Deepfakes Are Going To Wreak Havoc
On Society. We Are Not Prepared.
None of these people exist. These images were generated using deepfake technology.
THISPERSONDOESNOTEXIST.COM
Last month during ESPN’s hit documentary series The Last Dance, State Farm
debuted a TV commercial that has become one of the most widely
discussed ads in recent memory. It appeared to show footage from 1998 of an
ESPN analyst making shockingly accurate predictions about the year 2020.
As it turned out, the clip was not genuine: it was generated using cutting-edge
AI. The commercial surprised, amused and delighted viewers.
What viewers should have felt, though, was deep concern.
The State Farm ad was a benign example of an important and dangerous new
phenomenon in AI: deepfakes. Deepfake technology enables anyone with a
computer and an Internet connection to create realistic-looking photos and
videos of people saying and doing things that they did not actually say or do.
A combination of the phrases “deep learning” and “fake”, deepfakes first
emerged on the Internet in late 2017, powered by an innovative new deep
learning method known as generative adversarial networks (GANs).
Several deepfake videos have gone viral recently, giving millions around the
world their first taste of this new technology: President Obama using an
expletive to describe President Trump, Mark Zuckerberg admitting that
Facebook's true goal is to manipulate and exploit its users, Bill
Hader morphing into Al Pacino on a late-night talk show.
PROMOTED
The amount of deepfake content online is growing at a rapid rate. At the
beginning of 2019 there were 7,964 deepfake videos online, according to a
report from startup Deeptrace; just nine months later, that figure had jumped
to 14,678. It has no doubt continued to balloon since then.
While impressive, today's deepfake technology is still not quite to parity with
authentic video footage—by looking closely, it is typically possible to tell that a
video is a deepfake. But the technology is improving at a breathtaking pace.
Experts predict that deepfakes will be indistinguishable from real images
before long.
“In January 2019, deep fakes were buggy and flickery,” said Hany Farid, a UC
Berkeley professor and deepfake expert. “Nine months later, I’ve never seen
anything like how fast they’re going. This is the tip of the iceberg.”
Today we stand at an inflection point. In the months and years ahead,
deepfakes threaten to grow from an Internet oddity to a widely destructive
political and social force. Society needs to act now to prepare itself.
When Seeing Is Not Believing
The first use case to which deepfake technology has been widely applied—as is
often the case with new technologies—is pornography. As of September 2019,
96% of deepfake videos online were pornographic, according to the Deeptrace
report.
A handful of websites dedicated specifically to deepfake pornography have
emerged, collectively garnering hundreds of millions of views over the past
two years. Deepfake pornography is almost always non-consensual, involving
the artificial synthesis of explicit videos that feature famous celebrities or
personal contacts.
From these dark corners of the web, the use of deepfakes has begun to spread
to the political sphere, where the potential for mayhem is even greater.
It does not require much imagination to grasp the harm that could be done if
entire populations can be shown fabricated videos that they believe are real.
Imagine deepfake footage of a politician engaging in bribery or sexual assault
right before an election; or of U.S. soldiers committing atrocities against
civilians overseas; or of President Trump declaring the launch of nuclear
weapons against North Korea. In a world where even some uncertainty exists
as to whether such clips are authentic, the consequences could be
catastrophic.
Because of the technology’s widespread accessibility, such footage could be
created by anyone: state-sponsored actors, political groups, lone individuals.
In a recent report, The Brookings Institution grimly summed up the range of
political and social dangers that deepfakes pose: “distorting democratic
discourse; manipulating elections; eroding trust in institutions; weakening
journalism; exacerbating social divisions; undermining public safety; and
inflicting hard-to-repair damage on the reputation of prominent individuals,
including elected officials and candidates for office.”
Given the stakes, U.S. lawmakers have begun to pay attention.
“In the old days, if you wanted to threaten the United States, you needed 10
aircraft carriers, and nuclear weapons, and long-range missiles,” U.S. Senator
Marco Rubio said recently. “Today....all you need is the ability to produce a
very realistic fake video that could undermine our elections, that could throw
our country into tremendous crisis internally and weaken us deeply.”
