Description

• DAmageNet is generated in paper "Universal Adversarial Attack on Attention and the Resulting Dataset DAmageNet", IEEE TPAMI.
• DAmageNet is a massive dataset containing universal adversarial samples generated from ImageNet.
• DAmageNet contains 50000 224*224 images, whose original images have been centrally cropped and resized.
• DAmageNet images have an average root mean square deviation of around 7.32 from original samples.
• DAmageNet can fool pretrained models in ImageNet to have error rate up to 85%.
• DAmageNet can fool adversariral-trained models in ImageNet to have error rate up to 70%.

Test

• Prepare DAmageNet, unzip to this folder as 'DAmageNet' and test by

python test.py DAmageNet VGG19,ResNet50,DenseNet121 [gpu_id]

• Each file in DAmageNet has the same name as in ILSVRC2012_img_val.

Generation

• Prepare ImageNet validation set (2012), place in folder 'ILSVRC2012_img_val'
• Prepare the environment as in test.py
• Copy base.py to the path in iNNvestigate

• run

python damagenet.py 0 50000 [gpu_id]

• See details in attack, run

python damagenet.py 0 100 [gpu_id]

Demo

• Reproduce the result of Fig. 4 in the paper, run

python lrp.py

Authors

• Sizhe Chen, [email protected]
• Zhengbao He, [email protected]
• Chengjin Sun, [email protected]
• Jie Yang, [email protected]
• Xiaolin Huang*, [email protected]
• Institute of Image Processing and Pattern Recognition, Shanghai Jiao Tong University.