Skip to content
/ TPA Public

Official implementation of paper "Threatening Patch Attacks on Object Detection in Optical Remote Sensing Images".

License

Notifications You must be signed in to change notification settings

plpl2019/TPA

Repository files navigation

Threatening Patch Attacks on Object Detection in Optical Remote Sensing Images

Datasets

The download link of these datasets can be accessed at AAOD-ORSI.

Training details of victim detectors

A total of four kinds of detectors are utilized in this paper for the evaluations including Faster R-CNN, RetinaNet, FCOS, and Yolo-v4. For Faster R-CNN, RetinaNet, and FCOS, we use MMDetection(we use mmcv-full=1.4.7=pypi_0, mmdet=2.3.0=dev_0 in this paper) as the main framework. For Yolo-v4, we use the same framework as RPAttack.

The main body of our method is based on the framework of RPAttack. Since I have graduated with a master's degree and started to work, the collation of the code may be delayed for a while. I will do my best to update the code as soon as possible. Below are the training details of the victim detectors leveraged in this paper. Thanks again for browsing this repository and hoping this can help you.

model datesets epoch learning rate batch decay decay rate mAP recall
Faster R-CNN+Resnet50 DIOR(train+val) 24 0.01 4 [16,22] 0.1 88.3 90.3
Faster R-CNN+Resnet50 DOTA(train) 12 0.001 2 [8,10] 0.1 68.7 77.7
Faster R-CNN+Resnet101 DIOR(train+val) 24 0.01 4 [16,22] 0.1 88.6 90.9
Faster R-CNN+Resnet101 DOTA(train) 12 0.001 2 [8,10] 0.1 68.4 76.1
FCOS+Resnet50 DIOR(train+val) 24 0.01 4 [16,22] 0.1 87.3 91.3
FCOS+Resnet50 DOTA(train) 12 0.001 2 [8,10] 0.1 65.7 79.1
FCOS+Resnet101 DIOR(train+val) 24 0.01 4 [16,22] 0.1 87.6 91.6
FCOS+Resnet101 DOTA(train) 12 0.001 2 [8,10] 0.1 66.8 80.0
RetinaNet+Resnet50 DIOR(train+val) 24 0.01 4 [16,22] 0.1 87.3 92.8
RetinaNet+Resnet50 DOTA(train) 12 0.001 2 [8,10] 0.1 62.2 79.5
RetinaNet+Resnet101 DIOR(train+val) 24 0.01 4 [16,22] 0.1 87.3 92.8
RetinaNet+Resnet101 DOTA(train) 12 0.001 2 [8,10] 0.1 64.8 81.3
YOLOv4 DIOR(train+val) 30000 0.001 32 [20000,25000] 0.1 89.5 90.0
YOLOv4 DOTA(train) 30000 0.001 32 [20000,25000] 0.1 69.7 76.8

Usage

Preparing your environment via requirements.txt, Then, to run the attack against the detector, e.g., faster r-cnn, with iteration = 10, you can run

REPO_ROOT$ python attack_faster.py --iters 10 --save_name xxxx

If you want to attack YOLO, please edit the file yolov4/eval_code/models/yolov4.cfg to set the parameters, then run

REPO_ROOT$ python attack_yolo.py --iters 10 --save_name xxxx

Citation

If you use these repository, please cite the following:

@article{sun2023threatening,
  title={Threatening Patch Attacks on Object Detection in Optical Remote Sensing Images},
  author={Sun, Xuxiang and Cheng, Gong and Pei, Lei and Li, Hongda and Han, Junwei},
  journal={IEEE Trans. Geosci. Remote Sens.},
  volume={61},
  pages={1-10},
  year={2023}
}

Notation

As I have graduated and am working on a new job position, I am currently unable to dedicate time to maintaining this codebase. Therefore, this repository is no longer actively maintained.

I want to extend my gratitude to everyone who has shown interest in this project, and to all colleagues who have contributed to its development. Thank you for your support and understanding!

I wish everyone success in their research endeavors and hope for smooth progress in your work.

About

Official implementation of paper "Threatening Patch Attacks on Object Detection in Optical Remote Sensing Images".

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published