OOD-Adv

On the Adversarial Robustness of Out-of-distribution Generalization Models

This code is used to reproduce the experiments the main paper of:

On the Adversarial Robustness of Out-of-distribution Generalization Models Xin Zou, Weiwei Liu 37th Conference on Neural Information Processing Systems (NeurIPS), 2023 [Website] [Paper]

Our code is based on DomainBed, for more details about the benchmark DomainBed, please refer to DomainBed.

Reproduce Experimental Results

1 Download the datasets

python -m domainbed.scripts.download --data_dir=./domainbed/data

2 Lunch the sweeps

Lunch the sweep for the algorithms that does not use adversarial training:

python -m domainbed.scripts.sweep launch --data_dir=./domainbed/data \
            --output_dir=./sweep/output/path \
            --command_launcher multi_gpu \
            --algorithms ERM MLDG CDANN VREx RSC \
            --datasets RotatedMNIST ColoredMNIST VLCS PACS OfficeHome \
            --n_hparams 20 --n_trials 1 --single_test_envs

Lunch the sweep for MAT and LDAT:

python -m domainbed.scripts.sweep launch --data_dir=./domainbed/data \
            --output_dir=./sweep/output/path \
            --command_launcher multi_gpu \
            --algorithms MAT LDAT \
            --datasets ColoredMNIST VLCS PACS OfficeHome \
            --n_hparams 20 --n_trials 1 --single_test_envs \
            --steps 8000

Lunch the sweep for AERM and RDANN:

python -m domainbed.scripts.sweep launch --data_dir=./domainbed/data \
            --output_dir=./sweep/output/path \
            --command_launcher multi_gpu \
            --algorithms AERM RDANN \
            --datasets RotatedMNIST ColoredMNIST VLCS PACS OfficeHome \
            --n_hparams 20 --n_trials 1 --single_test_envs \
            --steps 8000

3 find the best parameters

python -m domainbed.scripts.find_best_hparam --input_dir=./sweep/output/path \
            --dataset=RotatedMNIST

python -m domainbed.scripts.find_best_hparam --input_dir=./sweep/output/path \
            --dataset=ColoredMNIST

python -m domainbed.scripts.find_best_hparam --input_dir=./sweep/output/path \
            --dataset=VLCS
            
python -m domainbed.scripts.find_best_hparam --input_dir=./sweep/output/path \
            --dataset=PACS
            
python -m domainbed.scripts.find_best_hparam --input_dir=./sweep/output/path \
            --dataset=OfficeHome

4 Retrain the models with the best parameters

python -m domainbed.scripts.resweep launch --command_launcher multi_gpu \
            --datasets RotatedMNIST ColoredMNIST VLCS PACS OfficeHome \
            --algorithms ERM MLDG CDANN VREx RSC MAT LDAT AERM RDANN \
            --sweep_base_dir ./sweep/output \
            --selection_methods IIDAccuracySelectionMethod

5 Attack the trained models

python -m domainbed.scripts.sweep_attack launch --command_launcher multi_gpu \
            --datasets RotatedMNIST ColoredMNIST VLCS PACS OfficeHome \
            --sweep_base_dir ./sweep/output \
            --selection_methods IIDAccuracySelectionMethod \
            --train_methods ST \
            --attacks FGSM PGD \
            --algorithms ERM MLDG CDANN VREx RSC MAT LDAT AERM RDANN

python -m domainbed.scripts.sweep_attack launch --command_launcher multi_gpu \
            --datasets RotatedMNIST VLCS PACS OfficeHome \
            --sweep_base_dir ./sweep/output \
            --selection_methods IIDAccuracySelectionMethod \
            --train_methods ST \
            --attacks AutoAttack \
            --algorithms ERM MLDG CDANN VREx RSC MAT LDAT AERM RDANN

6 Collect the results

Run the following command to collect the results for FGSM and PGD

python -m domainbed.scripts.collect_attack_results --input_dir ./sweep/output \
            --datasets RotatedMNIST ColoredMNIST VLCS PACS OfficeHome \
            --selection_methods IIDAccuracySelectionMethod \
            --train_method ST \
            --attacks FGSM PGD \
            --algorithms ERM MLDG CDANN VREx RSC MAT LDAT AERM RDANN \
            --attack_all_in_one --latex

Then, you can find the .tex result file in the "./OOD-Adv/sweep/output/attacks_results/" folder.

Run the following command to collect the results for AutoAttack

python -m domainbed.scripts.collect_attack_results --input_dir ./sweep/output \
            --datasets RotatedMNIST VLCS PACS OfficeHome \
            --selection_methods IIDAccuracySelectionMethod \
            --train_method ST \
            --attacks AutoAttack \
            --algorithms ERM MLDG CDANN VREx RSC MAT LDAT AERM RDANN \
            --attack_all_in_one --latex

Then, you can find the .tex result file in the "./OOD-Adv/sweep/output/attacks_results/" folder.

Note that the result file will be overwritten if you run the two commands sequentially.