This code implements the adversarial sparsity metric introduced in the paper : How many perturbations break this model? Evaluating robustness beyond adversarial accuracy.
In order to run this code you must have installed RobustBench (https://github.com/RobustBench/robustbench), and Robustness (https://github.com/MadryLab/robustness). To run all experiments you should also have downloaded the DiffJPEG repository (https://github.com/mlomnitz/DiffJPEG) and the generated CIFAR10 data in https://github.com/deepmind/deepmind-research/tree/master/adversarial_robustness. Please specify the local paths to these repository and data, and other datasets, in paths.py
The scripts to run for a large part of our results are in the run.sh file. To run custom experiments, more details on argument options can be found in the scripts/scripts_utils.py file.