Query efficient black-box adversarial attack on deep neural networks

Summary: In this paper, the authors propose Cyclical Adversarial Attack (CA2), a method that uses cyclical optimization algorithm to generate adversarial examples and enhance the transferability between models. Extensive experiments demonstrate the effectiveness and superiority of this method.

Summary: This paper delves into the issue of adversarial examples in medical image analysis, finding that medical DNN models are more vulnerable to adversarial attacks compared to models for natural images. Furthermore, it is noted that medical adversarial attacks can be easily detected, with simple detectors achieving over 98% detection AUC.