777 assignment 2

“All Around Me Are Synthetic Faces”: The Mad World of AI-Generated Media

Lucas Whittaker Queensland University of Technology

Tim C. Kietzmann Radboud University & Cambridge University

Jan Kietzmann University of Victoria

Amir Dabirian KTH Royal Institute of Technology & California

State University

Abstract—Advances in artificial intelligence and deep neural networks have led to a rise in

synthetic media, i.e., automatically and artificially generated or manipulated photo, audio,

and video content. Synthetic media today is highly believable and “true to life”; so much so

that we will no longer be able to trust what we see or hear is unadulterated and genuine.

Among the different forms of synthetic media, the most concerning forms are deepfakes

and general adversarial networks (GANs). For IT professionals, it is important to

understand what these new phenomena are. In this article, we explain what deepfakes

and GANs are, how they work and discuss the threats and opportunities resulting from

these forms of synthetic media.

ANTICIPATING A “MAD WORLD” & BARACK OBAMA’S PUBLIC service announce- ment starts with the usual backdrop of American

flags within the Oval Office. His distinctive vocal

pauses and hand gestures lend credibility to

his address about the modern threat of digital

technologies and artificial intelligence (AI). But

suddenly, his own address starts to take a

strange turn, culminating in an alarming and out-

of-character statement: “President Trump is a

total and complete dipsh%t.” Wait, what? Obama

pauses to clarify, “See, now I would never say

these things, at least not in a public address. But

someone else would.”

Digital Object Identifier 10.1109/MITP.2020.2985492

Date of current version 11 September 2020.

Department: IT TrendsDepartment: IT Trends

90 1520-9202 � 2020 IEEE Published by the IEEE Computer Society IT Professional

This video then introduces comedian Jordan

Peele, who imitates Barack Obama’s voice. Peele’s

facial expression and mouth movements mor-

phed into Obama’s using FakeApp, a free tool. In

the video, Peele leverages the credibility of the

44th U.S. President’s face and voice to warn the

viewer that “how we move forward in the age of

information is going to be the difference between

whether we survive or whether we become some

kind of [expletive] dystopia.” 1

Published by Buzzfeed, the video reveals how

so-called “deepfakes” can be created to imper-

sonate public figures by others, including those

with a potentially harmful agenda. But these

deepfakes are not the only highly precarious

media that test what we should believe to be

true. Contrary to deepfakes and their ability to

generate images that combine features of a

source and target image (e.g., Peele’s and

Obama’s faces), generative adversarial net-

works, or GANs for short, can create entirely

new and lifelike, naturalistic content (e.g., faces

of people who do not exist).

The recent increase of convincing, highly

believable, and “true to life” deepfakes and GAN-

generated media prepares us for a mad world in

which we will no longer be able to trust that

what we see or hear is unadulterated and genu-

ine—a world in which “synthetic media” will

force us to re-evaluate our very perception of

reality.

SYNTHETIC MEDIA The examples above illustrate a world where

we not only live with the common digital content

that surrounds us, for instance advertisements

enhanced with apps like Photoshop or movies

with computer-generated imagery (CGI) like spe-

cial effects, but also with synthetic media. With

the term “synthetic media,” we refer to all auto-

matically and artificially generated or manipu-

lated media.

It is important to keep in mind that syn-

thetic media is an umbrella terms for a host of

other media. Types of synthetic media, for

instance, include synthesized audio (e.g., Goo-

gle Duplex), virtual reality, and even advanced

digital-image creation (beyond CGI, where

expert systems are increasingly capable of

automatically producing realism on a vast

scale). The two branches that cause the big-

gest concern currently though, are deepfakes

and GANs.

WHAT ARE DEEPFAKES AND HOW DO THEY WORK?

The deepfake phenomenon initially gained

public awareness in late 2017 within online hob-

byist communities. A Reddit user with the user-

name “deepfakes” used AI algorithms to insert

the faces of famous actresses into adult videos.

