phibenz/double-targeted-uap.pytorch

Double Targeted Universal Adversarial Perturbations

This is the repository for our ACCV 2020 paper titled Double Targeted Universal Adversarial Perturbations

Abstract

Despite their impressive performance, deep neural networks (DNNs) are widely known to be vulnerable to adversarial attacks, which makes it challenging for them to be deployed in security-sensitive applications, such as autonomous driving. Image-dependent perturbations can fool a network for one specific image, while universal adversarial perturbations are capable of fooling a network for samples from all classes without selection. We introduce a double targeted universal adversarial perturbations (DT-UAPs) to bridge the gap between the instancediscriminative image-dependent perturbations and the generic universal perturbations. This universal perturbation attacks one targeted source class to sink class, while having a limited adversarial effect on other nontargeted source classes, for avoiding raising suspicions. Targeting the source and sink class simultaneously, we term it double targeted attack (DTA). This provides an attacker with the freedom to perform precise attacks on a DNN model while raising little suspicion. We show the effectiveness of the proposed DTA algorithm on a wide range of datasets and also demonstrate its potential as a physical attack.

Setup

We performed our experiments with PyTorch v.0.4.1

Config

Copy config/sample_config.py to config/config.py. Edit the paths in config/config.py according to your system.

Datasets

ImageNet

1. Follow the common setup to make ImageNet compatible with pytorch as described in here
2. Set the path to the pytorch ImageNet dataset folder in the config file

GTSRB

1. Download GTSRB-Training_fixed.zip, GTSRB_Final_Test_GT.zip, GTSRB_Final_Test_Images.zip from here.
2. Extract the files into the following Folder structure:

GTSRB
    ∟- Training
    ∟- Final_Test
    ∟- GT-final_test.csv

3. Set the path to the GTSRB folder in the config file
4. Run python3 ./dataset_utils/gtsrb_preparation.py. This should generate a "Testing" folder in your GTSRB folder. The dataset is now ready to be used.

EuroSAT

1. Dowload the EuroSat data from here
2. Extract the zip file unzip EuroSAT.zip
3. Set the path to the EuroSAT folder in the config file

YCB

1. Download the YCB dataset via python3 dataset_utils/ycb_downloader.py
2. Make a correction in the dataset with mv ./ycb/home/bcalli/025_mug ./ycb/.

Mobilenet V2

1. The Mobilenet V2 checkpoint was downloaded from here and is placed in ./models/mobilenet_v2/mobilenet_v2-b0353104.pth.
2. Modify the path to the checkpont accordingly.

Experiments

The code can be directly tested with bash ./experiments/one2one_cifar10. To train additional models, have a look into the bash script ./experiments/train_model.sh

Citation

@inproceedings{benz2020double,
  title={Double targeted universal adversarial perturbations},
  author={Benz, Philipp and Zhang, Chaoning and Imtiaz, Tooba and Kweon, In So},
  booktitle={Proceedings of the Asian Conference on Computer Vision},
  year={2020}
}