [of AI] that can sing," he continued, explaining that in the future, we might be unable to decipher between “real” and simulated culture icons – an issue we’re already facing with online influencers.

This deepfake was created using an image of actress Jennifer Lawrence

It is widely assumed that the underlying technology behind most deepfakes and AI-powered synthetic media was invented in 2014 by the Montreal-based PhD student Ian Goodfellow, who was later offered a position at Google. Until Goodfellow’s invention of GAN, algorithms had been reasonably good at identifying images from mass quantities of data, but were unable to produce entirely manufactured images of an individual from front to back. Until now, that is. Researchers can now create software tools that permits users to edit the transcript of a video to alter the words on demand. At Stanford University, Deep Video Portraits – a system that can manipulate not only facial expressions but also 3D head positions, eye gaze, blinking, and head rotation – is already in operation.

Creating deepfakes “in the wild” requires a number of techniques. First, you must scan the target video to isolate phonemes, the essence of sounds and vocal tics spoken by the subject. These are then matched with corresponding visemes, the correct facial expressions that accompany each sound. Finally, a 3D simulation of the lower half of the subject’s face is created. All of this is then edited together to construct new footage, which is then seamlessly pasted onto the source video to create the final uncanny result.

The technology has proven fascinating for both programmers and the public. Recently an app created by MIT-IBM Watson AI Lab, which lets users upload photos they can transform into classical-style faux watercolors, oil paintings, or ink portraits based on famous art history movements, has already become a viral phenomenon. The tool’s creators used GAN models that fuse opposing neural networks to produce recognizable outcomes – a generator that looks at examples and tries to mimic them, and a discriminator, which deciphers if they are real by comparing them with existing images. For the app, GAN models scanned over 45,000 portrait images to train the program, including paintings by Van Gogh and Rembrandt. Using similar technology, Samsung’s AI lab made the Mona Lisa “smile” this spring, and created interactive portraits of Salvador Dali and Marilyn Monroe from a single image. With the popularity of Facetune and FaceApp, it’s easy to see why, despite the fears, consumers are still enamored of deepfakes.

“When anything can be faked, we need trusted journalists to determine sources and veracity,” said lauded author and documentarian Doug Rushkoff, who specializes in the study of “human autonomy in a digital age” and points to larger societal issues. “Deepfake videos are just computer animations of the same lies being tweeted today. They make it a little easier to believe untruths, but that effect should only last a couple of years,” he said, pointing to a future where, out of necessity, we must critically analyze images and video before posting or reposting. 

AI company Pinscreen presented research into deepfakes at the World Economic Forum

But there are those who challenge the technology. Giorgio Patrini, formerly a postdoc researcher at the University of Amsterdam working on deep learning, became obsessed with the possible destructive uses for GAN models and left his studies to co-found Deeptrace Labs, a Dutch startup that claims to be “the antivirus for deepfakes”. Deeptrace serves organizations by providing technology for monitoring and detection of deepfakes, and to date has uncovered more than 14,000 deepfake videos. “In 2019, we started to see how synthetic videos and audio can profoundly impact politics,” said Patrini, who points to cases of fake videos creating political tension in Gabon, Malaysia and India, and to a minor extent, in the US. “Deepfake voices are now a real worry for cybersecurity companies. Not unlikely, we might see soon high-profile cases of fraud and market manipulation.”

At UC Berkeley and the University of Southern California, researchers are crafting technologies that will influence the next generation of photorealistic humans, which can be used not only in movies but in video games, advanced telepresence (a 3D version of Skype chatting), and virtual reality. The team also published a recent paper that called for protecting world leaders from deepfakes, offering tools to help in the identification processes by analyzing “soft biometric signature” – a specific individual’s style of speech and movement. Using machine learning to distinguish the subtle ticks that make us “us”, from hand gestures to facial expressions, in experiments the team’s technique was 92% accurate in spotting deepfakes. This research, presented at a computer vision conference in California this year, was funded by Google and DARPA, a research wing of the Pentagon. 

“There will be a point where it’s possible to create any type of content that will be indistinguishable from reality,” warned Hao Li, a professor at the University of Southern California and the CEO of Pinscreen, which creates instant 3D avatars. While currently, experts can differentiate using complex tools, there are other, more simple ways to decipher whether something is or isn’t a deepfake: “fact-checking mechanisms,” he explained, or “if you have multiple observations that it occurred” – both essentials if you do serious journalism.

Pinscreen allows users to instantly generate 3D avatars from a single image

While AI is responsible for deepfakes, it has also become invaluable in modern life – and companies are taking notice. Recently, Adobe developed an AI-enabled tool that can spot deepfakes in digital images, and has already shared details on evolving software that lets users edit recordings of speech with the click of a mouse. Deepfakes, while they cannot be “stopped” in the traditional sense, also point to future technology with limitless possibility.

“I think it is best to start from a place where we consider that any decent machine learning programmer will be able to create a deepfake in the near future,” said Matthew Putman, cofounder and CEO at Nanotronics, which recently launched a platform that combines AI, automation, and sophisticated imaging for industrial inspection. “Since this is the case, let’s figure out what the best, most interesting deepfakes could be.” He believes deepfakes are actually a call to imagination for scientists, engineers, and programmers. “Here, the end goal is not the manipulation of biased minds, but the expansion of the perceived limits of our physical reality. We must continue searching for truth, even as we readjust our ideas about what reality is.” Ultimately, if we can’t “beat” deepfakes or deepfake technology, we must adapt and rise to the challenge.

“AI is only as smart as you train it,” said Stephen Leps, EIC at IBM. Leps explained that the secret sauce to all AI and, in effect, all technology, is how we use it. Rather than reining in the study and development of deepfakes, the future may lie in regulating and ensuring that technology is being used by the right actors. “It’s how we harness the technology that we have,” he said. After all, “it's humans and machines working together”.

Text by Laura Feinstein.