Skip to content

ChiaraBi/dzofw-adversarial-attacks

 
 

Repository files navigation

Universal Adversarial Perturbation starring Frank-Wolfe

The main goal of this project is to analyze different Stochastic Gradient Free Frank-Wolfe algorithms based on Decentralized and Distributed architectures for producing Universal Adversarial Perturbations. These perturbations are designed to fool advanced convolutional neural networks on the classification task performed over the MNIST dataset.

The algorithms we analyzed are based on Anit Kumar Sahu and Soummya Kar paper, and they are:

  • Decentralized Stochastic Gradient Free Frank-Wolfe
  • Decentralized Variance-Reduced Zeroth-Order Frank-Wolfe
  • Distributed Zeroth-Order Frank-Wolfe

In the document report - Bigarella_Poletti_Singh_Zen.pdf, we reported some key concepts about adversarial attacks and the developed algorithms; finally we discussed the experiments we carried out and our results.

Code guidelines

In order to install the required libraries, run the following code:

pip install -r requirements.txt

In the src folder, there are the main following scripts:

  • main_distributed.py
  • main_variance_reduced.py
  • main_decentralized.py

The final analyses are summarized in the following Jupyter notebooks:

  • demo.ipynb
  • alexnet_test.ipynb

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 91.6%
  • TeX 5.1%
  • Python 1.9%
  • PureBasic 1.4%