Invisible-Robust-Copyright-Protection

The official implementation of An invisible, robust copyright protection method for DNN-generated content, which is accpeted by Neural Network.

Dataset

In this work, we use two datasets: "summer2winter" released by the cycleGAN and a subset of ImageNet that random sampled from ImageNet training set, and the sample index can be found in here.

Style Transfer

In this work, we adopt the style transfer technique proposed by A Neural Algorithm of Artistic Style, and we adopt the implementation provided by the pytorch tutorial.

To perform style transfer, please config the data path and style image in style_transfer/generate_transfered_images.py, then run the following command:

python style_transfer/generate_transfered_images.py

Training

To train the network, you should first configure the data path in 'config.yaml', and then run the following command:

python train_res256.py

Later, we will provide some pretrained models.

Testing

To test the performance of the trained network, you should configure the related setting according to checkpoint_dict in test_res256.py, for example

    "summer2winter": {  # the mode checkpoint for what (e.g., trained on which dataset) 
        "ckpt_path": "checkpoints/Res256_copyright_image_07-04-13-30/copyright_image_140000.pth",  # the checkpoint path
        "data_path": "F:/DataSource/StyleTransfer/summer2winter_yosemite/testA/",                   # the test image path
        "copyright_path": [  # the copyright image used during training
            'copyright_image/peking_university.png',
            'copyright_image/stanford_university.png',
            'copyright_image/Tsinghua.jpg',
            'copyright_image/ucla_university.png',
            'copyright_image/zhejiang_university.png',
            'copyright_image/UN.png',
        ]
    }

After that, you can run the following command to generate corresponding images

python test_res256.py

Run the above command will spawn the following folder:

Results
     --summer2winter
        --Test_clean      # the test image
        --Test_encoded    # the encoded image with copyright image
        --Test_decoded    # the decoded copyright image
        --Test_cover_residual
        --Test_secret_residual
        --Test_copyright_image  # the ground truth copyright image

Metric Calculating

To calculate the metric, you should configure the evaluation_directories in evaluate_metrics_ignite.py, then run the following command

python evaluate_metrics_ignite.py

Citation information

@article{WANG2024106391,
title = {An invisible, robust copyright protection method for DNN-generated content},
journal = {Neural Networks},
volume = {177},
pages = {106391},
year = {2024},
issn = {0893-6080},
doi = {https://doi.org/10.1016/j.neunet.2024.106391},
url = {https://www.sciencedirect.com/science/article/pii/S0893608024003150},
author = {Donghua Wang and Wen Yao and Tingsong Jiang and Weien Zhou and Lang Lin and Xiaoqian Chen},
}

winterwindwang / Invisible-Robust-Copyright-Protection