Cheap Not Deep … This Time
Finally, the election season is over! Or is it?
The danger of falsified election content was highlighted by The Hill in July:
“Imagine if a nation-state or terrorist organization were able to hack President Trump’s Twitter feed and send out messages indicating that the U.S. was going to strike China imminently. … As the old adage goes, the lie is half-way around the world before the truth gets its boots on.”
Nothing Fake About this Real Threat
Given how “fake news” memorably altered the 2016 US elections, worries that fabricated “deepfake” videos could become an even bigger problem in coming elections seem well-founded. As previously covered by this series, “deepfakes” refer to digital representations produced by generative adversarial networks (GAN)s that yield seemingly realistic, but entirely synthetic images and sounds.
The danger becomes obvious given that deepfake technology can be used to create videos that seem to show politicians saying things they never said, or doing things they never have done. While it had been demonstrated before, the potential for technology to create havoc first gained widespread attention in April 2018, when comedian Jordan Peele created a video that appeared to show former President Barack Obama insulting President Donald Trump in a speech.
It’s worth noting that it currently is not a US federal crime to create and distribute deepfake videos. The first federal bill targeting deepfakes, the Malicious Deep Fake Prohibition Act, was introduced in December 2018, though it has not been signed into law. States however aren’t waiting for the federal government to act- both California and Texas have enacted laws that make deepfakes illegal when they’re used to interfere with elections.
In the run-up to the 2020 election, researchers in the synthetic media space ramped up defenses, training and releasing both novel detection models that used both somewhat obvious visual distortions called artifacts. As this DARPA explainer shows (below) these artifacts can manifest in several ways:
But with newer deepfake generation technology looking for even slight artifacts may not be enough. Can you guess how many of the following photos are of real people?
If you couldn’t spot any artifacts, you’re in good company. I couldn’t either. But none of those people are real. So the techniques have had to go beyond skin deep.
The Heart as a Fingerprint for the Soul
Researchers deployed an advanced detection model this election cycle that was based on the observation that the “sub-perception biological signal domain” (!!!) generated by the human heart creates a unique pattern that can be utilized for source detection. The hypothesis, later proven, was that this heart beat driven blood flow variance could be used to pinpoint synthetics even really good ones because, when resolved, since they literally don’t have a heart, under the hood they look as noisy as the neural nets that generated them.
But despite the fears, aside from a few notable exceptions (which we’ll cover below) sophisticated deepfakes do not appear to have been deployed during this fraught political season something that caused NPR to ask at one point: “where are the deepfakes?”
While significant advances in detection had an effect, they likely weren’t deployed because simple deceptions, such as selective editing, got the job done with less effort at lower cost. Unlike deepfakes these so-called “cheapfake” techniques don’t require artificial intelligence to create, leading to their rampant use this election season, deployed by campaigns and independent groups alike on every side.
Cheapfakes in 2020
Two altered videos of House Speaker Nancy Pelosi were released on Facebook this summer. In the first, the footage was altered to make the Speaker appear to be stumbling through a speech while in the second, the footage was altered to make the Speaker appear intoxicated. While both were labeled partly false by Facebook’s fact-checkers, they were not removed from the site. These followed an altered video that was released by the Trump campaign in February that appeared to show the President addressing a veteran as the Speaker tore up a copy of his State of the Union speech. Despite repeated requests from the Speaker’s Office, Facebook refused to remove any of these videos, once again demonstrating the power of the platform to disintermediate the flow of content and control the spread of these forms of disinformation if desired.
But it wasn’t just the President and independent groups aligned with him that were seeking to bend the truth. During the primary, both Mike Bloomberg and Pete Buttigieg campaigns received pushback for releasing ads that deceptively edited footage in favorable ways. Buttigieg's campaign added applause to CNN footage, while Bloomberg edited primary debate footage to make his questioning of other candidates look much more incisive than it was.
Deepfakes in the Campaign: Obvious and Detectable… for Now
There were two notable deepfakes released during the later stages of the campaign. The first, Sassy Justice, was a 15-minute video created by Trey Parker and Matt Stone, the minds behind South Park. Sassy Justice features several deepfakes, including Al Gore, Michael Caine, and most notably a deep fake of Donald Trump as Fred Sassy Cheyenne, Wyoming local consumer advocacy reporter who interviews various celebrities (all of whom are deepfakes) about the dangers of deepfakes. While extremely well done from a technical standpoint, the work is obviously satirical and not meant to deceive.
The second example was a short video of Vladamir Putin in which he appears to speak in english. He notes that while he is blamed for degrading America’s electoral process, the real fault is with Americans themselves for weaponizing the electoral process. The video was an ad for RepresentUS — a nonprofit organization that calls itself: “A powerful movement of independents, progressives, and conservatives … pursuing federal reform through the states.”
While not satire, the context, and topic purposefully used the synthetic nature of the ad to prominently illustrate a point.
What does the use of Synthetic Media in future elections mean for the United States and rest of the world? While cheapfakes continue to sow confusion for low cost, deepfakes could increase the stakes. As the cost of generating them continues to fall, it is likely that we will continue to see a rise in their use. Given the power of visual media to change voters’ minds and shape the outcome of elections, it seems that increased investment in synthetic media detection technology is prudent both here and abroad.
The threat of deepfakes is greatest where their impacts are felt the most. While half of the United States celebrated in the election of their preferred candidate, the other half protested and questioned the validity of the win. America remains a house divided, uniquely vulnerable to pernicious threats of deepfakes and other forms of political synthetic media.
Jeremy Bash and Michael Steed, Opinion Contributors. “Deepfakes Threaten the 2020 Election.” TheHill. The Hill, July 21, 2020. https://thehill.com/opinion/cybersecurity/508202-deepfakes-threaten-the-2020-election.
Mak, Tim, and Dina Temple-Raston. “Where Are The Deepfakes In This Presidential Election?” NPR. NPR, October 1, 2020. https://www.npr.org/2020/10/01/918223033/where-are-the-deepfakes-in-this-presidential-election.
Ciftci, Umur Aybars, Ilke Demir, and Lijun Yin. “How Do the Hearts of Deep Fakes Beat? Deep Fake Source Detection via Interpreting Residuals with Biological Signals,” August 26, 2020. https://arxiv.org/abs/2008.11363.
Wulfsohn, Joseph. “Pete Buttigieg Campaign Roasted for Adding Applause to CNN Town Hall Clips.” Fox News. FOX News Network, February 7, 2020. https://www.foxnews.com/media/pete-buttigieg-cnn-town-hall-applause.
Itzkoff, Dave. “The ‘South Park’ Guys Break Down Their Viral Deepfake Video.” The New York Times. The New York Times, October 29, 2020. https://www.nytimes.com/2020/10/29/arts/television/sassy-justice-south-park-deepfake.html.
Puutio, Alexander. “Deepfake Democracy: Here’s How Modern Elections Could Be Decided by Fake News.” World Economic Forum. Accessed November 9, 2020. https://www.weforum.org/agenda/2020/10/deepfake-democracy-could-modern-elections-fall-prey-to-fiction/.