Understanding adversarial attacks on deep learning based medical image analysis systems
作者：
Highlights：
• Medical image DNNs are easier to be attacked than natural non-medical image DNNs.
• Complex biological textures of medical images may lead to more vulnerable regions.
• State-of-the-art deep networks can be overparameterized for medical imaging tasks.
• Medical image adversarial attacks can also be easily detected.
• High detectability may be caused by perturbations outside the pathological regions.
摘要
•Medical image DNNs are easier to be attacked than natural non-medical image DNNs.•Complex biological textures of medical images may lead to more vulnerable regions.•State-of-the-art deep networks can be overparameterized for medical imaging tasks.•Medical image adversarial attacks can also be easily detected.•High detectability may be caused by perturbations outside the pathological regions.
论文关键词：Adversarial attack,Adversarial example detection,Medical image analysis,Deep learning
论文评审过程：Received 18 July 2019, Revised 3 March 2020, Accepted 12 March 2020, Available online 1 May 2020, Version of Record 1 November 2020.
论文官网地址：https://doi.org/10.1016/j.patcog.2020.107332
