Sizhe-Chen/RAD

Relevance attack on detectors

Environment

• This is the official implementation for "Relevance Attack on Detectors" published in Pattern Recognition.

conda create -n rad python=3.6.2
conda activate rad
conda install tensorflow-gpu==1.13.1
pip install pillow==6.2.1 opencv-python==4.1.2.30 scikit-image==0.15.0 matplotlib==2.2.2

• COCO/val2017 download
• COCO/annotations/instances_val2017_resized.json

Attact YOLOv3

• model_data/yolo.h5 download config convert

• Run default parameters on GPU 0

python yolov3.py 0

• Run customized parameters on GPU 1

python yolov3.py 1 --alpha=1 --epsilon=10 --num_iteration=5

Attact RetinaNet

• model_data/resnet50_coco_best_v2.1.0.h5 download
• install keras-retinanet
• Run default parameters on GPU 0

python retinanet.py 0

• Run customized parameters on GPU 1

python retinanet.py 1 --alpha=1 --epsilon=10 --num_iteration=5

Test

• pytorch 1.3.1
• install mmdetection 1.1.0
• models download
• Test on GPU0 for 2020-06-09-15-19-07_RAD_SI_YOLOv3_Index0to5000_Eps2_Iter10

python test.py 2020-06-09-15-19-07_RAD_SI_YOLOv3_Index0to5000_Eps2_Iter10 0

AOCO Dataset

• Adversarial Objects in COntext (AOCO) is the first adversarial dataset for object detection and instance segmentation.
• AOCO serves as a potential benchmark to evaluate the robustness of detectors, which is beneficial to network designers.
• AOCO is generated from the full COCO 2017 validation set with 5k samples.
• AOCO contains 5K adversarial samples for evaluating object detection and 5K for instance segmentation.
• All 10K samples in AOCO are crafted by our RAD.
• [download] password: gbug