Francesco Croce* (University of Tübingen), Maksym Andriushchenko* (EPFL), Vikash Sehwag* (Princeton University), Edoardo Debenedetti* (EPFL), Nicolas Flammarion (EPFL), Mung Chiang (Purdue University), Prateek Mittal (Princeton University), Matthias Hein (University of Tübingen)

Leaderboard: https://robustbench.github.io/

Paper: https://arxiv.org/abs/2010.09670

❗Note❗: if you experience problems with the automatic downloading of the models from Google Drive, install the latest version of RobustBench via pip install git+https://github.com/RobustBench/robustbench.git.

• May 2022: We have extended the common corruptions leaderboard on ImageNet with 3D Common Corruptions (ImageNet-3DCC). ImageNet-3DCC evaluation is interesting since (1) it includes more realistic corruptions and (2) it can be used to assess generalization of the existing models which may have overfitted to ImageNet-C. For a quickstart, click here. Note that the entries in leaderboard are still sorted according to ImageNet-C performance.

• May 2022: We fixed the preprocessing issue for ImageNet corruption evaluations: previously we used resize to 256x256 and central crop to 224x224 which wasn't necessary since the ImageNet-C images are already 224x224 (see this issue). Note that this changed the ranking between the top-1 and top-2 entries.

The goal of RobustBench is to systematically track the real progress in adversarial robustness. There are already more than 3'000 papers on this topic, but it is still often unclear which approaches really work and which only lead to overestimated robustness. We start from benchmarking the Linf, L2, and common corruption robustness since these are the most studied settings in the literature.

Evaluation of the robustness to Lp perturbations in general is not straightforward and requires adaptive attacks (Tramer et al., (2020)). Thus, in order to establish a reliable standardized benchmark, we need to impose some restrictions on the defenses we consider. In particular, we accept only defenses that are (1) have in general non-zero gradients wrt the inputs, (2) have a fully deterministic forward pass (i.e. no randomness) that (3) does not have an optimization loop. Often, defenses that violate these 3 principles only make gradient-based attacks harder but do not substantially improve robustness (Carlini et al., (2019)) except those that can present concrete provable guarantees (e.g. Cohen et al., (2019)).

To prevent potential overadaptation of new defenses to AutoAttack, we also welcome external evaluations based on adaptive attacks, especially where AutoAttack flags a potential overestimation of robustness. For each model, we are interested in the best known robust accuracy and see AutoAttack and adaptive attacks as complementary to each other.

RobustBench consists of two parts:

• a website https://robustbench.github.io/ with the leaderboard based on many recent papers (plots below 👇)
• a collection of the most robust models, Model Zoo, which are easy to use for any downstream application (see the tutorial below after FAQ 👇)

Q: How does the RobustBench leaderboard differ from the AutoAttack leaderboard? 🤔
A: The AutoAttack leaderboard was the starting point of RobustBench. Now only the RobustBench leaderboard is actively maintained.

Q: How does the RobustBench leaderboard differ from robust-ml.org? 🤔
A: robust-ml.org focuses on adaptive evaluations, but we provide a standardized benchmark. Adaptive evaluations have been very useful (e.g., see Tramer et al., 2020) but they are also very time-consuming and not standardized by definition. Instead, we argue that one can estimate robustness accurately mostly without adaptive attacks but for this one has to introduce some restrictions on the considered models. However, we do welcome adaptive evaluations and we are always interested in showing the best known robust accuracy.

Q: How is it related to libraries like foolbox / cleverhans / advertorch? 🤔
A: These libraries provide implementations of different attacks. Besides the standardized benchmark, RobustBench additionally provides a repository of the most robust models. So you can start using the robust models in one line of code (see the tutorial below 👇).

