The vast majority of the adversarial work on DNN’s is about using small input perturbations to push the system from one stable output state to a different stabl

The vast majority of the adversarial work on DNN’s is about using small input perturbations to push the system from one stable output state to a different stable output state, mostly because work is mostly done on images rather than video so repeatedly re-classifying the same image isn’t very interesting. That doesn’t mean stable state to stable state is the only form of adversarial input that can be found, it just means that’s all that has been looked for so far. The odds that there are adversarial perturbations that reliably push the system into an unstable state between two or more output states is probably quite high. It just hasnt hit the threshold at which it becomes interesting for postdocs to look for yet.