Slashdot Mirror

← Back to Stories (view on slashdot.org)

The Defense Department Has Produced the First Tools For Catching Deepfakes (technologyreview.com)

Posted by msmash on from the fighting-back dept.

Fake video clips made with artificial intelligence can also be spotted using AI -- but this may be the beginning of an arms race. From a report: The first forensics tools for catching revenge porn and fake news created with AI have been developed through a program run by the US Defense Department. Forensics experts have rushed to find ways of detecting videos synthesized and manipulated using machine learning because the technology makes it far easier to create convincing fake videos that could be used to sow disinformation or harass people. The most common technique for generating fake videos involves using machine learning to swap one person's face onto another's. The resulting videos, known as "deepfakes," are simple to make, and can be surprisingly realistic. Further tweaks, made by a skilled video editor, can make them seem even more real. Video trickery involves using a machine-learning technique known as generative modeling, which lets a computer learn from real data before producing fake examples that are statistically similar. A recent twist on this involves having two neural networks, known as generative adversarial networks, work together to produce ever more convincing fakes. The tools for catching deepfakes were developed through a program -- run by the US Defense Advanced Research Projects Agency (DARPA) -- called Media Forensics. The program was created to automate existing forensics tools, but has recently turned its attention to AI-made forgery.

20 of 45 comments (clear)

  1. Not just pasting faces, and not just video by Rick+Schumann · · Score: 1

    https://www.wnycstudios.org/st...
    The original image can be manipulated, and the audio even more convincingly manipulated to say pretty much anything you want.

    1. Re:Not just pasting faces, and not just video by admin7087 · · Score: 1

      That's one of the worse examples and wouldn't convince many people, but the problem is going to be massive once the techniques have been improved. The real problem is that people get their "information" from the most shady sources and tend to switch their brain of. Captain Disillusion's Youtube Channel is a great and entertaining way of becoming more skeptical.

    2. Re:Not just pasting faces, and not just video by rtb61 · · Score: 1

      People will believe what they want to believe in order to favour their own personal biases and even their own particular personalities. They will claim to believe it, even if they don't and they will claim disbelief even if they believe, belibers (no accident) will be beliebers, idiot just-ins to the world of thoughts.

      Here's a news flash for you, prove the photo or vid is fake, so the fuck what, they'll just claim yours is fake about the fake because yeah, it can be. Unless you legalise the truth, legalise news as a profession and prosecute for lying in the news, with news as a protected word, you are stuck in a fake loop. Your fake, their fake, your fake, their fake, endless loop, of fake news.

      So the professional paranoid now have a way of detecting fakes, well, you know exactly what they are going to do with that, make sure their fakes pass the fucking test ;D (you know that is true, all of your, know it is 100% true, believe me ;D).

      --
      Chaos - everything, everywhere, everywhen
  2. Scarier than it sounds by omnichad · · Score: 4, Insightful

    Any tool that can catch a deepfake can be used to help train the deepfake generator. This cover story just tells us that DARPA is working on training their own system for generating deepfakes. Dueling neural nets is the new normal and there's no reason to think they only developed one side of this.

    1. Re:Scarier than it sounds by Aighearach · · Score: 1

      I hope this is true, because without deepfakes how are slashdot users going to use dating sites?!

      Think of the innocent neckbeards, people, think of the neckbeards.

    2. Re:Scarier than it sounds by nitehawk214 · · Score: 1

      But any tool that can generate deepfakes can be used to train the deepfake detector.

      It will eventually conclude that it must nuke humanity.

      --
      I'm a good cook. I'm a fantastic eater. - Steven Brust
  3. I'm Convienced by Anonymous Coward · · Score: 1

    That this will be a cat and mouse game of trying to evade detection. What we will need and likely eventually settle on is video & audio that includes digital certificates embedded in the video encoding. That way there is a certified chain of proof from the moment the video is recorded to the point of distribution. This will likely matter in cases where legal burden of proof is required - security camera feeds, official court & governmental documentation, and news anchor releases.

  4. Re:let me guess by Anonymous Coward · · Score: 1

    This.

    The whole reason this near real-time video editing was released by the CIA and/or NSA was because someone somewhere has some video of some politician doing something so incredible heinous, there needed to be a plausible deniability already in the ether.

