"Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks"
Francesco Croce, Matthias Hein
ICML 2020
https://arxiv.org/abs/2003.01690

We propose to use an ensemble of four diverse attacks to reliably evaluate robustness:

• APGD-CE, our new step size-free version of PGD on the cross-entropy,
• APGD-DLR, our new step size-free version of PGD on the new DLR loss,
• FAB, which minimizes the norm of the adversarial perturbations (Croce & Hein, 2019),
• Square Attack, a query-efficient black-box attack (Andriushchenko et al, 2019).

Note: we fix all the hyperparameters of the attacks, so no tuning is required to test every new classifier.

• [Sep 2021]
  • We add automatic checks for potential cases where the standard version of AA might be non suitable or sufficient for robustness evaluation.
  • The evaluations of models on CIFAR-10 and CIFAR-100 are no longer maintained. Up-to-date leaderboards are available in RobustBench.
• [Mar 2021] A version of AutoAttack wrt L1, which includes the extensions of APGD and Square Attack (Croce & Hein, 2021), is available!
• [Oct 2020] AutoAttack is used as standard evaluation in the new benchmark RobustBench, which includes a Model Zoo of the most robust classifiers! Note that this page and RobustBench's leaderboards are maintained simultaneously.
• [Aug 2020]
  • Updated version: in order to i) scale AutoAttack (AA) to datasets with many classes and ii) have a faster and more accurate evaluation, we use APGD-DLR and FAB with their targeted versions.
  • We add the evaluation of models on CIFAR-100 wrt Linf and CIFAR-10 wrt L2.
• [Jul 2020] A short version of the paper is accepted at ICML'20 UDL workshop for a spotlight presentation!
• [Jun 2020] The paper is accepted at ICML 2020!

We here list adversarial defenses, for many threat models, recently proposed and evaluated with the standard version of AutoAttack (AA), including

• untargeted APGD-CE (no restarts),
• targeted APGD-DLR (9 target classes),
• targeted FAB (9 target classes),
• Square Attack (5000 queries).

See below for the more expensive AutoAttack+ (AA+) and more options.

We report the source of the model, i.e. if it is publicly available, if we received it from the authors or if we retrained it, the architecture, the clean accuracy and the reported robust accuracy (note that might be calculated on a subset of the test set or on different models trained with the same defense). The robust accuracy for AA is on the full test set.

We plan to add new models as they appear and are made available. Feel free to suggest new defenses to test!

To have a model added: please check here.

Checkpoints: many of the evaluated models are available and easily accessible at this Model Zoo.

The robust accuracy is evaluated at eps = 8/255, except for those marked with * for which eps = 0.031, where eps is the maximal Linf-norm allowed for the adversarial perturbations. The eps used is the same set in the original papers.
Note: ‡ indicates models which exploit additional data for training (e.g. unlabeled data, pre-training).

Update: this is no longer maintained, but an up-to-date leaderboard is available in RobustBench.

The robust accuracy is computed at eps = 8/255 in the Linf-norm, except for the models marked with * for which eps = 0.031 is used.
Note: ‡ indicates models which exploit additional data for training (e.g. unlabeled data, pre-training).

Update: this is no longer maintained, but an up-to-date leaderboard is available in RobustBench.

The robust accuracy is computed at eps = 0.3 in the Linf-norm.

The robust accuracy is computed at eps = 0.5 in the L2-norm.
Note: ‡ indicates models which exploit additional data for training (e.g. unlabeled data, pre-training).

Update: this is no longer maintained, but an up-to-date leaderboard is available in RobustBench.

pip install git+https://github.com/fra31/auto-attack

Import and initialize AutoAttack with

from autoattack import AutoAttack
adversary = AutoAttack(forward_pass, norm='Linf', eps=epsilon, version='standard')

where:

• forward_pass returns the logits and takes input with components in [0, 1] (NCHW format expected),
• norm = ['Linf' | 'L2' | 'L1'] is the norm of the threat model,
• eps is the bound on the norm of the adversarial perturbations,
• version = 'standard' uses the standard version of AA.

To apply the standard evaluation, where the attacks are run sequentially on batches of size bs of images, use

x_adv = adversary.run_standard_evaluation(images, labels, bs=batch_size)

To run the attacks individually, use

dict_adv = adversary.run_standard_evaluation_individual(images, labels, bs=batch_size)

which returns a dictionary with the adversarial examples found by each attack.

To specify a subset of attacks add e.g. adversary.attacks_to_run = ['apgd-ce'].

To evaluate models implemented in TensorFlow 1.X, use

from autoattack import utils_tf
model_adapted = utils_tf.ModelAdapter(logits, x_input, y_input, sess)

from autoattack import AutoAttack
adversary = AutoAttack(model_adapted, norm='Linf', eps=epsilon, version='standard', is_tf_model=True)

where:

• logits is the tensor with the logits given by the model,
• x_input is a placeholder for the input for the classifier (NHWC format expected),
• y_input is a placeholder for the correct labels,
• sess is a TF session.

If TensorFlow's version is 2.X, use

from autoattack import utils_tf2
model_adapted = utils_tf2.ModelAdapter(tf_model)

from autoattack import AutoAttack
adversary = AutoAttack(model_adapted, norm='Linf', eps=epsilon, version='standard', is_tf_model=True)

where:

• tf_model is tf.keras model without activation function 'softmax'

The evaluation can be run in the same way as done with PT models.

Examples of how to use AutoAttack can be found in examples/. To run the standard evaluation on a pretrained PyTorch model on CIFAR-10 use

python eval.py [--individual] --version=['standard' | 'plus']

where the optional flags activate respectively the individual evaluations (all the attacks are run on the full test set) and the version of AA to use (see below).

A more expensive evaluation can be used specifying version='plus' when initializing AutoAttack. This includes

• untargeted APGD-CE (5 restarts),
• untargeted APGD-DLR (5 restarts),
• untargeted FAB (5 restarts),
• Square Attack (5000 queries),
• targeted APGD-DLR (9 target classes),
• targeted FAB (9 target classes).

In case of classifiers with stochastic components one can combine AA with Expectation over Transformation (EoT) as in (Athalye et al., 2018) specifying version='rand' when initializing AutoAttack. This runs

• untargeted APGD-CE (no restarts, 20 iterations for EoT),
• untargeted APGD-DLR (no restarts, 20 iterations for EoT).

It is possible to customize the attacks to run specifying version='custom' when initializing the attack and then, for example,

if args.version == 'custom':
	adversary.attacks_to_run = ['apgd-ce', 'fab']
        adversary.apgd.n_restarts = 2
        adversary.fab.n_restarts = 2

It is possible to fix the random seed used for the attacks with, e.g., adversary.seed = 0. In this case the same seed is used for all the attacks used, otherwise a different random seed is picked for each attack.

To log the intermediate results of the evaluation specify log_path=/path/to/logfile.txt when initializing the attack.

@inproceedings{croce2020reliable,
    title = {Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks},
    author = {Francesco Croce and Matthias Hein},
    booktitle = {ICML},
    year = {2020}
}

@inproceedings{croce2021mind,
    title={Mind the box: $l_1$-APGD for sparse adversarial attacks on image classifiers}, 
    author={Francesco Croce and Matthias Hein},
    booktitle={ICML},
    year={2021}
}