High-tech deception of ‘deepfake’ videos
apnews.com
apnews.com
I bring you no solutions. Our best programmatic tool for these kinds of video analysis is currently the exact same technology that enables it in the first place. Furthermore, any neural network that can detect fake videos will only be used adversarially to train a better deep fake creator!
The only solution I see is the camera somehow cryptographically signing the video, and enabling users to easily check whether the signature is valid. However I don't think this is feasible, due to the amount of cooperation we'd need between camera makers and video distribution platforms, especially since it's not really in the businesses direct line of best interest help.
The world's going to start getting real surreal.
Or perhaps active tattoos that display private key signed messages that describe what they’re saying and doing.
Will the government be in the business of commenting on the veracity of whatever video(s) they anticipate coming? And if they say it's fake (or real), will the public believe them?
Start with something naturally unfocused and exists only as a one off original, and it would be extremely hard to disprove without comparing it to other video of the same event.
And in order to implement that, you'll be talking to people who'd sign a video by getting a pen, some paper, and a flatbed scanner.
I'm not entirely joking. My point is that this proves that there comes a point in every field where technical illiteracy goes from being annoying for you (you can't use the best tools anymore, and your workflow is now sub-par) to annoying for others (you can't collaborate effectively) to completely unworkable, as when your work product is now worthless because you can't demonstrate it isn't a cheap forgery. And some people are just not technical, through stubbornness or otherwise. This transition will hurt.
Maybe tools will follow the demand. I'm not entirely sanguine.
However I'd hope that we move to some agreed standard of signing of videos, so their source can be confirmed by the player. It doesn't matter if that source is the bbc, washington post, alex jones, national enquirer, the onion, or "Bob Joe Billy-Bob the Third", it means you know where the video came from, and can trust (or not) it based on that information.
A video with a bbc banner showing the president saying something could easily be faked. One that is signed by the BBC will only be fake if the BBC itself is publishing it as a fake, and that comes with a reputational impact.
Sites that reversion that video (recompress it etc) can simply sign it by themselves. They could claim that they took an original BBC item to do it, and then it's how much you trust that site.
People see someone, be it Donald Trump, or John Simpson, or Walter Mondale, and instinctively know that person said what they can see. That's changing, and it's going to be a long time before our brains change. Even now, you read some rubbish that you consciously know is rubbish, but part of it lodges in your subconcious and affects your thinking down the line. At least if you know who created the video, it gives you a fighting chance.
With this system anyone can still publish, even anonymously, but reputations are hard to gain and easy to lose. I don't like the daily mail, but I know if they publish a video of Obama saying "Killmonger was right, and President Trump is a total and complete dipshit", I'll probably believe he did. If he didn't, that will come out, and the daily mail is off the list of sources I trust to not make crap up (rather than just bend things), and it makes it far easier for reasonable people to dismiss it.
This moves the onus on news sources to be more discerning in their facts, but if they get it wrong then they lose their readers. By signing the news, they are backing it with their reputation. If Alex Jones has signed it, then he's putting his reputation on the line (if he has any reputation for saying something).
Nobody cares much about the reputational impact. Fringe tabloid fake-news, and mainstream 'fake'-news gets plenty of viewers.
BBC itself has manufactured video 'evidence' during the Libyan war. [1] Yet, most people think it's an accurate source of news.
[1] Their 'live broadcasts' showed video of people in India waving the Indian tri-color, which the BBC has claimed was rebels in Libya's Green Square waving the old Libyan tri-color. But, eh, who cares about honesty - when you've got a politically sensitive story to tell.
CBS and Dan Rather takes documents months before the 2004 election [0] and Rather still have 400k+ twitter followers. Even BBC didn’t cover the Iraq War properly when we really needed it to [1].
I read news for opinion and some awareness. But I’d trust a video of Trump strangling a hobo about the same from BBC as TMZ and would need a lot of verification.
[0] https://en.wikipedia.org/wiki/Killian_documents_controversy [1] http://news.bbc.co.uk/2/hi/entertainment/3890821.stm
Apple could single-handedly launch a "verified video" system.
Please build this.
— everyone
(As mentioned elsewhere in this thread, I go into a bit more detail here: https://www.washingtonpost.com/news/theworldpost/wp/2018/02/... )
Will try to write a HN friendly tldr/update soon.
