easyrobust1/easyrobust

• EasyRobust
• • Install
  • Training
  • Testing
  • Templates for Training
  • Supported Methods
  • • Augmentation-Based
    • Model-Based
    • Activation-Based
    • Pooling-Based
    • Norm-Based
    • Training Methods
    • Loss function
  • TODO

EasyRobust

EasyRobust is a toolkit for training your robust models, with a collection of adversarially robust / non-adversarially robust / standard training method on CNN / ViT architectures.

Install

git clone https://github.com/alibaba/easyrobust.git
cd easyrobust/easyrobust
pip install -e .

Training

sh train.sh train_configs/imagenet/resnet50_baseline.yaml

Testing

see test_scripts

Templates for Training

• train_configs/imagenet/deit_small_baseline.yaml: baseline training on deit_small
• train_configs/imagenet/resnet50_baseline.yaml: baseline training on resnet50
• train_configs/imagenet/advtrain/resnet50_advtrain.yaml: adversarial training on resnet50
• train_configs/imagenet/advtrain/deit_small_advtrain.yaml: adversarial training on deit_small

More training templates will be supported in future.

Supported Methods

Augmentation-Based

• All data augmentation in timm.

• StyleAugmentation Style Augmentation: Data Augmentation via Style Randomization

• CartoonAugmentation CartoonGAN: Generative Adversarial Networks for Photo Cartoonization

Model-Based

• All implemented models in timm.

• WaveCNet: WaveCNet: Wavelet Integrated CNNs to Suppress Aliasing Effect for Noise-Robust Image Classification

• RVT: Towards Robust Vision Transformer

• DrViT: Discrete Representations Strengthen Vision Transformer Robustness

Activation-Based

• kWTA: Resisting Adversarial Attacks by k-Winners-Take-All

• LP_ReLU: Robust Image Classification Using a Low-Pass Activation Function and DCT Augmentation

Pooling-Based

• BlurPool: Making Convolutional Networks Shift-Invariant Again

• GaussianPool: Gaussian-Based Pooling for Convolutional Neural Networks

Norm-Based

• SelfNorm: CrossNorm and SelfNorm for Generalization under Distribution Shifts

• pAdaIN: Permuted AdaIN: Reducing the Bias Towards Global Statistics in Image Classification

• CrossNorm: CrossNorm and SelfNorm for Generalization under Distribution Shifts

Training Methods

• Adversarial Training: Towards Deep Learning Models Resistant to Adversarial Attacks

• GNT: A simple way to make neural networks robust against diverse image corruptions

• Shape-Texture Debiased: Shape-Texture Debiased Neural Network Training

• AdvProp: Adversarial Examples Improve Image Recognition

Loss function

• ProbCompactLoss: Improving Adversarial Robustness via Probabilistically Compact Loss with Logit Constraints

• CEB: CEB Improves Model Robustness

TODO

• More training scripts
• Validation the results
• More robust methods