Fake videos of real people -- and how to spot them | Supasorn Suwajanakorn

翻译人员: jacks peng 校对人员: Kai Lu

Look at these images.

看看这些图像。

Now, tell me which Obama here is real.

现在，告诉我哪个是真的奥马巴。

(Video) Barack Obama: To help families refinance their homes,

巴拉克·奥巴马：帮助家庭对他们的房屋重做贷款，

to invest in things like high-tech manufacturing,

投资高科技制造业，

clean energy

清洁能源

and the infrastructure that creates good new jobs.

和带来良好就业机会的基础设施。

Supasorn Suwajanakorn: Anyone?

有人知道吗？

The answer is none of them.

答案是：都不是。

(Laughter)

（笑声）

None of these is actually real.

这些都不是真的。

So let me tell you how we got here.

那让我来告诉你们是怎么回事。

My inspiration for this work

我这个工作的灵感来自于

was a project meant to preserve our last chance for learning about the Holocaust

一个试图保存我们从幸存者那里 了解到的关于大屠杀

from the survivors.

的项目。

It's called New Dimensions in Testimony,

这个项目叫做证词新维度 (New Dimensions in Testimony)，

and it allows you to have interactive conversations

它可以让你与真实大屠杀幸存者的全息图

with a hologram of a real Holocaust survivor.

进行互动对话。

(Video) Man: How did you survive the Holocaust?

你是怎么在大屠杀中幸存下来的？

(Video) Hologram: How did I survive?

我怎么幸存下来？

I survived,

我幸存下来，

I believe,

我相信，

because providence watched over me.

是因为上帝眷顾我。

SS: Turns out these answers were prerecorded in a studio.

原来这些答案是预先在工作室录制的。

Yet the effect is astounding.

但效果令人吃惊。

You feel so connected to his story and to him as a person.

你会对他的故事， 他这个人感同身受。

I think there's something special about human interaction

我想人类互动的特别之处

that makes it much more profound

让它比图书，演讲或电影

and personal

告诉我们的

than what books or lectures or movies could ever teach us.

要更加深刻和真实。

So I saw this and began to wonder,

所以我就开始想，

can we create a model like this for anyone?

我们能不能为每个人做个模型？

A model that looks, talks and acts just like them?

这个模型的样子， 谈话和举止就跟真人无异。

So I set out to see if this could be done

于是我开始探索这个能不能搞定，

and eventually came up with a new solution

并最终找到了一个新的解决方案，

that can build a model of a person using nothing but these:

只需使用下面这些东西就能构建人的模型：

existing photos and videos of a person.

个人现存的照片和视频。

If you can leverage this kind of passive information,

如果你能利用这种被动信息，

just photos and video that are out there,

只需公开的照片和视频，

that's the key to scaling to anyone.

这是扩展到其他人的关键。

By the way, here's Richard Feynman,

顺便说一句，这是理查德·费曼，

who in addition to being a Nobel Prize winner in physics

他除了是诺贝尔物理学奖得主

was also known as a legendary teacher.

也是位传奇教师。

Wouldn't it be great if we could bring him back

这岂不是很棒？ 如果能够把他带回来

to give his lectures and inspire millions of kids,

讲课并激励成千上万的小孩，

perhaps not just in English but in any language?

用英语或者其他任何语言？

Or if you could ask our grandparents for advice and hear those comforting words

或者你也可以征求祖父母的意见， 听听那些让人宽慰的言语，

even if they're no longer with us?

即便他们已经离开我们了。

Or maybe using this tool, book authors, alive or not,

或者使用这个工具，图书的作者， 不管是活着的还是去世的，

could read aloud all of their books for anyone interested.

可以为任何有兴趣的人朗读他们的书本。

The creative possibilities here are endless,

这里的创意可能是无限的，

and to me, that's very exciting.

对我而言，这非常让人兴奋。

And here's how it's working so far.

这是目前它的工作原理。

First, we introduce a new technique

首先我们引入一种新的技术

that can reconstruct a high-detailed 3D face model from any image

可以从任何图像中 重建一个高细节的3D人脸模型，

without ever 3D-scanning the person.

而且无需经对真人进行3D扫描。

And here's the same output model from different views.

这是不同视角下的同一输出模型。

This also works on videos,

这也可以应用于视频，

by running the same algorithm on each video frame

通过对每一幅视频 使用同样的算法

and generating a moving 3D model.

产生移动的3D模型。

And here's the same output model from different angles.

这是不同视角下的同一输出模型。

It turns out this problem is very challenging,

这些问题富有挑战性，

but the key trick is that we are going to analyze

但关键技巧在于我们需要提前

a large photo collection of the person beforehand.

分析一个人的大量照片集。

For George W. Bush, we can just search on Google,

对乔治·沃克·布什， 我们只需要搜索谷歌，

and from that, we are able to build an average model,

这样，我们就能建立一个平均模型，

an iterative, refined model to recover the expression

一个迭代，精炼的模型来恢复表达的细节，

in fine details, like creases and wrinkles.

比如折痕和皱纹。

What's fascinating about this

迷人的是

is that the photo collection can come from your typical photos.

照片集可以来自你的特定照片。

It doesn't really matter what expression you're making

你做何表情或者你在哪里拍照

or where you took those photos.

并不那么关键。

What matters is that there are a lot of them.

关键的是数量要足够多。

And we are still missing color here,

这里我们仍然缺少肤色，

so next, we develop a new blending technique

所以下一步， 我们开发了一种新的混合技术

that improves upon a single averaging method

改善了平均模型，

and produces sharp facial textures and colors.

