Backdoored-COVID-Net

This repository contains the codes used in our study on Backdoor attacks to deep neural network-based system for COVID-19 detection from chest X-ray images.

Terms of use

MIT licensed. Happy if you cite our paper when using the codes:

Matsuo Y & Takemoto K (2021) Backdoor attacks to deep neural network-based system for COVID-19 detection from chest X-ray images. Appl. Sci. 11, 9556. doi:10.3390/app11209556

Usage

Requirements

• python 3.6.6
• tensorflow-gpu==1.15.0
• scikit-learn==0.23.2
• numpy==1.19.2
• opencv-python==4.4.0.44
• matplotlib==3.3.2
• Pillow==7.2.0

Image dataset and Model

See lindawangg/COVID-Net: Table of Contents for installation.

• Check the requirements
• Generate the COVIDx dataset
• Download the following datasets
  • covid-chestxray-dataset
    • Figure1-COVID-chestxray-dataset
    • rsna-pneumonia-detection-challenge
    • COVID-19-Radiography-Database
  • Use create_COVIDx.ipynb
• Download the COVID-Net models available
  • COVIDNet-CXR4-A

# Directories
.
├── models
│   └── COVIDNet-CXR4-A
│ 
├── data
│   ├── train
│   └── test
│
├── labels
│ 
├── train.py
├── eval.py
├── data.py
└── poison.py

Backdoored models

e.g., obtain the backdoored model for targeted attacks to COVID-19.

python train.py \
   --backdoor_attack = True \
   --attack_type = 'targeted' \
   --targeted_class = 2

Fine-tuned models from backdoored models

python train.py \
   --backdoor_attack = False