Content-based Graph-Privacy-Advisor

Graph Privacy Advisor (GPA) is an image privacy classifier that uses scene and object cardinality information to predict the privacy of an image. GPA refines the relevance of the information extracted from the image and determines the most informative features to be used for the privacy classification task.
paper: https://arxiv.org/abs/2210.11169

Installation

Requirements

• opencv-python==4.5
• torchvision=0.9.1
• python=3.9
• pytorch=1.8
• numpy=1.20
• sklearn
• matplotlib

Download datasets

• PicAlert: http://l3s.de/picalert/
• VISPR: https://tribhuvanesh.github.io/vpa/
• PrivacyAlert: https://zenodo.org/record/6406870#.Y2KtsdLP3ow

Instructions

1. Clone repository
   git clone https://github.com/dimitriStoidis/Graph-Privacy-Advisor

2. From a terminal or an Anaconda Prompt, go to project's root directory and run:
   conda create --name gpa python=3.9
conda activate gpa
   and install the required packages

3. Setup Object detection:
   Create a folder named /config inside /data_preprocess_ subdirectory
   Include the COCO object labels and Yolo model configuration files available here: https://github.com/pjreddie/darknet
   Download pre-trained weights for Yolo running: wget https://pjreddie.com/media/files/yolov3.weights in the /config directory

4. Setup scene recognition:
   Download ResNet model following instructions available here: https://github.com/CSAILVision/places365
   For ResNet-18 model run: wget http://places2.csail.mit.edu/models_places365/resnet18_places365.pth.tar

Demo

To run the demo script run
python demo.py --model_name GPA_scene_card --cardinality True --scene True --image_name your_image.jpg

Training example

To train the model run:
python main.py --model_name model1 --num_epochs 50 --batch_size 64 --cardinality True --scene True

References

The work is based on:

• https://github.com/guang-yanng/Image_Privacy
• https://github.com/HCPLab-SYSU/SR
• https://github.com/CSAILVision/places365
• https://pjreddie.com/darknet/yolo/

Cite

@misc{https://doi.org/10.48550/arxiv.2210.11169,
  doi = {10.48550/ARXIV.2210.11169},
  url = {https://arxiv.org/abs/2210.11169},
  author = {Stoidis, Dimitrios and Cavallaro, Andrea},
  keywords = {Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {Content-based Graph Privacy Advisor},
  publisher = {IEEE BigMM},
  year = {2022},
  copyright = {Creative Commons Attribution 4.0 International}
}

Contact

For any enquiries contact dimitrios.stoidis@qmul.ac.uk, a.cavallaro@qmul.ac.uk

Licence

This work is licensed under the MIT License.