TED2018
Supasorn Suwajanakorn: Fake videos of real people -- and how to spot them
Filmed:
Readability: 4.6
1,453,308 views
Do you think you're good at spotting fake videos, where famous people say things they've never said in real life? See how they're made in this astonishing talk and tech demo. Computer scientist Supasorn Suwajanakorn shows how, as a grad student, he used AI and 3D modeling to create photorealistic fake videos of people synced to audio. Learn more about both the ethical implications and the creative possibilities of this tech -- and the steps being taken to fight against its misuse.
Supasorn Suwajanakorn - Computer scientist
Supasorn Suwajanakorn works on ways to reconstruct, preserve and reanimate anyone -- just from their existing photos and videos. Full bio
Supasorn Suwajanakorn works on ways to reconstruct, preserve and reanimate anyone -- just from their existing photos and videos. Full bio
Double-click the English transcript below to play the video.
00:12
Look at these images.
0
876
1151
00:14
Now, tell me which Obama here is real.
1
2051
2635
00:16
(Video) Barack Obama: To help families
refinance their homes,
refinance their homes,
2
4710
2861
00:19
to invest in things
like high-tech manufacturing,
like high-tech manufacturing,
3
7595
2647
00:22
clean energy
4
10266
1159
00:23
and the infrastructure
that creates good new jobs.
that creates good new jobs.
5
11449
2779
00:26
Supasorn Suwajanakorn: Anyone?
6
14647
1484
00:28
The answer is none of them.
7
16155
1874
00:30
(Laughter)
8
18053
1114
00:31
None of these is actually real.
9
19191
1786
00:33
So let me tell you how we got here.
10
21001
1840
00:35
My inspiration for this work
11
23940
1578
00:37
was a project meant to preserve our last
chance for learning about the Holocaust
chance for learning about the Holocaust
12
25542
5411
00:42
from the survivors.
13
30977
1768
00:44
It's called New Dimensions in Testimony,
14
32769
2627
00:47
and it allows you to have
interactive conversations
interactive conversations
15
35420
3126
00:50
with a hologram
of a real Holocaust survivor.
of a real Holocaust survivor.
16
38570
2556
00:53
(Video) Man: How did you
survive the Holocaust?
survive the Holocaust?
17
41793
1966
00:55
(Video) Hologram: How did I survive?
18
43783
1668
00:57
I survived,
19
45912
1807
01:00
I believe,
20
48419
1527
01:01
because providence watched over me.
21
49970
3023
01:05
SS: Turns out these answers
were prerecorded in a studio.
were prerecorded in a studio.
22
53573
3454
01:09
Yet the effect is astounding.
23
57051
2452
01:11
You feel so connected to his story
and to him as a person.
and to him as a person.
24
59527
3619
01:16
I think there's something special
about human interaction
about human interaction
25
64011
3301
01:19
that makes it much more profound
26
67336
2757
01:22
and personal
27
70117
2198
01:24
than what books or lectures
or movies could ever teach us.
or movies could ever teach us.
28
72339
3485
01:28
So I saw this and began to wonder,
29
76267
2425
01:30
can we create a model
like this for anyone?
like this for anyone?
30
78716
2810
01:33
A model that looks, talks
and acts just like them?
and acts just like them?
31
81550
2975
01:37
So I set out to see if this could be done
32
85573
2007
01:39
and eventually came up with a new solution
33
87604
2310
01:41
that can build a model of a person
using nothing but these:
using nothing but these:
34
89938
3220
01:45
existing photos and videos of a person.
35
93747
2214
01:48
If you can leverage
this kind of passive information,
this kind of passive information,
36
96701
2617
01:51
just photos and video that are out there,
37
99342
2007
01:53
that's the key to scaling to anyone.
38
101373
2056
01:56
By the way, here's Richard Feynman,
39
104119
1777
01:57
who in addition to being
a Nobel Prize winner in physics
a Nobel Prize winner in physics
40
105920
3413
02:01
was also known as a legendary teacher.
41
109357
2453
02:05
Wouldn't it be great
if we could bring him back
if we could bring him back
42
113080
2198
02:07
to give his lectures
and inspire millions of kids,
and inspire millions of kids,
43
115302
3265
02:10
perhaps not just in English
but in any language?
but in any language?
44
118591
2992
02:14
Or if you could ask our grandparents
for advice and hear those comforting words
for advice and hear those comforting words
45
122441
4602
02:19
even if they're no longer with us?
