END^2: Robust Dual-Decoder Watermarking Framework Against Non-Differentiable Distortions

Overview

END^2 is a deep learning-based digital watermarking framework that enables end-to-end training against arbitrary non-differentiable distortions. It leverages self-supervised contrastive learning with a teacher-student architecture to achieve robust watermarking that can withstand various types of real-world noise and transformations.

The system uses a specialized encoder for message embedding and dual decoders (teacher and student) for extraction. The teacher network guides the student network through contrastive learning to handle non-differentiable distortions, enabling the framework to learn from operations where gradients cannot be directly computed.

Usage

To train the model:

python main.py

Configuration

It can be configured through the cfg.yaml file with parameters such as:

• Message length
• Image size
• Loss weights
• Training hyperparameters

Project Structure

• main.py: Entry point for training
• model.py: Contains the END^2 model architecture
• trainer.py: Implementation of the training framework
• T1Data.py: Dataset loading and preprocessing
• cfg.yaml: Configuration parameters