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

Stopping deepfake news with an AI algorithm that can tell when a face doesnt fit

spie.org

Post reply on HN
Print view View original

51–60 of 105 posts

Re: Stopping deepfake news with an AI algorithm that can tell when a face doesnt fit

#51

Post by elliekelly »

paulryanrogers wrote: post #33
ericlewis wrote: post #18

Earlier quoted context omitted.

I think religion might have proved that people will believe whatever they want to first. The book Sapiens really opened my eyes to people.

Well it took me 35 years but I escaped religion despite being indoctrinated from a young age. And leaving was a difficult choice but education and evidence were the keys that freed me. Piercing cognitive biases is hard, maybe harder if those most in need feel forced.

> Piercing cognitive biases is hard, maybe harder if those most in need feel forced.

I think about this article (post?) called “How to Change a Mind”[1] a lot. It’s about a woman whose husband was in a religious cult and how he finally realized it.

[1]https://forge.medium.com/how-to-change-a-mind-1774681b9369

Re: Stopping deepfake news with an AI algorithm that can tell when a face doesnt fit

#53

Post by dvtrn »

ben_w wrote: post #10
IgorPartola wrote: post #7

Earlier quoted context omitted.

I feel like it’s time to bring in Heinlein’s Fair Witness concept: https://www.urbandictionary.com/define.php?term=Fair%20Witne... If you think about it, when we had relatively few journalists and their reputation was on the line with every story they wrote they in a way were fair witnesses. Since we have moved to a hybrid model of professional journalists and a huge caste of YouTubers, podcasters, bloggers, and twit…

We need something, but Fair Witnesses seem both exploitable and unachievable given the way human memories work.

“Expert witnesses” then?

Re: Stopping deepfake news with an AI algorithm that can tell when a face doesnt fit

#54

Post by gruez »

xiphias2 wrote: post #48
gruez wrote: post #37

Earlier quoted context omitted.

Digital signatures would only match of the contents match bit for bit. This means that you can't recompress/remux/resize the video. This will probably provide a very poor ux for mobile or other bandwidth limited users. Also, if the video was used as part of another video (eg. TV broadcast), you'd either have to splice the digitally signed video into your existing stream (not modifying the bits at all), or provide the…

Fingerprinting can take care of the recompression problems. Remuxing is a hard problem, as it can easily take people out of context, at the same time it's very important to summarize the videos, as people have limited time. Right now what I see though is that Facebook/Google/Twitter aren't even trying to do the bare minimum end-to-end authentication that Whatsapp/Telegram/Signal already does (create a private key on…

> Fingerprinting can take care of the recompression problems.

That only opens another problem: adversarial attacks that look totally different, but has a similar fingerprint

>Right now what I see though is that Facebook/Google/Twitter aren't even trying to do the bare minimum end-to-end authentication that Whatsapp/Telegram/Signal already does (create a private key on the end devices, and sign the content to verify the authenticity of the publisher).

>Requiring HTTPS was a great first step for tech companies to protect people and from ISPs. But they do nothing to protect people from themselves being compromised (the Twitter incident was a great proof for this).

Mainly because to 99.9% of users, there's no difference between a message that's signed by the author, and a link to a tweet that's made by the author. Even if you're the 0.1% that do care about public key cryptography it doesn't matter because you're trusting the site to do the verification and key management. It's not like PGP where you can get the public keys and verify yourself.

There's also the problem that people simply don't care. Have you seen how many photoshopped tweets end up on social media? If people are willing believe a screenshot without a link (which is trivially easy to add and verify), what makes you think they won't believe a screenshot without a signature?

Re: Stopping deepfake news with an AI algorithm that can tell when a face doesnt fit

#55

Post by trott »

Looking at the diagram, this appears to be just an LSTM slapped on top of a CNN. If so, I'm failing to see any novelty in this approach. RNNs on top of CNNs have been used before, including for deepfake detection. See for example: https://arxiv.org/abs/1905.00582

The recent DeepFake Detection Challenge threw a lot of manpower at the problem earlier this year, BTW.

Re: Stopping deepfake news with an AI algorithm that can tell when a face doesnt fit

#56

Post by dijksterhuis »

sildur wrote: post #17
amelius wrote: post #12

Earlier quoted context omitted.

Antibiotics resistance is also a cat and mouse game. It's still useful to keep playing the game though.

The difference is that all the bacteria can become resistant to the antibiotic soon after it is created.

As can deepfakes. All it takes is an additional step to optimise wrt the model that tries to catch it.

Re: Stopping deepfake news with an AI algorithm that can tell when a face doesnt fit

#57

Post by justforyou »

ericlewis wrote: post #18
yes_man wrote: post #16

The problem will increasingly be "whose algorithm do we believe". Internet has revealed that people believe mostly what they want to. We have seen that a large subset of people are believing Bill Gates is behind the pandemic. Why would masses somehow be more rational in picking the most rigorous and objective neural network to recognize deepfakes, than they are in making sense of the world in general? In the end we w…

I think religion might have proved that people will believe whatever they want to first. The book Sapiens really opened my eyes to people.

>> people will believe whatever they want to first.

Including Harari. There's more wrong in that book than there is right, aside from basic facts that people should already know.

Chriatopher Ryan (Civilized to Death & Sex at Dawn) explores the few compelling points made by Sapiens in depth and without all of the disingenuity and naive philosphy present in Sapiens.

Re: Stopping deepfake news with an AI algorithm that can tell when a face doesnt fit

#58

Post by puranjay »

travisoneill1 wrote: post #47
fny wrote: post #19

Earlier quoted context omitted.

Because YouTube, Facebook will flag these videos before they become too widespread.

"I trust Google and FB, and I would like them to decide what is true or not for me instead of deciding myself"

Amazes me how people are okay with giving these massive corporations so much power and control over our lives - voluntarily. Most would be rightfully horrified if Walmart or Citibank was to decide what you get to watch or not tomorrow. But somehow, if Alphabet or Apple does it, it's all okay.

The tech world in general needs to wake up about the overreaching power and influence tech companies have on our lives.

Post reply on HN