After the computer code necessary to generate

deepfakes became publicly available, online

communities began to create more and more

deepfakes, including the abovementioned Peele/

Obama video. Before December of 2017, the

term “deepfake” did not even register on Google

Trends (a website that analyzes the popularity

of top search queries in Google Search), but ever

since it has been on a steady incline. Not even

two years after the first appearance, in the fall of

2019, CNN reported that there were almost

15 000 deepfakes, nearly all of them were non-

consensual instances of deepfake porn. 2 How-

ever, there are many other applications of the

technology used to create deepfakes, at this

point mainly (but not exclusively) for entertain-

ment purposes.

At first, these meme-like videos created by

hobbyists were usually crudely constructed

with obvious AI manipulation—but still demon-

strated the immense potential of deepfake tech-

nology. For instance, a curious obsession with

planting Nicolas Cage in popular movies arose,

with viral videos depicting Nicolas Cage in roles

such as Indiana Jones in Raiders of the Lost Ark

and Lois Lane in Batman versus Superman. The

past two years have seen deepfakes move

beyond online communities and become created

with increasing sophistication across an array of

formats. 3

� Photo deepfakes can be created by swapping

out faces and bodies within images.

� Audio deepfakes can create, alter and imitate

voices, from audio sources or from text.

� Video content can be (deep)faked, with a mov-

ing face being able to be swapped or morphed

September/October 2020 91

with another, or the bodily movement of a per-

son being replaced by someone else’s.

� To create the most perhaps sophisticated

deepfakes, the above approaches can be

combined to manipulate mouth movements

and facial expressions alongside audio mate-

rials to have someone convincingly say

things which they have never said.

At their core, deepfakes are the product of AI

and the machine learning technique of “deep

learning,” which is used to train deep neural net-

works (DNNs). On a very abstract level, DNNs

resemble some computational principles also

found in the brain. Their synthetic neurons, bet-

ter known as “units,” perform simple nonlinear

operations. Yet, when setup as a network of

thousands and millions of units, these simple

functions combine to perform complex feats,

such as object recognition, language translation,

or robot navigation. Importantly, the function of

a whole network of units is determined by the

pattern of its unit connections. To drive a net-

work to perform the desired function, the con-

nection strengths between units are adjusted via

training on large sets of example data. In the

case of deepfakes, such DNNs are trained as a

central part of AI systems that automatically

merge, combine, replace, and superimpose

images and video over a targeted video to create

a hyperrealistic, yet fake alternative content. 3

Due to the requirement for large sets of training

data that enable DNNs to adjust the millions of

unit connections, individuals in the public eye,

such as celebrities, actors, and politicians are

particularly susceptible to becoming targeted by

deepfakes as vast amounts of audio and visual

content of them are widely available online.

To date, most available deepfakes are cre-

ated by using a specific network architecture,

known as an “autoencoder,” which specializes in

generating real-looking, yet fake, facial images of

a target person (please see the work done by

Kietzmann et al. 3 for an in-depth explanation of

the underlying technology). It is important to

keep in mind that autoencoders not only work

with faces, though, and can be trained on a vari-

ety of different content (e.g., voices of people).

To use autoencoders for deepfakes, they are fed

with large numbers of facial images of a given

person. Their task is to recreate the same image

that they are presented with. While seemingly

simple, the task is complicated by the design of

the network architecture, which requires the

image information to pass through an informa-

tion bottleneck that applies significant compres-

sion. To solve this complex task, autoencoders

learn to first extract more abstract facial charac-

teristics and emotional expression from the

input image (known as the “encoder” part of the

autoencoder), and from there generate the out-

put image (known as the “decoder”) that

matches the input. As the geometry of faces is

quite stereotypical, they lend themselves partic-

ularly well for such compression. As a result, the

decoder part of the autoencoder can generate

any image of the person it was trained on,

whether such image previously existed or not.

To create deepfakes with autoencoders, two

separate networks are trained, one for each per-

son. To stick with our introductory example of

Obama and Peele, deepfake artists would train

two autoencoders based on image sets of either

person. The trick that enables a transition from

the image of one person, say Peele, to another,

say Obama, is that the two networks are linked

such that the connection strengths of the

encoder part of both networks are kept identical.

