如何识别伪造的AI照片 | How To Spot Fake AI Photos

✨ Key Takeaways

📋 TED演讲大纲：如何识别伪造的AI照片

I. 引言：信任危机

开场情境：设想你收到一张模糊的照片，声称绑架了你的士兵/亲人。你该如何第一时间辨别真伪？

演讲者背景：Hany Farid，应用数学家和计算机科学家，拥有30年数字图像取证经验。

现状升级：从前一月一案，现在几乎每天都有。原因在于生成式AI的普及和社交媒体对谎言的放大。

核心论点：我们正处于一场全球性的“真相之战” 。

II. 生成式AI的工作原理与缺陷

统计学过程：AI不是像相机那样记录光线，而是通过学习数十亿张图，将“噪音”逆转为符合文本描述的图像。

核心弱点：AI本质是概率统计，它不懂物理世界、几何学和光学原理，这导致了它会留下破绽。

III. 识别AI伪造的三大技术手段

1. 残留噪声分析 (Residual Noise)：

自然照片和AI图像在噪点模式上有本质区别。

AI生成的噪点在频域（傅里叶变换）下呈现出独特的“星状图案” 。

2. 消失点检测 (Vanishing Points)：

原理：现实世界中的平行线（如铁轨）在远处会汇聚于一点（消失点）。

AI破绽：AI生成的场景（如地下室墙壁）中，平行线往往无法汇聚于同一点，几何透视混乱。

3. 阴影一致性 (Shadows)：

原理：阴影的投射路径应与光源位置在几何上一致。

AI破绽：AI生成的阴影（如士兵腿部的影子）延伸线往往无法交汇于同一个光源点，违反物理规律。

IV. 结论与行动建议

不要做受害者：虽然难以辨别，但区分真假是可能的。

技术对策：国际标准的“内容凭证”（Content Credentials）即将推出，从源头认证内容。

给普通人的建议：

戒掉坏习惯：不要把社交媒体作为主要新闻来源，那是充满了谎言和AI垃圾的“垃圾食品” 。

停止传播：在分享信息前深呼吸，不要有意无意地成为欺骗亲友、污染信息生态的帮凶。

V. 问答环节 (Q&A)

某些平台上，假图片的比例可能接近50% 。

目前没有靠谱的网站供外行检测图片真伪（小心被骗），主要是依靠专家工具。

像CSI美剧里那样“增强”模糊照片清晰度在技术上是可行的。

📝 Notes

现在的AI诈骗真的太可怕了！😱 不管是通过“绑架照”勒索，还是冒充CEO开视频会议骗走几千万，AI生成的图像已经到了以假乱真的地步。甚至连专家都说，我们正处于一场“全球真相之战”中！⚔️

最近看了一场TED演讲，主讲人是研究了30年图像取证的大佬 Hany Farid。他揭秘了AI的底层逻辑：AI只是在玩概率统计，它根本不懂物理！ 🧠🚫

利用这一点，我们可以通过观察这 3个破绽 来“找茬”：

1️⃣ 看“消失点” (几何学) 📐 在真实照片里，平行的线条（比如墙壁、铁轨）延伸到远处一定会汇聚成一个点。 ❌ AI破绽：如果你把背景里的线条画延长线，发现它们乱七八糟汇聚不到一起，那八成是假的！

2️⃣ 看“阴影” (物理学) 💡 影子的方向必须指向光源。 ❌ AI破绽：把物体和对应的影子连线，如果几条线指出的光源位置不一样，那就是违反物理定律的假图！

3️⃣ 看“噪点” (玄学/技术流) 🌌 这是专家用的方法。正常照片的噪点是自然的，而AI生成的图片在特殊分析下会有奇怪的“星状图案” 。虽然肉眼难看，但这告诉我们：假的总有痕迹！

⚠️ 大佬的终极建议 (必须听！)

远离社交媒体新闻：那些是为了偷走你的注意力而制造的“垃圾食品”，充斥着谎言和AI垃圾。
不要手滑转发：如果你不确定真假，千万别转！转发虚假信息，你也是“共犯” 。

保护好自己和家人的钱包，从学会质疑每一张图开始！💪

🖊 Highlights

0:00.430

You are a senior military officer and you've just received a chilling message on social media. Four of your soldiers have been taken, and if demands are not met in the next ten minutes, they will be executed. All you have to go on is this grainy photo, and you don't have the time to figure out if four of your soldiers are, in fact, missing. What's your first move?

