Authors: Rundong Luo*, Yifei Wang*, Yisen Wang from Peking University

*: Equal contribution

Paper

Recent works have shown that self-supervised learning can achieve remarkable robustness when integrated with adversarial training (AT). However, the robustness gap between supervised AT (sup-AT) and self-supervised AT (self-AT) remains significant. Motivated by this observation, we revisit existing self-AT methods and discover an inherent dilemma that affects self-AT robustness: either strong or weak data augmentations are harmful to self-AT, and a medium strength is insufficient to bridge the gap. To resolve this dilemma, we propose a simple remedy named DynACL (Dynamic Adversarial Contrastive Learning). In particular, we propose an augmentation schedule that gradually anneals from a strong augmentation to a weak one to benefit from both extreme cases. Besides, we adopt a fast post-processing stage for adapting it to downstream tasks. Through extensive experiments, we show that DynACL can improve the state-of-the-art self-AT robustness by 8.84% under Auto-Attack on the CIFAR-10 dataset, and can even outperform vanilla supervised adversarial training. We demonstrate that self-supervised AT can attain even better robustness than supervised AT for the first time.

A standard pytorch environment (>=1.0.0) with basic packages (e.g., numpy, pickle) is enough. Besides, to evaluate under the Auto-Attack benchmark, the autoattack package is required. Run the following code to install:

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

Additionally, we need one package for generating pseudo labels. We recommend the kmeans-pytorch package (can be directly installed by pip), and sklearn is an alternative.

CIFAR10, CIFAR100, and STL10 dataset are required. You may manually download them and put in the ./data folder, or directly run our provided scripts to automatically download these datasets.

We assume you have a GPU with no less than 16GB memory (e.g., 3090). Evaluation require fewer memory (less than 8GB).

For DynACL, run

python train_DynACL.py --experiment EXP_PATH --dataset DATASET_NAME --data PATH_TO_DATASET

All training results (checkpoints and loggers) will be stored in ./experiments/EXP_PATH. Training takes around one day on a single RTX 3090 GPU.

For DynACL++, we need to first generate pseudo labels:

python assign_label.py --checkpoint PATH_TO_PRETRAINED_MODEL --dataset DATASET_NAME --data PATH_TO_DATASET

Pseudo labels will be stored in the same folder as the checkpoint.

To perform DynACL++ finetuning, we provide two scripts. Both scripts will first perform DynACL++ training, then perform evaluation on these trained models.

If you want to train the model for linear evaluation (SLF & ALF), run

python LPAFT.py --experiment EXP_PATH --label_path PATH_TO_LABELS --checkpoint PATH_TO_PRETRAINED_MODEL --dataset DATASET_NAME --data PATH_TO_DATASET

The output model model.pt only contains the finetuned adversarial route. This script also includes the evaluation for SLF and ALF settings, you may specify your desired evaluation settings by --evaluation_mode.

If you want to train the model for adversarial full evaluation (AFF) or semi-supervised evaluation, run

python LPAFT_AFF.py --experiment EXP_PATH --label_path PATH_TO_LABELS --checkpoint PATH_TO_PRETRAINED_MODEL --dataset DATASET_NAME --data PATH_TO_DATASET

The output model model.pt contains both route of the backbone. This result will also automatcally perform adversarial adversarial full finetuning (AFF) after training ends.

All training results (checkpoints and loggers) will be stored in ./experiments/EXP_PATH. Standard accuracy (SA) and robust accuracy results are stored in log.txt, while the auto-attack (AA) results are stored in robustness_results.txt.

python test_LF.py --experiment EXP_PATH --checkpoint PATH_TO_PRETRAINED_MODEL --dataset DATASET_NAME --data PATH_TO_DATASET --cvt_state_dict --bnNameCnt 1 --evaluation_mode EVALUATION_MODE

python test_AFF.py --experiment EXP_PATH --checkpoint PATH_TO_PRETRAINED_MODEL --dataset DATASET_NAME --data PATH_TO_DATASET

Note that for DynACL++ evalution, the above training scripts (LPAFT.py and LPAFT_AFF.py) already contain the evaluation part. You may also use test_LF.py and test_AFF.py for evaluation. Be aware that LPAFT.py will give you a single-BN checkpoint, so there's no need to specify cvt_state_dict.

We borrow the semi-supervised evaluation code from ACL. You may clone their code and follow their semi-supervised evaluation steps. Note that you need to use LPAFT_AFF.py to train a model with dual BN.

We provide the pretrained weights for DynACL and DynACL++ on CIFAR10.

Pretrained after DynACL CIFAR10, CIFAR100, STL10.

Finetuned after DynACL++ (single-BN) CIFAR10, CIFAR100, STL10.

Finetuned after DynACL++ (dual-BN for AFF evaluation) CIFAR10.

If you find our paper inspiring, please cite

@inproceedings{DynACL,
    title = {Rethinking the Effect of Data Augmentation in Adversarial Contrastive Learning},
    author = {Luo, Rundong and Wang, Yifei and Wang, Yisen},
    booktitle = {ICLR},
    year = {2023},
}

Some of our code are borrowed from ACL. Thanks for their great work!