DeepFaceLab: A tool that utilizes ML to replace faces in videos
github.com
github.com
Considering how EV is essentially dead, I would assume it hasn't been very effective.
Because everybody isn't going to pay for the trust indicator and we're going to consume plenty of things that don't have it, the indicator is basically just noise. It's why EV failed and we're about to see it repeat with this BIMI technology in the email world IMO.
On the other hand, if you get a big giant danger warning that stands out because everything else doesn't have it...it gets attention.
( E.g., see California cancer warnings on just about everything -- makes it hard to distinguish between the <thing that will kill you now> vs the <thing that may increase your chance of a curable cancer by 1/10^6>, so it all gets ignored. )
A blue check mark only says this account/person has our imprimatur. Nothing more.
There are people who don’t have who are better disseminators than those who have them.
That would be terrible. Basically it would turn media into versions of Xinhua or former Pravda. Only approved voices and opinions get the “authenticity” mark.
The near future world signs everything with a personal identifier so they can prove they what people are seeing is genuine.
I imagine these identifiers could be extended to things like security cameras and things too so that there could be some verification that the video footage hasn't been doctored.
All of this of course relies on the general public getting the knowledge and tools to seamlessly do this verification on a daily basis. In the book it's mostly taken care of using google glass style wearables.
That’s because people are apt to believe those sources and not critically question them.
The exceptions are local news about local events that has no political angle, and occasionally investigative journalism.
The most common method that legitimate news sources use to bias news is by omission, and second by using a misleading context. Neither is strictly a lie, but the result is as much fake news as something fabricated.
It's not the primary focus of the story, but it does at least quickly attempt to answer these problems to advance the narrative in a reasonable way.
Overall, I just have a hard time imagining the details of how it would apply in today's world.
Pizza-gate is bonkers, but that doesn't mean there isn't a healthy segment of the population that believes it as a way to channel their collective hatred for someone. Which is itself just a way to solidify their group identity, much the same as the resurgence of flat-earthers who feed off of their own ignorance while proclaiming their supposed objectivity.
I think the reason people are downvoting your comment (aside from the obvious reason, which is that it's a little low-effort) is that people didn't like pizzagate. It had an altright, Trumpy 'beware the deep state satan pedophile illuminati' feel to it that we oppose being true for identity reasons (ie, we are not those people who say trollish, unintellectual things like that!).
Although pizzagate is false, it's not as false (in a kind of Bayesian way) as it was before the Epstein island entered the public conscience.
The NY Times seems to like it. https://www.nytimes.com/2019/06/14/books/review/fall-or-dodg...
To worry about fake photos/videos seems out of touch with the current state of the internet. People will upvote and bicker about a screenshot of the text of a tweet. Why even bother with ML when an image of an outrageous sentence is enough?
Photo editing became "household" at least a decade ago. So if you think of that, I can't see why the perception of video's in 3 years would be different than photo's today.
It's ironic that we're essentially being pushed back to a pre-technology, pre-media time. "If you didn't see it with your own two eyes, you can't believe it"
https://jamespublishing.com/2013/objecting-video-audio-evide...
https://www.esquiresolutions.com/federal-rules-catch-digital...
As the technology improves and becomes accessible soon anybody will be able to edit videos convincingly without having to invest a massive amount of time or money. That's going to have a large impact I think.
So in that sense it could have still be done pretty cheap.
A better example would probably be the moon landing footage or 9/11. And then it's really only questioned by conspiracy theorists and specifically because a state actor might have wanted to hide the evidence of a conspiracy.
Tomorrow with this new technology I might not convince you that Bigfoot exists but I might be able to show you extremely convincing footage of Margaret Thatcher and Mao Zedong frolicking in the Swiss Alps while Tupac is watching.
I had no idea Mao and Tupac were friends.
Nothing changes.
It was always possible, but it used to be far more difficult.
Given what we know of human memory and visual processing, you'd be a damned fool to trust your own eyes any better.
