joelma1/public

To understand adversarial attacks on medical deep learning systems for various imaging modalities (MRI, CT, mammogram) and use adversarial training to defend against attacks

Adversarial Attacks on Medical Deep Learning Models

In this project, we will investigate the robustness of existing medical DL models by testing their performance on adversarial examples from various medical imaging datasets. We will use adversarial training to strengthen them against adversarial attacks. Some questions we will explore are 1) whether some imaging modalities are more susceptible to adversarial attacks than others; 2) use PGD-based adversarial training to increase model robustness

Code Flow

- configs.py # Designates parameters for adversarial attack and training experiments
- utils.py # Defines useful functions used in experiments
- data_loader.py # Loads in desired training and test set, preprocesses data
- data_generator.py # Performs data augmentations (flips, rotations) on medical datasets
- models.py # Defines models (VGG-16 model)
- train.py # Trains model on data
- attacks.py # Creates adversarial attacks
- main.py # Applies attack on desired dataset and evaluates model accuracy on adversarial examples
- adv_trainer.py # Applies PGD-based adversarial training to models

How to Run

1. Clone Repository from Github
2. Install necessary dependencies python pip3 install -r Requirements.txt
3. Edit config.py to customize parameters
4. Run data_generator.py to augment medical datasets
5. Run train.py to train DNN models
6. Run main.py for adversarial attacking experiments
7. Run adv_trainer.py for adversarial training experiments