Supplementary code to Selecting Models based on the Risk of Damage Caused by Adversarial Attack

This folder contains all the scripts needed to reproduce the plots and tables contained in the paper Selecting Models based on the Risk of Damage Caused by Adversarial Attack.

The plots and tables can be reproduced in two phases:

1. Simulate an attack on four models by running src/simulate_attack.py. This simulation uses RobustBench to attack the models with different adversarial attack and writes the distance of the closest found adversarial example for each observation to data/minimal_distances.csv
2. Create the plots and tables by running create_experiment_plots.py. This python script uses the distances from data/minimal_distances.csv to create the plots as .pdf files and the tables as .tex files in out/.

As data/minimal_distances.csv is shipped along with the code, one can run phase (2) using the existing simulation data.

The out/logistic_regression_dma.pdf plot does not require any simulation and can be reproduced by running src/logistic_regression.py.