The winning submission for NIPS 2017: Defense Against Adversarial Attack of team TSAIL

Defense against Adversarial Attacks Using High-Level Representation Guided Denoiser

• prepare_data.ipynb: generate dataset

• Originset, Originset_test: the folder for original image

• toolkit: the program running the attack in batch

• Attackset: the attacks

• Advset: the adversarial images

• checkpoints: the models checkpoint used, download here

• Exps: the defense model

• GD_train, PD_train: train the defense model using guided denoise or pixel denoise

the attacks are stored in folder Attackset the script is in the toolkit folder. in the run_attacks.sh file: modify models to the attacks you want to generate, separate by comma, or use "all" to include all attacks in Attackset. use the command to run:

bash run_attacks.sh $gpuids

where gpuids is the id of the gpus you want to use, they are number separated by comma. It will generate the training set. Then change the line DATASET_DIR="${parentdir}/Originset" to DATASET_DIR="${parentdir}/Originset_test", and run the command bash run_attacks.sh $gpuids again.

Then specify a model you want to use, the models are stored in Exp folder, there is a sample folder, it refers to a model named "sample", let's use it. Then go to GD_train if you want to use guided denoiser, run

python main --exp sample

The program will load Exp/sample/model.py as a model to train. and also you can specify other parameters defined in the GD_train/main.py