5 ms·

Maybe... photoshop detection hasn’t gone down that path as far as I know. I would imagine the problems of hiding detection in video to be a magnitude harder to

Maybe... photoshop detection hasn’t gone down that path as far as I know. I would imagine the problems of hiding detection in video to be a magnitude harder to hide?

bogwog 6y ago
True, but photoshop is usually done by hand. An AI-generated deepfake could maybe eventually be trained so that it outputs videos that are indistinguishable from the real thing.
- thebean11 6y ago
  The detection side also gets the benefits of AI though. I can't see the generation side getting that far ahead.
  - Gh0stRAT 6y ago
    From what I understand though, generators always have an advantage because the generator is allowed to "see" the discriminator's gradients during training. [0] >The model for the discriminator is usually more complex than the generator (more filters and more layers) and a good discriminator gives quality information. In many GAN applications, we may run into bottlenecks where increasing generator capacity shows no quality improvement. Until we identify the bottlenecks and resolve them, increasing generator capacity does not seem to be a priority for many partitioners. [1] Put another way: GAN training ends when the discriminator can no longer meaningfully distinguish real from fake. By definition then, the best generator will have no useful discriminator that can distinguish its output from real data. (conversely, if you did have such a discriminator, you could use it to train a better generator) [0] https://developers.google.com/machine-learning/gan/generator https://developers.google.com/machine-learning/gan/generator [1] https://towardsdatascience.com/gan-ways-to-improve-gan-performance-acf37f9f59b https://towardsdatascience.com/gan-ways-to-improve-gan-perfo...
  - AnthonyMouse 6y ago
    > I can't see the generation side getting that far ahead. How far ahead do they need to be? Suppose that it's cat and mouse, at least initially. Every six months someone comes up with a new way to detect the best known deepfakes, then six months after that there is a new way to evade that means of detection as well. Someone drops a deepfake five weeks before an election.

Maybe... photoshop detection hasn’t gone down that path as far as I know. I would imagine the problems of hiding detection in video to be a magnitude harder to hide?

bogwog 6y ago

True, but photoshop is usually done by hand. An AI-generated deepfake could maybe eventually be trained so that it outputs videos that are indistinguishable from the real thing.

thebean11 6y ago
The detection side also gets the benefits of AI though. I can't see the generation side getting that far ahead.
- Gh0stRAT 6y ago
  From what I understand though, generators always have an advantage because the generator is allowed to "see" the discriminator's gradients during training. [0] >The model for the discriminator is usually more complex than the generator (more filters and more layers) and a good discriminator gives quality information. In many GAN applications, we may run into bottlenecks where increasing generator capacity shows no quality improvement. Until we identify the bottlenecks and resolve them, increasing generator capacity does not seem to be a priority for many partitioners. [1] Put another way: GAN training ends when the discriminator can no longer meaningfully distinguish real from fake. By definition then, the best generator will have no useful discriminator that can distinguish its output from real data. (conversely, if you did have such a discriminator, you could use it to train a better generator) [0] https://developers.google.com/machine-learning/gan/generator https://developers.google.com/machine-learning/gan/generator [1] https://towardsdatascience.com/gan-ways-to-improve-gan-performance-acf37f9f59b https://towardsdatascience.com/gan-ways-to-improve-gan-perfo...
- AnthonyMouse 6y ago
  > I can't see the generation side getting that far ahead. How far ahead do they need to be? Suppose that it's cat and mouse, at least initially. Every six months someone comes up with a new way to detect the best known deepfakes, then six months after that there is a new way to evade that means of detection as well. Someone drops a deepfake five weeks before an election.