DocSafe

Toward Practical Print-Proof Image Steganography via Frequency Decomposition and Covariance Alignment

🌐 Project Website · 📄 Paper (IEEE Xplore) · 📦 PyPI

Deep-learning-based watermark encoder/decoder for images and documents, with a full training pipeline (M1/M2/M3 model variants).

Figure 7. Representative samples of encoded images without spatial masking, where the message is uniformly embedded across the entire image. Model M-1 produces high-quality encoded images, but its decoder performs well only on the face image dataset. Models M-2 and M-3 generate encoded images with lower visual quality than M-1, but demonstrate improved and acceptable decoder performance on the object image dataset.

Install

pip install DocSafe

Or from source:

git clone https://github.com/farhadsh1992/DocSafe.git
cd DocSafe
pip install .

Pretrained weights

Model weights are not bundled in the PyPI package (they're hundreds of MB, well past PyPI's practical size limits). Download pre_trained_models/ separately and point path_model= at it — see the release/host you're distributing weights from.

Quickstart: encode / decode a watermark

from DocSafe import encoder, decoder

encoder_router = encoder(model="M1", path_model="pre_trained_models/", secret_size=100)
encoder_router.load_network(device="cpu")

images = encoder_router.read_image(path=["test_images/original_images.jpg"])
image_batch = encoder_router.preprocess_images(images)
encoded = encoder_router(original_images=image_batch, messages="viste", mask=None)
encoder_router.save_encoded_image("encoded.png")

decoder_router = decoder(model="M1", path_model="pre_trained_models/", secret_size=100)
decoder_router.load_network(device="cpu")

encoded_images = decoder_router.read_image(path=["encoded.png"])
encoded_batch = decoder_router.preprocess_images(encoded_images)
messages = decoder_router(encoded_images=encoded_batch, mask=None)
print(messages)

See test_main.py for a complete runnable example.

Training

train_main.py shows the intended DocSafe.Trainer API. Two things to know before running it:

• configs/paths_config.py hardcodes paths from the original training machine (pretrained_models/..., dataset directories) — edit it for your own environment.
• Trainer.train() and Trainer.load_network() are still under active development (see the inline NOTE: comments in train_main.py).

Known limitations

• DocSafe.loss_functions_lib.models.stylegan2 (and everything that depends on it — hyperstyle, psp, e4e, restyle_e4e_encoders, w_encoder, the hypernetworks) JIT-compiles CUDA/C++ extensions on import via torch.utils.cpp_extension.load(). It requires an actual CUDA GPU + nvcc + ninja (pip install DocSafe[stylegan2]) and will not import on CPU-only machines.
• DocSafe.networks.net_StampOne is superseded/legacy — it isn't used by encoder/decoder/ Trainer (which all use DocSafe.networks.networks_M1/M2/M3) and has an unresolved Deformable_Conv2D dependency in one file.
• DocSafe.networks.DETR_NET.segmentation's panoptic-segmentation helpers need pip install DocSafe[detr-extras].

License

MIT — see LICENSE.