This implies that the network learns to not spe-

cialize on one person, but learns to recognize

more general facial features that are common

across the two people.

But how does this setup help to create deep-

fakes? The linked encoder parts of the autoen-

coder allow users to input an image of Peele into

the encoder, and subsequently generates a

matching image of Obama as the output. Impor-

tantly, this newly generated image of Obama will

have the same facial and emotional expression

as the Peele input image. This renders it particu-

larly easy to take the newly created image and

copy it into the input image, thereby swapping

the two people.

As per Figure 1 (adapted from the work by

Kietzmann et al. 3 ), to create a deepfake image of

a target person (Obama), step one consists of

extracting the face from an image showing an

actor that performs the wanted expression. As

step two, the autoencoder then translates this

source image into a novel image of the target

IT Trends

92 IT Professional

that directly matches the input actor in facial

and emotional expression and head orientation.

Importantly, the resulting generated image will

not just copy over the pixels, say the smiling

lips, of the actor, but generate a smile that is

unique to the target person (e.g., the exact way

Obama would smile). Step three consists of tak-

ing the generated image and copying it back into

the original image. Instead of the actor, now the

target person is shown in the same pose and

expression.

WHAT ARE GENERATIVE ADVERSARIAL NETWORKS AND HOW DO THEY WORK?

While DNN architectures for deepfakes

specialize in exchanging one person for another,

DNNs can also be used to generate entirely novel

content. This is the case for GANs, which,

like autoencoders, originate from the family

of “generative models” (as opposed to

“discriminative” networks that can take images

as input but are tasked to describe or classify

what is being shown). Generative models can be

used to create novel data that resembles the

data the networks were trained on. As general as

this statement may seem, the possibilities for

creating content from GANs are endless. GANs

were introduced in 2014, 4 and while the first

results, generating faces and written digits, were

low in resolution, they clearly demonstrated the

powerful capabilities of this new style of training

networks. Today, GANs are used to generate

high-resolution facial images of nonexistent

people, to create pictures of imaginary breeds of

dogs, to assist artists in their paintings, for

example, by filling in colored details into line

drawings, and can take existing images of low

quality and generate high-resolution versions of

them by “fantasizing” details that were not pres-

ent in the original. By design of the underlying

network architectures, the speciality of GANs is

to generate entirely novel content that is strik-

ingly similar to original, real images of people

and things (see Figure 2 for examples of GAN-

generated content). This renders them equally

powerful and dangerous.

The key to understanding DNNs that enable

the creation of deepfakes is that they in fact con-

sist of two interlinked networks, each of which is

trained to become an expert at generating one of

the two people to be swapped later on. GANs,

too, consist of two deep networks. However,

instead of working together, they perform oppos-

ing roles. One network, the “generator,” is used

to generate fake content from random input, say,

images showing human faces. A second network,

known as the “discriminator,” is fed with both

fake and real content, and it is trained to be suc-

cessful at determining whether a given input is

fake. Akin to an arms race between money coun-

terfeiters and the police, the generator improves

if its generated fake image was detected as such.

If instead the discriminator was fooled into mis-

taking a fake image as real, it learns from this mis-

take to prevent future error. The interplay of

these two networks, both of which want to out-

play the respective other, leads to incremental

improvements on both sides. Once the networks

Figure 1. Three step procedure for creating deepfakes (adapted from the work by Kietzmann et al).3

September/October 2020 93

are trained, the discriminator is commonly dis-

carded, and the generator is used to generate

new content originating from previously unused

random noise as input.

THREATS AND OPPORTUNITIES OF SYNTHETIC MEDIA

It is abundantly clear to the reader of this

article that the technologies that enable syn-

thetic media are very powerful. Synthetic media,

and especially deepfakes and GANs, provide

numerous threats and opportunities that we

ought to be aware of—as we will need to ques-

tion whether what we see and hear is authentic

and unadulterated by others. Below is a nonex-

haustive list of some threats and opportunities

of synthetic media—a list that will certainly

change and grow as deepfakes and GANs find

wider application.

