Towards evaluating the robustness of deep diagnostic models by adversarial attack

Mengting Xu, Tao Zhang, Zhongnian Li, Mingxia Liu, Daoqiang Zhang

Abstract

Highlights•A framework for robustness evaluation of deep diagnostic models under adversarial attack has been proposed.•A novel benchmark dataset and a robustness evaluation metric are proposed to evaluate the common perturbation robustness of three representative deep diagnostic models.•Two new defense methods are designed to handle adversarial examples in deep diagnostic models, called Multi Perturbations Adversarial Training (MPAdvT) and Misclassification Aware Adversarial Training (MAAdvT) MAAdvT), respectively.•Experimental results have shown that the proposed defense method scan effectively improve the robustness of deep diagnostic models.
BibTeX
@article{xu2021towards,
  author = {Mengting Xu and Tao Zhang and Zhongnian Li and Mingxia Liu and Daoqiang Zhang},
  title = {{Towards evaluating the robustness of deep diagnostic models by adversarial attack}},
  journal = {Medical Image Analysis},
  year = {2021},
  doi = {10.1016/j.media.2021.101977},
}
Figure