你是一名高级军官，刚刚在社交媒体上收到了一条令人不寒而栗的消息。你的四名士兵被俘，如果你不能在接下来的十分钟内满足对方的要求，他们将被处决。你仅有的就是这张模糊照片，而且你没有时间弄清楚你的四名士兵是否真的失踪了。你的第一步是什么？

0:26.823

If I may be so bold, your first move is to contact somebody like me and my team. I am by training an applied mathematician and computer scientist. And I know that seems like a very strange first call at a moment like this, but I have spent the last 30 years developing technologies to analyze and authenticate digital images and digital videos.

请允许我冒昧地提议，你的第一步是联系像我和我的团队这样的人。我是一位应用数学家和计算机科学家。我知道在这种时刻第一通电话打给我似乎很奇怪，但是在过去的 30 年里，我一直在开发分析和验证数字图像和数字视频的技术。

0:51.248

Along the way, we've worked with journalists, with courts and with governments on a range of cases from a damning photo of a cheating spouse, gut-wrenching images of child abuse, photographic evidence in a capital murder case, and of course, things that we just can't talk about.

在此过程中，我们与记者、法院和政府合作处理了一系列案件，从触目惊心的配偶出轨的照片，令人痛心的虐待儿童的图像，到死刑谋杀案的照片证据，当然还有我们无法谈论的事件。

1:12.502

It used to be a case would come across my desk once a month. And then it was once a week. Now, it's almost every day.

以前的情况，是每个月会有这样一个案件出现在我的办公桌上。然后是每周一次。现在，几乎每天都如此。

1:23.447

And the reason for this escalation is a combination of things. One, generative AI. We now have the ability to create images that are almost indistinguishable from reality. Two, social media dominates the world and is largely unregulated and actively promotes and amplifies lies and conspiracies over the truth. And collectively, this means that it is becoming harder and harder to believe anything that we read, see or hear online.

而这种升级的原因是由多种因素共同造成的。第一，生成式人工智能。现在，我们能够创建与现实几乎没有区别的图像。第二，社交媒体在世界中霸占着主导地位而且几乎不受管控，并积极宣传和放大着谎言和阴谋，压过了真相。所有因素集合起来，意味着人们越来越难以相信我们在网上读到、看到或听到的任何东西。

1:57.414

I contend that we are in a global war for truth with profound consequences for individuals, for institutions, for societies, and for democracies.

我认为，我们正处于一场全球范围的为真相而战的斗争，这对个人、机构、社会和民主都会有深远的后果。

2:10.927

And I'd like to spend a little time talking today about what my team and I are doing to try to return some of that trust to our online world and in turn, our offline world.

今天我想花一点时间谈谈我和我的团队在做些什么，是如何将这种信任归还我们的线上世界，进而也归还于我们的线下世界。

2:22.873

For 200 years, it seemed reasonable to trust photographs. But even in the mid 1800s, it turns out the Victorians had a sense of humor. They manipulated images. Or you could alter history. If you fell out of favor with Stalin, for example, you may be airbrushed out of the history books.

200年以来，信任照片似乎是合理的。但是，即使在 1800 年代中期，事实证明维多利亚时代的人也是有幽默感的。他们篡改图像，或者你可以改变历史。例如，如果你失去了斯大林的青睐，你可能会被从历史书籍中抹去。

2:43.894

But then, in the turn of the millennium, with the rise of digital cameras and photo-editing software, it became easier and easier to manipulate reality. And now, with generative AI, anybody can create any image of anything, anywhere, at a touch of a button. From four soldiers tied up in a basement to a giraffe, trying on a turtleneck sweater.

但是，在千年之交，随着数码相机和照片编辑软件的兴起，修改现实变得越来越容易。现在，借助生成式人工智能，任何人都可以随时随地创建任何东西的图像，只需按一下按钮。从四名被绑在地下室的士兵到正在试穿高领毛衣的长颈鹿。

millennium /mɪˈlɛnɪəm/

千年

giraffe /dʒɪˈɹɑːf/

长颈鹿

turtleneck /ˈtɜːrtlnek/

n. 圆翻领；高翻领毛衣 adj. 圆翻领的

3:12.422

(Laughter)

（笑声）

3:16.893

It's not fun and games, of course, because generative AI is being used to supercharge past threats and create entirely new ones. The creation of nudes of real women and children used to humiliate or extort them. Fake videos of doctors promoting bogus cures for serious illnesses. A Fortune 500 company losing tens of millions of dollars because an AI impersonator of their CEO infiltrated a video call. Those threats are real, they are here, and we are all vulnerable.

