Researcher Explains Deepfake Videos

Not every video on the internet is real,

and the fake ones are multiplying.

That's thank to the spread of Deepfakes.

Deepfakes are videos that have been altered

using machine learning, a form of artificial intelligence,

to show someone saying or doing something

that they did not in fact do or say.

The results can be great fun.

Take for example these hilarious clips

of Nicholas Cage starring in movies that he was never in,

but Deepfakes can also be a tool for harassment,

and a way to spread political misinformation.

To learn more about the Deepfakes era we live in,

I spoke with Sam Gregory, who tracks these videos

at the human rights non-profit Witness.

What is a Deepfake and where do they come from?

Why are we talking about them all of a sudden?

What Deepfakes are are the next generation

of video and audio manipulation, and sometimes images,

they're based on artificial intelligence,

and they make it much easier to do a range of things.

So, what people think of as a Deepfake is typically

the face swap, right?

You take the face of one person and you transfer it

onto another person.

But we might also think within the same category

of other forms of synthetic media manipulation,

like the ability to manipulate someone's lips,

and perhaps sync them up with a fake or real audio track,

or the ability to make somebody's body move,

or appear to move, in a way that is realistic

but is in fact computer generated.

And all of this is driven

by advances in artificial intelligence,

particularly the use of what are known as

generative adversarial networks.

And in these adversarial networks, they have the capacity

to set two artificial intelligence networks

competing against each other, one producing forgeries,

the other competing to detect the forgeries.

And as the forgeries improve, they do it based on

this competition between two networks.

So this is one of the big challenges underlying Deepfakes is

that often they are improving because of the nature

of the inputs.

There are so many different ways you could use

that technology.

What are we seeing out there in the wild?

For the moment,

they're primarily non-consensual sexual images.

Probably up to 95% of the Deepfakes out there

are images of celebrities,

or they're non-consensual images of ordinary people

being shared on porn sites,

or being shared in closed messaging.

We have started to see some other cases

of Deepfakes being used in other contexts,

targeting women journalists or civic activists

with images that appear to show them in

sexual situations.

We've also started to hear people using the

it's a Deepfake excuse.

So in the small number of political level cases

where there was potentially a Deepfake,

you see people weaponizing the phrase, it's a Deepfake

and it almost in that case it's really a version

of the same phrase, it's fake news.

And Sam, tell us how easy this technology has become

to access?

You mentioned that it's improved.

Can anyone do this?

It's still not at the point that anyone can do

a really convincing face swapping fake.

There's code available online,

there are websites you can go to that will allow you

to create a Deepfake.

You know, some of those Deepfakes will be imperfect,

but we also know that imperfect Deepfakes

can still cause harm.

So it's getting more accessible

because it's getting commercialized, monetized,

and what's become clear in the last six months is that

Deepfakes, and also other synthetic media

like audio generation, is getting better and better,

and requiring less training data, less examples

you need to generate the data,

all of which mean we're gonna get more and more

of this content, and it's probably gonna be

of better and better quality.

In Congress there has been concern

about Deepfakes being used to distort political campaigns,

maybe even the 2020 Presidential Campaign.

Right, there's clearly vulnerabilities

for political candidates

for the last minute surprise of the compromising video.

A lot of attention goes to political candidates,

there are detection methods being developed

for those political candidates

to protect them from Deepfakes.

And the reason people worry about the advances in Deepfakes

and in other synthetic media,

is we've really seen quite significant progress

in the last six to 12 months,

we've seen a decline in the amount of training data needed

down to a few images

for some of the face expression modification.

We've seen people started to combine video manipulation,

like the lips, with simulation of audio.

And we're starting to see the commercialization of this

into apps.

And as things go to mobile,

that increases them as they become apps,

they obviously get much more available.

And this is why it puts the pressure to be saying

how are we making sure that as these get more available

they're detectable,

and that app makers also think about detection

at the same time as they think about creation

because we have a Pandora's Box there,

and we've already seen how a Pandora's Box like that

can be unleashed.

What possible solutions are people talking about?

You mentioned the idea of a technical solution,

I guess the ideal thing

would be something like a spam filter,

spam filtering is pretty good these days,

you don't see much spam,

could we do that for Deepfakes, just block them out?

We could, but we'd have to define what we think is

a malicious Deepfake, right?

Because Deepfakes and all this genre of synthetic media

really they're related to computational photography,

doing a funny face filter on an app.

Now you might say that's fun, that's my granny,

or you might say that's great,

I think it's great that that's a satire of my president,

or you might look and say I wanna check this

against another source.

What we're not doing actually at the moment

is telling people how to detect Deepfakes

with technical clues.

And the reason for that is,

each of those glitches is the current algorithmic

kind of Achilles's Heel, right?

It's the problem of the current version of the algorithm

but as we put different data into the algorithm

and as we recognize that's a problem,

it's not gonna do that.

So for example, a year ago people thought that Deepfakes

didn't really blink, and now you see Deepfakes that blink.

Now there are technical solutions.

They are all gonna be partial solutions,

and we should want them to be partial solutions.

There's a lot of investment in detection,

using advanced forms of media forensics.

The problem with all those approaches is that

they always are at a disadvantage,

the attacker has the advantage there with the new technique,

and can learn from the previous generations of creation,

and forgery, and forgery detection.

Substituting a kind of technical check mark

for human reasoning is not a great idea.

Systems like that get broken,

they're a absolute honeypot for hackers

and people who wanna disrupt it,

and also because these things are complex, right?

Something may look real, and that may not matter to us

that it has had some manipulation

and you don't wanna give that a cross,

and something may have a tick mark but in fact

the context is all wrong.

I tend to think of detection as being the thing

that at least gives us some signals,

some signals that might help us say

actually there's something suspicious here,

I'm gonna need to use my media literacy,

I'm gonna have to think about it.

Well, that's interesting.

You mention the question of

how people should think differently

now that we're in the Deepfake era, you might call it.

I guess it was never a good idea to believe everything

you saw on the internet,

and now you can't believe anything you see?

What's the right mindset to have?

I think it's also a problem generally

with the misinformation disinformation discussion,

is we've convinced people they can't believe anything online

when the reality is much of what's shared online

is true, or true enough.

It does increase the pressure on us to recognize

that photos and text are not necessarily trustworthy,

we need to use our media literacy on them

to assess where it came from, is there corroboration,

and what's complicated about video and audio

is we have a different cognitive reaction,

we don't have the filters

we've either built or cognitively have

around text and photo,

so I think a real onus on both platforms

who have the capacity to be looking for this,

but also people who've built the tools

that are starting to create this

to feel a responsibility to yes,

build create tools for creation,

but also to build tools for detection,

and then we can plug that in to a culture that

where we're really saying you do need media literacy,

you do need to look at content and assess it,

and I don't think that's the same

as saying it's the end of truth.

I think it's saying we have to be skeptical viewers,

how do we give them technical signals,

how do we build the media literacies

that will deal with this latest generation of manipulation.

Well Sam, thank you very much for you help

understanding Deepfakes.

Thank you Tom, I appreciate the interview.