IEEE Technology and Society Magazine - December 2019 - 7

deepfake technique of image synvideo2 of U.S. House of Reprethesis uses a GAN to superimpose
sentatives speaker Nancy Pelosi
existing images and videos of huappearing to slur drunkenly was
man subjects onto corresponding
v iewe d m i l l i o n s of times and
source images or videos, so that not
tweeted by the U.S. President, and
only is the discriminative network
although the video is demonstrably
unable to distinguish between real
a hoax, the tweet remains undeletand synthetic data, but neither can
ed. Secondly, as has already been
a human observer.
demonstrated, on mainstream media,
Disturbingly, but not altogether
where fact checking is secondary
unsurprisingly, deepfake pornoto the pressures of "churnalism"
graphic videos have been made,
[9] and is ruthlessly exploited by the
where images of famous people's
merchants of doubt [10], misinforfaces have been superimposed
mation can be even more rapover the original actors and actressidly spread by social media, where
es (depressingly, it is nearly always
veracity is a poor third to the willthe actresses). The software tools
ingness to believe and the desire
have also been made more readfor likes, sha res, and retweets.
ily available, such that with access
Thirdly, whatever national laws are
to a sufficiently powerful compassed do nothing to prevent transputer and enough data about the
national interference, whether it
target, it is possible for such vidis simply a financial incentive based
eos to be routinely made
by "amateur" developers
about less-famous people.
For some, this involves (for
Freud-knows-what reason)
If what is fake can be made to be
superimposing themselves
on the video; for others,
seen as real, then what is real can
it involves super impos ing another person (e.g.,
be made to be seen as fake.
for reasons of "revenge"),
which can be very difficult
on "clickbait" or the willful subto counteract, and very damagversion of a foreign power. Timing
ing to personal and professionhere is everything: the release of
al reputations.
deepfake videos (like the release of
However, distressing as this (mis)
fake news in the last few weeks of
use of the technology is, a number
the U.K. Brexit referendum in 2016)
of writers have argued that democcould flip an election before the
racy itself is under (yet another) threat
debunking catches up. Finally, the
from deepfake videos.
technolog y could even be used
The Potential Damage to
for "false flag" operations: make
Democracy
and relea se a fake video, then
It has been argued that deepfake
blame (and so discredit) the political
videos present a number of
opposition for so doing.
pos sible threats to politics, poliAfter the fall of the communist
ticians and democratic political
regimes and the end of the cold war,
regimes [8]. Firstly, deepfake vid2
Although this particular video does not seem
eos could be used to create comto have been created using deepfake techpromising material of politicians:
niques, it is symptomatic of what could be
achieved by a deepfake video.
for example, the digitally-altered

or has financial motivations that are
not openly revealed.
However, recent advances in artificial intelligence and the continually
decreasing cost of increasing computer power have made it possible
to develop software that can "do" to
video what photoshop could "do" to
photographs. The advance in artificial intelligence is based on a class of
machine learning algorithm called a
generative adversarial network (GAN)
[7]. In a GAN, two artificial neural
networks compete with each other
in a "game," where one network, the
generative network, generates data
that is evaluated by the second network, the discriminative network.
The discriminative network is trained
on a known dataset until it achieves
acceptable accuracy. Then, the generative network learns to generate
new data with the same statistical
characteristics as the original (training) data, depending on whether or
not the generated data succeeds in
being classified as "real" by the discriminative network. The outcome
of the discriminative network's decision is used to fine tune the training
of both networks. Eventually, both
networks will plateau, in the sense
that they cannot improve any further,
at which point, in the preferred outcome, the discriminative network is
unable to differentiate between real
data and synthetic data produced by
the generative network (the synthetic
data are sufficiently "realistic" to be
indistinguishable from real data).
While GAN have already achieved
significant beneficial results, a
potentially more insidious ap plication is the use of G A N to
generate deepfake1 videos. The
1

The "deep" in "deep learning" refers to the
number of layers in the artificial neural network.
Since both the generative and discriminative
networks in a GAN tend to have several layers, they are deemed "deep," and since they
are being used to produce counterfeit images
or videos, the term "deepfake" was used to
describe the products of this technique.

DECEMBER 2019

∕

IEEE TECHNOLOGY AND SOCIETY MAGAZINE

IEEE Technology and Society Magazine - December 2019

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - December 2019