Facebook, academics think they've cracked spotting deepfakes by spotting how they're generated

Computer scientists have built prototype software capable of not only detecting fake images forged by neural networks but also estimating the properties of the model used to generate these so-called deepfakes. The project is a collaboration between academics at Michigan State University (MSU) in America and Facebook AI …

[macjules]

"spotting deepfakes with 99% accuracy"

My deepfake spotter has a 100% success rate: it simply blocks access to Facebook.

[Anonymous Coward]

Photos all the way down

A lot of tools to detect whether an image has been manipulated can be tricked simply by photographing a print of the manipulated image.

[bombastic bob]

Re: Photos all the way down

I would expect that generating the deepfake in a very high resolution, and then using a standard method of shrinking it down to something with less resolution (cubic interpolation, let's say), or using JPEG vs PNG even, might be just enough to fool the deep-fake spotter-bots.

I like use 'gimp' and a hand-done "fuzzy" technique around the borders of, let's say, a face, surrounded by transparency, that melds right into another photo when proper re-sizing and perspective is done (no pr0n though, just funny things).

in 'Little Nicky' (Adam Sandler) Nicky's brother pasted his face over Al Pacino in a clip from 'Scarface', and did it poorly. But often times it's funnier if poorly done, which is then obvious to everyone (including algorithms I hope).

[CrackedNoggin]

So a deepfaker needs to check their work against the most popular deepfake detectors and fiddle until undetected.

[Blank Reg]

That may be feasible for images but is likely too time consuming for video. Though so far I've yet to see a deep fake video that I couldn't detect was a fake fairly easily

[DS999]

Sure, today

Deepfake videos are in their infancy. I doubt you'll be able to say the same in a decade.

[NoneSuch]

Re: Sure, today

You'll be saying that in less than a day, if someone wants it.

[Anonymous Coward]

Re: Sure, today

> Deepfake videos are in their infancy. I doubt you'll be able to say the same in a decade.

The problem with spotting a deepfake in 10 years time is that by then all videos will be fakes.

All TV presenters will be computer generated because they'll be cheaper than paying real people. Soap operas will have CGI cast so you only have to pay script writers. And politicians will be sending video "clones" of themselves to give tailored speeches to tailored audiences while the actual human only talks to donors rich enough to buy time in his/her presence.

So spotting a fake will be difficult because everything will be a fake - the only question being whether it is an authorised fake or an unauthorised fake.

[DS999]

Re: Sure, today

All TV presenters will be computer generated because they'll be cheaper than paying real people

So the hosts on Fox News and MSNBC will be out of a job? Sounds like a win for everyone!

[NoneSuch]

My word...

...I think he's got it...

[iron]

Surely fake accounts don't need deepfake avatars, they can just use real pictures stolen from another website.

[ThatOne]

Indeed, came here to say the same thing: Why bother when Internet is full of millions of peoples' pictures? Just grab one from a different website, or, better, crop one out of somebody's vacation pictures so it is kind of unique.

[Michael Wojcik]

Well, the toolchain is simpler if you generate a fake picture. If you're automating the creation of fake profiles, that's a win.

[KittenHuffer]

The war has begun!

What will happen now is that the deepfakers will introduce something like a post-processing layer to obfuscate the artifacts that this technique is picking up on.

As soon as you develop a deepfake detector the deepfakers will start working on better deepfakes.

.

And what do I have to do to stop all my posts going in for moderation?

[ThatOne]

Re: The war has begun!

> And what do I have to do to stop all my posts going in for moderation?

Happens to me too from time to time: No matter if you're an old member with hundreds/thousands of posts, at some random point some wacky algorithm decides you're all of a sudden utterly suspect. Then it usually reverts back to normal.

I haven't managed to see any pattern (time, content, topic), it seems to happen totally randomly.

[elsergiovolador]

PR

I doubt that this is being done to tackle bots. I think that more likely they do it to have plausible deniability. When another elections get swayed, they can say "hey! we have this 99% system here that detects bots, so whatever happened was genuine!".

[Short Fat Bald Hairy Man]

Looks like Kafka (the author)?

What is the false positive rate? I do not see any reference to that?

If this is used, will those poor bastards be eternally fucked trying to prove that their photo/image/whatever is genuine?

[ThatOne]

Re: Looks like Kafka (the author)?

"Sorry mister but you look like a deepfake, we can't let you in..."

[Anonymous Coward]

I doubt it matters does it? People will still believe what fits in with their preconceptions, however ridiculous, and however much pesky evidence or facts to the contrary are provided.