How zhenglisec/blind-watermark-for-dnn works
Project type and purpose
• This is an open-source machine-learning research framework for embedding and verifying a blind watermark in deep neural networks (DNNs).
• It uses PyTorch (Python 3.7) to train a pair of networks (a HidingUNet and a Discriminator) alongside a host model (LeNet / VGG / ResNet variants / etc.) on MNIST or CIFAR-10.
File structure
• Top-level:
• main.py – orchestrates training and evaluation
• utils.py – common helper functions (argument parsing, logging, checkpointing)
• data/ – raw datasets and watermark logo
• models/ – definitions of all network architectures (Host models, HidingUNet, Discriminator)
• results/ – per-model subfolders containing generated images, trained checkpoints, and run scripts
• Patterns:
• Modular model definitions in models/ for easy swapping of architectures
• Clear separation between data ingestion (data/), model code (models/), orchestration (main.py), and outputs (results/)
README for insights
• Framework supports two datasets (mnist, cifar10)
• Hosts a set of pre-defined network backbones
• Contains instructions on setting hyperparameters, seeds, and where outputs land
Main components and their interactions
a. Data Ingestion
– Dataset Loader: MNIST or CIFAR-10 reader (auto-downloads to data/)
– Logo Loader: samples a watermark image (mnist sample or data/IEEE/logo)
b. Host Model
– One of the classification networks (LeNet1/3/5, VGG11/13/16/19, ResNet18/34/101, PreActResNet18/34, GoogLeNet, DPN26, MobileNetV2)
c. Embedding Network (HidingUNet)
– Takes host model’s feature maps (or input image) + logo → produces a watermarked image
d. Discriminator
– Attempts to extract/recover the embedded logo from the watermarked image
e. Training Orchestrator (main.py)
– Initializes models, optimizers, and losses
– In each training loop:
• Forward pass: host model + HidingUNet + Discriminator
• Compute classification loss (host task), reconstruction loss (logo), adversarial loss
• Backpropagate and update HidingUNet & Discriminator (and sometimes host)
– Checkpoints final watermarked host model and networks into results/
f. Utilities
– Argument parser, logging, checkpoint save/load, visualization utilities (utils.py)
Architectural patterns and design principles
• Pipeline pattern: Data → Preprocessing → Embedding → Discrimination → Loss → Update
• Modular model registry: models are interchangeable, easy to extend with new backbones
• Separation of concerns:
• Data & configuration (main/data/)
• Model code (models/)
• Experiment orchestration (main.py + utils.py)
• Results & artifacts (results/)
Detailed element checklist
• Version note (PyTorch, Python)
• Directory mapping (e.g., show that model definitions live in models/, outputs in results/)
• Interaction notes (e.g., “HidingUNet consumes feature maps from Host Model”)