Zhaoyang Jia, Han Fang, Weiming Zhang (from University of Science and Technology of China)

[arXiv] [PDF]

This is the source code of paper MBRS : Enhancing Robustness of DNN-based Watermarking by Mini-Batch of Real and Simulated JPEG Compression, which is received by ACM MM' 21 (oral). Please contact me in issue page or email [email protected] if you find bugs. Thanks!

Because of the training process for models with diffusion model (details about diffusion model is in the paper) is not stable, we update the training process for a more stable process.

• Raw training is for 128x128 images and 30 bits message, and it's 256 dimensions after the Fully Connection embedding. Batch size is 16, and learning rate is 1e-3. We train it for 300 epochs and apply early stopping at 110 epoch, decided by the validation results. In this way we get the pretrained model and the result in the paper is based on it.

• However, this training process is not stable for crop robustness (Crop attack · Issue #2), that is, the validation result for Crop varies from BER=2% to BER=25% for Crop(p=3.5%), and it's hard to guarantee the as good results as we get.

• To solve it, we update the training process in an easy but useful way.

  • First we train the model like in raw training for 100 epochs and apply early stop at 92 epoch to get a suboptimum model (BER = 20% and PSNR=29.75 for Crop(p=3.5%)).
  • Then we finetune the model for 50 epochs with the same settings but learning rate=1e-4, and apply early stop at 13 epoch to get the optimum model (BER = 1.85% and PSNR=30.89 for Crop(p=3.5%)).

Wish this can help you :)

We used these packages/versions in the development of this project.

• Pytorch 1.5.0
• torchvision 0.3.0a0+ec20315
• kornia 0.3.0
• numpy 1.16.4
• Pillow 6.0.0
• scipy 1.3.0

Please download ImageNet or COCO datasets, and push them into datasets folder like this :

├── datasets
│   ├── train
│   │   ├── xxx.jpg
│   │   ├── ...
│   ├── test
│   │   ├── xxx.jpg
│   │   ├── ...
│   ├── validation
│   │   ├── xxx.jpg
│   │   ├── ...
├── ...
├── results

For more details about the used datasets, please read the original paper.

Please download pretrained models in Google Drive and put the in path results/xxx/models/. (xxx is the name of the project, e.g. MBRS_256_m256)

Change the settings in json file train_settings.json, then run :

python train.py

The logging file and results will be saved at results/xxx/

Change the settings in json file test_settings.json, then run :

python test.py

The logging file and results will be saved at results/xxx/

Please cite our paper if you find this repo useful!

@inproceedings{jia2021mbrs,
  title={MBRS: Enhancing Robustness of DNN-based Watermarking by Mini-Batch of Real and Simulated JPEG Compression},
  author={Zhaoyang  Jia, Han Fang and Weiming Zhang},
  booktitle={arXiv:2108.08211},
  year={2021}
}

Contact: [email protected]