Adversarial learning in quantum artificial intelligence

Quantum artificial intelligence exploits the interplay between and quantum physics: on one hand, a plethora of tools ideas from can be adopted to tackle intricate problems; other computing could also bring unprecedented opportunities enhance, speed up, or innovate intelligence. Yet, learning systems, similar classical ones, may suffer adversarial attacks: adding tiny carefully-crafted perturbation legitimate input data would cause systems make incorrect predictions at notably high confidence level. In this paper, we introduce basic concepts learning, as well some recent advances along line. First, basics both learning. Through concrete examples, involving classifications phases two-dimensional Ising model three-dimensional chiral topological insulators, reveal vulnerability machine matter. addition, demonstrate classifiers with example classifying hand-written digit images. We theoretically elucidate celebrated no free lunch theorem perspectives, discuss universality properties attacks in classifiers. Finally, possible defense strategies. The study uncovers notable potential risks for which have far-reaching consequences future interactions two areas.