Deepfake detection by human crowds, machines, and machine-informed crowds | Hacker News Reader

Deepfake detection by human crowds, machines, and machine-informed crowds

pnas.org

27 pointsinfodocket14 comments

Super interesting. Short term this will be useful for detecting deepfakes. Long term this will be useful in training deepfake models to make undetectable deepfakes.

I foresee a day when deepfake will be so easy and widespread that people (authorities) will acquire the habit of digitally signing what they say.

Well, they might digitally sign the things they say that they think reflect well on them. Why would they sign their gaffes? And then we're back where we started: someone shares a video of someone making a gaffe. There's no signature. Is it missing because it's fake, or because the speaker knew it was a gaffe? "Hey, I recorded you saying bad stuff in a donor meeting, will you sign it for me to prove it's not a deepfake?"

That to me reads like a future where the right to be forgotten is finally restored.

Anyone can then claim fake news about anything that wasnt enumerated on their resume.

An unlimited "right to be forgotten" would serve criminals, conmen and crocked politicians perfect.

Almost everything a teenager does is cringey and awkward. People are all imperfect creatures that learn from their mistakes through feedback loops.

The digital age is scarey in that it tends to preserve the past in a way that allows a mistake from five years ago seem like it happened just today.

For example, Mel Gibson. Two damaging recordings got leaked of him a decade or more ago and those cringe moments of his life will be preserved forever.

It wasn't like that before and this is not a cool feature of the modern age.

For example, Mel Gibson. Two damaging recordings got leaked of him a decade or more ago and those cringe moments of his life will be preserved forever.

Mel Gibson is right-wing antisemite involved in a dubious cult. It's hardly surprising a concrete record of this appeared at one point[1].

You're not selling the idea that the "right to be forgotten" is for people who made just one slip-up. But there are certainly ways the right to be forgotten is reasonable - not for stars but for average people. My only point was it probably shouldn't be unlimited. That seems to strike a nerve but apparently among defenders of those who want long records of bad behavior concealed.

[1] Headline: "Absolutely every anti-Semitic thing Mel Gibson has ever said" (probably incomplete but a long list) https://forward.com/culture/449521/mel-gibson-anti-semitism-...

Ignoring how little sense it makes to say spoken word would be... digitally signed???

We've been at the point where a determined amateur could learn the skills to handcraft a false statement from a political figure.

The content is not the issue, the trust of that content is.

(In theory) the more authoritative the source, the more they'll require to treat a statement as real.

Right now you could get a good impersonator to say something stupid and send it to one of those half-trying tabloids and they'd run it even if they knew it was fake...

On the other hand a place like the NY Times might require more meat to the story, how did you get it, what was the chain of custody, etc.

The OP means cryptographically signed for proof of provenance, not bundled with a scanned copy of a signature.

Oh trust me I know... and it's more ridiculous than bundling a scanned copy of a signature, and at best only as functional as that!

I’m not trying to @ you here, but… are you sure you really know?

Cryptographic signatures offer gold-standard, definitive proof of authenticity/non-repudiation. “Trust of content”, as you put it, is the problem they solve.

I suppose it depends on who they're addressing, but you could also imagine 'legitimate' fakes being issued by politicians and CEOs making announcements without the need to actually appear in front a crowd to disseminate them.

Reply on news.ycombinator.com