Q: Why is Lp-robustness still interesting? 🤔
A: There are numerous interesting applications of Lp-robustness that span transfer learning (Salman et al. (2020), Utrera et al. (2020)), interpretability (Tsipras et al. (2018), Kaur et al. (2019), Engstrom et al. (2019)), security (Tramèr et al. (2018), Saadatpanah et al. (2019)), generalization (Xie et al. (2019), Zhu et al. (2019), Bochkovskiy et al. (2020)), robustness to unseen perturbations (Xie et al. (2019), Kang et al. (2019)), stabilization of GAN training (Zhong et al. (2020)).

Q: What about verified adversarial robustness? 🤔
A: We mostly focus on defenses which improve empirical robustness, given the lack of clarity regarding which approaches really improve robustness and which only make some particular attacks unsuccessful. However, we do not restrict submissions of verifiably robust models (e.g., we have Zhang et al. (2019) in our CIFAR-10 Linf leaderboard). For methods targeting verified robustness, we encourage the readers to check out Salman et al. (2019) and Li et al. (2020).

Q: What if I have a better attack than the one used in this benchmark? 🤔
A: We will be happy to add a better attack or any adaptive evaluation that would complement our default standardized attacks.

The goal of our Model Zoo is to simplify the usage of robust models as much as possible. Check out our Colab notebook here 👉 RobustBench: quick start for a quick introduction. It is also summarized below 👇.

First, install the latest version of RobustBench (recommended):

pip install git+https://github.com/RobustBench/robustbench.git

or the latest stable version of RobustBench (it is possible that automatic downloading of the models may not work):

pip install git+https://github.com/RobustBench/robustbench.git@v1.0

Now let's try to load CIFAR-10 and some quite robust CIFAR-10 models from Carmon2019Unlabeled that achieves 59.53% robust accuracy evaluated with AA under eps=8/255:

from robustbench.data import load_cifar10

x_test, y_test = load_cifar10(n_examples=50)

from robustbench.utils import load_model

model = load_model(model_name='Carmon2019Unlabeled', dataset='cifar10', threat_model='Linf')

Let's try to evaluate the robustness of this model. We can use any favourite library for this. For example, FoolBox implements many different attacks. We can start from a simple PGD attack:

!pip install -q foolbox
import foolbox as fb
fmodel = fb.PyTorchModel(model, bounds=(0, 1))

_, advs, success = fb.attacks.LinfPGD()(fmodel, x_test.to('cuda:0'), y_test.to('cuda:0'), epsilons=[8/255])
print('Robust accuracy: {:.1%}'.format(1 - success.float().mean()))

>>> Robust accuracy: 58.0%

Wonderful! Can we do better with a more accurate attack?

Let's try to evaluate its robustness with a cheap version AutoAttack from ICML 2020 with 2/4 attacks (only APGD-CE and APGD-DLR):

# autoattack is installed as a dependency of robustbench so there is not need to install it separately
from autoattack import AutoAttack
adversary = AutoAttack(model, norm='Linf', eps=8/255, version='custom', attacks_to_run=['apgd-ce', 'apgd-dlr'])
adversary.apgd.n_restarts = 1
x_adv = adversary.run_standard_evaluation(x_test, y_test)

>>> initial accuracy: 92.00%
>>> apgd-ce - 1/1 - 19 out of 46 successfully perturbed
>>> robust accuracy after APGD-CE: 54.00% (total time 10.3 s)
>>> apgd-dlr - 1/1 - 1 out of 27 successfully perturbed
>>> robust accuracy after APGD-DLR: 52.00% (total time 17.0 s)
>>> max Linf perturbation: 0.03137, nan in tensor: 0, max: 1.00000, min: 0.00000
>>> robust accuracy: 52.00%

Note that for our standardized evaluation of Linf-robustness we use the full version of AutoAttack which is slower but more accurate (for that just use adversary = AutoAttack(model, norm='Linf', eps=8/255)).

