AI Image Fakery

[lpetrich]

Fake Obama created using AI video tool - BBC News - YouTube

Startups Are Racing to Commercialize DeepFakes’ Powerful, Internet-Breaking AI - "A slew of startups are using AI to push the boundaries of media manipulation—and they all promise to have ethics in mind"

For more than a decade, son-father team Eric and Albert Yang have run a small but successful software firm called Topaz Labs, which builds various stand-alone image-editing tools along with plug-ins for software like Adobe’s After Effects. They spent years developing their algorithms, hiring image enhancement PhD researchers to do laborious, highly specialized work to craft their software — which can pull exceptionally sharp images from video, or add an electrical glow to a moving image.

Then the machine learning revolution happened. Suddenly, they could train neural nets to do the hard work of sharpening images: “One of our biggest surprises . . . was seeing we could just throw away our 10 years of hard work instantly.” So, over the last year, the studio has more or less rebuilt itself around machine learning. “It’s almost like a new startup,” says Eric.

Topaz Labs – Photography software powered by machine learning -- offers noise removal, sharpening, and enlarging with AI.

The AI is trained on what real-world mages tend to look like, so it adds the extra details by using what a real-world image would likely have.

Related to this is style transfer, giving images the stylistic features of other images: Prepare Yourself For Super-Powerful Photoshop AI and Watch This “Van Gogh” Filter Turn A Time-Lapse Into An Impressionist D - make pictures look like Van Gogh paintings.

There is now an audio version called Modulate. It can take one's voice and give it the features of other people's voices. The company attempts to address the issue of fraudulent audio by watermarking its output, by putting in very low-level audio information.

Meograph, Inc. does yet another variation, projecting 2D images onto 3D models, thus giving a 3D-model head a skin from a portrait picture. Thus, one can have a computer-game avatar that looks realistic from all directions, and that is realistic-looking when it talks.

 

[DrZoidberg]

The Running Man is finally a documentary

 

[lpetrich]

House holds hearing on "deepfakes" and artificial intelligence amid national security concerns - CBS News -- the audio gets out of sync with the video in the last half of it, but it does show some successful recent fakery.

Someone having his computer make an imitation of his voice for having a conversation with his mother.

Transfer of facial expressions from some video onto a picture. Demonstrated on Donald Trump and Vladimir Putin.

A clip from Saturday Night Live where Kate McKinnon satirizes Elizabeth Warren, with a version where KMK's face is replaced by EW's.

Then four portraits from the site This Person Does Not Exist They look real, but they are all fake. How it works: Style-based GANs – Generating and Tuning Realistic Artificial Faces | Lyrn.AI
[1812.04948] A Style-Based Generator Architecture for Generative Adversarial Networks by Tero Karras | Research
[1406.2661] Generative Adversarial Networks by Ian Goodfellow (@goodfellow_ian) / Twitter
The Illustrated Transformer – Jay Alammar – Visualizing machine learning one concept at a time

There is a companion site that works on cats: This cat does not exist -- its fakery has flaws here and there.

Some recent video fakery, but done with some video editing: Video of Speaker Nancy Pelosi's 'Slurred' Speech Is Fake (VIDEO)

Rush Limbaugh - I'm A Nazi Video - YouTube -- it's the audio that's the fakery -- snippets of Rush Limbaugh's radio broadcasts assembled to make it seem like he is describing himself with some of his numerous insults.

It may now be possible to take some video of Rush Limbaugh speaking and make a fake video of him seemingly singing that song.

 

[lpetrich]

Why ‘deepfake’ videos are becoming more difficult to detect | PBS NewsHour - has some stuff on detecting deepfakes by looking for imperfections, like editing out parts of videos, lips not moving as expected for speech, and incorrect lighting.

 

[lpetrich]

Top AI Researchers Race to Detect ‘Deepfake’ Videos: ‘We Are Outgunned’ from The Washington Post:

In a world where video has played a pivotal role in shaping modern history, researchers said it’s nevertheless critical to find a way to spot the fakes — and some fear what could happen if the authority of video slips away.

“As a consequence of this, even truth will not be believed,” Memon said. “The man in front of the tank at Tiananmen Square moved the world. Nixon on the phone cost him his presidency. Images of horror from concentration camps finally moved us into action. If the notion of not believing what you see is under attack, that is a huge problem. One has to restore truth in seeing again.”

 

[bilby]

It's impossible to restore truth in seeing again.

If it's technologically possible for something to be a fake, then it's always going to be actually possible that it's a fake.

Whether that matters, or how much, is an interesting question. It's mostly going to depend on how much trust we are required to place in systems that are untrustworthy, as technology outstrips law.

Fifteen years ago, I got a call from a freight forwarder, who told me that the faxed documents I had sent him were only valid with a signature. I told him that as it was a fax, a signature was meaningless; But he refused to accept this.

So I re-sent the paperwork, with his signature on it.

 

[Jimmy Higgins]

Well I'm glad the United States has a Government that is savvy enough to bite this unbelievably serious problem before something terrible happens. That the Government consists of people willing to listen to experts and govern responsibly and create constitutional legislation that bans this shit.