The worst solution is to panic and attempt to ban technology like this insuring that only the bad actors will use it.
I want everyone down to kids to be able to make fake videos to their little heart's content so we can all internalize this inevitable Black Mirrorization of reality.
And your proposed solution is fantastic BTW. Unfortunately, there are more than enough bad actors among us to social engineer those signatures out of otherwise trustworthy agencies and a general lack of intellectual curiosity on the part of most people to even bother to check for such things. Imagine the day we have a video where HRC confesses to the pedophile colony on Mars and coming up next, Trump tells everyone where he buried the bodies, and if people look, there will be bodies there (never mind that they're stolen cadavers). How will people cope with a world like this?
It is almost like we need pop art and pop culture to drive this point home before political mass media does it in the worst possible way.
Agreed, but I think it's been long proven most people do. People hear what they want to hear. Maybe that will change, but I'd say it seems unlikely at this point.
| he worst solution is to panic and attempt to ban technology like this insuring that only the bad actors will use it.
Very true.
| I want everyone down to kids to be able to make fake videos to their little heart's content so we can all internalize this inevitable Black Mirrorization of reality.
Increasingly getting there. It only takes one person to make a nice GUI, post it to github.
| How will people cope with a world like this?
Probably poorly if history has taught us anything.
| It is almost like we need pop art and pop culture to drive this point home before political mass media does it in the worst possible way.
Very intriguing. Traditionally, art and culture change as a reaction to current affairs, as opposed to a pre-emption. I'm not sure how to pragmatically flip that paradigm.
Everyone knows that a still image can be altered to (convincingly) depict something that never happened. So, a still image is not considered good evidence of anything unless it has a verifiable provenance, chain of custody etc. Presumably that wasn't always the case, but now that photo editing tech is universally accessible, we all know we can't trust a still image unless its authenticity can be corroborated somehow.
The deepfake tech merely puts video into the same category of (un)trustworthiness. The potential for abuse depends on people's trust that video is always authentic. So the "solution" is simply to make it as widely-known as possible that video is now easy to fake.
Previously it took at least one CGI artist to convincingly fake video, probably several. This inherently limits how much of a problem it can be, because it doesn't scale. AI scales.
I mean- either deep learning AI automation is going to be important and useful, or it isn't. If replacing all drivers with robots is important, then automated turnkey fakery has the potential for disrupting something, isn't it?
There's a shockingly large number of people who aren't even convinced that election meddling took place in 2016, or that propaganda was disseminated to many people on social media, despite unanimity from US intelligence.
If we as a society can't all agree that these things even happened, how can we hope to properly educate people on the existence/prevalence of fake video? It seems like common sense to many people, but we can't assume that everyone will accept these truths based on the political rhetoric we've seen in the last few years.
And yet, fake news does seem to be a problem. Checking provenance of facts and statistics is often skipped, even by journalists, let alone on Twitter.
This suggests that people will have lots of fun resharing deep fakes and won't care much that they're fake, as long as they express some emotion that they agree with.
Give people a fake that they want to believe in, and it's going to be very hard to talk them out of it.
Deepfakes' destruction of the credibility of videos is still worse because it totally removes the ability of a random citizen witness to capture evidence. This has been useful in the past, e.g. for documenting police brutality or just-developing events like shootings.
In a way this is a very undemocratic development -- putting the power of deepfakes in the hands of ordinary people is going to make everybody a less credible witness, no?
The whole Cambridge Analytica debacle started blowing up after deepfakes gained notoriety and started getting banned on most sites. The whole time I was watching the undercover documentary Channel 4 released I kept thinking, "this 'found footage' could easily be a hoax"
The police/etc have the same problem.
> putting the power of deepfakes in the hands of ordinary people is going to make everybody a less credible witness, no?
"Ordinary people" have a major advantage: parallelism and redundancy. Making one fake and disseminating it to many people is efficient and cheap. Faking many videos in a way that avoids easily discovered contradictions is very hard. Evidence always needs to be corroborated.
Photos, and by corollary video, haven't been incontrovertible for at least two decades, and Hollywood movies are a clear demonstration of that fact.
We won't really be able to rely on pictures as proof, at some point in the immediate future.
Unfortunately, in an ever-more terrifying far-flung future, where eye surgery becomes sufficiently advanced, we won't even be able to trust our eyes. After that, I think it's probably just game over.
