3 ms·

Even less likely to succeed if they wholesale fake it from nothing. Essentially the problem is very similar to detecting artifacts in video and audio. Reverse

Even less likely to succeed if they wholesale fake it from nothing. Essentially the problem is very similar to detecting artifacts in video and audio. Reverse steganography, if you prefer.

jobigoud 7y ago
It's not similar because it's based on adversarial networks. Any improvements in detection methods can be incorporated to improve the fakes.
- kuzehanka 7y ago
  Not only this, but the discriminator component of the adversarial net requires the original output from the generative component. Run the output of the generator through some noise, lens distortion, video compression, downscale it to a 480p surveillance style video, and the discriminator is at a huge disadvantage.
- toxik 7y ago
  This is an oversimplification. I can have ways of detecting artifacts that is not based on a differentiable loss function, so no, not true.
parksy 7y ago
Interesting train of thought, a potentially basic idea to combat this could be having a cryptographic algorithm used to sign each frame, basically set a specific value in one or more pseudo-random pixels each frame of a video, forcing a 1 or 0 in the least significant bit in a known pattern along with the signing identity of the original creator - an attacker without the private key wouldn't be able to successfully re-sign the video as they'd have no way of really knowing the original sequence, and they wouldn't have the signing key to be able to regenerate the original sequence or recreate it from the tampered frames. Unfortunately I am not a crypto expert but maybe a codec could check for signature pixels every frame in playback and provide a warning if no key signature is detected in realtime for any of the loaded frames if the user has a matching verification key. Either an attacker would maliciously blank or randomise the LSB which would show the entire video as unsigned, or the tampered portions would show up if and when the signature pixel chain is broken. I guess the issue would be around securely distributing verification keys that can verify both the signing identity and verify the correct sequence? But then that would put the ability to recreate the sequence in the hands of the attackers? Oh and I guess variable bitrates could cause issues. Damn this stuff is harder than I thought. There's probably a billion ways this wouldn't work and a thousand existing solutions?

Even less likely to succeed if they wholesale fake it from nothing. Essentially the problem is very similar to detecting artifacts in video and audio. Reverse steganography, if you prefer.

jobigoud 7y ago

It's not similar because it's based on adversarial networks. Any improvements in detection methods can be incorporated to improve the fakes.

kuzehanka 7y ago
Not only this, but the discriminator component of the adversarial net requires the original output from the generative component. Run the output of the generator through some noise, lens distortion, video compression, downscale it to a 480p surveillance style video, and the discriminator is at a huge disadvantage.
toxik 7y ago
This is an oversimplification. I can have ways of detecting artifacts that is not based on a differentiable loss function, so no, not true.

parksy 7y ago

Interesting train of thought, a potentially basic idea to combat this could be having a cryptographic algorithm used to sign each frame, basically set a specific value in one or more pseudo-random pixels each frame of a video, forcing a 1 or 0 in the least significant bit in a known pattern along with the signing identity of the original creator - an attacker without the private key wouldn't be able to successfully re-sign the video as they'd have no way of really knowing the original sequence, and they wouldn't have the signing key to be able to regenerate the original sequence or recreate it from the tampered frames. Unfortunately I am not a crypto expert but maybe a codec could check for signature pixels every frame in playback and provide a warning if no key signature is detected in realtime for any of the loaded frames if the user has a matching verification key. Either an attacker would maliciously blank or randomise the LSB which would show the entire video as unsigned, or the tampered portions would show up if and when the signature pixel chain is broken. I guess the issue would be around securely distributing verification keys that can verify both the signing identity and verify the correct sequence? But then that would put the ability to recreate the sequence in the hands of the attackers? Oh and I guess variable bitrates could cause issues. Damn this stuff is harder than I thought. There's probably a billion ways this wouldn't work and a thousand existing solutions?