Threats

Unsurprisingly, the ability for users to create

artificial realities using deepfakes and GANs is

fraught with malicious potential. In particular,

the believability and accessibility of deepfakes

play significant roles within the growing threat

that the technology represents. Deepfakes are

being increasingly difficult to distinguish from

authentic video, and the barriers to creation

have lowered significantly with widely available

mobile apps such as FakeApp and Zao enabling

unskilled individuals to create their own deep-

fakes. Disconcertingly, the growing simplicity

involved in creating convincing deepfakes, com-

bined with our increasingly digitally documented

Figure 2. Images created via GANs.5,6

IT Trends

94 IT Professional

lives, will heighten the potential for them to be

used for malicious purposes such as blackmail,

intimidation, sabotage, harassment, defamation,

revenge porn, identify theft, and bullying. 3, 7

The generation of nonconsensual videos is a

disturbing application of deepfakes and GANs. It

all started with actresses such as Natalie Port-

man and Gal Gadot in addition to other female

public figures such as Michelle Obama, Ivanka

Trump, and Kate Middleton who became victims

of nonconsensual deepfake insertion into adult

film scenes. To date, women remain the main vic-

tims of deepfakes. Content distributors such as

Reddit, Pornhub, and Gyfcat have banned artifi-

cially generated pornography from their plat-

forms, however, sexually explicit and degrading

deepfake content can still circumvent these

mainstream platforms and be widely distributed.

This occurred in the case of Rana Ayyub, a self-

described antiestablishment investigative jour-

nalist in India, who was the victim of a viral por-

nographic deepfake maliciously designed to

humiliate her in the public eye which was widely

circulated via Indian WhatsApp groups. 8 Even

those outside the public spotlight can be tar-

geted, such as the case of an 18-year-old woman

who reverse image-searched her photo out of

curiosity only to find hundreds of images of her

face inserted into pornographic scenes. After

trying to remove the images and becoming an

advocate against image-based sexual abuse, she

became the victim of a deepfake pornographic

video which was uploaded to numerous web-

sites. 9 Deepfakes are the next sinister breed of

revenge porn which can make everyone poten-

tial targets even if they have never taken explicit

photographs or videos of themselves (!).

Deepfakes and GANs represent the next gen-

eration of fake news and threaten to further

erode trust in online information. Artificially gen-

erated video poses a greater threat of disinfor-

mation compared to text or image-based fake

news due to how difficult the digital alteration

can be to spot. People are inclined to believe

what they see, and when combined with factors

such as confirmation biases and social media

echo chambers which facilitate the propagation

of fake news, artificially generated content will

further fuel the fake news crisis with their ability

to undermine truth and confuse viewers. As

malicious deepfakes become more common-

place, the public may even begin to lose trust

within news and factual information. 7

Disturbingly still, deepfakes can further

erode democracy and truth by creating a “liar’s

dividend,” acting as a plausible scapegoat to gen-

uine recordings of misbehavior and corruption.

One particularly concerning scenario is the dis-

missal of authentic video evidence documenting

human rights abuses, arguing that the content

has been faked. Without proper detection tech-

nology, every video, even those that are 100%

accurate recordings of real events, can be

passed up as fake—everyone will have plausible

deniability for any event caught on video.

The political sphere is particularly vulnerable

to deepfakes and GANs intentionally created to

deceive. The United States Senate recently

passed legislation to better understand how for-

eign governments and domestic groups use

deepfakes to damage national security and

threaten American democracy. The ability for

GANs to generate the faces of people who do not

exist can result in political astroturfing. Face-

book recently removed more than 900 fake

accounts using GAN-created profile photos

which circulated pro-Trump messages via social

media. 10

Deepfakes will be yet another tool within

state-sponsored disinformation campaigns to

interfere with elections and create civil unrest.