并产生尖锐的面部纹理和肤色。

And this can be done for any expression.

这可以用于做任何表情。

Now we have a control of a model of a person,

现在我们可以 对一个人的模型进行控制，

and the way it's controlled now is by a sequence of static photos.

它现在被控制的方式是 一系列静态的照片。

Notice how the wrinkles come and go, depending on the expression.

注意皱纹是如何产生和消失的， 这取决于你的表情。

We can also use a video to drive the model.

我们也可以使用视频来驱动模型。

(Video) Daniel Craig: Right, but somehow,

丹尼尔·克雷格：没错，但不管怎样，

we've managed to attract some more amazing people.

我们能够吸引到更多优秀的人才。

SS: And here's another fun demo.

这是另一个有趣的演示。

So what you see here are controllable models

所以你们看到的是 我使用人们的互联网图像

of people I built from their internet photos.

建立的个人控制模型。

Now, if you transfer the motion from the input video,

现在，如果你从视频中传递表情动作，

we can actually drive the entire party.

我们可以让整个派对动起来。

George W. Bush: It's a difficult bill to pass,

布什：这是个难以通过的法案，

because there's a lot of moving parts,

因为有太多可供商榷的部分，

and the legislative processes can be ugly.

立法过程可能让人奔溃。

(Applause)

（鼓掌）

SS: So coming back a little bit,

那么回到正题，

our ultimate goal, rather, is to capture their mannerisms

我们的最终目标， 不如说，是捕捉他们的言谈举止，

or the unique way each of these people talks and smiles.

或者每一个人交谈或微笑的独特之处。

So to do that, can we actually teach the computer

所以这样， 我们能不能只向电脑展示这个人的录像

to imitate the way someone talks

就能教会电脑

by only showing it video footage of the person?

去模仿人们谈话的方式？

And what I did exactly was, I let a computer watch

而我做的事情是，我让电脑

14 hours of pure Barack Obama giving addresses.

看了14个小时的奥巴马演讲。

And here's what we can produce given only his audio.

这是我们只通过他的音频生产出来的内容。

(Video) BO: The results are clear.

结果非常明显。

America's businesses have created 14.5 million new jobs

在过去75个月中，美国企业已经创造了

over 75 straight months.

1450万新的工作机会。

SS: So what's being synthesized here is only the mouth region,

所以这里合成的只是嘴巴部分，

and here's how we do it.

这是我们做的方法。

Our pipeline uses a neural network

我们的处理系统使用神经网络

to convert and input audio into these mouth points.

来转换和输入音频到这些嘴巴的位置。

(Video) BO: We get it through our job or through Medicare or Medicaid.

我们通过我们的工作或者医疗保险 或补助来实现这一目标。

SS: Then we synthesize the texture, enhance details and teeth,

然后我们合成纹理， 增强细节和牙齿，

and blend it into the head and background from a source video.

并将其与源视频中的 头部和背景混合在一起。

(Video) BO: Women can get free checkups,

女性可以获得免费的检查，

and you can't get charged more just for being a woman.

你不会因为是女性而需要支付更高的费用。

Young people can stay on a parent's plan until they turn 26.

年轻人可以在父母计划中呆到26岁。

SS: I think these results seem very realistic and intriguing,

我觉得这些结果看起来非常真实和有趣，

but at the same time frightening, even to me.

但同时，也让我担忧，即便是我。

Our goal was to build an accurate model of a person, not to misrepresent them.

我们的目标是构建人的精准模型， 而非歪曲他们。

But one thing that concerns me is its potential for misuse.

但让我担忧的是它被错误使用的可能。

People have been thinking about this problem for a long time,

人们思考这个问题很长时间了，

since the days when Photoshop first hit the market.

从Photoshop进入市场那天就开始了。

As a researcher, I'm also working on countermeasure technology,

作为一名研究人员， 我也在研究对抗技术，

and I'm part of an ongoing effort at AI Foundation,

我是人工智能基金会持续努力的一份子，

which uses a combination of machine learning and human moderators

它结合了机器学习和人工模型

to detect fake images and videos,

来识别假图像和视频，

fighting against my own work.

与我们自己的工作做斗争。

And one of the tools we plan to release is called Reality Defender,

我们打算发布的一个工具叫做真相卫士，

which is a web-browser plug-in that can flag potentially fake content

是个浏览器插件 可以用来自动标记潜在假内容，

automatically, right in the browser.

在浏览器中就可以使用。

(Applause)

（掌声）

Despite all this, though,

此外，

fake videos could do a lot of damage,

假视频可以带来很大危害，

even before anyone has a chance to verify,

甚至在人们有机会验证它之前，

so it's very important that we make everyone aware

所以让大家意识到这可能是什么

of what's currently possible

非常重要，

so we can have the right assumption and be critical about what we see.

这样我们才能得到正确的推断， 并对看到的保持谨慎。

There's still a long way to go before we can fully model individual people

在个人完全建模 以及确保技术的安全性方面，

and before we can ensure the safety of this technology.

仍有很长的路要走。

But I'm excited and hopeful,

但我兴奋且充满希望，

because if we use it right and carefully,

因为如果我们正确地使用它，

this tool can allow any individual's positive impact on the world

这个工具可以让 每个人对世界积极的影响

to be massively scaled

得到大规模的普及

and really help shape our future the way we want it to be.

并真正帮助塑造我们想要的未来。

Thank you.

谢谢。

(Applause)

（掌声）