MedicalAI-UAP

This repository contains the codes used in our study on Universal adversarial attacks on deep neural networks for medical image classification.

Terms of use

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

Hirano H, Minagi A & Takemoto K (2021) Universal adversarial attacks on deep neural networks for medical image classification. BMC Med. Imaging 21, 9. doi:10.1186/s12880-020-00530-y

Usage

# Directories
.
├── MedicalAI-UAP
└── dataset
    ├── CellData
    └── ISIC2018_Task3_Training_Input

1. Check the requirements.

• Python 3.6.6
• tensorflow-gpu 1.12.0
• Keras 2.2.4
• Keras-Applications 1.0.8
• numpy 1.18.2
• pandas 1.0.5
• scikit-learn 0.22.2.post1
• imbalanced-learn 0.6.2
• matplotlib 2.0.2
• Pillow 4.2.1
• tqdm 4.43.0

2. Download the following datasets.

• Skin lesion images
  • ISIC2018 Task 3: Lesion Diagnosis: Training
• OCT and Chest X-ray images
  • Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning

3. Generate the dataset.

# malanoma
python make_data.py --dataset melanoma --img_dir ../dataset/ISIC2018_Task3_Training_Input

# oct
python make_data.py --dataset oct --img_dir ../dataset/CellData/OCT

# chestx
python make_data.py --dataset chestx --img_dir ../dataset/CellData/chest_xray

4. Train

python train_model.py --dataset melanoma --model inceptionv3

# `--dataset` argument indicates the dataset: melanoma, oct or chestx (default).
# `--model` argument indicates the model: inceptionv3 (default), vgg16, vgg19, resnet50, inceptionresnetv2, densenet121 or densenet169.

5. Install the methods for generating universal adversarial perturbations (UAPs).

• pip install git+https://github.com/hkthirano/adversarial-robustness-toolbox

6. Generate UAPs.

# non-targeted UAP
python generate_nontargeted_uap.py --dataset melanoma

# UAP for targeted attacks to MEL
python generate_targeted_uap.py --dataset melanoma --target MEL
# `--target` argument indicates the target class:
#   when dataset is melanoma, the target class: MEL, NV, BCC, AKIEC, BKL, DF or VASC.
#   when dataset is oct     , the target class: CNV, DME, DRUSEN or NORMAL.
#   when dataset is chestx  , the target class: NORMAL or PNEUMONIA (default).

# random UAP
python generate_random_uap.py --dataset melanoma

7. Results

The targeted UAP causes the Inception-v3 models to classify most skin lesion images into melanoma.