zfancy / auto-attack

Code relative to "Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks"

AutoAttack

"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.

News

• [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!

Adversarial Defenses Evaluation

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.

CIFAR-10 - Linf

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).

#	paper	model	architecture	clean	report.	AA
1	(Gowal et al., 2020)‡	authors	WRN-70-16	91.10	65.87	65.88
2	(Gowal et al., 2020)‡	authors	WRN-28-10	89.48	62.76	62.80
3	(Wu et al., 2020b)‡	available	WRN-28-10	88.25	60.04	60.04
4	(Wu et al., 2020a)‡	available	WRN-34-15	85.60	59.78	59.78
5	(Carmon et al., 2019)‡	available	WRN-28-10	89.69	62.5	59.53
6	(Gowal et al., 2020)	authors	WRN-70-16	85.29	57.14	57.20
7	(Sehwag et al., 2020)‡	available	WRN-28-10	88.98	-	57.14
8	(Gowal et al., 2020)	authors	WRN-34-20	85.64	56.82	56.86
9	(Wang et al., 2020)‡	available	WRN-28-10	87.50	65.04	56.29
10	(Wu et al., 2020b)	available	WRN-34-10	85.36	56.17	56.17
11	(Alayrac et al., 2019)‡	available	WRN-106-8	86.46	56.30	56.03
12	(Hendrycks et al., 2019)‡	available	WRN-28-10	87.11	57.4	54.92
13	(Pang et al., 2020c)	available	WRN-34-20	86.43	54.39	54.39
14	(Pang et al., 2020b)	available	WRN-34-20	85.14	-	53.74
15	(Zhang et al., 2020b)	available	WRN-34-10	84.52	54.36	53.51
16	(Rice et al., 2020)	available	WRN-34-20	85.34	58	53.42
17	(Huang et al., 2020)*	available	WRN-34-10	83.48	58.03	53.34
18	(Zhang et al., 2019b)*	available	WRN-34-10	84.92	56.43	53.08
19	(Qin et al., 2019)	available	WRN-40-8	86.28	52.81	52.84
20	(Chen et al., 2020a)	available	RN-50 (x3)	86.04	54.64	51.56
21	(Chen et al., 2020b)	available	WRN-34-10	85.32	51.13	51.12
22	(Sitawarin et al., 2020)	available	WRN-34-10	86.84	50.72	50.72
23	(Engstrom et al., 2019)	available	RN-50	87.03	53.29	49.25
24	(Kumari et al., 2019)	available	WRN-34-10	87.80	53.04	49.12
25	(Mao et al., 2019)	available	WRN-34-10	86.21	50.03	47.41
26	(Zhang et al., 2019a)	retrained	WRN-34-10	87.20	47.98	44.83
27	(Madry et al., 2018)	available	WRN-34-10	87.14	47.04	44.04
28	(Pang et al., 2020a)	available	RN-32	80.89	55.0	43.48
29	(Wong et al., 2020)	available	RN-18	83.34	46.06	43.21
30	(Shafahi et al., 2019)	available	WRN-34-10	86.11	46.19	41.47
31	(Ding et al., 2020)	available	WRN-28-4	84.36	47.18	41.44
32	(Atzmon et al., 2019)*	available	RN-18	81.30	43.17	40.22
33	(Moosavi-Dezfooli et al., 2019)	authors	WRN-28-10	83.11	41.4	38.50
34	(Zhang & Wang, 2019)	available	WRN-28-10	89.98	60.6	36.64
35	(Zhang & Xu, 2020)	available	WRN-28-10	90.25	68.7	36.45
36	(Jang et al., 2019)	available	RN-20	78.91	37.40	34.95
37	(Kim & Wang, 2020)	available	WRN-34-10	91.51	57.23	34.22
38	(Wang & Zhang, 2019)	available	WRN-28-10	92.80	58.6	29.35
39	(Xiao et al., 2020)*	available	DenseNet-121	79.28	52.4	18.50
40	(Jin & Rinard, 2020)	available	RN-18	90.84	71.22	1.35
41	(Mustafa et al., 2019)	available	RN-110	89.16	32.32	0.28
42	(Chan et al., 2020)	retrained	WRN-34-10	93.79	15.5	0.26

CIFAR-100 - Linf

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

#	paper	model	architecture	clean	report.	AA
1	(Gowal et al. 2020)‡	authors	WRN-70-16	69.15	37.70	36.88
2	(Gowal et al. 2020)	authors	WRN-70-16	60.86	30.67	30.03
3	(Wu et al., 2020b)	available	WRN-34-10	60.38	28.86	28.86
4	(Hendrycks et al., 2019)‡	available	WRN-28-10	59.23	33.5	28.42
5	(Chen et al., 2020b)	available	WRN-34-10	62.15	-	26.94
6	(Sitawarin et al., 2020)	available	WRN-34-10	62.82	24.57	24.57
7	(Rice et al., 2020)	available	RN-18	53.83	28.1	18.95

MNIST - Linf

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

#	paper	model	clean	report.	AA
1	(Gowal et al., 2020)	authors	99.26	96.38	96.34
2	(Zhang et al., 2020a)	available	98.38	96.38	93.96
3	(Gowal et al., 2019)	available	98.34	93.78	92.83
4	(Zhang et al., 2019b)	available	99.48	95.60	92.81
5	(Ding et al., 2020)	available	98.95	92.59	91.40
6	(Atzmon et al., 2019)	available	99.35	97.35	90.85
7	(Madry et al., 2018)	available	98.53	89.62	88.50
8	(Jang et al., 2019)	available	98.47	94.61	87.99
9	(Wong et al., 2020)	available	98.50	88.77	82.93
10	(Taghanaki et al., 2019)	retrained	98.86	64.25	0.00

CIFAR-10 - L2

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).

#	paper	model	architecture	clean	report.	AA
1	(Gowal et al., 2020)‡	authors	WRN-70-16	94.74	-	80.53
2	(Gowal et al., 2020)	authors	WRN-70-16	90.90	-	74.50
3	(Wu et al., 2020b)	available	WRN-34-10	88.51	73.66	73.66
4	(Augustin et al., 2020)‡	authors	RN-50	91.08	73.27	72.91
5	(Engstrom et al., 2019)	available	RN-50	90.83	70.11	69.24
6	(Rice et al., 2020)	available	RN-18	88.67	71.6	67.68
7	(Rony et al., 2019)	available	WRN-28-10	89.05	67.6	66.44
8	(Ding et al., 2020)	available	WRN-28-4	88.02	66.18	66.09

How to use AutoAttack

Installation

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

PyTorch models

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'] 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'].

TensorFlow models

To evaluate models implemented in TensorFlow 1.X, use

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

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

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).

Other versions

AutoAttack+

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).

Randomized defenses

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).

Custom version

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

Other options

Random seed

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.

Log results

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

Citation

@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}
}