Awesome

Comdefend

The code for CVPR2019 (ComDefend: An Efficient Image Compression Model to Defend Adversarial Examples) paper The pure tensorflow of code is released in the link

Environmental configuration

tensorflow>=1.1 </br> python3 </br> canton(pip install canton) </br>

Train

Torch: Train_comdefend_Torch.py </br> TF: Train_comdefend_TF.py

Description

clean_image: we select 7 clean images which are classified correctly by the classifier </br> attack_image: we select 7 adversarial images which are attacked by the FGSM method </br> defend_image: we use the Comdefend to deal with 7 adversarial images</br> chackpoints: the model parameters </br> com_imagenet_temp, temp_imagenet: the temporary files of the Comdefend</br> dev.csv: correspondence between images and labels</br> Resnet_imagenet.py: the classifier </br> compression_imagenet.py: the Comdefend for Imagenet</br> compression_mnist.py: the Comdefend for fashion_mnist</br>

In addition

E-mail: jiaxiaojun@iie.ac.cn