Here Come the Fake Videos, Too
By Kevin Roose, NY Times, March 4, 2018
The scene opened on a room with a red sofa, a potted plant and the kind of bland modern art youâd see on a therapistâs wall.
In the room was Michelle Obama, or someone who looked exactly like her. Wearing a low-cut top with a black bra visible underneath, she writhed lustily for the camera and flashed her unmistakable smile.
Then, the former first ladyâs doppelgĂ€nger began to strip.
The video, which appeared on the online forum Reddit, was whatâs known as a âdeepfakeâ--an ultrarealistic fake video made with artificial intelligence software. It was created using a program called FakeApp, which superimposed Mrs. Obamaâs face onto the body of a pornographic film actress. The hybrid was uncanny--if you didnât know better, you might have thought it was really her.
Until recently, realistic computer-generated video was a laborious pursuit available only to big-budget Hollywood productions or cutting-edge researchers. Social media apps like Snapchat include some rudimentary face-morphing technology.
But in recent months, a community of hobbyists has begun experimenting with more powerful tools, including FakeApp--a program that was built by an anonymous developer using open-source software written by Google. FakeApp makes it free and relatively easy to create realistic face swaps and leave few traces of manipulation. Since a version of the app appeared on Reddit in January, it has been downloaded more than 120,000 times, according to its creator.
Deepfakes are one of the newest forms of digital media manipulation, and one of the most obviously mischief-prone. Itâs not hard to imagine this technologyâs being used to smear politicians, create counterfeit revenge porn or frame people for crimes. Lawmakers have already begun to worry about how deepfakes could be used for political sabotage and propaganda.
Even on morally lax sites like Reddit, deepfakes have raised eyebrows. Recently, FakeApp set off a panic after Motherboard, the technology site, reported that people were using it to create pornographic deepfakes of celebrities. Pornhub, Twitter and other sites quickly banned the videos, and Reddit closed a handful of deepfake groups, including one with nearly 100,000 members.
Some users on Reddit defended deepfakes and blamed the media for overhyping their potential for harm. Others moved their videos to alternative platforms, rightly anticipating that Reddit would crack down under its rules against nonconsensual pornography. And a few expressed moral qualms about putting the technology into the world.
Then, they kept making more.
The deepfake creator community is now in the internetâs shadows. But while out in the open, it gave an unsettling peek into the future.
âThis is turning into an episode of Black Mirror,â wrote one Reddit user. The post raised the ontological questions at the heart of the deepfake debate: Does a naked image of Person A become a naked image of Person B if Person Bâs face is superimposed in a seamless and untraceable way? In a broader sense, on the internet, what is the difference between representation and reality?
The user then signed off with a shrug: âGodspeed rebels.â
After lurking for several weeks in Redditâs deepfake community, I decided to see how easy it was to create a (safe for work, nonpornographic) deepfake using my own face.
I started by downloading FakeApp and enlisting two technical experts to help me. The first was Mark McKeague, a colleague in The New York Timesâs research and development department. The second was a deepfake creator I found through Reddit, who goes by the nickname Derpfakes.
Because of the controversial nature of deepfakes, Derpfakes would not give his or her real name. Derpfakes started posting deepfake videos on YouTube a few weeks ago, specializing in humorous offerings like Nicolas Cage playing Superman. The account has also posted some how-to videos on deepfake creation.
What I learned is that making a deepfake isnât simple. But itâs not rocket science, either.
The first step is to find, or rent, a moderately powerful computer. FakeApp uses a suite of machine learning tools called TensorFlow, which was developed by Googleâs A.I. division and released to the public in 2015. The software teaches itself to perform image-recognition tasks through trial and error. The more processing power on hand, the faster it works.
To get more speed, Mark and I used a remote server rented through Google Cloud Platform. It provided enough processing power to cut the time frame down to hours, rather than the days or weeks it might take on my laptop.
Once Mark set up the remote server and loaded FakeApp on it, we were on to the next step: data collection.
Picking the right source data is crucial. Short video clips are easier to manipulate than long clips, and scenes shot at a single angle produce better results than scenes with multiple angles. Genetics also help. The more the faces resemble each other, the better.
