TECHNOLOGY
Deepfakes: More Frightening Than Photoshop
on Steroids
By Judge Herbert B. Dixon Jr. (Ret.)
would use the term “shallow fake” or “cheap
fake” to describe a minor adjustment such
as slowing down or speeding up a video,
but this type of video alteration can still
deceive. Do you recall the spring 2019
video circulating of House Speaker Nancy
Pelosi slurring her words? Many people who
saw the video spread misinformation on
their social media accounts of their belief
that Pelosi was drunk, was the victim of a
stroke, or had suffered some other loss of
mental acuity.1
A 2018 incident that went viral on
social media involves a video of CNN
White House Reporter Jim Acosta moving
his body and arms to avoid a White House
intern’s effort to physically take away a
microphone from Acosta’s possession.
ASSOCIATED PRESS PHOTO Some enterprising video editor, apparently
not a fan of Acosta, deleted Acosta’s state-
ment of “Pardon me, ma’am” and sped up
the video during that portion of the inci-
S
eeing is believing, so they say. How- successfully stitched or imposed the faces of dent, making it appear that Acosta’s
ever, any truism associated with this celebrities onto the bodies of people in por- actions were physically aggressive against
ancient statement may disappear nographic videos. The resulting videos were the intern.2
from our ethos. Have you heard about shockingly realistic. Although a few discrim-
deepfakes? inating viewers could spot imperfections or
Judge Herbert
A deepfake is video created or altered telltale signs of fraud, the resulting videos
B. Dixon Jr.
using digital means with the aid of artificial were nevertheless unsettling—especially to
retired from the
intelligence (AI). With deepfakes, persons the victims whose identities were falsely
Superior Court
appear to do or say things that did not hap- depicted in the pornographic productions.
of the District
pen. It is sometimes difficult for an expert The AI-driven software used to create the
of Columbia
to distinguish a deepfake from an unaltered phony videos employed a concept known as
after 30 years of
video. In other words, the world has turned deep learning to accomplish the deception.
service. He is a
on its head with a new governing princi- Let me give a few simplified descriptions of
former chair of
ple—“Do not necessarily believe what you the technology used to create these decep-
both the National Conference of State
see!” The dangers presented by deepfakes tive videos.
Trial Judges and the ABA Standing
are more frightening than Photoshop on
Committee on the American Judicial
steroids. How did we get here? Merely Slowing or System and a former member of the
The term deepfakes comes from the name Speeding a Video Does Techshow Planning Board. You can reach
of a Reddit contributor who surprised the Not Create a Deepfake him at [email protected]. Follow
technology community in 2017 when, using It takes more than a rudimentary modifi- Judge Dixon on Twitter @Jhbdixon.
publicly available AI-driven software, he cation of a video to create a deepfake. Some
SUMMER 2019 • The
Judges’ Journal 35
Published in The Judges' Journal, Volume 58, Number 3, Summer 2019. © 2019 by the American Bar Association. Reproduced with permission. All rights reserved. This information or any portion thereof
may not be copied or disseminated in any form or by any means or stored in an electronic database or retrieval system without the express written consent of the American Bar Association.
Bringing a Single Image to technology is a new paradigm—a game AI-driven software package detects forger-
Life Shows the Capability of changer. The product of today’s deepfake ies. Through a reiterative process, the first
Deepfake Technology technology has become unnervingly real- software package creates and modifies fake
Researchers in Russia (where else but Rus- istic. Because of the advances in deepfake videos until the other software package can
sia?) at Moscow’s Samsung AI Center and technology, members of the U.S. Congress no longer detect the forgery. This is known
Skolkovo Institute of Science and Tech- have requested a formal report from the as “unsupervised learning”—when
nology used AI-driven software to animate Director of National Intelligence.5 In addi- machine-language models teach them-
single images, including the Mona Lisa tion, the military, through the Defense selves. The larger the database of source
painting by the Italian Renaissance artist Advanced Research Projects Agency images, the easier it is for the deepfake cre-
Leonardo da Vinci, the 1665 oil paint- (DARPA), has embarked on research ator to produce a convincing deepfake. For
ing of Girl with a Pearl Earring by Dutch efforts with numerous academic and cor- this reason, videos of high-profile politi-
Golden Age artist Johannes Vermeer, and porate institutions to develop technology cians and Hollywood celebrities have been
the 1883 Portrait of an Unknown Woman by that can detect deepfakes.6 A concern used frequently to create deepfakes, as there
Russian artist Ivan Kramskoi. These efforts often expressed by the researchers is that is a substantial supply of publicly available
went far beyond the work of cartoon ani- continuing advances in deepfake technol- video footage to train the deepfake creat-
mators. The researchers produced a paper ogy may render it impossible for one ing system.
entitled “Few-Shot Adversarial Learning AI-driven system to detect a video created Said another way, the deepfake-creating
of Realistic Neural Talking Head Models”3 or modified by another AI-driven system. software studies the statistical patterns in a
describing the technology. The researchers Creating a deepfake involves modifying data set, such as a set of images or videos,
also created a video4 that shows the results images or videos using a machine-learning and then generates convincing fake videos.7
of their work making the iconic still images technique called a “generative adversarial The deepfake detection software then tries
move and talk as if they were real people. network” (GAN). The AI-driven software to distinguish between real and fake exam-
detects the way a subject moves his or her ples. Multiple iterations of the process enable
Creating an Ultra-Realistic mouth and face from the source images and the deepfake-creating software to produce
Deepfake duplicates those movements on the subject a result that is increasingly difficult for the
The phrase “ultra-realistic deepfake” is a of another video. To understand how a deepfake detection software to detect.
misnomer because, even absent a generally deepfake product is improved, imagine that Because the creating AI-driven system is
accepted definition, a deepfake is a decep- a deepfake creator has a substantial data- designed to produce a deepfake that the
tive video showing a subject doing or saying base of videos and two AI-driven software detecting AI-driven system is unable to dis-
things that did not occur. Although Hol- packages. One software package trains tinguish, some commentators suggest the
lywood for years has used highly skilled itself on the database of source images and eventuality that the technology to detect
artists to alter video footage, deepfake creates video forgeries. The second deepfakes may always be in catch-up mode.8
36 The
Judges’ Journal • VOL. 58 NO. 3
Published in The Judges' Journal, Volume 58, Number 3, Summer 2019. © 2019 by the American Bar Association. Reproduced with permission. All rights reserved. This information or any portion thereof
may not be copied or disseminated in any form or by any means or stored in an electronic database or retrieval system without the express written consent of the American Bar Association.
