Black-Box Adversarial Attacks on Deep Neural Networks: A Survey

Deep neural networks are capable of performing many challenging tasks, such as image classification, speech recognition and game playing. However, recent research shows that deep neural networks can be tricked by adversarial attacks, which craft adversarial examples by adding subtle perturbations to the clean examples. Adversarial attacks can be broadly divided into white-box attacks and black-box attacks. This paper presents a comprehensive survey on black-box attacks. The black-box attacks are divided into three categories: transfer-based attacks, score-based attacks and decision-based attacks. We also introduce some existing defenses against black-box attacks.