What about other types of perturbations? Is Lp-robustness useful there? We can evaluate the available models on more general perturbations. For example, let's take images corrupted by fog perturbations from CIFAR-10-C with the highest level of severity (5). Are different Linf robust models perform better on them?

from robustbench.data import load_cifar10c
from robustbench.utils import clean_accuracy

corruptions = ['fog']
x_test, y_test = load_cifar10c(n_examples=1000, corruptions=corruptions, severity=5)

for model_name in ['Standard', 'Engstrom2019Robustness', 'Rice2020Overfitting',
                   'Carmon2019Unlabeled']:
 model = load_model(model_name, dataset='cifar10', threat_model='Linf')
 acc = clean_accuracy(model, x_test, y_test)
 print(f'Model: {model_name}, CIFAR-10-C accuracy: {acc:.1%}')

>>> Model: Standard, CIFAR-10-C accuracy: 74.4%
>>> Model: Engstrom2019Robustness, CIFAR-10-C accuracy: 38.8%
>>> Model: Rice2020Overfitting, CIFAR-10-C accuracy: 22.0%
>>> Model: Carmon2019Unlabeled, CIFAR-10-C accuracy: 31.1%

As we can see, all these Linf robust models perform considerably worse than the standard model on this type of corruptions. This curious phenomenon was first noticed in Adversarial Examples Are a Natural Consequence of Test Error in Noise and explained from the frequency perspective in A Fourier Perspective on Model Robustness in Computer Vision.

However, on average adversarial training does help on CIFAR-10-C. One can check this easily by loading all types of corruptions via load_cifar10c(n_examples=1000, severity=5), and repeating evaluation on them.

3D Common Corruptions (3DCC) is a recent benchmark by Kar et al. (CVPR 2022) using scene geometry to generate realistic corruptions. You can evaluate robustness of a standard ResNet-50 against ImageNet-3DCC by following these steps:

1. Download the data from here using the provided tool. The data will be saved into a folder named ImageNet-3DCC.

2. Run the sample evaluation script to obtain accuracies and save them in a pickle file:

import torch 
from robustbench.data import load_imagenet3dcc
from robustbench.utils import clean_accuracy, load_model

corruptions_3dcc = ['near_focus', 'far_focus', 'bit_error', 'color_quant', 
                   'flash', 'fog_3d', 'h265_abr', 'h265_crf',
                   'iso_noise', 'low_light', 'xy_motion_blur', 'z_motion_blur'] # 12 corruptions in ImageNet-3DCC

device = torch.device("cuda:0")
model = load_model('Standard_R50', dataset='imagenet', threat_model='corruptions').to(device)
for corruption in corruptions_3dcc:
    for s in [1, 2, 3, 4, 5]:  # 5 severity levels
        x_test, y_test = load_imagenet3dcc(n_examples=5000, corruptions=[corruption], severity=s, data_dir=$PATH_IMAGENET_3DCC)
        acc = clean_accuracy(model, x_test.to(device), y_test.to(device), device=device)
        print(f'Model: {model_name}, ImageNet-3DCC corruption: {corruption} severity: {s} accuracy: {acc:.1%}')

In order to use a model, you just need to know its ID, e.g. Carmon2019Unlabeled, and to run:

from robustbench import load_model

model = load_model(model_name='Carmon2019Unlabeled', dataset='cifar10', threat_model='Linf')

which automatically downloads the model (all models are defined in model_zoo/models.py).

Reproducing evaluation of models from the Model Zoo can be done directly from the command line. Here is an example of an evaluation of Salman2020Do_R18 model with AutoAttack on ImageNet for eps=4/255=0.0156862745:

python -m robustbench.eval --n_ex=5000 --dataset=imagenet --threat_model=Linf --model_name=Salman2020Do_R18 --data_dir=/tmldata1/andriush/imagenet --batch_size=128 --eps=0.0156862745

The CIFAR-10, CIFAR-10-C, CIFAR-100, and CIFAR-100-C datasets are downloaded automatically. However, the ImageNet datasets should be downloaded manually due to their licensing:

• ImageNet: Obtain the download link here (requires just signing up from an academic email, the approval system there is automatic and happens instantly) and then follow the instructions here to extract the validation set in a pytorch-compatible format into folder val.
• ImageNet-C: Please visit here for the instructions.
• ImageNet-3DCC: Download the data from here using the provided tool. The data will be saved into a folder named ImageNet-3DCC.

