This is the repository for the NeurIPS 2020 paper titled UDH: Universal Deep Hiding for Steganography, Watermarking, and Light Field Messaging.

Neural networks have been shown effective in deep steganography for hiding a full image in another. However, the reason for its success remains not fully clear. Under the existing cover (C) dependent deep hiding (DDH) pipeline, it is challenging to analyze how the secret (S) image is encoded since the encoded message cannot be analyzed independently. We propose a novel universal deep hiding (UDH) meta-architecture to disentangle the encoding of S from C. We perform extensive analysis and demonstrate that the success of deep steganography can be attributed to a frequency discrepancy between C and the encoded secret image. Despite S being hidden in a cover-agnostic manner, strikingly, UDH achieves a performance comparable to the existing DDH. Beyond hiding one image, we push the limits of deep steganography. Exploiting its property of being universal, we propose universal watermarking as a timely solution to address the concern of the exponentially increasing amount of images/videos. UDH is robust to a pixel intensity shift on the container image, which makes it suitable for challenging application of light field messaging (LFM). This is the first work demonstrating the success of (DNN-based) hiding a full image for watermarking and LFM.

We performed our experiments with PyTorch v.0.4.1 for the main UDH, DDH, and watermarking, and PyTorch v.1.0.0 as well as torchgeometry v.0.1.2 for LFM.

1. Follow the common setup to make ImageNet compatible with pytorch as described in here.
2. Set the path to the pytorch ImageNet dataset folder in the main file.

To train the main UDH model, run the script bash ./scripts/train_main_udh.sh. To train the main DDH model, run the script bash ./scripts/train_main_ddh.sh. To train the UDH model for watermarking, run the script bash ./scripts/train_watermarking.sh. To train the UDH model for LFM, run the script bash ./scripts/train_lfm.sh.

Weights and other details of the training instances are saved into the ./training/ folder.

To get the main qualitative and quantitative results, run the script bash ./scripts/test_main.sh. To get the results for watermarking, run the script bash ./scripts/test_watermarking.sh.

For test scripts, training folders should be specified with the argument --test. For DDH training folder in the test_main.sh script, the argument is --test_diff.

Utility code is provided in the ./util/ folder for LFM processing: align_images.py and cr2png.py.

Pre-trained weights for different models can be accessed here. Download the folder training and put it inside the project folder. Now the test scripts can be run without training.

Update 21.06.21: LFM, hiding 6 secret images in 3 cover images, hiding 2 colored images in 1 grayscale image, and hiding with multiple encoders and decoders weights are availabe here. These weights were retrained for your convenience (not original weights used in the paper).

Code for the pseudo-differentiable JPEG used for watermarking is available here.

This code is strictly for non-commercial academic use only.

@inproceedings{zhang2020udh,
  title={UDH: Universal Deep Hiding for Steganography, Watermarking, and Light Field Messaging},
  author={Zhang, Chaoning and Benz, Philipp and Karjauv, Adil and Sun, Geng and Kweon, In-So},
  booktitle={34th Conference on Neural Information Processing Systems, NeurIPS 2020},
  year={2020},
  organization={Conference on Neural Information Processing Systems}
}