News for hackers. Stuff that matters.
Re:New deepfake detector can spot a real/fake video based on... by cudgy on Friday November 18, 2022 @09:05PM (#33665958)
So, basically a NOT operator on a neural network. Does this even require a differentiation?
Re:New deepfake detector can spot a real/fake video based on... by nl on Friday November 18, 2022 @11:37PM (#33666980)
A NN is trained to make something indistinguishable from reality as possible - so it can't tell the difference. The inverse of that will just claim reality is fake too.

What you need a a deep fake and a real video of the same thing, then train on the difference. Clearly this is impossible - which is what makes the problem hard.

Re:New deepfake detector can spot a real/fake video based on... by sigmoid10 on Saturday November 19, 2022 @05:32AM (#33668935)
You actually don't need that. You only need a set of real videos and a generator for fake ones. Then train the discriminator to tell these two classes apart and make use of its differentiability to update the generator in tandem with the discriminator.
Re:New deepfake detector can spot a real/fake video based on... by anankaie on Sunday November 20, 2022 @09:05AM (#33680964)
Would you run into class balance issues if you do not have a growing pool of initial videos?
Re:New deepfake detector can spot a real/fake video based on... by sigmoid10 on Sunday November 27, 2022 @01:00PM (#33764976)
Not if you have a balanced dataset. If not, you're already getting into trouble with GANs, which is why they were superseded by diffusers on image generation tasks.
Re:New deepfake detector can spot a real/fake video based on... by rusticpenn on Saturday November 19, 2022 @02:34AM (#33667954)
Not really a not operator. It’s more like tuning the generated models until the fake detector cannot detect it.
Re:New deepfake detector can spot a real/fake video based on... by anankaie on Sunday November 20, 2022 @09:03AM (#33680953)
It depends on how you set it up. You can use a very non-expressive valuation that assigns a score, but then you need to use reinforcement learning techniques to transform that score to a model update. Alternatively, if your valuation represents a differentiable metric function from your target you have a way of going directly from your output to a model update.

The second way requires dramatically fewer update steps (usually) than the first. Thus - having your adversarial target be differentiable definitely helps, though it is possible to do even absent such a criterion.