Seems to me a bit like the adversarial networks used to train AI.

Of course, you want an adversarial network that's difficult to trick, otherwise the training will produce suboptimal results.