yuezunli / BMVC2018R-AP

Robust Adversarial Perturbation on Deep Proposal-based Models

DenseFoolbox

DenseFoolbox is a python repository for white-box attacking object detectors, instance segmentation. This repository is a simple implementation of our paper Robust Adversarial Perturbation on Deep Proposal-based Models, BMVC2018.

Content

1. Overview
2. Requirements
3. Demo

Overview

We target Region Proposal Network (RPN) as the bottleneck of Deep-proposal based networks. The detections can be disrupted by breaking object proposal generation. To do so, we disturb the predicted class score as well as offset regression of object proposals.

Requirements

• Pytorch 0.4.0
• Ubuntu 16.04
• CUDA 8.0
• Python 2.7
• opencv3

Demo

Attacking Faster-RCNN

1. We use Faster-RCNN detector based on pytorch framework pytorch-faster-rcnn. We make modifications to this repository which can be downloaded herehere[code:rqic].
2. Unzip the repository to object_detectors.
3. Look into attack_wrapper/object_detectors_v2 and run run.py.

   python run.py \
   --net=faster-rcnn \  # faster-rcnn or ssd (update later)
   --base=vgg16 \
   --data_dir=demo/ \
   --res_dir=res/
   

Attacking Mask-RCNN

Update later

Citation

If you find this implementation helpful, please cite:

@inproceedings{li2018rap,
author={Li, Yuezun and Tian, Daniel and Chang, Mingching and Bian, Xiao and Lyu, Siwei},
title={Robust Adversarial Perturbation on Deep Proposal-based Models},
booktitle={BMVC},
year={2018}}