Official implementation of "Adversarial attacks and defenses using feature-space stochasticity"

What's included in this repository?

• Attack to create feature-space adversarial examples (Section 4.1)
• Geometrical analysis (Section 4.2)
• Defense experiments (Section 4.3)

All experiments are performed on VGG16 and ResNet50 classifier using STL-10 dataset.

Requirements

PyTorch >= 1.7

Run

Code	Description
attack_vgg16.sh	Create feature-space adversarial examples on VGG16 classifier
attack_resnet50.sh	Create feature-space adversarial examples on ResNet50 classifier
geometry.sh	Run geometry analysis
smoothing_predict.sh	Run defense experiments

Code References

• Classifier implementation: torchvision package
• GAN implementation: PyTorch-GAN
• Randomized smoothing: Official implementation
• C&W attack: Random self-ensemble

Citation

Adversarial attacks and defenses using feature-space stochasticity. Jumpei Ukita and Kenichi Ohki, Neural Networks, 2023. Paper link