5 ms·

Example from a news story. After seeing the video, I must admit I’m not sure if I believe the guy or not, which is scary. Edit: Better link from deadmutex bel

by jcims 5y ago

Example from a news story. After seeing the video, I must admit I’m not sure if I believe the guy or not, which is scary. Edit: Better link from deadmutex below - https://www.youtube.com/watch?v=vqr0oER03SE https://www.youtube.com/watch?v=vqr0oER03SE https://www.wfla.com/8-on-your-side/better-call-behnken/instagram-account-taken-over-by-imposter-who-posted-deepfake-video-of-tampa-man/ https://www.wfla.com/8-on-your-side/better-call-behnken/inst...

k4runa 5y ago
It's exactly the same situation with my friend. “I just invested $300 into Bitcoin and got $10,000 back. Gotta try it,” except the numbers are higher... but she actually says it in the video too.
deadmutex 5y ago
Another link, same story: https://www.youtube.com/watch?v=vqr0oER03SE https://www.youtube.com/watch?v=vqr0oER03SE
lelandfe 5y ago
Telling that a major news outlet can't get anything more than a "we're looking into it" response from IG support.
- stjohnswarts 5y ago
  I mean what if they are looking into it though? Would you prefer a half baked response?
echelon 5y ago
I work really closely on deep fake tech [1] and I'd say I'm relatively current with the state of the art in the literature. This was not deepfaked. The person recorded it themselves and is lying. The video quality is too good. The lighting and movements lack mistakes. It can't be first order model, wav2lip, or any of the relatively new audio to video models. The audio doesn't suffer from spectral noise, and it matches the lip movements close enough to not be TTS. Voice conversion (VC) introduces pitch issues that are readily apparent, and it's incredibly hard to train VC models without a ton of parallel audio data from source and target speakers. This is absolutely a lie (not a deepfake) and I'd bet money on it. [1] I created https://fakeyou.com https://fakeyou.com cartoon and celebrity TTS, real time voice to voice mapping for VTubers, and am currently working on ML blendshapes.
- jcims 5y ago
  The perfection in the audio is precisely what made me skeptical.
- blueblisters 5y ago
  I agree. I think the most telling sign is the hand gesture for the number 3 when he says "just invested 300 bucks". Don't think Deepfake models can understand intent yet.
  - idontwantthis 5y ago
    Wouldn’t it be a real person making the gesture and his face is deepfaked on?
    - yonixw 5y ago
      I think you are right, if you find a similar person with similar voice (even commission them 5$ in fiverr) and only deep-fake the face in a low quality video, this is very much achievable in today's state of the art.
  - whoisjuan 5y ago
    What do you mean? It’s not like a deepfake is a model that produces a new video from scratch in a completely random setting. It needs a base video to be modified. They need a video of a person talking to do the deepfake.
    - quitit 5y ago
      I think the issue is that there would be insufficient audio sources to generate new spoken language.
      - pyinstallwoes 5y ago
        That is not an issue. You can generate new spoken language with not much more than a word.
pangolinplayer 5y ago
Yeah. It's almost like the post is a scam and we are all being duped into promoting the non-deepfake crypto scam.

by jcims 5y ago

k4runa 5y ago

It's exactly the same situation with my friend. “I just invested $300 into Bitcoin and got $10,000 back. Gotta try it,” except the numbers are higher... but she actually says it in the video too.

deadmutex 5y ago

Another link, same story: https://www.youtube.com/watch?v=vqr0oER03SE https://www.youtube.com/watch?v=vqr0oER03SE

lelandfe 5y ago

Telling that a major news outlet can't get anything more than a "we're looking into it" response from IG support.

stjohnswarts 5y ago
I mean what if they are looking into it though? Would you prefer a half baked response?

echelon 5y ago

I work really closely on deep fake tech [1] and I'd say I'm relatively current with the state of the art in the literature. This was not deepfaked. The person recorded it themselves and is lying. The video quality is too good. The lighting and movements lack mistakes. It can't be first order model, wav2lip, or any of the relatively new audio to video models. The audio doesn't suffer from spectral noise, and it matches the lip movements close enough to not be TTS. Voice conversion (VC) introduces pitch issues that are readily apparent, and it's incredibly hard to train VC models without a ton of parallel audio data from source and target speakers. This is absolutely a lie (not a deepfake) and I'd bet money on it. [1] I created https://fakeyou.com https://fakeyou.com cartoon and celebrity TTS, real time voice to voice mapping for VTubers, and am currently working on ML blendshapes.

jcims 5y ago
The perfection in the audio is precisely what made me skeptical.
blueblisters 5y ago
I agree. I think the most telling sign is the hand gesture for the number 3 when he says "just invested 300 bucks". Don't think Deepfake models can understand intent yet.
- idontwantthis 5y ago
  Wouldn’t it be a real person making the gesture and his face is deepfaked on?
  - yonixw 5y ago
    I think you are right, if you find a similar person with similar voice (even commission them 5$ in fiverr) and only deep-fake the face in a low quality video, this is very much achievable in today's state of the art.
- whoisjuan 5y ago
  What do you mean? It’s not like a deepfake is a model that produces a new video from scratch in a completely random setting. It needs a base video to be modified. They need a video of a person talking to do the deepfake.
  - quitit 5y ago
    I think the issue is that there would be insufficient audio sources to generate new spoken language.
    - pyinstallwoes 5y ago
      That is not an issue. You can generate new spoken language with not much more than a word.

pangolinplayer 5y ago

Yeah. It's almost like the post is a scam and we are all being duped into promoting the non-deepfake crypto scam.