I really just don't have a strategy to countermand that one.
That being said, movements like the anti-vaxxers and the flat-Earthers are testament to people's ability to believe whatever the hell they want.
Video of the Prime Minister/President taking bribes? Eh, you can't trust these videos these days!
In a court of law, presumably video will now be subject to the same scrutiny as photos. IANAL but I imagine photos do not "stand alone" - the side that uses them has to make a convincing argument that they are trustworthy.
It's the court of public opinion that worries me. That court is very important to the functioning of a liberal democracy. If the facts themselves are always open to suspicion, a society cannot defend itself very well against tyranny.
Being able to dismiss credible video evidence as 'deepfake' is equally powerful.
In any case, it would be extremely risky to hang a case on a single confession video that the accused claims is fake, given the current pattern of prosecution the DOJ would never bring that case; it would be embarrassing in the public eye and probably overturned on appeal, if it even made it to trial.
The issue is that it will be really hard to prove as fake so even if an adversary didn’t just anonymously leak the material. In a PR situation you don’t need authenticity, just confusion.
I expect that you could use this pretty effectively for stock scams (eg, video posted on Friday that Bezos is launching electric car tanks Tesla, etc). Or even more sinister fake videos of things that may be true (video of Musk assaulting some woman).
What I’m actually scanning for are small markets with small $10-100k upside potential like a fake video of Tom Cruise that impact Mission Impossible opening that can be shorted through mediapredict or other markets that aren’t SEC regulated.
Not to do it of course, but to see trends in how these videos are developed.
[0] http://www.dmlp.org/legal-guide/california-recording-law
I don't really think reputation has a better chance of solving this either, people are all-too-happy to believe anything that confirms their personal beliefs and ignore whatever contradicts them regardless of source.
It's interesting that a lot of the comments here look at how to prove a certain video is real (and assume that people will accept that any video without a certain trait is fake) rather than proving directly that a certain video is fake.
It's a first step, and my thinking has also evolved since—expect more updates in this space soon.
> AI researchers and platform technologists need similar in-house or third-party societal impact review boards to help them evaluate the potential unintended consequences of their work.
As a machine learning researcher myself (albeit one who is probably unlikely to run afoul of one of these boards), this suggestion for research gives me pause. In the context of deepfakes, are you suggesting that it would have been more prudent to stop the research that made it possible before it had a chance to do so? This seems like a good way to ensure that less scrupulous parties harness destructive technologies before more scrupulous parties (the kind who'd adopt review boards).
I imagine this debate to some extent has already played out with nuclear weapons and bioweapons. I'm much less familiar with these two spheres, but as I recall research continued in both.
I'm not suggesting stopping research on this. But the way in which that research is communicated, what the focus is, and the way which particular sorts of replication is enabled or discouraged has an impact.
An example that the HN crowd likely knows better is vulnerability/exploit disclosure. There are good ways to do it and harmful ways to do it.
You can perhaps think about deepfakes as an exploit for the human sensors/cognition that handle authentication.
I do think most of it will come from people lower down the political food chain (the meme creators, etc) who seek to convince others of their POV out of desperation and/or maliciousness.
Some of these techniques can be used to make fixes in motion picture where it would be far too expensive to reshoot.
But there is the opposite where this technology can be used to falsify a video that can potentially be used to make false confessions or change what a person said for nefarious purposes.
Welcome to the world of what was once thought to be science fiction.
It could also raise awareness of the issue, so that if people see a video that seems "too good to be true" on a different platform, they would know that it might not be real.
People will believe what they want to believe. We're also seeing that they don't like mega corps acting as arbiters of truth. So, I'm not sure what, if any, effect this moderation/curation effort would have.
I don't currently have any better ideas that "getting deepfaked". Suggestions?
You should be extremely selective about what analog media you trust.
If you weren't there to see it happen yourself, all you've got to go on is somebody else's perception. Be careful who you trust.
I think that we may want to consider using cryptographic signature methods as a way to mark materials as coming from the source they claim to be from. Though this is not a perfect solution.
Though I still work with analog film, though it is becoming more and more cost prohibitive everyday. Analog can just as easily be altered. If you looked at a silver gelatin print, if the printmaker was talented you probably would never be able to detect the changes to the image that were made during the printing process. Is the resulting image a fake, or is it just an artists vision?