A Implementation of Transferable Adversarial Attacks for Image and Video Object Detection

1. Introduction

• This code is the implementation of the paper Transferable Adversarial Attacks for Image and Video Object Detection
• This paper has been received by IJCAI-19.

2. Pretrain

https://drive.google.com/open?id=1PpffJyBNuF8jw6tetea3HVGij6CbrjbH

You can dowmload them as your pretrained model of fasterRCNN, thanks to the author Yun Chen for sharing.

3. Demo

After you modify the relevant configuration, just run this ipynb in colab.

test.ipynb

4. Train

Just run this ipynb in colab.

feature_and_attack_train.ipynb

5. Results

• Training this network cost a lot of time, evary iteration cost around 7s(running in colab), while once training is comleted, the speed of generating adversarial examples is quite fast.
• Due to the slow upload speed, we only show 7 examples in 'sdd' folder.

Acknowledgement

This work was just slightly modified from this awesome github project:

• Adversarial-Attacks-for-Image-and-Video-Object-Detection .