4 ms·

I have a feeling this cat and mouse chase game will never end, and trying to tackle using tech is a meaningless journey. Wouldn’t all these techniques just end

I have a feeling this cat and mouse chase game will never end, and trying to tackle using tech is a meaningless journey. Wouldn’t all these techniques just end up help training the deep fake algorithms to avoid detection so it becomes even closer to perfect?

brundolf 6y ago
"never end" and "meaningless" aren't necessarily the same thing. Plain old cybersecurity is a cat and mouse game that never ends, but it's far from meaningless.
- cocktailpeanuts 6y ago
  Yes I understand that. But I’m specifically talking about this in the context of machine learning, more specifically deep learning. This tech by design is supposed to learn on its own. So while the traditional cyber security is a game between humans, this is a battle between humans and machines that learn faster than humans.
  - karmakaze 6y ago
    Deep learning doesn't entirely learn on its own, it takes labelled input. All that has to happen now is to feed in samples of good deepfakes but with a fake pulse labelled as such so that they aren't generated as being plausible.
- AnthonyMouse 6y ago
  Most cybersecurity isn't really a cat and mouse game. Yes, the bad guys find a vulnerability and then the good guys patch it. But then there is one less vulnerability. In principle you eventually find all of the vulnerabilities. But deep fakes are cat and mouse the other way. A new way of detecting them is like a vulnerability in their creation. Then it gets patched and there is one less vulnerability.
  - jjk166 6y ago
    In cybersecurity you asymptotically approach being secure, but you never reach a point where it's safe to say "there are no more vulnerabilities" With deepfakes you'll asymptotically approach indistinguishability, but there will always potentially be some lurking tell. I've seen cases where photosphopped images were identified because the shadows on the ground were incorrect for the time of day of the image's timestamp, one could imagine similarly that a pulse detector could be sensitive enough to not only detect that a pulse exists but to evaluate if the pulse is right for a person's age and body type, or even compare it to known videos of the same person under similar circumstances. Anything that didn't really happen will have to be inconsistent with what did, and that inconsistency, no matter how slight, will always be potentially detectable.

brundolf 6y ago

"never end" and "meaningless" aren't necessarily the same thing. Plain old cybersecurity is a cat and mouse game that never ends, but it's far from meaningless.

cocktailpeanuts 6y ago
Yes I understand that. But I’m specifically talking about this in the context of machine learning, more specifically deep learning. This tech by design is supposed to learn on its own. So while the traditional cyber security is a game between humans, this is a battle between humans and machines that learn faster than humans.
- karmakaze 6y ago
  Deep learning doesn't entirely learn on its own, it takes labelled input. All that has to happen now is to feed in samples of good deepfakes but with a fake pulse labelled as such so that they aren't generated as being plausible.
AnthonyMouse 6y ago
Most cybersecurity isn't really a cat and mouse game. Yes, the bad guys find a vulnerability and then the good guys patch it. But then there is one less vulnerability. In principle you eventually find all of the vulnerabilities. But deep fakes are cat and mouse the other way. A new way of detecting them is like a vulnerability in their creation. Then it gets patched and there is one less vulnerability.
- jjk166 6y ago
  In cybersecurity you asymptotically approach being secure, but you never reach a point where it's safe to say "there are no more vulnerabilities" With deepfakes you'll asymptotically approach indistinguishability, but there will always potentially be some lurking tell. I've seen cases where photosphopped images were identified because the shadows on the ground were incorrect for the time of day of the image's timestamp, one could imagine similarly that a pulse detector could be sensitive enough to not only detect that a pulse exists but to evaluate if the pulse is right for a person's age and body type, or even compare it to known videos of the same person under similar circumstances. Anything that didn't really happen will have to be inconsistent with what did, and that inconsistency, no matter how slight, will always be potentially detectable.