This is the official implementation of our ECCV paper "Adversarial Robustness on In- and Out-Distribution Improves Explainability".
Maximilian Augustin, Alexander Meinke, Matthias Hein

We have updated the source code to reflect the findings of the newer survey paper Uncovering the Limits of Adversarial Training against Norm-Bounded Adversarial Examples. The changes to our original paper include:

1. Use of a larger model architecture (WideResNet34x10 instead of ResNet50)
2. Fixed number of attack steps (10/20 instead of increasing schedule)
3. Proper Min Max optimization of the adversarial training objective instead of logit based-loss for inner maximization.
4. Averaging of model weights (EMA)
5. Optional: Additional unlabeled data from 500k-TI
6. Optional: Use ASAM optimization (related to Adversarial Weight Perturbation)

For CIFAR10 only training. we found AutoAugment to improve robustness whereas it decreases performance when combined with 500k-TI. All models with additional data are thus trained with default (crop + mirror) augmentation only. We also include the option to use Auto-PGD (APGD) during training which can help to increase robustness even further.

Results for the new models trained with an in-distribution radius of 0.5 and an out-distribution radius of 1.0 are given in the following table, robustness values are evaluated using the AutoAttack package. The model checkpoints are available here. To use these models, simply extract the .zip file in the project directory.

Before running the code, you will have to update the base dataset path variable in Line 6 in:

utils.datasets.paths

Cifar10/100 will be downloaded automatically. For training your own models, you will have to manually place the 80 Million Tiny Images .bin file in a subdirectory: "80M Tiny Images"

If you want to train models on 500k-TI, download "ti_500K_pseudo_labeled.pickle" from the authors Github and place it into a subdirectory "cifar10_ti_500k" in your datasets folder.

The required packages can be installed from the included yml file using conda:

conda env create -f environment.yml

To create counterfactual explanations for a trained model, you have to run:

cifar10_visualize_failure.py

By default, we create 100 counterfactual explanations for our pretrained models, however new models can be added by changing the "model_descriptions" variable.

The results can be found in the "Cifar10Eval" directory.

To evaluate our models on the CIFAR10 test set, run:

cifar10_robustness_test.py

Note that this requires the AutoAttack package.

To train a new model on CIFAR10, use

run_training_cifar10.py

To replicate above models, use one of the following commands:

CIFAR10 PGD:

python run_training_cifar10.py --gpu 0 1 --net wideresnet34x10 --augm autoaugment_cutout --id_steps 10 --od_steps 20 --train_type advacet --epochs 300 --ema True --ema_decay 0.999 --test_epochs 10 --dataset cifar10 --schedule cosine --eps 0.5 --od_eps_factor 2

CIFAR10 APGD:

python run_training_cifar10.py --gpu 0 1 --net wideresnet34x10 --augm autoaugment_cutout --id_steps 10 --od_steps 20 --id_pgd apgd --od_pgd apgd --train_type advacet --epochs 300 --ema True --ema_decay 0.999 --test_epochs 10 --dataset cifar10 --schedule cosine --eps 0.5 --od_eps_factor 2

500k-TI PGD:

python run_training_cifar10.py --gpu 0 1 --net wideresnet34x10 --augm default --id_steps 10 --od_steps 20 --train_type advacet --epochs 300 --ema True --ema_decay 0.999 --test_epochs 10 --dataset semi-cifar10 --schedule cosine --eps 0.5 --od_eps_factor 2

500k-TI APGD:

python run_training_cifar10.py --gpu 0 1 --net wideresnet34x10 --augm default --id_steps 10 --od_steps 20 --id_pgd apgd --od_pgd apgd --train_type advacet --epochs 300 --ema True --ema_decay 0.999 --test_epochs 10 --dataset semi-cifar10 --schedule cosine --eps 0.5 --od_eps_factor 2

500k-TI APGD ASAM:

For ASAM training, we only use a single GPU with BS 128 as typically we do not want to average weight gradients over too many datapoints (see the discussion of m-sharpness in the original SAM-paper) and the PyTorch DataParallel implementation does not support non-synchronized per-device computations.

python run_training_cifar10.py --gpu 0 --net wideresnet34x10 --augm default --id_steps 10 --od_steps 20 --id_pgd apgd --od_pgd apgd --train_type advacet --epochs 250 --ema True --ema_decay 0.999 --test_epochs 10 --dataset semi-cifar10 --schedule cosine --eps 0.5 --od_eps_factor 2 --nesterov False --optim SAM --sam_adaptive True --sam_rho 1.0