- 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%.
- 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.
- 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]
- Reproduce the result of Fig. 4 in the paper, run
python lrp.py
- 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.