This repo is the official code for

• DeepMIH: Deep Invertible Network for Multiple Image Hiding    [论文速览] [VALSE] [Bilibili]
  • Zhenyu Guan¹, Junpeng Jing¹(co-first), Xin Deng, Mai Xu, Lai Jiang, Zhou Zhang, Yipeng Li.

Published on IEEE Transactions of Pattern Analysis and Machine Intelligence (TPAMI 2022). @ Beihang University.

• Python 3 (Recommend to use Anaconda).
• PyTorch = 1.0.1 .
• See environment.yml for other dependencies.

• In this paper, we use the commonly used dataset DIV2K, COCO, and ImageNet.
• For train or test on your own path, change the code in config.py:
  line50: TRAIN_PATH_DIV2K = ''
line51: VAL_PATH_DIV2K = ''
line54: VAL_PATH_COCO = ''
line55: TEST_PATH_COCO = ''
line57: VAL_PATH_IMAGENET = ''

1. Here we provide a trained model.
2. Download and update the MODEL_PATH and the file name suffix before testing by the trained model.
   For example, if the model name is model_checkpoint_03000_1.pt, model_checkpoint_03000_2.pt, model_checkpoint_03000_3.pt,
   and its path is /home/usrname/DeepMIH/model/,
   set:
   PRETRAIN_PATH = '/home/usrname/DeepMIH/model/',
   PRETRAIN_PATH_3 = '/home/usrname/DeepMIH/model/',
   file name suffix = 'model_checkpoint_03000'.
3. Check the dataset path is correct.
4. Create an image path to save the generated images. Update TEST_PATH.
5. Run test_oldversion.py.

1. Create a path to save the trained models and update MODEL_PATH.
2. Check the optim parameters in config.py is correct. Make sure the sub-model(net1, net2, net3...) you want to train is correct.
3. Run train_old_version.py. Following the Algorithm 1 to train the model.
4. Note: DeepMIH may be hard to train. The model may suffer from explosion. Our solution is to stop the training process at a normal node and abate the learning rate. Then, continue to train the model.

In the train_old_version.py at line 223:
rev_secret_dwt_2 = rev_dwt_2.narrow(1, 4 * c.channels_in, 4 * c.channels_in) # channels = 12,
the recovered secret image_2 is obtained by spliting the middle 12 channels of the varible rev_dwt_2. However, in the forward process_2, the input is obtained by concatenating (stego, imp, secret_2) together. This means that the original code train_old_version.py has a bug on recovery process (the last 12 channels of the varible rev_dwt_2 should be splited to be the recovered secret image_2, instead of the middle 12 one). We found that in this way the network is still able to converge, thus we keep this setting in the test process.
We also offer a corrected version train.py (see line 225) and test.py. You can also train your own model in this way.

Feel free to contact: [email protected].

If you find this repository helpful, you may cite:

@ARTICLE{9676416,
  author={Guan, Zhenyu and Jing, Junpeng and Deng, Xin and Xu, Mai and Jiang, Lai and Zhang, Zhou and Li, Yipeng},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={DeepMIH: Deep Invertible Network for Multiple Image Hiding}, 
  year={2022},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TPAMI.2022.3141725}}