We listened to Sirius 117 thru our ride and they always tell crime stories. It was a death row story and when the narrator said that the only evidence they had against a suspect was a DNA but it was enough to put him on electric chair, she chuckled a little. It was odd so I asked her why. As a court reporter she witnessed many horrors written by life, but this one was particularly scary for another reason. She was reporting on a case where the only evidence of a murder was a DNA, even though a suspect was in New York on a day where victim was murdered in Los Angeles. The DA rested case explaining that DNA evidence is enough and holds up in courts for many years and that they do not need to explain anything else, how he go from NY to LA at the night of murder or really nothing else mattered "because of DNA." So defended lawyers ran a suspect DNA via one of popular DNA-kit websites and to their surprise they found a good enough match. She was lucky to be called in when the case was heard again (most of the time different court reporter would be called in depending of availability) and they showed that result to DA/prosecutors asking if they would consider prosecuting based on that person DNA. They look at it and said it was the same person DNA. So defended lawyers revealed that it was different person. To judge shock, then DA came to say that DNA match doesn't have to be 100% and over the span of human existence some 100 billion people lived so its quite possible for extremely similar DNA to be found in wild. She recalls looking at judge face at that moment and he turned pale blue - you couldn't help to see this man wondering if over the span of last 2 decades, he may have sentenced hundreds of people based on inadequate evidence. The case did not ended there but shortly after when she was checking on it, the whole case all of sudden got sealed and even year later was still locked down. All she knows is that most likely he wasn't found guilty, as checking on his name recently did not yield any record on Cali's death row list.
So so much of DNA being of 100% proof.
Fake news on high octane race gas.
* Everyone can access it. Not just sophisticated actors anymore
* It might make us rethink the entire "truth" chain, on how we source our information
Of course, the two go hand-in-hand, but the latter point is overdue: while video is perhaps more glaring, it's something that's needed in a lot of other areas as well (text -- news articles, messages, mail; sound -- phone calls, etc; image -- photoshop, though we start to get used to it).
Why do you like this?
Basically, a (good) deepfake was out of reach for all but a few until recently, so if some bad actor with money and resources wanted to fake a video they could do it and few people would even think it could be faked.
Nowadays, the bar is way higher for this, as every video can be suspect.
But how many will suspect them? If 10 million people watch a video but only (generously) 1 million people think or are at least willing to consider it may be fake, that is very effective disinformation. We have to remember that the average person (especially in older generations) is likely unaware that this sort of technology is possible and easily accessible.
I agree that provenance could become more important, but I don't see it changing how memes spread. For a lot of people it's just entertainment and they don't care whether it's true.
I think parent was making a point that once this tech is in the hands of the common man, it's value for sophisticated players might diminish strongly. Same way as an undisclosed 0day in a high value target (say IOS), is extremely valuable. Once it gets disclosed and everyone knows about it, people can come up with workarounds and eventual fixes, rendering the threat basically defanged.
I'm leaning towards the same school of thought. I'd rather see this tech widely accessible and hence all video/picture material henceforth considered completely untrustworthy, than living in the misguided notion that video evidence is trustworthy.
This is why I am all for the current hype and the apps. Everyone that has access to electronic media should now that videos are now more easily to manipulate than ever. This undermines the attack vector of supporting fake news with fake videos. (At least it should, I do not know about the psychological side of it. Maybe even knowledge of a video's falsehood does not diminish its impact that much. Still I think it is best to make video fakeability common knowledge.)
On the flip side, this enables revenge porn deepfakes and person to person harm on a much, much more common level. If you are someone who becomes briefly famous on the internet, expect that a TON of harmful or disturbing videos of "you" will show up in short order. I'm not sure we have good regulations and laws in place to protect people against the misuse of this tech yet.
There's not a single thing in the chain of tools and knowhow required to produce this that can be kept out of the hands of malicious actors.
The best course of action now is to rapidly educate people of the implications and hope that the initial wave of abuse will be without too many casualties.
Now, when you see a before and after photo, or an advertisement, the default is to assume it has been manipulated. That change came about with broad access to once-rare tools.
So to do this pro/con analysis you probably need to include the pros of these related technologies as well, which are certainly more than trivial amusement.