United States lawmakers have expressed con-

cern that deepfakes will soon be used by mali-

cious foreign actors, with Senator Marco Rubio

stating that deepfakes would be used in “the

next wave of attacks against America and West-

ern democracies.” 11

These deepfakes targeting

politicians could be legitimized using similar tac-

tics to those allegedly employed by Russia in

their interference within the 2016 U.S. Presiden-

tial Election by their use of troll-farms to circu-

late disinformation. Alarmingly, deepfakes and

GANs could be used by rogue agents such as ter-

rorist groups. Using minimal resources, artificial

realities could potentially be created of (for

example) United States soldiers discussing or

engaging in war-crimes while stationed in the

Middle East, to serve the agendas of rogue

agents attempting to radicalize and recruit indi-

viduals to join their cause. 7

September/October 2020 95

Set in the context of the private sector, syn-

thesized voice, created via deepfake technology,

can be used for fraudulent activity. Using AI-

based software to mimic the voice of a CEO,

fraudsters were able to successfully request the

transfer of $US243,000 into a fraudulent account

during a phone call with the firm’s CFO, who was

tricked to recognize the slight German accent

and vocal melody of his boss during the fraudu-

lent phone conversation. 12

Deepfakes also represent a major threat to

organizations and brands, adding further weight

to fake news stories already being fabricated to

target specific companies. Fake news can cause

significant economic impact, such as the fake

news article which intentionally misquoted

Pepsi’s CEO Indra Nooyi as saying that support-

ers of Donald Trump should “take their business

elsewhere,” resulting in calls for boycotts and a

3.75% decrease in Pepsi’s share price. 13

The

weight of realism that deepfakes can contribute

toward malicious agendas means that there is a

need to plan for reputational damage sustained

from, for example, a senior executive or figure-

head being targeted by deepfake to say

compromising statements.

Opportunities

Despite the insidious potential of deepfakes

and GANs, they have exciting potential to do

good. For instance, deepfakes can aid in the

removal of language barriers which can con-

strain cross-cultural video content distribution,

which might typically require supplementation

with subtitles. Social intervention campaigns

such as the Malaria Must Die campaign demon-

strates how deepfakes can transcend language,

with former English footballer David Beckham

seemingly speaking nine different languages in a

voice petition to end malaria. 14

Deepfakes are also being used to transcend

language barriers within political spheres. Manoj

Tiwari, President of India’s ruling Bharatiya Janata

Party, recently embraced deepfake technology to

directly speak to Hindi-speaking constituents.

Originally a recording in English that criticized his

political opponent, his video was consensually

translated via deepfake into Haryanvi, a Hindi

dialect widely spoken by his target voters. The

deepfake had reportedly reached approximately

15 million people in 5800 WhatsApp groups. 15

Deepfake technology is also giving a voice to

those who have lost theirs due to medical condi-

tions such as motor neuron diseases. Using simi-

lar deep learning principles employed to create

video deepfakes, Project Revoice (www.

projectrevoice.org) uses generative AI from

voice samples provided by vocally-paralyzed cli-

ents to create personalized synthetic voices.

The film industry can also greatly benefit from

the implementation of deepfake technology,

which can be used to de-age actors to a compara-

tive level of costly CGI effects. Netflix’s The Irish-

man used CGI to de-age actors such as Robert De

Niro, Al Pacino, and Joe Pesci to appear decades

younger than their current selves, which report-

edly drove the budget of the film to as high as

$US175 million. Using free software called DeepFa-

ceLab and only seven days, one YouTuber recre-

ated Netflix’s de-aging effects and released a

video comparing the CGI of the actors within

scenes from The Irishman to the deepfake version

of the actors. The deepfake recreation of the

scenes was highly convincing and has even been

hailed as superior to the costly CGI effects, which

were reported as being “distractingly bad” and

like “some hellish uncanny valley.” 16

Digital revival is an already established, yet a

controversial practice within the world of cin-

ema. CGI revivals include resurrecting Peter

Cushing as Grand Moff Tarkin in Rogue One: A

Star Wars Story and Paul Walker as Brian

O’Conner in Furious 7. Though an ethical and

potentially legal minefield, deepfakes can give

filmmakers a cost-effective alternative to CGI res-

urrection by utilizing the vast audio and visual

material of deceased actors. The Golden Age of

Hollywood may yet rise again.