当然，这不是玩乐，因为生成式人工智能不仅被用来使过去的威胁变得如虎添翼，还创造出了全新的威胁。真实的妇女和儿童的裸照被制造出来用以羞辱或勒索他们。医生推销假冒的重病治疗方法的虚假视频。一家财富 500 强公司损失了数千万美金，只因人工智能冒充了其首席执行官的形象渗透到视频通话中。这些威胁是真实存在的，它们就在这里，而面对这些威胁我们都一样脆弱。

nude /nuːd/

adj. 裸体的，赤裸的；肉色的；光秃的，无装饰的 n. 裸体人像，人物裸体作品；肉色，裸色

extort /ɪkˈstɔː(ɹ)t/

v.强取 ,敲诈,勒索

impersonator /ɪmˈpɜːrsəneɪtər/

n. 模仿他人（或名人）的演员

infiltrated /ˈɪnfɪltreɪtɪd/

adj. 渗透的；浸润的 v.（使）透过，（使）浸入（infiltrate 的过去式）

3:55.098

Before we talk about how we would analyze this image to determine if it's real or not, it's useful to understand how generative AI works. Starting with billions of images with a descriptive caption like this, each image is degraded until nothing but visual noise is left. A random array of pixels.

在我们讨论如何分析这张图像以确定它的真伪之前，我们有必要了解生成式人工智能的工作原理。从数十亿张带有这样的描述性文本的图像开始，每张图像都会被降级，直到只留下视觉噪点。只剩下随机像素阵列。

4:15.418

And then the AI model learns how to reverse that process by essentially turning that noise back into the original image. And when this process is done, not once, not twice, but billions of times on a diverse set of images, the machine has learned how to convert noise into an image that is semantically consistent with anything you type.

然后，人工智能模型将学习如何逆转这一个过程本质上是通过将这些噪点变回原始图像。当对不同图像的转化重复不仅一次、不仅两次，而是重复数十亿次后，机器就学会了如何将噪点转换为图像。并且能保证图像与你键入的任何内容在语义上一致。

4:41.878

And it's incredible. But it is decidedly not how a natural photograph is taken, which is the result of converting light that strikes an electronic sensor into a digital representation. And so one of the first things we like to look at is whether the residual noise in an image looks more like a natural image or an AI-generated image.

这确实非常不可思议。但这绝不是自然拍摄照片的方式，自然的拍摄过程应当是将照射到电子传感器上的光转换为数字表现的结果。因此，我们首先要观察的一点就是图像中残留的噪声看起来更像是自然图像，还是人工智能生成的图像。

5:03.233

Here, for example, is our real dog and our AI dog. And here is the residual noise that I've extracted. And if you look at this, it's not at all obvious that there's any difference between those two patterns. But here, in this visualization of the noise, you can see a decidedly different pattern between the natural and the artificial. Those star-like patterns are a telltale sign of generative AI.

例如，这里是我们真正的狗和由人工智能生成的狗。这是我提取的残留噪声。如果你只看这个，会发现这两种模式之间的区别并不是很明显。但是在这里，在这张噪声的可视化中，你可以清晰地看到自然噪声和人造噪声之间截然不同的纹路。这些星状图案是生成式人工智能的明显标志。

5:28.491

Now, for the mathematicians and the physicists in the audience, that is the magnitude of the Fourier transform of the noise residual. For everybody else, that detail doesn't matter, but you definitely should have taken more math in college.

现在，对于听众中的数学家和物理学家来说，这是噪声残差的傅立叶变换量的体现。对于其他人来说，这个细节并不重要，但你确实应该在大学里多学点数学。

5:40.170

(Laughter)

（笑声）

5:41.538

Professors can't help themselves.

当教授的都很难忍住不这么说。

5:44.374

So let's apply this analysis to this image. Here's the noise residual that I've extracted. And there is that star-like pattern that you see in the bottom right. Our first suggestion that something may be wrong here.

现在，让我们用这个方法对这张图片进行分析。这是我提取的噪声残差。在右下角你能看到星状的纹路。我们的初步建议是这里可能有问题。

5:57.854

But no forensic technique is perfect. And so you don't stop after one thing, you keep going.

但是没有任何取证技术是完美的。因此你不会在确定一件事后就停下来，而是继续前进。

6:03.460

So let's go on to our next one, the vanishing points. If you image parallel lines in the physical world, they will converge to a single point, what's called the vanishing point. A good intuition for that, the railroad tracks. When I took this photo, the railroad tracks are obviously parallel, but you can see that they narrow as they recede away from me and intersect at a single vanishing point. This is a phenomenon that artists have known for centuries.

