DPatch: An Adversarial Patch Attack on Object Detectors

This is a PyTorch implementation of DPATCH. Please refer to the paper https://arxiv.org/abs/1806.02299

Installation

1. Please install PyTorch following the instuctions on the official website. The version here should be 0.4.0

    conda install pytorch torchvision -c pytorch

2. The yolo codes are referred to https://github.com/longcw/yolo2-pytorch. Download the pretrained yolo model yolo-voc.weights.h5

3. Download the dataset VOCdevkit

   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar

4. Extract all of these tars into one directory named VOCdevkit

   tar xvf VOCtrainval_06-Nov-2007.tar
   tar xvf VOCtest_06-Nov-2007.tar
   tar xvf VOCdevkit_08-Jun-2007.tar

5. It should have this basic structure

   $VOCdevkit/                           # development kit
   $VOCdevkit/VOCcode/                   # VOC utility code
   $VOCdevkit/VOC2007                    # image sets, annotations, etc.
   # ... and several other directories ...

6. Since the program loading the data in DPatch/data by default, you can set the data path as following.

   cd DPatch
   mkdir data
   cd data
   ln -s $VOCdevkit VOCdevkit2007

Train a DPatch

The trained DPATCH are saved in trained_patch/${target_class}/ bash python train.py

Test

Download a trained DPATCH https://drive.google.com/open?id=1_G5xXWIJWNGuss4KZbBQ9pMvuthmL_wc , or you can test your trained DPatch and set the path in cfgs/config.py (Line 102) bash python test.py –-attack untargeted python test.py –-attack targeted

Demo

The demo pictures are saved in demo/patch, the detection results are saved in demo/ bash python demo.py

Click the following picture to watch the video demo.