The potential for deepfake resurrection has

already been realized beyond film. Tourist

attractions such as The Dal�ı Museum in

St. Petersburg, Florida have adopted deepfake

technology to breathe life into the Spanish surre-

alist who passed away in 1989. After pressing a

button next to a life-sized screen, Dal�ı leaves his

easel and approaches the visitor, talking to them

about his artwork and the museum. Upon leav-

ing the museum, Dal�ı appears to the visitor once

more to ask whether they would like a selfie with

IT Trends

96 IT Professional

him, where he takes a photo of himself with the

visitor which is delivered via SMS. 17

When

employed respectfully, such application of deep-

fake technology has incredible power to estab-

lish deeper connections with the deceased to

create emotional experiences and provide richer

insights into their lives.

Deepfakes offer boundless potential for per-

sonalized media creation. The highly popular

Chinese app Zao enables users to swap their

face over the actor within scenes from popular

movies and TV shows, allowing anyone to be a

star and share their creation with the world.

This deep level of personalization combined

with the increasingly accessible nature of

deepfakes offers organizations new ways of

engaging with customers. For example, a brand

could release an app allowing customers to

deepfake themselves into purpose-built adver-

tisements or scenarios to share with their

social networks.

Companies such as Artificial Talent (https://

artificialtalent.co/) specialize in the creation of

AI promotion models using GANs, with business

clients able to customize the physical character-

istics of the fashion models and the clothes they

wear to create promotional campaigns without

the need to cast human models. GANs can also

be used to generate images of humans condi-

tioned to form a specified pose. 18

Brands are

now able to generate models which are perfectly

cast to reflect their brand image or promotional

campaign without the need for a human model,

circumventing scheduling issues, costly photo-

shoots, controversies the model may find them-

selves within, and even aging.

Deepfakes and synthetic AI models generated

by GANs could be merged for the ultimate

personalization of online customer experiences,

such as online clothes shopping. Customers

could provide basic physical characteristics to

an online fashion store which generates their

own personalized avatar to use for online fash-

ion shopping. Further personalization could

even occur with the customer being able to

deepfake themselves as the model for a true

visualization only rivalled by the in-store dress-

ing room.

GANs can be used to stylistically transfer imag-

ery. By being trained on landscape photographs

and mimic collections of art styles (e.g., the entire

works of Monet and Van Gogh) GANs can generate

the original photographs in the style of Monet and

Van Gogh, offering new ways to replicate artistic

styles and generate new artwork. 19

Anyone can

become a GAN artist using Artbreeder (https://

www.artbreeder.com/), which allows users to cre-

ate and merge images to generate completely new

artistic pieces constructed by GANs.

GANs can also be used to create superresolu-

tion imagery from low resolution inputs. Though

the creation of high-resolution imagery from

lower resolution input is not new, the technol-

ogy can still struggle to remove noise and com-

pression artifacts. GANs can optimize this

process by creating a higher quality image than

one that ever existed—“fantasizing” details onto

the low resolution image. 20

GANs could there-

fore be used to augment current applications of

superresolution technology, which is used in

contexts such as satellite and aerial imaging,

medical image processing, facial image detec-

tion, text image improvement, and compressed

image and video enhancement.

There are a multitude of further GAN applica-

tions which we cannot fully explore within the

scope of this article, however, others include

image creation using only text input, context-

based pixel prediction to repair images with

entire sections removed, detecting online spam

reviews, and even music composition.

CONCLUSION The “good old days” of purely analog and

digital content creation and modification of

media are numbered. Synthetic media are mak-

ing fast inroads, and as technologies around

deepfakes and GANs keep improving, we will see

more and more applications that will challenge

our current understanding of what is real and

what is not, and that will put our sense of truth

and trust to a test.