Computer forensics experts say that judges and
litigators are not thinking enough about how
to address deepfake evidentiary issues.
AI-Driven Software Can Create platform or speaking of the candidate’s provides no definitive method to detect
Images of Fake People, Change deceptive embrace of a policy he does not deepfakes? Will our society evolve to a new
Facial Expressions, and Imitate support. The video could cause serious paradigm in which no video is accepted as
Voices harm to the candidate’s electability. proof of what it purports to show? Will the
Although the term deepfake video has been Consider the possible adverse influence phrase “seeing is believing” cease to exist?
used throughout this column, audio forger- on a shareholder’s meeting called to con- “The proof is in the pudding”—whatever
ies are often a part of the process. An AI sider a major business acquisition if, just that means. n
startup in Montreal, Canada, has demon- before the meeting, a deepfake video sur-
strated the ability to synthesize a person’s faces of the CEO musing about his Endnotes
voice with just a one-minute recording of upcoming riches from the merger and the 1. Sarah Mervosh, Distorted Videos of Nancy
the original. Adobe has come up with the sucker shareholders who do not understand Pelosi Spread on Facebook and Twitter, Helped by
unique name of “Photoshop for audio.” what is going on. Trump, N.Y. Times (May 24, 2019), https://nyti.
Its software technology, “VoCo,” requires Lastly, judges and lawyers, consider ms/2ZvJ8V5.
merely a 20-minute sample of someone’s what actions should be taken during a pre- 2. Drew Harwell, White House Shares Doctored
voice, which the AI-driven software ana- trial conference where (1) a party offers an Video to Support Punishment of Journalist Jim Acosta,
lyzes and learns to mimic. After that, the exhibit of a cell phone video disclosed dur- Wash. Post (Nov. 8, 2018), https://wapo.
Photoshop AI-driven software will pro- ing discovery that supports the offering st/2x6iaXW.
duce the words in the targeted person’s party’s position of an agreement reached 3. Egor Zakharov et al., Few-Shot Adversarial
voice from a typed statement.9 Even more, by the two parties, (2) the offering party Leaning of Realistic Neural Talking Head Models,
AI-driven software using machine-learn- will testify affirmatively concerning the https://bit.ly/30NtiGW.
ing algorithms has created fake images of authenticity and accuracy of the video, and 4. Egor Zakharov et al., Few-Shot Adversarial
faces that don’t belong to real people and (3) the opposing party will testify that he Learning of Realistic Neural Talking Head Models,
modified images by changing a frown into never said the words portrayed in the video. YouTube (May 21, 2019), https://bit.ly/2WnVRer.
a smile, night into day, and summer into With the constant advancements in 5. Donie O’Sullivan, When Seeing Is No Longer
winter. deepfake technology, all of the above Believing, CNN Bus. (2019), https://cnn.
examples are plausible. Numerous experts it/2DCzQhJ.
Using Deepfakes to on this subject agree that deepfakes created 6. Will Knight, The US Military Is Funding an
Create Mischief by AI-driven software can be nearly impos- Effort to Catch Deepfakes and Other AI Trickery,
Political pundits say that deepfakes may sible to detect and the number of fake MIT Tech. Rev. (May 23, 2018), https://bit.
cause chaos during the 2020 presidential videos will grow with advances in free, ly/2IYzLIL.
elections. Business consultants say deep- readily available tools to create them. 7. Id.
fakes will undermine future business 8. See, for example, statements by David Gun-
ventures. Computer forensics experts say Final Thoughts ning, DARPA project manager, reported in id., and
that judges and litigators are not thinking In today’s climate of “fake news” claims and in J.M. Porup, How and Why Deepfake Videos
enough about how to address deepfake evi- increasing awareness of forged signatures Work—and What Is at Risk, CSO (Apr. 10, 2019),
dentiary issues when they show up in a few and “photoshopped” images, it is possible https://bit.ly/2Pmk9jc.
yea r s. C on sider t he follow i n g that the public may be naturally skeptical 9. Harmon Leon, Why AI Deepfakes Should
possibilities. of certain types of evidence and fully pre- Scare the Living Bejeezus Out of You, Observer
What would happen if, just before an pared to consider claims that a video does (June 12, 2019), https://bit.ly/2Zp4kMu; Scott
election, a deepfake video surfaces of a not accurately reflect what the subject of Amyx, Using Artificial Intelligence for Forgery: Fake
political candidate supporting a policy the video actually said and did. What is Could Be Eerily Real, IoT Agenda (July 26, 2017),
opposed to the candidate’s primary the answer to this dilemma if technology https://bit.ly/2X146cV.
SUMMER 2019 • The
Judges’ Journal 37
Published in The Judges' Journal, Volume 58, Number 3, Summer 2019. © 2019 by the American Bar Association. Reproduced with permission. All rights reserved. This information or any portion thereof
may not be copied or disseminated in any form or by any means or stored in an electronic database or retrieval system without the express written consent of the American Bar Association.
