This repository contains the source code of the paper entitled:

A PAC-Bayes Analysis of Adversarial Robustness
Paul Viallard, Guillaume Vidot, Amaury Habrard, Emilie Morvant
NeurIPS 2021, 2021

Running the experiments

The datasets can be downloaded and preprocessed by running the following command in your bash shell.

./generate_data

Then, to run the experiments, you have to run the following command in your bash shell. (The outputs of the experiment are in data/merge)

./run_all

To generate the plots of the paper, you have to run the following command in your bash shell.

cd plot/ 
python generate_bar.py
python generate_plot.py
cd ../

Dependencies

The code was tested on GNU Bash 4.4.20 and Python 3.6.10 with the packages

• h5py (2.10.0)
• matplotlib (3.3.4)
• numpy (1.18.1)
• pandas (1.1.5)
• scikit_learn (0.23.1)
• seaborn (0.11.1)
• torch (1.6.0)
• torchvision (0.7.0)

These dependencies can be installed (using pip) with the following command.

pip install h5py==2.10.0 matplotlib==3.3.4 numpy==1.18.1 pandas==1.1.5 scikit_learn==0.23.1 seaborn==0.11.1 torch==1.6.0 torchvision==0.7.0