Like the time my uncle (by marriage) sent a JibJab to the whole family featuring 4 recently-dead family members... real big pro, I'm sure... /s
Could also make for some interesting low-budget book commercials/trailers where authors (or I guess anyone making a video starring people) can design the characters in a video without worrying about what actors are available and/or what CGI costs.
https://www.nextgov.com/emerging-tech/2019/08/darpa-taking-d...
https://techcrunch.com/2018/04/30/deepfakes-fake-videos-darp...
...The consensus seems to be that ongoing detection is going to be a challenge as the tech evolves. I suppose that is like anything though....
If you think this won't attract huge investments from VCs, you are likely too optimistic.
Can you think of many recent examples of video evidence was the sole evidence in a news report?
Bereavement. It won’t be for everyone, but perhaps some people who have lost loved ones might like to feel that they are still alive by bringing them into modern day videos.
Fantasies. You hear about how some people like to live out fantasy lives in video games. I heard one about a severely disabled teen or young man who gets to be a big strapping strongman in his favorite video game. Imagine an old lady who can create a video of herself at age 25 free soloing El Capitan by deepfaking herself into an Alex Honnold video.
No, this tech will be used to amuse and/or influence the intellectually challenged.
https://github.com/deepfakes/faceswap
With a big development community and interesting results.
Punking your friends too.
In a productive context in filmmaking you can have fun with doubles and stuntmen and stunt women. Or if you have the rights to license the face and likeness of an actor actress and they can be in five films, ten films in a year (forgetting about brand dilution). John Wayne can be in a new Western movie,
"Punking" does not qualify as non-malicious.
Next question: does it make CCTV footage obsolete as a form of permissible evidence in court?
Or going the other way, use a real video of your friend and replace their face with a famous actor.
But who should they look like?
Maybe that spot could be sold...
I like your idea about selling spots though, definitely something I have never thought of before.
But thinking about it some more, I at least don't mind clothing ads using me as the model. (Assuming there's opt in and all)
It's better than mentally swapping out the skin tone and body type to see if it was really the clothes that looked good or the attractive model.
I am not a lawyer, though.
Why use a real face? You could generate a fake one using a GAN. You may even be able to build a fully automatic anonymize tool that can identify the broad characteristics of each face present (skin tone, age, male/female etc) then use that to generate replacement faces.
Often English language films are poorly dubbed for local (mostly non-English speaking) markets. Having a local/recognisable actor filling the role would be a big improvment.
I can see them modifying the actual original actor to speak in a local language however to sync with the localized dialogue.
That said, your proposition is also interesting. It raises some questions though. Imagine an indie French film where the protagonist (or whoever) gets replaced by a Hollywood actor. Or an indie Angolan film which gets the treatment. There would be lots of purists (for many reasons) protesting this kind of option.
If an actor dies while filming a movie, it can be required in the contract that they allow their face to be generated for scenes that haven't been filmed yet.
They could also use this to use a younger face as they age.
Travel industry can have uses as well - these days many places offer video recordings of their clients doing some dangerous activity; for those that can't do them yet still want the video they could offer pre-baked scenes where taking a selfie could kill somebody and then replace their face/body with some green screen footage of their client using this app and some pose estimator/body switcher. Or just show you how much would you enjoy some beach property in some hot location with a video featuring you etc. This would make it useful for custom ads featuring you as well.
Reshooting scenes of movies without having to get key actors back on location.
Immersive storytelling where you get to be placed inside a movie.
Actor safety where they can go through transformations for a shoot, such as extreme weight loss or weight gain.
I can really only think of it in terms of entertainment. There may be some therapy benefits which are yet to get discovered, but I would imagine that's a whole new level of complexity.
Not to mention you'll just come off as a huge weirdo as soon as you're found out.
>Ha! People who wear power suits, makeup, expensive haircuts and shoes are already doing all this. Technology just means you don't have to spend time in the stylist's chair to accomplish it any more!
Did you forget what you wrote?
They are for reinforcing it.
This isn't imprecision of language.
That side issue does not have much to do with the video issues under discussion, though.
I shudder at the thought of a world in which always-on lies about your appearance become standard practice.
Have physical reality and truth really become irrelevant?