In order to use the models from the Model Zoo, you can find all available model IDs in the tables below. Note that the full leaderboard contains a bit more models which we either have not yet added to the Model Zoo or their authors don't want them to appear in the Model Zoo.

Note: the values (even clean accuracy) might have small fluctuations depending on the version of the packages e.g. torchvision.

We host all the notebooks at Google Colab:

• RobustBench: quick start: a quick tutorial to get started that illustrates the main features of RobustBench.
• RobustBench: json stats: various plots based on the jsons from model_info (robustness over venues, robustness vs accuracy, etc).

Feel free to suggest a new notebook based on the Model Zoo or the jsons from model_info. We are very interested in collecting new insights about benefits and tradeoffs between different perturbation types.

Contributions to RobustBench are very welcome! You can help to improve RobustBench:

• Are you an author of a recent paper focusing on improving adversarial robustness? Consider adding new models (see the instructions below 👇).
• Do you have in mind some better standardized attack? Do you want to extend RobustBench to other threat models? We'll be glad to discuss that!
• Do you have an idea how to make the existing codebase better? Just open a pull request or create an issue and we'll be happy to discuss potential changes.

In case you have some new (potentially, adaptive) evaluation that leads to a lower robust accuracy than AutoAttack, we will be happy to add it to the leaderboard. The easiest way is to open an issue with the "New external evaluation(s)" template and fill in all the fields.

The easiest way to add new models to the leaderboard and/or to the model zoo, is by opening an issue with the "New Model(s)" template and fill in all the fields.

In the following sections there are some tips on how to prepare the claim.

The claim can be computed in the following way (example for cifar10, Linf threat model):

import torch

from robustbench import benchmark
from myrobust model import MyRobustModel

threat_model = "Linf"  # one of {"Linf", "L2", "corruptions"}
dataset = "cifar10"  # one of {"cifar10", "cifar100", "imagenet"}

model = MyRobustModel()
model_name = "<Name><Year><FirstWordOfTheTitle>"
device = torch.device("cuda:0")

clean_acc, robust_acc = benchmark(model, model_name=model_name, n_examples=10000, dataset=dataset,
                                  threat_model=threat_model, eps=8/255, device=device,
                                  to_disk=True)

In particular, the to_disk argument, if True, generates a json file at the path model_info/<dataset>/<threat_model>/<Name><Year><FirstWordOfTheTitle>.json which is structured in the following way (example from model_info/cifar10/Linf/Rice2020Overfitting.json):

{
  "link": "https://arxiv.org/abs/2002.11569",
  "name": "Overfitting in adversarially robust deep learning",
  "authors": "Leslie Rice, Eric Wong, J. Zico Kolter",
  "additional_data": false,
  "number_forward_passes": 1,
  "dataset": "cifar10",
  "venue": "ICML 2020",
  "architecture": "WideResNet-34-20",
  "eps": "8/255",
  "clean_acc": "85.34",
  "reported": "58",
  "autoattack_acc": "53.42"
}

The only difference is that the generated json will have only the fields "clean_acc" and "autoattack_acc" (for "Linf" and "L2" threat models) or "corruptions_acc" (for the "corruptions" threat model) already specified. The other fields have to be filled manually.

If the given threat_model is corruptions, we also save unaggregated results on the different combinations of corruption types and severities in this csv file (for CIFAR-10).

