zhengziqiang/CAIS

The implementation of "Component Aware Image Steganography via Adversarial Global-and-Part Checking"

Introduction

CAIS is a Tensorflow-based framework for training and testing of our paper: Component Aware Image Steganography via Adversarial Global-and-Part Checking.

News

2022.06.09 - Our paper is accepted by IEEE Transactions on Neural Networks and Learning Systems (TNNLS)

Installation

1. We use Anaconda3 as the basic environment. If you have installed the Anaconda3 in path Conda_Path, please create a new virtual environment by conda create -n tf114, then source activate tf114. Install tensorflow-gpu using the command conda install tensorflow-gpu==1.14.0.
2. Install the dependencies by pip install -r requirements.txt (if necessary). The requirements.txt file is provided in this package.
3. Please download the pre-trained VGG19 model imagenet-vgg-verydeep-19.mat (passwd:a8yv), then place it at the current path.

Data preparation

102Flowers (Within-domain)

Please download the original image files from this. Decompress this file and prepare the training and testing image files as follows:

mkdir datasets
cd datasets
mkdir flower
# The directory structure of flower should be this:
├──flower
   ├── train_cover
       ├── cover1.jpg 
       └── ...
   ├── train_message
       ├── message1.jpg 
       └── ...
   ├── test_cover
       ├── test_a.jpg (The test cover image that you want)
       └── ... 
   ├── test_message
       ├── test_b.jpg (The test message image that you want)
       └── ... 

We also provide a simple prepare.py file to randomly split the images. Please edit the img_path to specify the image path before running this file.

102Flowers and Caricature dataset (Cross-domain)

Please download the caricature image dataset. We follow the training/testing split of this dataset. And prepare the traing/testing images as follows:

mkdir flowercari
# The directory structure of flower should be this:
├──flowercari
   ├── train_cover
       ├── cover1.jpg (the same 7000 flower images)
       └── ...
   ├── train_message (the train caricature images)
       ├── message1.jpg 
       └── ...
   ├── test_cover
       ├── test_a.jpg (The rest flower images)
       └── ... 
   ├── test_message
       ├── test_b.jpg (The test caricature images)
       └── ... 

Train

sh scripts/train_flower.sh. You can also edit the default parameters referring the main.py.

Test

sh scripts/test_flower.sh. Generate the steganographic images by using random two images: one cover image and one message image.

Reconstruction

sh scripts/recon_flower.sh. Reconstruct the message images by using the steganographic images. Please specify the stegano_dir and recon_dir while running this procedure.

Losses

• Perceptual loss.
• LSGAN: Least Square GAN.

Image steganography description

The detailed CAIS model

Pre-trained model

• Flower: Google Drive; BaiduYun (1u8c)
• Flowercari: Google Drive; BaiduYun (ykc3)
• ImageNet: Google Drive; BaiduYun (9fto)

All the models are trained at the resolution 256*256. Please put them at the path ./check and unzip them.

TODO list

• Add the analysis tools.

Citation

If you find our work useful in your research, please consider citing:

@article{zheng2022composition,
  title={Composition-Aware Image Steganography Through Adversarial Self-Generated Supervision},
  author={Zheng, Ziqiang and Hu, Yuanmeng and Bin, Yi and Xu, Xing and Yang, Yang and Shen, Heng Tao},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2022},
  publisher={IEEE}
}

Acknowledgments

Code borrows from CycleGAN and DCGAN. The network architecture design is modified from DCGAN. The generative network is adopted from neural-style with Instance Normalization.