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

譯者: Lilian Chiu 審譯者: 易帆 余

Look at these images.

看看這些影像。

Now, tell me which Obama here is real.

告訴我，當中哪一個 歐巴馬是真的？

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

（影片）歐巴馬： 協助家庭做房款再融資，

to invest in things like high-tech manufacturing,

投資於高科技製造業之類、

clean energy

乾淨能源

and the infrastructure that creates good new jobs.

及能夠創造新工作機會的基礎建設。

Supasorn Suwajanakorn: Anyone?

講者：有人知道嗎？

The answer is none of them.

答案是：通通不是。

(Laughter)

（笑聲）

None of these is actually real.

這些通通不是真的。

So let me tell you how we got here.

讓我告訴各位， 我們如何走到這一步。

My inspiration for this work

我做出這項作品的靈感

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

是一項計畫，計畫原來的目的是要 保存我們從大屠殺生還者那裡

from the survivors.

了解大屠殺的最後機會。

It's called New Dimensions in Testimony,

它叫做「證詞的新維度」，

and it allows you to have interactive conversations

它讓你可以和大屠殺生還者的

with a hologram of a real Holocaust survivor.

立體投影進行互動式交談。

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

（影片）男子：你如何 從大屠殺活下來？

(Video) Hologram: How did I survive?

（影片）投影：我如何活下來？

I survived,

我能活下來，

I believe,

我相信，

because providence watched over me.

是因為有老天保佑我。

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

講者：其實這些答案都是 在攝影棚預先錄影好的。

Yet the effect is astounding.

但效果仍然很驚人。

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

你會覺得和他的故事 及他本人很有連結。

I think there's something special about human interaction

我認為，人類互動中有著某種特性，

that makes it much more profound

它會讓使用者的體驗相當深切，

and personal

相當個人化，

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

而這些是書本、課程，或電影 沒有辦法教導我們的。

So I saw this and began to wonder,

我看到這個之後，開始納悶，

can we create a model like this for anyone?

我們能不能為任何人 創造這樣的模型？

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

和他們有相似外表、 說話方式，和行為的模型？

So I set out to see if this could be done

所以，我打算試試能不能辦到，

and eventually came up with a new solution

最終，我找出了一個新的解決方案，

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

只要用下列這些東西， 就能建造出一個人的模型：

existing photos and videos of a person.

一個人既有的照片和影片。

If you can leverage this kind of passive information,

若你能發揮這類被動資訊的功效，

just photos and video that are out there,

只要既有的照片和影片，

that's the key to scaling to anyone.

那就是能夠將規模擴大 到所有人的關鍵。

By the way, here's Richard Feynman,

順道一提，這是理察費曼，

who in addition to being a Nobel Prize winner in physics

他不只是諾貝爾物理獎的得主，

was also known as a legendary teacher.

也是一位傳奇性的老師。

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

如果我們能找他回來教課，

to give his lectures and inspire millions of kids,

鼓舞數百萬名孩子，不是很棒嗎？

perhaps not just in English but in any language?

且不只是用英文， 還可以用任何語言？

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

或是，我們的祖父母已經不在了， 還能問問他們意見，

聽他們說些撫慰的話，如何？

even if they're no longer with us?

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

或許能用這樣的工具， 讓不論還活著或已故的作家

could read aloud all of their books for anyone interested.

出來朗讀他們的書 給想要聽的讀者聽。

The creative possibilities here are endless,

這個工具有著無限的創意可能性，

and to me, that's very exciting.

我對此感到十分興奮。

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

目前，它的運作方式如下：

First, we introduce a new technique

首先，我們先採用一項新技術，

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

它能從任何影像，重新建構出 一個人非常細節的 3D 面部模型，

without ever 3D-scanning the person.

不需要對他做實際的 3D 掃瞄。

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

這是從不同視角，輸出的類似模型。

This also works on videos,

它也能用在影片上，

by running the same algorithm on each video frame

對每一格影片進行同樣的演算法，

and generating a moving 3D model.

產生出會動的 3D 模型。

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

這是從不同的角度，輸出的類似模型。

It turns out this problem is very challenging,

結果發現，這個問題非常有挑戰性，

but the key trick is that we are going to analyze

但關鍵的技巧在於，我們得要

a large photo collection of the person beforehand.

事先分析同一個人的大量照片集。

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

如果要做小布希總統， 我們用 Google 搜尋就可以了，

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

這樣，我們就可以建立 一個平均的模型，

an iterative, refined model to recover the expression

一個經過反覆運算、改良後的模型， 它能夠很精密地重現出表情，

in fine details, like creases and wrinkles.

如皺褶和皺紋等。

What's fascinating about this

很炫的一點是，

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

只要用你一般的照片集 就可以做到了。

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

你在照片中做什麼表情都無所謂，

or where you took those photos.

在任何地方拍的照片都行。

What matters is that there are a lot of them.