所以让我们继续讨论下一个要点：消失点。如果你在物理世界中对平行线进行成像，它们会聚到一个点，即所谓的消失点。铁轨就是一个很直观的例子。当我拍这张照片时，铁路轨道显然是平行的，但你可以看到它们离我越远就变得越窄，并在一个消失点相交。这是艺术家几个世纪以来都知道的现象。

6:28.919

But here's the great thing. AI doesn't know this. Because AI is fundamentally, as I just described, a statistical process. It doesn't understand the physical world, the geometry and the physics. So if we can find physical and geometric anomalies, we can find evidence of manipulation or generation.

不过很棒的一点是生成式人工智能不懂这个。因为正如我刚才描述的那样，从根本上讲，生成式人工智能是一个统计过程。它不了解物理世界、几何和物理学。因此，如果我们能找到物理和几何异常，我们就能找到操纵或生成的证据。

6:48.805

Here in this image, I've annotated four parallel lines on the parallel sides of the wall in our basement photo, and you can see a lack of a coherent vanishing point. That suggests a physically implausible scene. Evidence number two.

在这张地下室照片中我在平行的两侧墙壁上标注了四条平行线，你可以看到这些平行线无法聚成一个消失点。意味着这一场景在物理逻辑上很牵强。这就是第二项证据。

7:05.322

Alright, what else can we learn? Surprisingly, shadows have a lot in common with vanishing points. Here, what I've done is I've annotated a point on a shadow with the corresponding part on the bottom of the rail that is casting that shadow. And I've extended those lines outwards. And they intersect, not at a vanishing point, but at the light that is casting that shadow. And again, this is a physical phenomena that you expect in natural images. And because AI fundamentally doesn't model the physics and the geometry of the world, it tends to violate these physics.

好吧，我们还能学到什么？令人惊讶的是，阴影与消失点有很多共同之处。我的做法是，在阴影上标注一个点，并在栏杆底部标注出会投射出阴影的部分与其对应。然后将这些线向界限外延伸。当它们相交时呈现的不是消失点，而是投射出阴影的光源。这再一次证明，在自然图像中你会预料看到这种物理现象。而由于人工智能的运作本质不是对世界进行物理和几何建模，因此它往往会违反这些物理规律。

7:41.391

Let's apply this analysis to our image. Here I've annotated four shadows on the bottom from the soldiers' shadows to their legs. And you can see that the lines aren't even close to intersecting.

让我们把这个分析应用到我们的图像上。在这里，我标注了阴影就是从士兵们的影子到他们的腿部的部分。这里你可以看到，这些线条距离交汇相差甚远。

7:53.903

Not one, not two, but three anomalies. We now have a very good indication that this image is not authentic.

不是一个，不是两个，而是三个异常。我们现在有很强的迹象表明这张图片不是真实的。

8:05.415

The most important thing I want you to take away from this is that while it may not be easy, it is possible to distinguish what is real from what is fake.

我希望你从这些分析中获得的最重要的一点是，尽管这也许并不容易，但区分真实的和虚假是可能的。

8:16.960

I think this image is a bit of a metaphor for how a lot of us feel. We feel like hostages. We don't know what to trust anymore. We don't know what is real. What is fake. But we don't have to be hostages. We don't have to succumb to the worst human instincts that pollute our online communities. We have agency, and we can effect change.

我认为这张照片在某种程度上隐喻了我们许多人的感受。我们感觉自己像人质。我们已经不知道该相信什么了。我们不知道什么是真实的。什么是假的。但是我们不必成为人质。我们不必屈服于那些污染我们线上社区的最坏的人类本能。我们拥有自我，我们可以促成变革。

8:45.688

Now, I can't turn you all into digital forensics experts in ten minutes. But I can leave you with a few thoughts.

显然，我无法在十分钟内把你们全部变成数字取证专家。但我可以给你留下一些想法。

8:55.365

One, take comfort in knowing that the tools that I've described and that my team and I are developing are being made available to journalists, to institutions, to the courts to help them tell what's real and fake, which in turn helps you.

第一，我和我的团队正在开发的工具正在逐渐被提供给记者、机构、法院，以帮助他们辨别真伪，这反过来也会对你有所帮助。这一点希望能给你带来宽慰。

9:10.680

Two, there is an international standard for so-called content credentials that can authenticate content at the point of creation. As these credentials start to roll out, they will help you, the consumer, figure out what is real and what is fake online. And while they won't solve all of our problems, they will absolutely be part of a larger solution.

第二，所谓的内容凭证有一项国际标准，可以在创建时即对内容进行真伪验证。随着这些凭证开始推出，它们将帮助作为消费者的各位，弄清楚什么是真实的，什么是假的。尽管它们无法解决我们的所有问题，但它们绝对会成为更大的解决方案的一部分。

9:33.803

Three, please understand that social media is not a place to get news and information.