46
127067
1770
02:21
Or maybe using this tool,
book authors, alive or not,
book authors, alive or not,
47
129683
3396
02:25
could read aloud all of their books
for anyone interested.
for anyone interested.
48
133103
2937
02:29
The creative possibilities
here are endless,
here are endless,
49
137199
2437
02:31
and to me, that's very exciting.
50
139660
1713
02:34
And here's how it's working so far.
51
142595
2002
02:36
First, we introduce a new technique
52
144621
1667
02:38
that can reconstruct a high-detailed
3D face model from any image
3D face model from any image
53
146312
4572
02:42
without ever 3D-scanning the person.
54
150908
2119
02:45
And here's the same output model
from different views.
from different views.
55
153890
2642
02:49
This also works on videos,
56
157969
1502
02:51
by running the same algorithm
on each video frame
on each video frame
57
159495
2852
02:54
and generating a moving 3D model.
58
162371
2222
02:57
And here's the same
output model from different angles.
output model from different angles.
59
165538
2772
03:01
It turns out this problem
is very challenging,
is very challenging,
60
169933
2534
03:04
but the key trick
is that we are going to analyze
is that we are going to analyze
61
172491
2525
03:07
a large photo collection
of the person beforehand.
of the person beforehand.
62
175040
2966
03:10
For George W. Bush,
we can just search on Google,
we can just search on Google,
63
178650
2539
03:14
and from that, we are able
to build an average model,
to build an average model,
64
182309
2499
03:16
an iterative, refined model
to recover the expression
to recover the expression
65
184832
3111
03:19
in fine details,
like creases and wrinkles.
like creases and wrinkles.
66
187967
2336
03:23
What's fascinating about this
67
191326
1403
03:24
is that the photo collection
can come from your typical photos.
can come from your typical photos.
68
192753
3423
03:28
It doesn't really matter
what expression you're making
what expression you're making
69
196200
2603
03:30
or where you took those photos.
70
198827
1885
03:32
What matters is
that there are a lot of them.
that there are a lot of them.
71
200736
2400
03:35
And we are still missing color here,
72
203160
1736
03:36
so next, we develop
a new blending technique
a new blending technique
73
204920
2348
03:39
that improves upon
a single averaging method
a single averaging method
74
207292
2836
03:42
and produces sharp
facial textures and colors.
facial textures and colors.
75
210152
2818
03:45
And this can be done for any expression.
76
213779
2771
03:49
Now we have a control
of a model of a person,
of a model of a person,
77
217485
2499
03:52
and the way it's controlled now
is by a sequence of static photos.
is by a sequence of static photos.
78
220008
3795
03:55
Notice how the wrinkles come and go,
depending on the expression.
depending on the expression.
79
223827
3126
04:00
We can also use a video
to drive the model.
to drive the model.
80
228109
2746
04:02
(Video) Daniel Craig: Right, but somehow,
81
230879
2593
04:05
we've managed to attract
some more amazing people.
some more amazing people.
82
233496
3771
04:10
SS: And here's another fun demo.
83
238021
1642
04:11
So what you see here
are controllable models
are controllable models
84
239687
2246
04:13
of people I built
from their internet photos.
from their internet photos.
85
241957
2444
04:16
Now, if you transfer
the motion from the input video,
the motion from the input video,
86
244425
2904
04:19
we can actually drive the entire party.
87
247353
2152
04:21
George W. Bush:
It's a difficult bill to pass,
It's a difficult bill to pass,
88
249529
2172
04:23
because there's a lot of moving parts,
89
251725
2303
04:26
and the legislative processes can be ugly.
90
254052
5231
04:31
(Applause)
91
259307
1630
04:32
SS: So coming back a little bit,
92
260961
1837
04:34
our ultimate goal, rather,
is to capture their mannerisms
is to capture their mannerisms
93
262822
3191
04:38
or the unique way each
of these people talks and smiles.
of these people talks and smiles.
94
266037
3045
04:41
So to do that, can we
actually teach the computer
actually teach the computer
95
269106
2313
04:43
to imitate the way someone talks
96
271443
2222
04:45
by only showing it
video footage of the person?
video footage of the person?
97
273689
2420
04:48
And what I did exactly was,
I let a computer watch
I let a computer watch
98
276898
2577
04:51
14 hours of pure Barack Obama
giving addresses.
giving addresses.
99
279499
3277
04:55
And here's what we can produce
given only his audio.
given only his audio.
100
283443
3516
04:58
(Video) BO: The results are clear.