For ImageNet benchmarks, the users should specify what preprocessing should be used (e.g. resize and crop to the needed resolution). There are some preprocessings already defined in robustbench.data.PREPROCESSINGS, which can be used by specifying the key as the preprocessing parameter of benchmark. Otherwise, it's possible to pass an arbitrary torchvision transform (or torchvision-compatible transform), e.g.:

transform = transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor()
    ])
clean_acc, robust_acc = benchmark(model, model_name=model_name, n_examples=10000, dataset=dataset,
                                  threat_model=threat_model, eps=8/255, device=device,
                                  to_disk=True, preprocessing=transform)

In case you want to add a model in the Model Zoo by yourself, then you should also open a PR with the new model(s) you would like to add. All the models of each <dataset> are saved in robustbench/model_zoo/<dataset>.py. Each file contains a dictionary for every threat model, where the keys are the identifiers of each model, and the values are either class constructors, for models that have to change standard architectures, or lambda functions that return the constructed model.

If your model is a standard architecture (e.g., WideResNet), does not apply any normalization to the input nor has to do things differently from the standard architecture, consider adding your model as a lambda function, e.g.

('Cui2020Learnable_34_10', {
    'model': lambda: WideResNet(depth=34, widen_factor=10, sub_block1=True),
    'gdrive_id': '16s9pi_1QgMbFLISVvaVUiNfCzah6g2YV'
})

If your model is a standard architecture, but you need to do something differently (e.g. applying normalization), consider inheriting the class defined in wide_resnet.py or resnet.py. For example:

class Rice2020OverfittingNet(WideResNet):
    def __init__(self, depth, widen_factor):
        super(Rice2020OverfittingNet, self).__init__(depth=depth, widen_factor=widen_factor,
                                                     sub_block1=False)
        self.mu = torch.Tensor([0.4914, 0.4822, 0.4465]).float().view(3, 1, 1).cuda()
        self.sigma = torch.Tensor([0.2471, 0.2435, 0.2616]).float().view(3, 1, 1).cuda()

    def forward(self, x):
        x = (x - self.mu) / self.sigma
        return super(Rice2020OverfittingNet, self).forward(x)

If instead you need to create a new architecture, please put it in robustbench/model_zoo/archietectures/<my_architecture>.py.

You should also add your model entry in the corresponding <threat_model> dict in the file robustbench/model_zoo/<dataset>.py. For instance, let's say your model is robust against common corruptions in CIFAR-10 (i.e. CIFAR-10-C), then you should add your model to the common_corruptions dict in robustbench/model_zoo/cifar10.py.

The model should also contain the Google Drive ID with your PyTorch model so that it can be downloaded automatically from Google Drive:

    ('Rice2020Overfitting', {
        'model': Rice2020OverfittingNet(34, 20),
        'gdrive_id': '1vC_Twazji7lBjeMQvAD9uEQxi9Nx2oG-',
})

In case you want to keep your checkpoints private for some reasons, you can also submit your claim by opening an issue with the same "New Model(s)" template, specifying that the submission is private, and sharing the checkpoints with the email address adversarial.benchmark@gmail.com. In this case, we will add your model to the leaderboard but not to the Model Zoo and will not share your checkpoints publicly.

By default, the models are released under the MIT license, but you can also tell us if you want to release your model under a customized license.

In order to run the tests, run:

• python -m unittest discover tests -t . -v for fast testing
• RUN_SLOW=true python -m unittest discover tests -t . -v for slower testing

For example, one can test if the clean accuracy on 200 examples exceeds some threshold (70%) or if clean accuracy on 10'000 examples for each model matches the ones from the jsons located at robustbench/model_info.

Note that one can specify some configurations like batch_size, data_dir, model_dir in tests/config.py for running the tests.

Would you like to reference the RobustBench leaderboard or you are using models from the Model Zoo?
Then consider citing our whitepaper:

@inproceedings{croce2021robustbench,
  title     = {RobustBench: a standardized adversarial robustness benchmark},
  author    = {Croce, Francesco and Andriushchenko, Maksym and Sehwag, Vikash and Debenedetti, Edoardo and Flammarion, Nicolas and Chiang, Mung and Mittal, Prateek and Matthias Hein},
  booktitle = {Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year      = {2021},
  url       = {https://openreview.net/forum?id=SSKZPJCt7B}
}

Feel free to contact us about anything related to RobustBench by creating an issue, a pull request or by email at adversarial.benchmark@gmail.com.