VITA-Group / 3D_Adversarial_Logo

[Preprint] "Can 3D Adversarial Logos Cloak Humans?"

Can 3D Adversarial Logos Clock Humans?

Can 3D Adversarial Logos Clock Humans?

Yi Wang*, Jingyang Zhou*, Tianlong Chen, Sijia Liu, Shiyu Chang, Chandrajit Bajaj, Zhangyang Wang

Overview

Examples of our 3D adversarial logo attack on different3D object meshes to fool a YOLOV2 detector.

Methodology

Prerequisites and Installation

• Python 3

• Pytorch 1.1.0

• CUDA 9.2 (lower versions may work but were not tested)

• TensorboardX

  pip install tensorboardX tensorboard

• Neural renderer

  pip install neural-renderer-pytorch

• Clone this repo

  git clone https://github.com/TAMU-VITA/3D_Adversarial_Logo.git
  cd 3D_Adversarial_Logo

• Yolov2 weights

  mkdir weights
  curl https://pjreddie.com/media/files/yolov2.weights -o weights/yolo.weights

Usage

Prepare the dataset

• Datasets and our prediction images can be found here.
• Our pre-trained 3d adversarial logos can be found here.

Remark. After downloading them, put the data and weight under the 3D_Adversarial_Logo folder.

Arguments

• --width : the width of the universal logo.
• --height : the height of the universal logo.
• --depth : the depth of the universal logo (default is 1 to form an image).
• --angle_range : the angle range of the camera view for training (eg. 10 for +/- 10 degree)
• --logonum : the total faces of the 3D logo.
• --train_mesh : which mesh to train (the 123 denotes mesh 1&2&3, which can be further upgraded by adding nargs='+' in parsing args).
• --test_mesh : which mesh to test (the 123 denotes mesh 1&2&3, which can be further upgraded by adding nargs='+' in parsing args).
• --consistent : whether to share the same universal logo among all the meshes.
• --logo_ref : which logo to train with.
• --train_patch : whether to train 2D patch.
• --restore_model : whether to restore the saved universal logo.
• --save_model : whether to save the universal logo.

Run 3D adversarial logos

• logo: X

  python train_patch1.py --width=100 --height=150 --depth=1 --angle_range=0 --logonum=25 --train_mesh=123 --consistent --logo_ref=X --save_model >logo-records/logoX.out 2>&1 &

• logo: X (medium)

  python train_patch1.py --width=100 --height=150 --depth=1 --angle_range=0 --logonum=10 --train_mesh=123 --consistent --logo_ref=X --save_model >logo-records/logoX.out 2>&1 &

• logo: X (small)

  python train_patch1.py --width=100 --height=150 --depth=1 --angle_range=0 --logonum=5 --train_mesh=123 --consistent --logo_ref=X --save_model >logo-records/logoX.out 2>&1 &

• logo: H

  python train_patch1.py --width=100 --height=150 --depth=1 --angle_range=0 --logonum=25 --train_mesh=123 --consistent --logo_ref=H --save_model >logo-records/logoH.out 2>&1 &

• logo: T

  python train_patch1.py --width=100 --height=150 --depth=1 --angle_range=0 --logonum=25 --train_mesh=123 --consistent --logo_ref=T --save_model >logo-records/logoT.out 2>&1 &

• logo: G

  python train_patch1.py --width=100 --height=100 --depth=1 --angle_range=0 --logonum=17 --train_mesh=123 --consistent --logo_ref=G --save_model >logo-records/logoG.out 2>&1 &

• logo: C

  python train_patch1.py --width=100 --height=100 --depth=1 --angle_range=0 --logonum=17 --train_mesh=123 --consistent --logo_ref=C --save_model >logo-records/logoC.out 2>&1 &

• logo: O

  python train_patch1.py --width=100 --height=100 --depth=1 --angle_range=0 --logonum=17 --train_mesh=123 --consistent --logo_ref=O --save_model >logo-records/logoO.out 2>&1 &

• logo: tw

  python train_patch1.py --width=100 --height=100 --depth=1 --angle_range=0 --logonum=25 --train_mesh=123 --consistent --logo_ref=tw --save_model >logo-records/logotw.out 2>&1 &

• logo: drop

  python train_patch1.py --width=100 --height=150 --depth=1 --angle_range=0 --logonum=25 --train_mesh=123 --consistent --logo_ref=drop --save_model >logo-records/logodrop.out 2>&1 &

Train 2D adversarial patch

• logo: G

  python train_patch1.py --width=100 --height=100 --depth=1 --angle_range=0 --logonum=17 --train_mesh=123 --consistent --logo_ref=G --restore_model --save_model --train_patch

Citation

If you are use this code for you research, please cite our paper.

@article{chen2020can,
  title={Can 3D Adversarial Logos Cloak Humans?},
  author={Chen, Tianlong and Wang, Yi and Zhou, Jingyang and Liu, Sijia and Chang, Shiyu and Bajaj, Chandrajit and Wang, Zhangyang},
  journal={arXiv preprint arXiv:2006.14655},
  year={2020}
}