DO DEEPFAKES HAVE A POSITIVE USE?Discuss deepfakesatwww.facebook.com/webusermagazine 2 - 15 October 2019^39Deepfakesbrought the crazeto light in 2017
(on a now-banned subreddit), used
thesetools within his self-created AI
software. AnotherReddit user called
‘deepfakeapp’went onto develop an
age- and gender-changingtool called
Fake App, which proved very popular.
Again, thiswas underpinnedby
TensorFlow, and made the deep-faking
process quick and easy.
Doesn’t this open acan of
worms?
Absolutely.The technologywas initially
seen as a bit of fun, when actor Nicolas
Cage’sface was used in avariety of
different movies he didn’t originally
appear in. But a more sinister sidewas
quickto follow. Videos beganto appear
featuring theface s of femalecelebrities–
including that ofWonderWomanstar Gal
Gadot –swapped onto the bodies of
porn actresses.What made iteven worse
was that thousands flockedto view.
Since then, there have been instances
of revenge porn, political tinkering and
more skullduggery using theface s of
celebrities such asTaylor Swift, Maisie
Williams and Scarlett Johansson.Even if
a deepfake can be identified as such,
anyone taking a cursory glance may be
unaware that they’re being tricked and
the damage is done.
Are deepfakes easy to spot?
In many cases,yes. There may be some
flickering or slight blurring around aface.
There can also be instances of the
‘uncannyvalley’ – that eeriefeelingyou
get whenyou’re lookingat a humanface
but you just know something isn’t quite
right, usually because theeyes are not
behaving asexpected.Low-resolution
rendering is also a problem andyet some
videos are pretty advanced.The sheer
number of genuinerecordings of
celebrities and politicians provides
enough datato crea te co nvincing
deepfakes of such people.But I’ll be safe,won’t I?
Maybe. Butconsider how many selfies
many of us take and uploadto social
media, puttingvast image sources within
the reach of deepfakers.The Reddit user
‘deepfakes’ said he had grabbed videos
and pictures fromYouTube,stock-
photography sites and Google Images.
It’s also interestingto note the
controversy over a face -swapping app
called Zao, which allows usersto
insinuate theirface s in place of
characters fromtelevision and film.
But while members of the publicwere
innocently enjoyingstarring in their
favourite movie clips, Zao’sterms and
conditions handed the developers thepermanent global rightsto use any
image created on the app, and allowed
those rightsto be transferredto a third
party without permission.Far from
having our identitiesstealthilystolen,
we’re building these databases ourselves.So, what is being done?
In the case of Zao – which is only
available in China – the developers did an
about-turn and said that biometric
informationand detailed user information
would not bestored. Meanwhile, the
Chinese authorities said itwas looking
to tighten theregulations aroundface-
stitchingtechnology. One of the big
concernswas that theresulting videos–
based onstill images –could be usedto
bypassfacial-recognitionsystems. Alipay
had to assure users that its SmileTo Pay
method of authorisation, which asks users
to look into a camerato verify a purchase,
could not be cheated using Zao.Can technologycombat
deepfakes?
It seems so. A group involvingFacebook,
Microsoft, thePartnership on AIcoalition
and academics from seven universities
has launched the Deepfake Detection
Challenge (deepfakedetectionchallenge
.ai), which is setto run until spring of
next year. It is lookingfor participantsto
crea te a solution, using acollection of
deepfake videosreleasedby Facebook
for the purpose ofresearch.The hope is
to come up with approaches that have
so far not beenconsidered.Does any detection software
currently exist?
Well, a team at ImperialCollege in
London is developing software that will
check if filmfootage has been altered.
It is currentlyteaching machine-learning
software called CheckFaketo pick up on
doctoredfootageby analysing 1,000
examples offakes. According to Maja
Pantić, professor ofAffective and
BehaviouralComputing, the software
could be madeavailable as a download
for all, allowingfootageto be input and
checked.The fight, it seems, is on.The film,television and gaming
industries arecertainlycottoning onto
the potential of deepfaketechnology.
You could,for example, play as
yourself in a game (Facebook’s AI
Researchteam has created Vid2Play,
whichextracts playable characters
fromreal-world videos) orrevive dead
actorsto star in current movies –a
controversialtopic in itself.There are
also suggestions that deepfake AI
could be usedfor better dubbing, so
whatyou see and whatyou hear are
better aligned when a different
language is being spoken. But theimportant element isconsent, and the
technology will alwayscourt
controversy when placed in the hands
of people hell-bent on causing trouble.Zao pr oved controversialfor claiming the
rightsto any image created using the appNicolas Cagewas the subjectof manyof
the earliestexamplesof deepfake videos
The Deepfake Detection Challenge aimsto
find a solutionto the problemof deepfakes