Page 48 - Payout Magazine Online Volume 10.01

depictions of Al Pacino, Tom Cruise, and Arnold During its research, Adobe displayed pairs
Schwarzenegger when doing impressions of
the celebrities on a popular talk show.
These have all been entertaining curiosities.
Still, deepfake has also been used to create
more disturbing videos of politicians and
influential individuals, which highlight the
dangers should the technology be used for
more nefarious purposes.
Mark Zuckerberg appeared on camera
and supposedly said: “Imagine this for a
second: One man, with total control of billions
of people’s stolen data, all their secrets, their
lives, their futures.”
House Speaker Nancy Pelosi has also been
the victim of a video that was manipulated into
making her appear intoxicated.
What Can AI Software Do?
Adobe has been developing techniques for
spotting image editing for many years now.
However, most of their efforts were focused
on human editors using Photoshop tools to
manipulate an image directly.
For example, splicing images, removing
objects, and copying one part of an image
to use in another area are well-known photo
manipulation strategies. Human editors using
these techniques will disrupt the digital noise
captured by the camera’s digital sensor. The
editor of the image unwittingly leaves a digital
trail the software can detect to reveal the
forgery.
Adobes renewed AI efforts are focused
more on reverse engineering edits which have
been initially made by AI. The main focus for
Adobe and the research team is to restore the
public’s faith in digital media, in an age where
fakes and touch-ups are commonplace.
How the Research Team Trained
the AI
To create the AI, the team studied the Face
Away Liquify feature available in Photoshop.
The primary function of this tool is its ability to way of forged imagery and videos. As head
change the faces, eyes, and mouths easily.
A convolutional neural network (CNN) was
trained to analyze the visual data and pick up
the changes made to the faces in the image.
48 PAYOUTMAGAZINE
of images to the human subjects and the
neural network. One copy was the original, but
the other was a manipulated version.
The neural network proved itself incredibly
adept at picking up fakes. In tests, it managed
a score of 99 percent accuracy. In comparison,
the best human eyes could only maintain a
53% success rate.
It’s one thing to be able to detect the
forgeries, but the team went one step further
and gave the AI the ability to revert the images
to their original form. While the results are
promising, the reverted images are a close
approximation to the original. The software is
not yet able to remove all traces of the warping
effect introduced by Photoshop’s Liquify tool.
Fake images and videos are getting easier
to create and require nothing more than an
app, a couple of still photos, and a sufficiently
powerful smartphone. There is growing
importance in developing technology that can
more easily spot these manipulated images
and expose them for the frauds they are.
Unfortunately, the availability of such
sophisticated photo editing tools means
that the number of people trying to detect
such fakes are outnumbered 100 to 1 against
those creating fake images and videos. While
one side is busy working out how to identify
faked images, the other team is hard at work
developing techniques to create more realistic
and convincing forgeries.
The challenge is more significant than
just creating AI that can successfully root
out fake images. An even more substantial
problem may lie in getting people to trust that
the AI is accurate enough to be meaningful
– especially on content that is political, or in
some way designed to polarize the public.
Abode firmly believes that AI will play an
increasingly important role in validating digital
media, while also making it more challenging
for those wishing to spread fake news by
of Adobe Research, Gavin Miller says, “This
is an important step in being able to detect
certain types of image editing.”