Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Creating a deepfake took two weeks and cost $552

arstechnica.com

Post reply on HN
Print view View original

51–60 of 119 posts

Re: Creating a deepfake took two weeks and cost $552

#51

Post by muttled »

aianus wrote: post #45
ryanwaggoner wrote: post #8

Earlier quoted context omitted.

I don't really understand this argument at all. We've been in this "post truth" world for a long time already for every form of media other than video. Text quotes, photos, and audio can all be easily faked. If I post a ridiculous quote here and say it's from Obama, you won't believe me. But if the NYTimes does the same thing, it carries a lot more weight. We've been here for a long time already, now those standards…

> How good will your deep fake have to be in order to fool a deep fake detection AI? The best generator and the best detector are actually part of the same model! If you create a better detector, you are making the generator better at the same time so you have not accomplished anything. "GANs are a clever way of training a generative model by framing the problem as a supervised learning problem with two sub-models: t…

Is the generator always assumed to be able to keep up when the discriminator is improved algorithmically? I've always wondered that when I hear this counter-argument.

Re: Creating a deepfake took two weeks and cost $552

#52

Post by turk73 »

The whole "deep fake" controversy is only going on because the elite politicians all have blackmail videos that are probably in the hands of people that will soon make them public and the media is preparing the big lie that these are just "deep fakes, nothing to see here, move along." Meanwhile, the truth is that these people actually did do the heinous things we all think they did.

Just look at what Prince Andrew tried to do: He was out whoring around and had photos taken--with the whore--and he has attempted to deny it citing that the photos are faked. But honestly, his own behavior and track record makes it clear that he is lying and that in all likelihood, he did in fact have sex with a 17 year old woman and all the allegations are true.

The worst part of all this is that Prince Andrew's transgressions are the ones they WANT you to see because they are the ones that are more easy for the public to write off (she was 17 and consenting) and it was bad but just his bad judgement or what have you.

My sense is that these "simple" cases are the are the tip of the iceberg, however. The ones they DON'T want you to see haven't been made public yet. And those are truly, pitchforks and torches, bad.

Re: Creating a deepfake took two weeks and cost $552

#53

Post by kitd »

Can content providers not include a digital signature in the video metadata? Then you can authenticate the video against the original provider, a bit like the green padlock next to the URL bar in Chrome.

Even better (though I'm not sure this technology exists yet), some some kind of "rolling signature", available running along side the video in metadata, that a media player can use to validate the authenticity of the video at that point in time. Then you don't need the full video to be able to verify the source.

Re: Creating a deepfake took two weeks and cost $552

#54

Post by Forge36 »

gdubs wrote: post #50

The central technique used in DeepFakes is fascinating. Initially I assumed they were using Generative Adversarial Networks (GANs). But — and someone can correct if I’m wrong — Deep Fakes use two Autoencoders. So essentially you’re training a network how to compress an image down to a very tiny representation, and then uncompress it as accurately to the original as possible. You train two of these: one for the origin…

That is a fascinating way to describe this.

Re: Creating a deepfake took two weeks and cost $552

#55

Post by nojvek »

gdubs wrote: post #50

The central technique used in DeepFakes is fascinating. Initially I assumed they were using Generative Adversarial Networks (GANs). But — and someone can correct if I’m wrong — Deep Fakes use two Autoencoders. So essentially you’re training a network how to compress an image down to a very tiny representation, and then uncompress it as accurately to the original as possible. You train two of these: one for the origin…

Makes sense. You want to boil down the face to some sort of abstract representation of movement of lips, face alignment etc. Then put in different parameters and have it be photorealistic on the other side. Could do wonders for video chat.

Re: Creating a deepfake took two weeks and cost $552

#56

Post by tralarpa »

gdubs wrote: post #50

The central technique used in DeepFakes is fascinating. Initially I assumed they were using Generative Adversarial Networks (GANs). But — and someone can correct if I’m wrong — Deep Fakes use two Autoencoders. So essentially you’re training a network how to compress an image down to a very tiny representation, and then uncompress it as accurately to the original as possible. You train two of these: one for the origin…

How do you guarantee that the two trained autoencoders speak the "same language", i.e. the code learned by the autoencoder of the original face gives a reasonable output when uncompressed by the target autoencoder?

Re: Creating a deepfake took two weeks and cost $552

#57

Post by PeterisP »

muttled wrote: post #51
aianus wrote: post #45

Earlier quoted context omitted.

> How good will your deep fake have to be in order to fool a deep fake detection AI? The best generator and the best detector are actually part of the same model! If you create a better detector, you are making the generator better at the same time so you have not accomplished anything. "GANs are a clever way of training a generative model by framing the problem as a supervised learning problem with two sub-models: t…

Is the generator always assumed to be able to keep up when the discriminator is improved algorithmically? I've always wondered that when I hear this counter-argument.

Well, from the pure GAN theory perspective, it is theoretically possible for a perfect generator to exist which generates samples that perfectly match the 'source data distribution' (the only question is how do we obtain one or close to it), and thus a perfect discriminator can't exist - a fake image could be realistic and indistinguishable in all aspects.

And it does seem that the human-visible weaknesses of many currently published deepfake examples are those that would be hard for the current discriminator architectures to detect, not those which would be harder to generate than any other part of the fake.

Re: Creating a deepfake took two weeks and cost $552

#58

Post by panarky »

trianglem wrote: post #37
thepete2 wrote: post #35

Earlier quoted context omitted.

What I find most worrying is the 2nd order consequence of not being able to trust pictures anymore.

I feel like that ship sailed a while ago with Photoshop being around. I think the problem now is you can do the same with videos and the technology is more commoditized so you don’t have to be very knowledgeable to do these things.

People predicted that Photoshop was the end of truth, that journalism could no longer be trusted, that photographic evidence at criminal trials would be inadmissible.

None of that actually happened because people learned to be skeptical, and to care about who took the photo, the chain of custody, and the controls to prevent digital alteration.

Maybe we need something similar but more rigorous for video. The camera itself could encode a PKI signature in every frame of the original. If subsequent edits preserve frames as is, then the viewer could determine how many frames were cut or duplicated.

If subsequent edits crop, recolor or otherwise alter the frames, that editor could encode its own digital signature so the viewer could judge how much to trust their edits.

Re: Creating a deepfake took two weeks and cost $552

#59

Post by claytonius »

gdubs wrote: post #50

The central technique used in DeepFakes is fascinating. Initially I assumed they were using Generative Adversarial Networks (GANs). But — and someone can correct if I’m wrong — Deep Fakes use two Autoencoders. So essentially you’re training a network how to compress an image down to a very tiny representation, and then uncompress it as accurately to the original as possible. You train two of these: one for the origin…

I think you're right that most DeepFakes use Variational AutoEncoders (As described by the article) where the latent space is jointly distributed between the two domains, allowing for translation. For additional context: AutoEncoders are also frequently adversarially regularized so you're not wrong about GANs being involved. My understanding is that the most accurate implementations also have some manual feature engineering for more tractable problems like facial orientation.

One of my favorite related papers: https://arxiv.org/pdf/1703.00848.pdf

Re: Creating a deepfake took two weeks and cost $552

#60

Post by skizm »

gdubs wrote: post #50

The central technique used in DeepFakes is fascinating. Initially I assumed they were using Generative Adversarial Networks (GANs). But — and someone can correct if I’m wrong — Deep Fakes use two Autoencoders. So essentially you’re training a network how to compress an image down to a very tiny representation, and then uncompress it as accurately to the original as possible. You train two of these: one for the origin…

The article describes this process in detail.
Post reply on HN