LAAT

The official implementation of LAAT (Language-Driven Anchors for Zero-Shot Adversarial Robustness, CVPR 2024)

Prerequisite

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

• python == 3.9.12
• torch == 1.12.1
• torchvision == 0.13.1
• autoattack == 0.1

Usage

Craft anchors

To craft anchors with expansion algorithm, in anchors/,

python get_anchors.py --classes cifarfs_classes.txt --to cifarfs_anchors.npy
python convert_anchor.py cifarfs 64

Note that 64 means the first 64 classes are training classes, others are testing classes, which are not used to construct expansion mapping.

Train & Test

Here we give an exmaple.

The following command can train a zero-shot robust classidier on CIFAR-FS with the Conv4-512 backbone:

python train.py \
--data_dir /path/to/dataset --exp_name /path/to/save \
--seed 3407 \
--n_support 0 \
--dataset CIFAR100FS \
--model Conv4-512 \
--use_linear --head cos-span \
--train_type TRADES-cos \
--suffix 1

Test FGSM attack:

python train.py \
--data_dir /path/to/dataset --exp_name /path/to/save \
--seed 3407 \
--n_support 0 \
--dataset CIFAR100FS \
--model Conv4-512 \
--use_linear --head cos-span \
--train_type TRADES-cos \
--suffix 1 \
--eval --load_best \
--attack FGSM

Test AutoAttack:

python train.py \
--data_dir /path/to/dataset --exp_name /path/to/save \
--seed 3407 \
--n_support 0 \
--dataset CIFAR100FS \
--model Conv4-512 \
--use_linear --head cos-span \
--train_type TRADES-cos \
--suffix 1 \
--eval --load_best \
--attack AA --n_test 100

To train & test 1-shot, change --n_support 0 to --n_support 1.

Datasets

• CIFAR-FS
• miniImageNet

If you find this useful in your research, please cite this work:

@inproceedings{li2024languagedriven,
  author = {Li, Xiao and Zhang, Wei and Liu, Yining and Hu, Zhanhao and Zhang, Bo and Hu, Xiaolin},
  title = {Language-Driven Anchors for Zero-Shot Adversarial Robustness},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2024}
}