I remember watching a TV show years ago about a woman who had had lots of plastic surgery, found the man of her dreams, and was now pregnant. She was afraid the baby would come out looking totally different than she did, because she had physically changed herself so much, and her husband would leave her over the deceit.
At the same time, with cosmetic surgery, you are changing physical reality. Someone has sliced up your body and reconfigured it to be more what you want.
So it's not necessarily a "lie", as such.
Like I said, though, I see what you're getting at, and it's a valid point.
* A demo video of what life in a new house would be like. Pre-recorded video featuring actors waking up, eating breakfast, and living life in a home could have the actors replaced with the potential home buyers.
* Security tests... for facilities that have to look for "banned" people, occasionally a face could be replaced on the video to see if the security guard notices. This is similar to the TSA sometimes testing workers by putting in fake images of guns or bombs into the X-Ray display to make sure their workers are actually paying attention.
* Retroactively correcting visual media that is no longer appropriate. I would love a version of "The Cosby Show" where Bill Cosby was replaced by another famous black actor just because it would trigger my liberal guilt less. This could allow studios to craft contracts where if a star committed a grievous act or if they die, their likeness could be replaced throughout the series and they would stop receiving royalties for their visual appearance. Dialog could be re-dubbed as well to complete the retroactive continuity.
* Updating low-res video... If there is a very poor and grainy interview with someone, the image might be able to be enhanced more completely by replacing the subjects face with a higher resolution copy of their own face!
* Smart therapy mirrors... people who have had disfigurements or have body image problems could have a smart mirror that displays a modified image for either self-esteem purposes or to help them psychologically.
* Same applies for funhouse/haunted house mirrors... seeing a mirror image of yourself with a zombie-ified head would be a great illusion.
* Real-world facial masking... You know how people can 3d map the exterior of a building and then project an image onto it using that 3d map taken into account to pull off some spectacular illusions? [1]
Something similar, sans holographic technology, could be done to make someone appear like someone else. A camera and 3d scanner environment to track the facial contours and location, with a projector projecting the image of another face onto the real face. Similar to the scene in Bladerunner 2049 where the AI girlfriend rents a real-world avatar for a personal encounter with her boyfriend. [2]
If this gets mainstream enough, "revenge porn" would also lose a lot of its current value, being easily dismissible as "fake". (And so would genuine photographical evidence...)
Journalists could use it to hide their sources from the public. If this gets reaaly good, it would become a nice alternative to motion capture for cinematography (pose estimation, including facial expressions). Then swap in the 3D model, or concept art of the person you want to appear on screen. Significantly reduce the costs for makeup, and enable more in post-prod. That was probably already available for big teams with big budgets, but this makes it more mainstream.
Personalized entertainment.
1. including "Face off"
[1] https://www.dezeen.com/2019/05/24/salvador-dali-deepfake-dal...
[EDITED]
probably only funny if you watch these folks regularly
https://github.com/deepfakes/faceswap#faceswap-has-ethical-u...
Celebrities could sell the 'rights' to their faces and then ordinary people could 'act' in commercials and later on even movies and then just stick the face of the celebrities?! xD Celebrities will become cheaply available to 'star' in your next Netflix series.
(But unfortunately, even if celebrities don't sell their faces, they'll be stolen!)
What I think Zao does is preprocess the videos (manually or highly accurate facepoint detection). They pre-calculate the transforms (standard facemorphing algorithms with opacity/alpha tweaks), and shading depending on the scene lighting. Then they just need a good frontal selfie, or do some frontalization, and keypoint detection, and the rest can be rendered/computed without much resources, following a pre-defined script.
If more advanced than facemorphing, then perhaps something more like: https://github.com/facebookresearch/supervision-by-registrat... (pre-fitting a 3D face mask, then texturing it with your selfie)
From the results, I think you are exactly right in how they are accomplishing those videos.
As for the political implications, go watch "Wag the Dog" again.
The natural conclusion goes past revenge porn to revenge videos of beheadings, murders, etc that can be sent to friends or family members who aren't technically savvy enough to recognize fakes for what they are. Not to mention the existing phenomenon of blaming random civilians for murders and terror attacks - this will be far more convincing when it's paired with convincing fakes. It'll probably claim some lives.