PyTorch implementaion for Proactive Privacy-preserving Learning for Cross-modal Retrieval
pip install -r requriements.txt
h5py==3.8.0
numpy==1.23.5
opencv_python==4.7.0.68
Pillow==9.4.0
scipy==1.10.1
torch==1.13.1
torchvision==0.14.1
tqdm==4.64.1
Download the MIRFLICKR25K(password:8dub) and put it into ./datasets.then run
python train.pyThe training process costs about 30min on 1*GeForce RTX 3090.
JDSH is selected as the targeted fooling retrieval system to test the validity of PPCL.
Download the checkpoint file of PPCL and well-trained JDSH model based on JDSH repo and put them into ./models,then run
python eval.py
| Categories | MAP |
|---|---|
| MAP of Image to Text | 0.775 |
| MAP of Image to Image | 0.824 |
| MAP of Text to Image | 0.847 |
| MAP of Text to Text | 0.833 |
| MAP of Image to Noise Image | 0.703 |
| MAP of Noise Image to Image | 0.684 |
| MAP of Noise Image to Noise Image | 0.683 |
| MAP of Text to Noise Image | 0.724 |
| MAP of Noise Image to Text | 0.663 |
The retrieval model takes Alexnet as the backbone instead of VGG16,which is slightly different from the paper.