• Clone repo and install submodules

git clone --recurse-submodules https://github.com/tancik/StegaStamp.git
cd StegaStamp

• Install tensorflow (tested with tf 1.13)
• Download dependencies

pip install -r requirements.txt

• Set dataset path in train.py

TRAIN_PATH = DIR_OF_DATASET_IMAGES

• Train model

bash scripts/base.sh EXP_NAME

The training is performed in train.py. There are a number of hyperparameters, many corresponding to the augmentation parameters. scripts/bash.sh provides a good starting place.

Run the following in the base directory to download the trained network used in paper:

wget https://people.eecs.berkeley.edu/~tancik/stegastamp/saved_models.tar.xz
tar -xJf saved_models.tar.xz
rm saved_models.tar.xz

The training code for the detector model (used to segment StegaStamps) is not included in this repo. The model used in the paper was trained using the BiSeNet model released here. CROP_WIDTH and CROP_HEIGHT were set to 1024, all other parameters were set to the default. The dataset was generated by randomly placing warped StegaStamps onto larger images.

The exported detector model can be downloaded with the following command:

wget https://people.eecs.berkeley.edu/~tancik/stegastamp/detector_models.tar.xz
tar -xJf detector_models.tar.xz
rm detector_models.tar.xz

To visualize the training run the following command and navigate to https://localhost:6006 in your browser.

tensorboard --logdir logs

The script encode_image.py can be used to encode a message into an image or a directory of images. The default model expects a utf-8 encoded secret that is <= 7 characters (100 bit message -> 56 bits after ECC).

Encode a message into an image:

python encode_image.py \
  saved_models/stegastamp_pretrained \
  --image test_im.png  \
  --save_dir out/ \
  --secret Hello

This will save both the StegaStamp and the residual that was applied to the original image.

The script decode_image.py can be used to decode a message from a StegaStamp.

Example usage:

python decode_image.py \
  saved_models/stegastamp_pretrained \
  --image out/test_hidden.png

The script detector.py can be used to detect and decode StegaStamps in an image. This is useful in cases where there are multiple StegaStamps are present or the StegaStamp does not fill the frame of the image.

To use the detector, make sure to download the detector model as described in the installation section.

python detector.py \
  --detector_model detector_models/stegastamp_detector \
  --decoder_model saved_models/stegastamp_pretrained \
  --video test_vid.mp4

Add the --save_video FILENAME flag to save out the results.