Adversarially-Aware Robust Object Detector (RobustDet)

Introduction

This repo is the official PyTorch implementation of ECCV2022 oral paper "Adversarially-Aware Robust Object Detector".

RobustDet is an approach to improve the adversarial robustness of CNN-based object detectors. It aims to defends adversarial attacks specifically for object detectors. For more details, please refer to our paper.

Usage

Requirements

Python 3.9
PyTorch >= 1.8
numpy
cv2

pip install -r requirements.txt  # install requirements

Data Preparation

Download the PASCAL VOC and MS-COCO dataset and unpack them. The data structure should look like this

VOC

VOCdevkit
|-- VOC2007
|      |-- Annotations
|      |-- ImageSets
|      |-- JPEGImages
|-- VOC2012
       |-- Annotations
       |-- ImageSets
       |-- JPEGImages

COCO

coco2017
|-- 2017_clean  # merge train2017 and val2017
|-- annotations
       |-- instances_train2017.json
       |-- instances_val2017.json

RobustDet

We provide codes to reproduce the results in our paper.

Training

Download pretrained backbone and clean models, and put them intoweights/ before training.

To train RobustDet model on VOC dataset:

python train_robust.py --cfg cfgs/RobustDet_voc.yaml --adv_type mtd --data_use clean --multi_gpu False \
    --basenet weights/ssd300_mAP_77.43_v2.pth --dataset_root <path_to_your_VOC_root>

Training on COCO dataset:

python train_robust.py --cfg cfgs/RobustDet_coco.yaml --adv_type mtd --data_use clean --multi_gpu False \
    --basenet weights/ssd300_COCO_clean_final_300000.pth --dataset_root <path_to_your_COCO_root>

Evaluation

VOC

python eval_attack.py --cfg cfgs/RobustDet_voc.yaml --trained_model <path_to_your_trained_model> \
    --data_use clean --adv_type cls \ # attack type, choice in [clean, cls, loc, cwat, dag]
     --dataset_root <path_to_your_VOC_root>

COCO

python eval_attack.py --cfg cfgs/RobustDet_coco.yaml --trained_model <path_to_your_trained_model> \
    --data_use clean --adv_type cls \ # attack type, choice in [clean, cls, loc, cwat, dag]
     --dataset_root <path_to_your_COCO_root>

Baseline Methods (MTD and CWAT)

Our baseline method "Towards Adversarially Robust Object Detection"(MTD) and "Class-Aware Robust Adversarial Training for Object Detection github"(CWAT) both not provide source code.

But in this repo we provide the code that we reproduced.

Training

MTD

python train_adv.py --cfg cfgs/MTD_voc.yaml --adv_type mtd --data_use clean --multi_gpu False \
    --dataset_root <path_to_your_VOC_root>

CWAT

python train_adv.py --cfg cfgs/MTD_voc.yaml --adv_type cwat --data_use clean --multi_gpu False \
    --dataset_root <path_to_your_VOC_root>

Evaluation

python eval_attack.py --cfg cfgs/MTD_voc.yaml --trained_model <path_to_your_trained_model> --data_use clean --adv_type cls

Pretrained Models

Citation

@InProceedings{dong2022robustdet,
  author = {Ziyi Dong, Pengxu Wei, Liang Lin},
  title = {Adversarially-Aware Robust Object Detector},
  booktitle = {Proceedings of the European Conference on Computer Vision},
  year = {2022}
}

7eu7d7/RobustDet