• Gradient Concealment: Free Lunch for Defending Adversarial Attacks
• • About the proposed Gradient Concealment Module
  • • Installation
    • Highlights of Gradient Concealment Module
    • Some Attack Robustness Results on ImageNet
  • About the CVPR Robust Classification Challenge
  • • Conclusion
    • Datasets
    • Data Augmentation Schemes
    • Image Pre-Processing
    • Adversarial Defense Schemes
    • Architectures
    • Training Details
    • DDP Training
  • Reference
  • Citation
  • Acknowledgement

Gradient Concealment: Free Lunch for Defending Adversarial Attacks

---

Paper |
Checkpoints |
Homepage

Official PyTorch Implementation

Sen Pei, Jiaxi Sun, Xin Zhang, Qing Li

Institute of Automation, Chinese Academy of Sciences

---

Table of Contents

1. About the proposed Gradient Concealment Module
   • Installation
   • Highlights of Gradient Concealment Module
   • Some Attack Robustness Results on ImageNet
2. About the CVPR Robust Classification Challenge
   • Conclusion
   • Datasets
   • Data Augmentation Schemes
   • Image Pre-Processing
   • Adversarial Defense Schemes
   • Architectures
   • Training Details
   • DDP Training
3. Reference
4. Citation
5. Acknowledgement

---

About the proposed Gradient Concealment Module

Installation

• pytorch, no strict version constraint.
• set GradientConcealment() in model/robust_layer.py as the top layer of your model in forward().

Highlights of Gradient Concealment Module

• Parameter-free, training-free, plug-and-play.
• Promising performance in both classification task and adversarial defense.
• Superior generalization across different model architectures, including CNN-based models and attention-based models.

Some Attack Robustness Results on ImageNet

• Download pre-trained models here:
  • RobustART
  • RobustBench
• Values below are AR (Attack Robustness) metric.

Model	Method	Top 1 Acc	FGSM Linf=8/255	PGD L1=1600	PGD L2=8.0	PGD Linf=8/255	C&W L2=8.0
ResNet-50	Vanilla	77.89	31.77	0.01	0.01	0.00	0.21
ResNet-50	GCM	78.57	95.18	94.82	94.41	97.38	95.11
WideResNet-50	Vanilla	78.21	20.88	0.36	0.61	0.50	0.21
WideResNet-50	GCM	78.08	96.06	94.46	94.51	97.69	95.66
DenseNet-121	Vanilla	74.86	16.82	0.04	0.05	0.06	0.12
DenseNet-121	GCM	74.71	94.98	94.31	94.08	97.16	95.49
EfficientNet-B4	Vanilla	71.52	1.23	0.36	0.28	0.20	1.88
EfficientNet-B4	GCM	71.76	94.68	89.95	90.87	97.97	93.07
ViT-B/16	Vanilla	79.46	15.86	0.00	0.00	0.00	0.90
ViT-B/16	GCM	79.47	92.24	94.94	95.07	98.24	93.31
Swin-Transformer-S	Vanilla	82.93	16.93	0.20	0.00	0.00	0.76
Swin-Transformer-S	GCM	82.79	94.38	90.71	91.04	98.77	92.31

---

About the CVPR Robust Classification Challenge

Conclusion

• Backbone does matter, ConvNext is better than SeResNet.
• Randomization is efficient for defending adversarial attacks.
• Data augmentation is vital for improving the classification performance, reducing overfitting.
• Gradient concealment dramatically improves AR metric of classifiers in presence of perturbed images.

Datasets

• train_phase1/images/ : 22987 images for training
• train_phase1/label.txt : ground-truth file
• track1_test1/ : 20000 images for testing
• train_p2 : 127390 images within 100 categories

Data Augmentation Schemes

data_aug.py supports the following operations currently:

• PepperSaltNoise
• ColorPointNoise
• GaussianNoise
• Mosaic in black / gray / white / color
• RGBShuffle / ColorJitter
• Rotate
• HorizontalFlip / VerticalFlip
• RandomCut
• MotionBlur / GaussianBlur / ConventionalBlur
• Rain / Snow
• Extend
• BlockShuffle
• LocalShuffle (for learning local spatial feature)
• RandomPadding (for defense of adversarial attacks)

Image Pre-Processing

• transforms.Resize(256)
• transforms.RandomResizedCrop(224)
• Data augmentation schemes

Adversarial Defense Schemes

• Adversarial training using fast gradient sign method.
• Resize and pad the input images for mitigating adversarial effects.
• Gradient concealment module for hiding the vulnerable direction of classifier's gradient.

Architectures

• ConvNext(tiny) + FC + FGSM regularization + GCM(Gradient Concealment Module) + Randomization
• ConvNext(tiny) + ML Decoder + FGSM regularization + GCM(Gradient Concealment Module) + Randomization

Training Details

• Training from scratch in a two-stage manner, we provide our checkpoints.
• The first stage: train ConvNext without ResizedPaddingLayer.
• The second stage: finetune ConvNext with ResizedPaddingLayer.

DDP Training

• python -m torch.distributed.launch --nproc_per_node=5 train.py --batch_size 64 --n_gpus=5
• If you have more GPUs, you can modify the nproc_per_node and n_gpus to utilize them.

---

Reference

• Explaining and Harnessing Adversarial Examples (NeurIPS, 2014)
• Deep Residual Learning for Image Recognition (CVPR, 2016)
• Squeeze-and-Excitation Networks (CVPR, 2018)
• Mitigating adversarial effects through randomization (ICLR, 2018)
• CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features (ICCV, 2019)
• A ConvNet for the 2020s (CVPR, 2022)

---

Citation

@article{Pei2022Grad,
  title={Gradient Concealment: Free Lunch for Defending Adversarial Attacks},
  author={Sen Pei, Jiaxi Sun, Xiaopeng Zhang and Gaofeng Meng},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  year={2022}
}

---

Acknowledgement

• This work is done during the author's internship at ByteDance.
• The author would like to thank Xiaojie Jin and Ye Yuan for their support in both provided computing resources and vital discussions.