Technologists agree. In the words of Hani Farid, one of the world's leading
experts on deepfakes: “If we can't believe the videos, the audios, the image, the
information that is gleaned from around the world, that is a serious national
security risk.”
This risk is no longer just hypothetical: there are early examples of deepfakes
influencing politics in the real world. Experts warn that these incidents are
canaries in a coal mine.
Last month, a political group in Belgium released a deepfake video of the
Belgian prime minister giving a speech that linked the COVID-19 outbreak to
environmental damage and called for drastic action on climate change. At
least some viewers believed the speech was real.
Even more insidiously, the mere possibility that a video could be a deepfake
can stir confusion and facilitate political deception regardless of whether
deepfake technology has actually been used. The most dramatic example of
this comes from Gabon, a small country in central Africa.
In late 2018, Gabon's president Ali Bongo had not been seen in public for
months. Rumors were swirling that he was no longer healthy enough for office
or even that he had died. In an attempt to allay these concerns and reassert
Bongo’s leadership over the country, his administration announced that he
would give a nationwide televised address on New Years Day.
In the video address (which is worth examining firsthand yourself), Bongo
appears stiff and stilted, with unnatural speech and facial mannerisms. The
video immediately inflamed suspicions that the government was concealing
something from the public. Bongo’s political opponents declared that the
footage was a deepfake and that the president was incapacitated or dead.
Rumors of a deepfake conspiracy spread quickly on social media.
The political situation in Gabon rapidly destabilized. Within a week, the
military had launched a coup—the first in the country since 1964—citing the
New Years video as proof that something was amiss with the president.
To this day experts cannot definitively say whether the New Years video was
authentic, though most believe that it was. (The coup proved unsuccessful;
Bongo has since appeared in public and remains in office today).
But whether the video was real is almost beside the point. The larger lesson is
that the emergence of deepfakes will make it increasingly difficult for the
public to distinguish between what is real and what is fake, a situation that
political actors will inevitably exploit—with potentially devastating
consequences.
“People are already using the fact that deepfakes exist to discredit genuine
video evidence,” said USC professor Hao Li. “Even though there’s footage of
you doing or saying something, you can say it was a deepfake and it's very
hard to prove otherwise.”
In two recent incidents, politicians in Malaysia and in Brazil have sought to
evade the consequences of compromising video footage by claiming that the
videos were deepfakes. In both cases, no one has been able to definitively
establish otherwise—and public opinion has remained divided.
Researcher Aviv Ovadya warns of what she terms “reality apathy”: “It’s too
much effort to figure out what’s real and what’s not, so you’re more willing to
just go with whatever your previous affiliations are.”
In a world in which seeing is no longer believing, the ability for a large
community to agree on what is true—much less to engage in constructive
dialogue about it—suddenly seems precarious.
A Game of Technological Cat-And-Mouse
The core technology that makes deepfakes possible is a branch of deep
learning known as generative adversarial networks (GANs). GANs were
invented by Ian Goodfellow in 2014 during his PhD studies at the University of
Montreal, one of the world's top AI research institutes.
In 2016, AI great Yann LeCun called GANs “the most interesting idea in the
last ten years in machine learning.”
Before the development of GANs, neural networks were adept at classifying
existing content (for instance, understanding speech or recognizing faces) but
not at creating new content. GANs gave neural networks the power not just to
perceive, but to create.
Goodfellow’s conceptual breakthrough was to architect GANs using two
separate neural networks—one known as the “generator”, the other known as
the “discriminator”—and pit them against one another.
Starting with a given dataset (say, a collection of photos of human faces), the
generator begins generating new images that, in terms of pixels, are
mathematically similar to the existing images. Meanwhile, the discriminator is
fed photos without being told whether they are from the original dataset or
from the generator's output; its task is to identify which photos have been
synthetically generated.
As the two networks iteratively work against one another—the generator
trying to fool the discriminator, the discriminator trying to suss out the
generator’s creations—they hone one another’s capabilities. Eventually the
discriminator’s classification success rate falls to 50%, no better than random
guessing, meaning that the synthetically generated photos have become
indistinguishable from the originals.