Iâm a brown-haired white man with a short beard, so Mark and I decided to try several other brown-haired, stubbled white guys. We started with Ryan Gosling. (Aim high, right?) I also sent Derpfakes, my outsourced Reddit expert, several video options to choose from.
Next, we took several hundred photos of my face, and gathered images of Mr. Goslingâs face using a clip from a recent TV appearance. FakeApp uses these images to train the deep learning model and teach it to emulate our facial expressions.
To get the broadest photo set possible, I twisted my head at different angles, making as many different faces as I could.
Mark then used a program to crop those images down, isolating just our faces, and manually deleted any blurred or badly cropped photos. He then fed the frames into FakeApp. In all, we used 417 photos of me, and 1,113 of Mr. Gosling.
When the images were ready, Mark pressed âstartâ on FakeApp, and the training began. His computer screen filled with images of my face and Mr. Goslingâs face, as the program tried to identify patterns and similarities.
About eight hours later, after our model had been sufficiently trained, Mark used FakeApp to finish putting my face on Mr. Goslingâs body. The video was blurry and bizarre, and Mr. Goslingâs face occasionally flickered into view. Only the legally blind would mistake the person in the video for me.
We did better with a clip of Chris Pratt, the scruffy star of âJurassic World,â whose face shape is a little more similar to mine. For this test, Mark used a bigger data set--1,861 photos of me, 1,023 of him--and let the model run overnight.
A few days later, Derpfakes, who had also been training a model, sent me a finished deepfake made using the footage I had sent and a video of the actor Jake Gyllenhaal. This one was much more lifelike, a true hybrid that mixed my facial features with his hair, beard and body.
Derpfakes repeated the process with videos of Jimmy Kimmel and Liev Schreiber, both of which turned out well. As an experienced deepfake creator, Derpfakes had a more intuitive sense of which source videos would produce a clean result, and more experience with the subtle blending and tweaking that takes place at the end of the deepfake process.
In all, our deepfake experiment took three days and cost $85.96 in Google Cloud Platform credits. That seemed like a small price to pay for stardom.
On the day of the school shooting last month in Parkland, Fla., a screenshot of a BuzzFeed News article, âWhy We Need to Take Away White Peopleâs Guns Now More Than Ever,â written by a reporter named Richie Horowitz, began making the rounds on social media.
The whole thing was fake. No BuzzFeed employee named Richie Horowitz exists, and no article with that title was ever published on the site. But the doctored image pulsed through right-wing outrage channels and was boosted by activists on Twitter. It wasnât an A.I.-generated deepfake, or even a particularly sophisticated Photoshop job, but it did the trick.
Online misinformation, no matter how sleekly produced, spreads through a familiar process once it enters our social distribution channels. The hoax gets 50,000 shares, and the debunking an hour later gets 200. The carnival barker gets an algorithmic boost on services like Facebook and YouTube, while the expert screams into the void.
Thereâs no reason to believe that deepfake videos will operate any differently. People will share them when theyâre ideologically convenient and dismiss them when theyâre not. The dupes who fall for satirical stories from The Onion will be fooled by deepfakes, and the scrupulous people who care about the truth will find ways to detect and debunk them.
âThereâs no choice,â said Hao Li, an assistant professor of computer science at the University of Southern California. Mr. Li, who is also the founder of Pinscreen, a company that uses artificial intelligence to create lifelike 3-D avatars, said the weaponization of A.I. was inevitable and would require a sudden shift in public awareness.
âI see this as the next form of communication,â he said. âI worry that people will use it to blackmail others, or do bad things. You have to educate people that this is possible.â
So, O.K. Here I am, telling you this: An A.I. program powerful enough to turn Michelle Obama into a pornography star, or transform a schlubby newspaper columnist into Jake Gyllenhaal, is in our midst. Manipulated video will soon become far more commonplace.
And thereâs probably nothing we can do except try to bat the fakes down as they happen, pressure social media companies to fight misinformation aggressively, and trust our eyes a little less every day.
