Moire-Backdoor-Attack (MBA)

Official Pytorch implementation for our ACM MM 2023 paper: Moiré Backdoor Attack (MBA): A Novel Trigger for Pedestrian Detectors in the Physical World

Samples

Preparation

• Python
• Pytorch
• YOLOv5
• MMDetection

Usage

Please download the COCO2017 dataset in this link and the OCHuman in this link.

• Step 0:
  Creating COCOPerson and OCHuman datasets in YOLO format, complete with mask labels, based on the aforementioned two datasets.

• Step 1:
  Generating poisoned data samples.

  python moire2img.py --source imageDir

• Step 2:
  Generating poisoned dataset and put in the dataset folder.

  python move2dataset.py

• Step 3:
  Train on the poisoned dataset using the corresponding scripts of the two libraries YOLOv5 and MMDetection.

Acknowledgements

We would like to acknowledge the YOLOv5 open-source library (https://github.com/ultralytics/yolov5) and MMDetection open-source library (https://github.com/open-mmlab/mmdetection).