Adaptive AutoAttack

"Automatic Discovery of Adaptive Attacks on Adversarial Defenses"

Chengyuan Yao, Pavol Bielik, Petar Tsankov, Martin Vechev

arxiv: https://arxiv.org/abs/2102.11860v1

Setup:

python version==3.8.X

1. pip install -r requirement.txt
2. pip install git+https://github.com/RobustBench/robustbench.git@v0.2.1
3. download folder for models from: https://drive.google.com/file/d/1ajCXluAPUPiyLe2ka9i41jj1YQ6u7YX-/view?usp=sharing (2.2GB)
4. unzip the folder and replace the folder zoo/saved_instances

Reproduce main result:

Models need to be selected in the script, and epsilon need to be adjusted.

python bmA_eval.py (A^3 on group A models)
python bmB_eval.py (A^3 on group B models)
python bmC_eval.py (A^3 on group C models)
python autoattack_eval.py (AutoAttack on defenses)
python aa_detector_eval.py (AutoAttack on defenses with detector)

Reproduce scalability result:

python bmA_scalability.py

Reproduce ablation study:

Run Loss formulation comparison:

python progression_comparison.py --algo 0 --dir=<result_dir_name, eg loss_noloss1>

Run TPE vs Random comparison:

python progression_comparison.py --algo 1 --dir=<result_dir_name, e.g tpe_random1>

Analyze result (need to update the result directory names in the script)

python progression_analysis.py --algo <bool> --dual <bool, to plot TPE vs Random and loss formulation together>

Reproduce the scatter plot on attack-scores: