Defense against Universal Adversarial Perturbations

This repository contains the Tensorflow implementation of "Defense against Universal Adversarial Perturbations"(CVPR2018)

Environment

• Python 2.7
• Tensorflow 1.4.0

Usage

TESTING

1. Download Inception PRN models from google drive.

2. Edit TESTING script, set --data_dir to the direcotry of clean images, set --perturb_dir to the base pertubation directory, and set --pert_test_dir to the TESTING perturbation folder.

Directory structure of -data_dir:

ILSVRC2012_img
├── 1
│   ├── ILSVRC2012_val_00000756.png
│   ├── ILSVRC2012_val_00001260.png
│   ├── ILSVRC2012_val_00006145.png
│   └── ...
├── 2
├── 3
├── ...
└── 1000

Directory structure of --pert_test_dir:

inception_L2_Pert
├── perturbation_map_1.npy 
├── perturbation_map_2.npy 
└── ...

4. Run TESTING script.

TRAINING

1. Edit TRAINING configuration. Set --data_dir to clean images, note that data_dir_1~data_dir_4 are the corner-cropped images used as data augmentation. perturb_dir and pert_train_dir have similar reference as in TESTING.

2. Run TRAINING script.

References

• LTS4 Universal project
• Taehoon Kim's simGAN code