Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (1): 152-157.DOI: 10.3778/j.issn.1002-8331.2007-0204

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Research of DNN Adversarial Attack on COVID-19 CT Image Dataset

HU Geng, CAI Yanguang   

  1. School of Automation, Guangdong University of Technology, Guangzhou 510006, China
  • Online:2022-01-01 Published:2022-01-06



  1. 广东工业大学 自动化学院,广州 510006

Abstract: In the research of applying deep learning to intelligent COVID-19 CT recognition, a large number of researchers train DNN models to understand the content of medical image, and assist in the diagnosis of the COVID-19. Firstly, this paper proposes the AMDRC-Net architecture, in which the residual structure solves the network degradation problem through identity mapping. However, the residual structure hinders the exploration of new features, and the long and short attention guidance mechanism is inspired by the latest research such as the attention mechanism. Afterwards, it focuses on the security of deep learning models and discusses the adversarial attack based on gradient ascent. In order to solve the problem of singularity, the long and short attention mechanism is used to increase effective counter disturbances while reducing redundant disturbances. Then, the counterattack attacks proposed algorithm A-IM-FGSM transforms the adversarial attack problem into an adaptive constraint problem, that is, the idea of micro-transformation can be used in iterative attacks to explore the relationship between the attention guidance mechanism and DNN adversarial attack. In the final experiments, the AMDRC-Net is used for model training on the COVID-19 CT dataset, the comparison experiments, visualization experiments, and adversarial attack experiments are completed.

Key words: COVID-19 CT image, attention mechanism, deep learning, DNN adversarial attacks

摘要: 在深度学习应用于新型冠状肺炎CT智能识别的研究中,大量研究人员通过构建深度神经网络训练模型,从而理解医学影像数据内容,辅助新冠肺炎诊断。提出AMDRC-Net架构,其中的残差结构,通过恒等映射解决了网络退化问题,与此同时,针对残差结构阻碍新特征探索的新问题,受到注意力机制等最新研究启发,研究了长短注意力引导机制。关注深度学习模型安全性问题,讨论基于梯度上升的对抗攻击方法;为了解决其单一性问题,通过长短注意力机制,增加有效对抗扰动的同时减少冗余扰动,紧接着,提出的对抗攻击算法A-IM-FGSM,将对抗攻击问题转化为自适应约束问题,即可微变换思想用于迭代攻击中,探究注意力引导机制与DNN对抗攻击的相互关系。最后进行的实验中,在新型冠状肺炎CT数据集上,通过AMDRC-Net进行模型训练,设计对比实验、可视化实验、对抗攻击实验。

关键词: 新冠肺炎CT影像, 注意力引导机制, 深度学习, DNN对抗攻击