101
286983
1777
05:00
America's businesses have created
14.5 million new jobs
14.5 million new jobs
102
288784
4349
05:05
over 75 straight months.
103
293157
2774
05:07
SS: So what's being synthesized here
is only the mouth region,
is only the mouth region,
104
295955
2905
05:10
and here's how we do it.
105
298884
1540
05:12
Our pipeline uses a neural network
106
300764
1826
05:14
to convert and input audio
into these mouth points.
into these mouth points.
107
302614
2936
05:18
(Video) BO: We get it through our job
or through Medicare or Medicaid.
or through Medicare or Medicaid.
108
306547
4225
05:22
SS: Then we synthesize the texture,
enhance details and teeth,
enhance details and teeth,
109
310796
3420
05:26
and blend it into the head
and background from a source video.
and background from a source video.
110
314240
3074
05:29
(Video) BO: Women can get free checkups,
111
317338
1905
05:31
and you can't get charged more
just for being a woman.
just for being a woman.
112
319267
2968
05:34
Young people can stay
on a parent's plan until they turn 26.
on a parent's plan until they turn 26.
113
322973
3306
05:39
SS: I think these results
seem very realistic and intriguing,
seem very realistic and intriguing,
114
327267
2952
05:42
but at the same time
frightening, even to me.
frightening, even to me.
115
330243
3173
05:45
Our goal was to build an accurate model
of a person, not to misrepresent them.
of a person, not to misrepresent them.
116
333440
4015
05:49
But one thing that concerns me
is its potential for misuse.
is its potential for misuse.
117
337956
3111
05:53
People have been thinking
about this problem for a long time,
about this problem for a long time,
118
341958
2971
05:56
since the days when Photoshop
first hit the market.
first hit the market.
119
344953
2381
05:59
As a researcher, I'm also working
on countermeasure technology,
on countermeasure technology,
120
347862
3801
06:03
and I'm part of an ongoing
effort at AI Foundation,
effort at AI Foundation,
121
351687
2942
06:06
which uses a combination
of machine learning and human moderators
of machine learning and human moderators
122
354653
3397
06:10
to detect fake images and videos,
123
358074
2144
06:12
fighting against my own work.
124
360242
1514
06:14
And one of the tools we plan to release
is called Reality Defender,
is called Reality Defender,
125
362675
3190
06:17
which is a web-browser plug-in
that can flag potentially fake content
that can flag potentially fake content
126
365889
4039
06:21
automatically, right in the browser.
127
369952
2533
06:24
(Applause)
128
372509
4228
06:28
Despite all this, though,
129
376761
1453
06:30
fake videos could do a lot of damage,
130
378238
1840
06:32
even before anyone has a chance to verify,
131
380102
3294
06:35
so it's very important
that we make everyone aware
that we make everyone aware
132
383420
2722
06:38
of what's currently possible
133
386166
2007
06:40
so we can have the right assumption
and be critical about what we see.
and be critical about what we see.
134
388197
3369
06:44
There's still a long way to go before
we can fully model individual people
we can fully model individual people
135
392423
5007
06:49
and before we can ensure
the safety of this technology.
the safety of this technology.
136
397454
2786
06:53
But I'm excited and hopeful,
137
401097
1587
06:54
because if we use it right and carefully,
138
402708
3539
06:58
this tool can allow any individual's
positive impact on the world
positive impact on the world
139
406271
4309
07:02
to be massively scaled
140
410604
2190
07:04
and really help shape our future
the way we want it to be.
the way we want it to be.
141
412818
2742
07:07
Thank you.
142
415584
1151
07:08
(Applause)
143
416759
5090
ABOUT THE SPEAKER
Supasorn Suwajanakorn - Computer scientistSupasorn Suwajanakorn works on ways to reconstruct, preserve and reanimate anyone -- just from their existing photos and videos.
Why you should listen
Can we create a digital avatar that looks, acts and talks just like our sweet grandma? This question has inspired Supasorn Suwajanakorn, a recent PhD graduate from the University of Washington, to spend years developing new tools to make it a reality. He has developed a set of algorithms that can build a moving 3D face model of anyone from just photos, which was awarded the Innovation of the Year in 2016. He then introduced the first system that can replicate a person's speech and produce a realistic CG-animation by only analyzing their existing video footage -- all without ever bringing in the person to a Hollywood capture studio.
Suwajanakorn is working in the field of machine learning and computer vision. His goal is to bring vision algorithms out of the lab and make them work in the wild.
Supasorn Suwajanakorn | Speaker | TED.com