The goal of the discriminator in a GAN is to distinguish between real inputs and inputs synthesized by the generator.
Suppose I train a GAN until the generator is good enough to fool the discriminator much of the time. Could I then use the discriminator as a classifier that tests whether an input belongs to a single class?
For instance, if I train StyleGAN to be able to synthesize photorealistic cats, could I use the trained discriminator to detect whether an image is a cat or not?
My thinking is that perhaps the discriminator would be more accurate than other classifier models because it has effectively trained on many, many more inputs thanks to the generator.
On the other hand, perhaps the discriminator is somehow worse because it has been trained overwhelmingly on cat-like images (assuming the generator has gotten pretty good), and hasn't seen a wide variety of negative examples. It is concerned less with "is this a cat?" than "what are the tell-tale signs of this being synthetic?"