Adversarial-Semantic-Contour

This is the official implementation code for the CVIU paper "To Make Yourself Invisible with Adversarial Semantic Contours".

Installation

Please refer to Installation for installation instructions.

Note that you should install it from source, since some of the source code in mmdetection have been modified to fit our need for ASC.

Run the demo

To check the clean detection

python test.py --imgdir data/demo/images --model frcn --task disappear

To generate adversarial example with Fixed Adeversarial Semantic Contour

python fasc.py --root data/demo --model frcn --method fasc

To generate adversarial example with Optimized Adeversarial Semantic Contour

python oasc.py --root data/demo --model frcn

Reproduce the result

Here, we provide the dataset of 1000 Person from MSCOCO2017 and the corresponding result of F-ASC for Faster RCNN. Download the zip file and place it under the directory data.

To check the clean detection rate (98.3%)

python test.py --imgdir data/coco_person/images --model frcn --task disappear

To reproduce the detection rate under attack by F-ASC (9.7%)

python test.py --imgdir data/coco_person/result/disappear/fasc_frcn --model frcn --task disappear

Citation

@article{zhang2023make,
  title={To make yourself invisible with Adversarial Semantic Contours},
  author={Zhang, Yichi and Zhu, Zijian and Su, Hang and Zhu, Jun and Zheng, Shibao and He, Yuan and Xue, Hui},
  journal={Computer Vision and Image Understanding},
  volume={230},
  pages={103659},
  year={2023},
  publisher={Elsevier}
}