The task for IT professionals is to improve

the technology of creating synthetic content

while at the same time ensuring that “fakes” will

be identifiable as such, the duty for lawmakers is

to update our laws about creating, distributing,

and sharing synthetic videos, images, or audio.

The big challenge for us as everyday consumers

September/October 2020 97

is to stop believing everything we see and hear

on social media, and instead to start thinking

critically again and to question the “evidence” in

front of us. All of this will take time, and until we

all arrive at a common understanding of how

synthetic media fit into our lives, it will be a mad

world!

& REFERENCES

1. D. Mack, “This PSA about fake news from Barack

Obama is not what it appears,” BuzzFeed News, 2018.

Accessed on: Mar. 26, 2020. [Online]. Available:

https://www.buzzfeednews.com/article/davidmack/

obama-fake-news-jordan-peele-psa-video-buzzfeed.

2. R. Metz, “The number of deepfake videos online is up

84 per cent in 10 months,” CTV News, 2019.

Accessed on: Mar. 28, 2020. [Online]. Available:

https://www.ctvnews.ca/sci-tech/the-number-of-

deepfake-videos-online-is-up-84-per-cent-in-10-

months-1.4627938.

3. J. Kietzmann, L.W. Lee, I.P. McCarthy, and

T.C. Kietzmann, “Deepfakes: Trick or treat?” Bus.

Horizons, vol. 63, no. 2, pp. 135–146, 2020.

4. I. J. Goodfellow et al., “Generative adversarial nets,” in

Proc. Adv. Neural Inf., 2014, pp. 2672–2680.

5. Thispersondoesnotexist. Accessed on: Mar. 28, 2020.

[Online]. Available: https://www.

thispersondoesnotexist.com.

6. A. Brock, J. Donahue, and K. Simonyan, “Large scale

GAN training for high fidelity natural image synthesis,”

in Proc. Int. Conf. Learn. Representations, 2019.

[Online]. Available: https://openreview.net/pdf?

id=B1xsqj09Fm

7. R. Chesney and D. Citron, “Deepfakes and the new

disinformation war: The coming age of post-truth

geopolitics,” Foreign Affairs, vol. 98, no. 1, pp. 147–

152, 2019.

8. R. Ayyub, “I was the victim of a deepfake porn plot

intended to silence me,” Huffington Post, 2018.

Accessed on: Mar. 28, 2020. [Online]. Available:

https://www.huffingtonpost.co.uk/entry/deepfake-

porn_uk_5bf2c126e4b0f32bd58ba316.

9. K. Melville, “The insidious rise of deepfake porn videos

— and one woman who won’t be silenced,” ABC

News, 2019. Accessed on: Mar. 28, 2020. [Online].

Available: https://www.abc.net.au/news/2019-08-30/

deepfake-revenge-porn-noelle-martin-story-of-image-

based-abuse/11437774.

10. M. Nu~nez, “Facebook removes hundreds of fake pro-

Trump accounts using AI-generated profile photos,”

Forbes, 2019. Accessed on: Mar. 28, 2020. [Online].

Available: https://www.forbes.com/sites/mnunez/2019/

12/20/facebook-removes-hundreds-of-fake-pro-

trump-accounts-using-ai-generated-profile-photos/

#170794956175.

11. D. O’Sullivan, “Lawmakers warn of ‘deepfake’ videos

ahead of 2020 election,” CNN Bus., 2019. Accessed

on: Mar. 27, 2020. [Online]. Available: https://edition.

cnn.com/2019/01/28/tech/deepfake-lawmakers/index.

html.

12. C. Stupp, “Fraudsters used AI to mimic CEO’s voice in

unusual cybercrime case,” Wall Street J., 2019.

Accessed on: Mar. 26, 2020. [Online]. Available:

https://www.wsj.com/articles/fraudsters-use-ai-to-

mimic-ceos-voice-in-unusual-cybercrime-case-

11567157402.

