Adversarial Weight Perturbation Helps Robust Generalization

Code for NeurIPS 2020 "Adversarial Weight Perturbation Helps Robust Generalization" by Dongxian Wu, Shu-Tao Xia, and Yisen Wang.

News

10/13/2020 - Our code and paper are released.

Requisite

This code is implemented in PyTorch, and we have tested the code under the following environment settings:

python = 3.7.3
torch = 1.2.0
torchvision = 0.4.0

What is in this repository

Codes for our AWP-based adversarial training (AT-AWP) are in at-awp, and those for AWP-based TRADES (TRADES-AWP) are in ./trades-awp:

In ./at-awp, the codes for CIFAR-10, CIFAR-100, and SVHN are in train_cifar10.py, train_cifar100.py, train_svhn.py respectively.
In ./trades-awp, the codes for CIFAR-10 and CIFAR-100 are in train_trades_cifar.py.

The checkpoints can be found in Google Drive or Baidu Drive(pw: 8tsv).

How to use it

For AT-AWP with a PreAct ResNet-18 on CIFAR-10 under L_inf threat model (8/255), run codes as follows,

python train_cifar10.py --data-dir DATASET_DIR

where $DATASET_DIR is the path to the dataset.

For TRADES-AWP with a WRN-34-10 on CIFAR10 under L_inf threat model (8/255), run codes as follows,

python train_trades_cifar.py --data CIFAR10 --data-path DATASET_DIR

The Leaderboard Under Auto Attack

To verify the effectiveness of AWP further, we evaluate the robustness under a stronger attack, auto-attack [3]. Here we only list Top 10 results on the leadboard (up to 10/13/2020) and our results. Compared with the leadboard results, AWP can boost the robustness of the AT and its variants (TRADES[2], MART[4], Pre-training[5], RST[6], etc.), ranking 1st on both with and without data. Even some AWP-based methods without additional data can surpass the results under additional data.

More results can be found in ./auto-attacks

#	method / paper	model	architecture	clean	report.	AA
-	RST-AWP (ours)‡	downloads	WRN-28-10	88.25	-	60.04
1	(Wu et al., 2020)‡	available	WRN-34-15	85.60	59.78	59.78
2	(Carmon et al., 2019) RST‡	available	WRN-28-10	89.69	62.5	59.53
-	Pre-training-AWP (ours)‡	downloads	WRN-28-10	88.33	-	57.39
3	(Sehwag et al., 2020)‡	available	WRN-28-10	88.98	-	57.14
4	(Wang et al., 2020)‡	available	WRN-28-10	87.50	65.04	56.29
-	TRADES-AWP (ours)	downloads	WRN-34-10	85.36	-	56.17
5	(Alayrac et al., 2019)‡	available	WRN-106-8	86.46	56.30	56.03
6	(Hendrycks et al., 2019) Pre-training‡	available	WRN-28-10	87.11	57.4	54.92
-	MART-AWP (ours)	downloads	WRN-34-10	84.43	-	54.23
-	AT-AWP (ours)	downloads	WRN-34-10	85.36	-	53.97
7	(Pang et al., 2020b)	available	WRN-34-20	85.14	-	53.74
8	(Zhang et al., 2020b)	available	WRN-34-10	84.52	54.36	53.51
9	(Rice et al., 2020) AT	available	WRN-34-20	85.34	58	53.42
10	(Huang et al., 2020)*	available	WRN-34-10	83.48	58.03	53.34

Citing this work

@inproceedings{wu2020adversarial,
    title={Adversarial Weight Perturbation Helps Robust Generalization},
    author={Dongxian Wu and Shu-Tao Xia and Yisen Wang},
    booktitle={NeurIPS},
    year={2020}
}