break-fix

We perform adversarial attacks to a CNN trained on medical images. Then, we apply counter measures to make the model robust and mantain high accuracy.

The project is part of the course "Ethics in Artificial Intelligence" at the University of Bologna.

Contributors:

• Alessandro Folloni
• Daniele Napolitano
• Marco Solime

Breaking

We perform the following attacks (code in the sandbox.ipynb notebook):

• Data poisoning
• Adversarial examples:
  • FGSM: Fast Gradient Sign Method
  • PGD: Projected Gradient Descent
• Biasing the loss:
  • NoisyLoss
  • FoolingLoss
• Manipulating Gradient direction

Fixing

We apply the following counter measures:

• Data Augmentation: makes the model robust and invariant to small changes in the input
• Adversarial Training: to contrast adversarial examples
• Pattern Detector: We train a model to detect the possible presence of adversarial examples, to be used in the early stages of the production pipeline

Requirements

To run the code, you need to install the libraries from the requirements.txt file. You can do it by running the following command:

pip install -r requirements.txt