13. I. Liffreing, “So your brand is the victim of fake news.

now what?” PR Week, 2016. Accessed on: Mar. 26,

2020. [Online]. Available: https://www.prweek.com/

article/1416264/so-brand-victim-fake-news-what.

14. M. M. Die, “David Beckham speaks nine languages to

launch malaria must die voice petition,” Apr. 8, 2019.

Accessed on: Mar. 26, 2020. [Online]. Available:

https://www.youtube.com/watch?v¼QiiSAvKJIHo. 15. C. Jee, “An Indian politician is using deepfake

technology to win new voters,” MIT Technol. Rev.,

2019. Accessed on: Mar. 27, 2020. [Online]. Available:

https://www.technologyreview.com/f/615247/an-

indian-politician-is-using-deepfakes-to-try-and-win-

voters/.

16. C. Wood, “A deepfake artist’s attempt to make Robert

de Niro look younger in ‘The Irishman’ is being hailed

as superior to Netflix’s CGI,” Bus. Insider Aust., 2020.

Accessed on: Mar. 27, 2020. [Online]. Available:

https://www.businessinsider.com.au/deepfake-netflix-

correcting-the-irishman-de-ageing-tech-2020-1?

r¼US&IR¼T. 17. The Dali Museum, Behind the Scenes: Dali Lives,

May 08, 2019. Accessed on: Mar. 26, 2020. [Online].

Available: https://www.youtube.com/watch?

v¼BIDaxl4xqJ4. 18. L. Ma, X. Jia, Q. Sun, B. Schiele, T. Tuytelaars, and

L. Van Gool, “Pose guided person image generation,”

in Proc. Adv. Neural Inf., 2017, pp. 406–416.

19. J. Y. Zhu, T. Park, P. Isola, and A.A. Efros, “Unpaired

image-to-image translation using cycle-consistent

adversarial networks,” in IEEE I nt . Conf. Comp ut . Vis

ion, 2017, pp. 2223–2232.

IT Trends

98 IT Professional

20. C. Thomas, “Deep learning based super resolution,

without using a GAN,” Towards Data Sci., 2019.

Accessed on: Mar. 28, 2020. [Online]. Available:

https://towardsdatascience.com/deep-learning-

based-super-resolution-without-using-a-gan-

11c9bb5b6cd5.

Lucas Whittaker is currently working toward the

Ph.D. degree with the School of Advertising, Market-

ing and Public Relations and Research Student with

the Centre for Behavioral Economics, Society, and

Technology, Queensland University of Technology,

Brisbane City, QLD, Australia. His research explores

how visual deepfakes can potentially be weaponized

to target brands and investigates the role behavioral

biases play within the acceptance of visual deep-

fakes. Contact him at [email protected].

Tim C. Kietzmann is currently an Assistant Pro-

fessor with the AI Department, Donders Institute for

Brain, Cognition and Behavior (Radboud University),

Nijmegen, Netherlands. In addition, he is a Research

Associate with the MRC Cognition and Brain Science

Unit (University of Cambridge). In his interdisciplin-

ary work, he combines neuroscience with AI technol-

ogies to help better understand principles of neural

information processing in the brain. Contact him at

[email protected].

Jan Kietzmann is currently an Associate Profes-

sor with the Management Information Systems,

University of Victoria, Victoria, BC, Canada. He is

also an Associate Editor with Business Horizons.

His research focuses on organizational and social

perspectives related to emerging technologies.

He received the Ph.D. degree in Innovation and

Information Systems from LSE. Contact him at

[email protected].

Amir Dabirian is currently the Vice President for

the Division of Information Technology and as a Fac-

ulty Member with the Department of Marketing, Cali-

fornia State University, Fullerton, CA, USA. He is

currently working toward the Ph.D. degree with the

Department of Industrial Economics and Manage-

ment, KTH Royal Institute of Technology, Stockholm,

Sweden, with research interest in employer brand-

ing. Contact him at [email protected].

September/October 2020 99