• Adversarial Texture for Fooling Person Detectors in Physical World
• • • • 1. Installation
    • Requirements
    • • 2. Preparation
      • 3. Evaluation
      • • RCA
        • TCA
        • EGA
        • TC-EGA
        • Plot multiple results together
      • 4. Train

Adversarial Texture for Fooling Person Detectors in Physical World

Official implementation for the CVPR 2022 paper Adversarial Texture for Fooling Person Detectors in Physical World

by Zhanhao Hu, Siyuan Huang, Xiaopei Zhu, Xiaolin Hu, Fuchun Sun, Bo Zhang.

Note: The code is based on the official implementation for the paper Fooling Automated Surveillance Cameras: Adversarial Patches to Attack Person Detection.

1. Installation

Requirements

All the codes are tested in the following environment:

• Linux (Ubuntu 18.04.4)
• Python 3.6
• PyTorch 1.7.1
• CUDA 11.0
• TensorboardX 2.2
• EasyDict 1.9

2. Preparation

You need to download the yolov2 and yolov3 weights by

wget -P ./data/models/ https://pjreddie.com/media/files/yolov2.weights
wget -P ./data/models/ https://pjreddie.com/media/files/yolov3.weights 

and prepare the Inria Dataset

curl ftp://ftp.inrialpes.fr/pub/lear/douze/data/INRIAPerson.tar -o inria.tar
tar xf inria.tar
mv INRIAPerson ./data

3. Evaluation

We provide the pre-trained parameters of 4 methods for implementing Adversarial Textures (AdvTexture) to attack YOLOv2. Use the following commands to output average precision (AP) for single run. Each command will additionally output a numpy data file and a png format image which are stored in directory "test_results/"

RCA

When the side length of the cloth is two times that of the patch

python evaluation_texture.py --method RCA --load_path pretrained/RCA2.npy --suffix yolov2_RCA2 --prepare_data

When the side length of the cloth is six times that of the patch

python evaluation_texture.py --method RCA --load_path pretrained/RCA6.npy --suffix yolov2_RCA6 --prepare_data

TCA

python evaluation_texture.py --method TCA --load_path pretrained/TCA.npy --prepare_data

EGA

python evaluation_texture.py --method EGA --load_path pretrained/EGA.pkl --prepare_data

TC-EGA

python evaluation_texture.py --method TCEGA --load_path pretrained/EGA.pkl --load_path_z pretrained/TCEGA_z.npy --prepare_data

Plot multiple results together

In addition, we provide a command to plot all the results together. For example, if one run all five evaluations above, then run:

python evaluation_texture.py --npz_dir ./test_result

It will output a precision v.s. recall curve located at "test_result/PR-curve.png".

4. Train

We provide the command to train for each method.

 python training_texture.py --net $NET_NAME --method $METHOD_NAME

One need to replace $NET_NAME to "yolov2" or "yolov3" to target different detectors. The $METHOD_NAME can be replaced to "RCA", "TCA", EGA" or "TCEGA" to use different methods. For example, one can use TC-EGA to attack YOLOv2 by

 python training_texture.py --net yolov2 --method TCEGA

The patterns and checkpoints will be saved in directory "results/".