第三，请明白社交媒体不是获取新闻和信息的地方。

9:42.145

(Applause)

（掌声）

9:48.885

It is a place that Silicon Valley created to steal your time, your attention, by delivering you the equivalent of junk food. And like -- thank you.

硅谷创造的这个地方，为你供应着垃圾食品质量的内容，以窃取你的时间，和你的注意力。就像 —— 谢谢。

10:01.498

(Applause)

（掌声）

10:03.433

And like any bad habit, you should quit.

就像任何坏习惯一样，你应该戒掉。

10:06.302

(Laughter)

（笑声）

10:07.470

And if you can't quit, at least do not let this be your primary source of information, because it is simply too riddled with lies and conspiracies and now AI slop, to be even close to being reliable.

如果你无法放弃，至少不要让它成为你的主要信息来源，因为那里遍布着谎言和阴谋，现在再加上人工智能垃圾，可靠性更是所剩无几。

10:21.284

Four. Understand that when you share false or misleading information, intentionally or not, you're all part of the problem. Don't be part of the problem. There are serious, smart, hard-working journalists and fact-checkers out there who work every day, because I talk to them every day, to sort out the lies from the truths. Take a breath before you share information, and don't deceive your friends and your families and your colleagues, and further pollute the online information ecosystem.

第四，要知道，当你有意或无意地分享虚假或误导性信息时，你都是问题的一部分。不要成为问题的一部分。在外边有很多认真、聪明、勤奋的记者和事实核查人员，我每天都与他们打交道。他们每天都在努力辨别真伪。在分享信息之前深吸一口气，不要欺骗你的朋友、家人和同事，也不要进一步污染线上的信息生态系统。

10:53.917

(Applause)

（掌声）

10:58.955

We're at a fork in the road. One path, we can keep doing what we've been doing for 20 years, allowing technology to rip us apart as a society, sowing distrust, hate, intolerance. Or we can change paths. We can find a new way to leverage the power of technology to work for us and with us, and not against us. That choice is entirely ours.

我们正处在岔路口。其中一条路，我们可以继续重复过去的20年，允许科技把我们的社会四分五裂，播下不信任、仇恨和不容忍。或者我们可以改变路径。我们可以找到一种新的方法来利用技术的力量为我们服务和助我们一臂之力，而不是对我们不利。这个选择完全取决于我们自己。

11:26.649

Thank you.

谢谢。

11:27.917

(Applause)

（掌声）

11:36.159

Latif Nasser: We're doing rapid-fire questions, you ready? Roughly what percent of images online do you believe to be fake?

拉蒂夫 · 纳赛尔（Latif Nasser）：快速提问环节，你准备好了吗？你认为网上的图片大概有多少是假的？

11:42.565

Hany Farid: Depends on the platform. Signal-to-noise ratio is getting close to one. Stay off of Twitter, of X, and stay off of everything else for that matter.

汉尼·法里德：取决于平台。信噪比接近一。远离 Twitter 和 X，也远离其他任何平台。

11:50.106

LN: So what do you think?

LN：那你怎么看？

11:51.341

HF: I would say we're getting close to 50 percent.

HF：我想说我们已经接近50％了。

11:53.710

LN: Can you differentiate between things that have, like, Instagram or TikTok filters or Photoshop versus fully AI generated images?

LN ：你能区分带有Instagram、 Tik Tok滤镜或经过Photoshop处理的事物与完全由人工智能生成的图像吗？

12:01.251

HF: Yes, but it's becoming increasingly more difficult.

HF：是的，但现在变得越来越困难了。

12:04.554

LN: Are there any websites a layperson can use to check?

LN：有没有外行人可以用来检查的网站？

12:07.190

HF: No. By the way, this is a secondary problem, which is now people are creating fake things, then going to fake sites to authenticate them. And it's all getting very weird, don't do it.

HF：没有。顺便说一句，这是一个次要问题，那就是现在人们在制造假东西，然后去虚假的网站验证这些假货。这一切都变得很奇怪，不要这样做。

12:17.800

LN: Last one, most important one. In CSI crime shows, when they say enhance, can you do that?

LN：最后一个，最重要的一个。在 CSI 犯罪节目中，他们会说增强图像清晰度，你能做到吗？

12:23.106

(Laughter)

（笑声）

12:24.274

HF: Yes.

HF：可以。

12:25.441

LN: OK, great. Hany Farid, everybody.

LN：好的，太好了。哈尼·法里德，大家。

How To Spot Fake AI Photos如何识别伪造的AI照片

📋 TED演讲大纲：如何识别伪造的AI照片

How To Spot Fake AI Photos
如何识别伪造的AI照片