One reason deepfakes have proliferated is the machine learning community’s
open-source ethos: starting with Goodfellow’s original paper, whenever a
research advance in generative modeling occurs, the technology is generally
made available for free for anyone in the world to download and make use of.
Given that deepfakes are based on AI in the first place, some look to AI as a
solution to harmful deepfake applications. For instance, researchers have
built sophisticated deepfake detection systems that assess lighting, shadows,
facial movements, and other features in order to flag images that are
fabricated. Another innovative defensive approach is to add a filter to an
image file that makes it impossible to use that image to generate a deepfake.
A handful of startups have emerged that offer software to defend against
deepfakes, including Truepic and Deeptrace.
Yet such technological solutions are not likely to stem the spread of deepfakes
over the long term. At best they will lead to an endless cat-and-mouse
dynamic, similar to what exists in cybersecurity today, in which breakthroughs
on the deepfake detection side spur further innovation in deepfake generation.
The open-source nature of AI research makes this all the more likely.
To give one example, in 2018 researchers at the University of Albany
published analysis showing that blinking irregularities were often a telltale
sign that a video was fake. It was a helpful breakthrough in the fight against
deepfakes—until, within months, new deepfake videos began to emerge that
corrected for this blinking imperfection.
“We are outgunned,” said Farid. “The number of people working on the video-
synthesis side, as opposed to the detector side, is 100 to 1.”
The Path Forward
Looking beyond purely technological remedies, what legislative, political, and
social steps can we take to defend against deepfakes’ dangers?
One tempting, simple solution is to pass laws that make it illegal to create or
spread deepfakes. The state of California has experimented with this
approach, enacting a law last year that makes it illegal to create or distribute
deepfakes of politicians within 60 days of an election. But a blanket deepfake
ban has both constitutional and practical challenges.
The First Amendment of the U.S. Constitution enshrines the freedom of
expression. Any law proscribing online content, particularly political content,
risks running afoul of these constitutional protections.
“Political speech enjoys the highest level of protection under U.S. law,” said
law professor Jane Kirtley. “The desire to protect people from deceptive
content in the run-up to an election is very strong and very understandable,
but I am skeptical about whether they are going to be able to enforce [the
California] law.”
Beyond constitutional concerns, deepfake bans will likely prove impracticable
to enforce due to the anonymity and borderlessness of the Internet.
Other existing legal frameworks that might be deployed to combat deepfakes
include copyright, defamation and the right of publicity. But given the broad
applicability of the fair use doctrine, the usefulness of these legal avenues may
be limited.
In the short term, the most effective solution may come from major tech
platforms like Facebook, Google and Twitter voluntarily taking more rigorous
action to limit the spread of harmful deepfakes.
Relying on private companies to solve broad political and societal problems
understandably makes many deeply uncomfortable. Yet as legal scholars
Bobby Chesney and Danielle Citron put it, these tech platforms’ terms-of-
service agreements are “the single most important documents governing
digital speech in today’s world.” As a result, these companies’ content policies
may be “the most salient response mechanism of all” to deepfakes.
A related legislative option is to amend the controversial Section 230 of the
Communications Decency Act. Written in the early days of the commercial
Internet, Section 230 gives Internet companies almost complete civil
immunity for any content posted on their platforms by third parties. Walking
these protections back would make companies like Facebook legally
responsible for limiting the spread of damaging content on their sites. But
such an approach raises complex free speech and censorship concerns.
In the end, no single solution will suffice. An essential first step is simply to
increase public awareness of the possibilities and dangers of deepfakes. An
informed citizenry is a crucial defense against widespread misinformation.
The recent rise of fake news has led to fears that we are entering a “post-truth”
world. Deepfakes threaten to intensify and accelerate this trajectory. The next
major chapter in this drama is likely just around the corner: the 2020
elections. The stakes could hardly be higher.
“The man in front of the tank at Tiananmen Square moved the
world,” said NYU professor Nasir Memon. “Nixon on the phone cost him his
presidency. Images of horror from concentration camps finally moved us into
action. If the notion of not believing what you see is under attack, that is a
huge problem. One has to restore truth in seeing again.”
