How cuihaoleo/cnn_watermark works
Project type and purpose
• This is an open-source research/tool repository that implements a CNN-based watermarking scheme in TensorFlow Keras, using the CIFAR-10 dataset.
• Purpose: train a watermark-aware CNN model (cnn_watermark.py), then evaluate watermark embedding, robustness (add_noise.py), and extraction (extract.py).
File structure
• Top level
• cnn_watermark.py – training pipeline that loads CIFAR-10, builds and trains the CNN, and serializes the model.
• eval/ – evaluation scripts that operate independently:
– embed.py: embeds a watermark into images using the trained model.
– add_noise.py: simulates attacks by adding noise to watermarked images.
– extract.py: extracts the watermark from noisy images.
• No frontend/backend split; this is a standalone Python workflow. No deployment scripts.
README for additional insights
• Confirms usage of TensorFlow Keras and CIFAR-10.
• Clear separation between “training” (cnn_watermark.py) and “evaluation” (eval/).
Detailed breakdown—don’t overthink
• Break the project into the six components above.
• Storage artifacts (dataset, model file, image files).
• Data dependencies: which files or outputs feed into which module.