JaFR_IJCAI2022

This is our Pytorch implementation of JaFR.

How Does Frequency Bias Affect the Robustness of Neural Image Classifiers against Common Corruption and Adversarial Perturbations?
Alvin Chan, Yew-Soon Ong, Clement Tan
https://arxiv.org/abs/2205.04533

TL;DR: We propose JaFR to more directly study the frequency bias of a model through the lens of its Jacobians and its implication to model robustness.

Requirements

• Python 3.8.1 on Linux
• PyTorch 1.4

Dataset

See instructions here to download the CIFAR-10-C and CIFAR-100-C corruption datasets.

Examples

Folder containing scripts to train and evaluate models, with the parameters reported in the paper: example_scripts/.

Training

Train model with JaFR and FGSM, with the parameters reported in the paper:

python train.py  --model_name=model_folder_name --dataset=cifar10 --eval_early_stopped_model --attack=fgsm --eps=8 --attack_init=zero --epochs=30 --grad_align_cos_lambda=0 --lr_max=0.30 --low_freq_bias_lambda=0.001  --seed=1

After training, the COCON block's weights will be saved in models/model_folder_name.

Training Key Arguments

--model_name : path of folder where model weights, logs and parameters are saved
--attack : type of adversarial training --dataset: dataset to train model on
--low_freq_bias_lambda: weight of JaFR in training
--epochs: Number of training epochs

Adversarial Example Evaluation

Evaluate model with adversarial examples (PGD with eps value of 8):

python eval.py --model_dir=models/model_folder_name --eps 8 --eval_results_filename=eval_results.txt

Evaluation output text file (eval_results.txt) will be saved in models/model_folder_name.

Evaluation Key Arguments

--model_dir : path of folder where model weights, logs and parameters are saved
--eps : eps of PGD/FGSM attack --dataset: type of image dataset
--pgd_loss_func: type of loss function used to compute adversarial examples

Corruption Evaluation

Evaluate model with CIFAR-10-C corruption examples:

python eval_cifar10c.py --model_dir=models/model_folder_name

Evaluate model with CIFAR-10-C corruption examples:

python eval_cifar100c.py --model_dir=models/model_folder_name

Evaluation output text file (eval_results.txt) will be saved in models/model_folder_name.

Citation

If you find our repository useful, please consider citing our paper:

@article{chan2022does,
  title={How Does Frequency Bias Affect the Robustness of Neural Image Classifiers against Common Corruption and Adversarial Perturbations?},
  author={Chan, Alvin and Ong, Yew-Soon and Tan, Clement},
  journal={arXiv preprint arXiv:2205.04533},
  year={2022}
}

Acknowledgements

Code is adapted from:

• https://github.com/tml-epfl/understanding-fast-adv-training