Introduction

This is a project for the Tianchi competition: adversarial attack for universal object detection. Here is the url: https://tianchi.aliyun.com/competition/entrance/531806/information. We obtain the second in this contest.

Data preparation and model checkpoint.

• Download 1000 pictures needed for the competition on the official website
• You can get data (images.zip) and the definition, weight and evaluation code of two white box models (eval_code.zip). We use yolov4 and faster_rcnn as whitebox models.
• Create two new folders, images and models, Unzip images.zip to images, and move all checkpoint and config files to models.

Requirements

This code is based on pytorch. Some basic dependencies are recorded in requirements.txt

• torch
• torchvision
• pillow
• numpy
• tqdm
• scipy
• scikit-image

You can run yolov4 now if all above requirements are satisfied.

Another faster rcnn model is implemented based on mmdetection. So, ensure that the mmdetection library has been installed and can be run on your machine. You can refer install guide of mmdetection to github

After installation, put the mmdetection directory into eval_code/ below. Alternatively, it is optional that using docker provided by mmdetection.

Usage

Unzip eval_code.zip，move and unzip images.zip to images, ensure the following structure:

|--images
    |-- XXX.png
    |-- XXX.png
    |-- XXX.png
    …
    |-- XXX.png

Move all checkpoints and config files to models as:

|--models
    |-- faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth
    |-- yolov4.cfg
    |-- yolov4.weights

Run Attack algorithm

python attack.py --patch_type grid --lines 3 --box_scale 1.0
python attack.py --patch_type grid --lines 2 --box_scale 1.0
python attack.py --patch_type grid --lines 1 --box_scale 1.0


python attack.py --patch_type astroid

Run ensemble algorithm

python ensemble.py