Won't this be used in the next deepfake as an adversarial network in order to produce more realistic results? It's an endless cat-and-mouse game.
This is often stated but I think it has to be obviously wrong. This isn't a traditional interactive game such as malware & anti-malware.
You have existing sensors which operate under the constraint of [ real world -> theoretic pixel space -> optics & aberrations & sensor noise -> compression ]. And a single adversary which attempts to fake this chain.
The detection of fakes isn't even an adversary in this game, it's merely a detection of deviation of the faking process.
At some point, probably soon, the faking process will reach a point where any deviation will be drowned out by the noise aspect of optics & sensors & compression.
What you need a a deep fake and a real video of the same thing, then train on the difference. Clearly this is impossible - which is what makes the problem hard.
The second way requires dramatically fewer update steps (usually) than the first. Thus - having your adversarial target be differentiable definitely helps, though it is possible to do even absent such a criterion.
Stable Diffusion produces faces- especially eyes that are malformed. You'll often see "restore faces" in online services which feeds the end result into a GFPGAN which is used to restore faces.
Link: http://readfrom.net/neal-stephenson/612819-jipi_and_the_para...
And that we accept that the only thing that's assured to be real is face-to-face, and live with that reality?
So we can say that either it's real or we can identify the specific people who might be lying.
People, and crowds, will respond to first impressions.
Rather, there'll need to be the emergence of credible and trustworthy channels which verify messaging and content before it's widely disseminated.
That largely means re-implementing the sort of media gatekeepers we've seen in the past. Though the challenge of bad-faith actors emerging in such roles at considerable scale points to further challenges, even with that model.
Under a regime in which free speech is considered a fundamental right, any sort of preemptive management by government mandate becomes intensely difficult.
I mean, people still believe we put men on the moon, and people still believe the Earth is round! /s
Don't breathe consciously.
Don't lose the game.
Unfortunately, he was the British Prime Minister by that point.
The overwhelming response from the market, investors, etc was “We like what you’re doing and this is all well and good but the ship has sailed”.
I think the real hot take here (and they danced around this one) is essentially the partisanship. For almost everywhere this really matters a good chunk of people are more likely to believe as real whatever aligns with their pre-conceived notions, political leanings, whatever.
Essentially, real is whatever my tribe says is real.
Get on Facebook and watch people swap around things that are so blatantly and obviously fake. In the words of one potential investor: “no one cares”. Or, to repeat a phrase you may have heard “fake news”.
“Falsehood flies, and the Truth comes limping after it” - Jonathan Swift, 1710
Usually a lot of context is missing with for instance a scary headline. In that sense context tools could also be great to have. Something that exposes connections between information.
I don’t believe no one cares tbh, but there are those who are resistant to change in bias.
Yes, this is the way with anything software based that can earn people money.
See: video game hacks, SEO manipulation, etc
Before deepfakes, if you wanted to claim a video was doctored in court, you'd find an expert on video editing and have them testify.
But the same knowledge that allowed them to identify a doctored video (like 50hz/60hz hum) could be used in an adversarial manner to create a very convincing video.
At most deepfakes democratize that "knowledge" in the form of a model, so it still works both ways.
I don't think it's automatically true that being good at spotting fakes means that you're good at generating fakes. For example, I suspect that I, like most humans, would be pretty good at spotting humanoid robots at the current level of terminology. However, that does not suggest that I would be particularly good at creating humanoid robots that would evade detection.
But my example also satisfies that criterion. I (and most humans) am a domain expert at identifying humans, but that doesn't mean I would be good at faking a human.
I said video editor not video watcher.
I don't think you're thinking about this enough. If you make one effort to create something that looks like a human, you'll do a terrible job unless you are already a skilled artist.
But that doesn't support your claims here. We have a scenario where you've prepared a fake human and attempted to pass it off as real. The only way this could actually happen is if you looked at your own handiwork and it passed quality tests. And since you're good at identifying humans, that necessarily means that your fake human is a high-quality imitation. If it were terrible, you'd look at it and realize it couldn't be passed off as real.