  5. where this technology will go now by slashmydots · · Score: 3, Funny

    Well time to give it up and make technology for "dank fakes" where you take a Shiba Inu's head and render it onto someone's body in a video.

  6. Futurama FTW by fuzznutz · · Score: 1

    I saw it once on a TV show. They told the computer that one guy always lies and then he tells the computer that he's lying.

    You need to update your references...

    Robot Santa: Nice try, but my head was built with paradox-absorbing crumple zones.

  7. Any tutorials to use it for not-porn? by Wulf2k · · Score: 1

    Here's probably as good a place as any to ask.

    Any time I've tried some simple searching on how-to, it's started going down porn paths. Does anybody have a porn-free tutorial on how to get started on creating a deepfake?

    I don't care to create anything that would fool anybody, I just think it'd be neat to screw around with over a few lunch hours at work.

    1. Re:Any tutorials to use it for not-porn? by Gilgaron · · Score: 1

      Other than the whole NSFW aspect the methodology should be about the same either way?

    2. Re:Any tutorials to use it for not-porn? by Kjella · · Score: 1

      Not really, but I've seen the results of people trying it out and it seems extremely fickle, needs a pretty big training set and you have to mask out other faces if there are any. People use it on celebs because there's tons of photos/videos and someone OCD enough to compile the set. And when they're trying to morph it into porn all the scenes that don't work and all the times it loses track doesn't matter so much. You just pick the clips the virtual mask sticks and glue it together. You'll notice that in a lot of the scenes the woman isn't actually moving her head much so it's a relatively straightforward face replacement, it has trouble with all sorts of half obscured faces, partial faces entering/leaving the camera, accessories like hats or glasses and so on.

      In the end it's a really cumbersome process that probably won't work on that cool clip you think it will. Sure you can probably morph someone you know into a Hollywood movie for a few scenes, but it's still a lot of work for a gag. If you just want to put yourself in there it's probably just easier to shoot some footage on green screen and overlay it rather than deep fake it. It's literally for when you want the person being deep faked to take over that actor/actress' performance... Oh also it just works for the face itself, so if the body or hair color/length/style is a mismatch it looks like your faked person in a wig or worse. Another reason why porn is easier, with so much to choose from you can find a similar body that just needs a new face to be a convincing fake.

      --
      Live today, because you never know what tomorrow brings
    3. Re:Any tutorials to use it for not-porn? by Wulf2k · · Score: 1

      Gag-level quality was pretty much what I was going for, but it sounds like the effort/reward ratio really isn't there.

      I'll probably still poke at it sometime but I'll keep my expectations low.

  8. SFW - Safe For Work Tutorials by Anonymous Coward · · Score: 1

    Safe For Work Tutorials:
    https://www.deepfakes.club/

    The easiest way to get started in deep fakes is to use the Python scripts.

    1. Re:SFW - Safe For Work Tutorials by Wulf2k · · Score: 1

      At a quick glance that seems to be ideal.

      Thank ye, noble AC.

  9. How to identify a deepfake? by PhunkySchtuff · · Score: 1

    At the moment it's pretty easy to identify a deepfake video. The AI's don't "know" that people regularly blink, they see this as a glitch in the data set.
    If you look at pretty much any deepfake video, no matter how realistic it is, the person who's been deepfaked will never blink.

    --
    Specialist Mac support for creative pros, Melbourne
  10. The thing about adversarial networks by goombah99 · · Score: 1

    If you build a better mousetrap an adversarial system builds a better mouse in response. Thus I wonder how they can make a detector that continues to detect. The whole idea of a GAN is to generate things that defeat the detector. So what's the strategy to make s detector the generator can't beat?

    --
    Some drink at the fountain of knowledge. Others just gargle.
  11. Re:I can't see this arms race lasting long by HiThere · · Score: 1

    I don't know whether the process is AI or not. Do you need to train it? If so, it's probably AI. This doesn't mean it isn't a rather specialized AI, of course. But the higher the bar gets, the more general the AI is going to need to be. Adding in proper eye blinking requires a significantly more powerful AI than just matching head positions....of course, you could, in principle, do the entire thing with image processing matches, but the required dataset would expand at least greater than linearly, and probably exponentially. Then there's matching skin textures under various light conditions. Then...

    --

    I think we've pushed this "anyone can grow up to be president" thing too far.
  12. adversarial by thygate · · Score: 1

    .. adversarial networks, work together ..

    what ?