重要的是，要有很多張照片。

And we are still missing color here,

這裡，我們還缺了顏色，

so next, we develop a new blending technique

所以，下一步，我們開發出 一種新的混和技術，

that improves upon a single averaging method

它是從單一平均化法改良而來的，

and produces sharp facial textures and colors.

能夠產生出精細的 面部質感和顏色。

And this can be done for any expression.

任何表情都可以使用。

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

現在我們可以控制一個人的模型，

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

控制它的方式，是透過 一系列靜態照片來進行。

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

注意看，根據不同表情， 皺紋有時會出現有時會消失。

We can also use a video to drive the model.

我們也可以用影片來驅動模型。

(Video) Daniel Craig: Right, but somehow,

（影片）丹尼爾克雷格： 對，但不知何故

we've managed to attract some more amazing people.

我們更能吸引到很了不起的人。

SS: And here's another fun demo.

講者：這是另一個有趣的展示。

So what you see here are controllable models

各位在這裡看到的是可控制的模型，

of people I built from their internet photos.

我用他們在網路上的照片建立的。

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

如果把輸入影像的動作做轉換，

we can actually drive the entire party.

我們就可以讓所有人跟著動起來。

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

喬治布希：要通過 這個法案很困難，

because there's a lot of moving parts,

因為有很多會變動的部分，

and the legislative processes can be ugly.

且立法過程有時是很醜陋的。

(Applause)

（掌聲）

SS: So coming back a little bit,

講者：先倒帶一下，

our ultimate goal, rather, is to capture their mannerisms

我們的終極目標， 其實是要捕捉他們的動作習性，

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

或是說每個人說話 和微笑的獨特方式。

So to do that, can we actually teach the computer

為了做到這一點，我們要教導電腦

to imitate the way someone talks

去模仿一個人說話的方式，

by only showing it video footage of the person?

做法是把那個人的影片給電腦看。

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

我所做的其實就是讓電腦觀看

14 hours of pure Barack Obama giving addresses.

十四小時的影片， 內容全是歐巴馬在演講。

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

只要提供他的聲音， 我們就能產生出這樣的成果。

(Video) BO: The results are clear.

（影片）歐巴馬：結果很清楚。

America's businesses have created 14.5 million new jobs

在 75 個月內， 美國的企業已經創造出了

over 75 straight months.

1450 萬個新工作機會。

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

講者：這裡合成的只有嘴部區域，

and here's how we do it.

我們的做法是這樣的。

Our pipeline uses a neural network

我們的指令傳遞途徑 使用的是類神經網路，

to convert and input audio into these mouth points.

將輸入的聲音轉換成那些嘴部的點。

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

（影片）歐巴馬：透過我們的運作 或醫療保險、醫療補助來達成。

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

講者：接著我們再合成肌理， 強化細節和牙齒，

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

將它結合到頭部 以及來源影片的背景上。

(Video) BO: Women can get free checkups,

（影片）歐巴馬： 女性能接受免費檢查，

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

你不會因為身為女性， 就要付比較多錢。

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

年輕人在 26 歲前都可以 掛在一位家長的方案底下。

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

講者：我認為這些結果 非常真實且有意思，

but at the same time frightening, even to me.

但同時，連我也覺得它們很嚇人。

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

我們的目標是要建造出很精確的 人類模型，而不是要故意歪曲他。

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

但我擔心的一件事情， 是這項工具可能被誤用。

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

從 Photoshop 剛出現 在市場上之後，

since the days when Photoshop first hit the market.

長久以來大家都一直 在想著這個問題。

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

身為研究者， 我也在研究反制的技術，

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

我參與了人工智慧 基金會的一項計畫，

which uses a combination of machine learning and human moderators

結合機器學習和人類調節

to detect fake images and videos,

來偵測出仿冒的影像和影片，

fighting against my own work.

及對抗我自己的作品。

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

我們打算要推出的工具之一， 叫做「真相守衛者」，

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

它是瀏覽器的插件， 能標出有可能是仿冒的內容，

automatically, right in the browser.

全自動化，就裝在瀏覽器上。

(Applause)

（掌聲）

Despite all this, though,

不過，儘管有這些努力，

fake videos could do a lot of damage,

仿冒影片仍能造成許多傷害，

even before anyone has a chance to verify,

甚至在任何人有機會 驗證之前，傷害就已造成，

so it's very important that we make everyone aware

所以，非常重要的是， 要讓大家都能意識到，

of what's currently possible

目前有什麼是可能的，

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

我們因此得以有正確的假設， 並對所看到的內容抱持批評的態度。

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

要能為個人完整地建模 還有很長的一段路要走，

and before we can ensure the safety of this technology.

也還要很多努力才能確保 這項科技的安全性。

But I'm excited and hopeful,

但我很興奮，也抱著希望，

because if we use it right and carefully,

因為如果我們能正確、 小心地使用它，

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

這工具能讓任何人正面地影響世界，

to be massively scaled

大規模地影響，

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

且能真正能用我們希望的方式 來型塑我們的未來。

Thank you.

謝謝。

(Applause)

（掌聲）