How adversarial attacks can disrupt seemingly stable accurate classifiers

Oliver J. Sutton, Qinghua Zhou, Ivan Yu. Tyukin, Alexander N. Gorban, Alexander Bastounis, Desmond J. Higham

Abstract: Highlights•A new theory for studying accuracy, adversarial attacks, and robustness is presented.•We present experiments confirming the theory on standard benchmarks.•The theory reveals when adversarial attacks affect seemingly stable classifiers.•Adding noise during training is inefficient for eradicating adversarial examples.