[ICLR 2026] THE SELF-RE-WATERMARKING TRAP: FROM EXPLOIT TO RESILIENCE

Official implementation of THE SELF-RE-WATERMARKING TRAP: FROM EXPLOIT TO RESILIENCE.

---

Table of Contents

• Requirements
• Setup & Run
• Usage
• Configuration
• Citation

---

Requirements

• Python 3.8+
• Conda (recommended)
• CUDA 11.8 if using GPU

All dependencies are captured in environment.yml for reproducibility.

---

Setup & Run

We provide a bash script that handles environment creation, checkpoint download, and running training/testing.

1. Clone the repository:

git clone --recurse-submodules https://github.com/SVithurabiman/SRW.git

2. Run the setup script:

bash run_all.sh [train|test]

This script will:

• Create a Conda environment from environment.yml

• Activate the environment

• Download pretrained checkpoints from HuggingFace into results/

• Run training if train is passed as an argument. After running, outputs and logs will be saved in results/.

• Run testing/evaluation if test is passed as an argument

Usage

If you prefer to run manually:

Training

    
python -m SRW.train --dataset_config ./SRW/configs/dataset.yaml --exp_config ./SRW/configs/train.yaml

Visualizing Training

This repository uses TensorBoard to log training progress.

Launch TensorBoard:

tensorboard --logdir results --port 6006

Then open in your browser:

http://localhost:6006

Evaluation / Testing

python -m SRW.test --dataset_config ./SRW/configs/dataset.yaml --exp_config ./SRW/configs/test.yaml

Configuration

All experiment settings are in the configs/ folder:

• configs/train.yaml – default training settings

• configs/test.yaml – default test settings

• configs/dataset.ymal - default Image and Message settings

Modify paths, hyperparameters, or model options in these YAML files as needed.

Citation

If you find this work useuful, please cite:

@inproceedings{senthuran2026selfrewatermarking,
  title={The Self‑Re‑Watermarking Trap: From Exploit to Resilience},
  author={Vithurabiman Senthuran and Yong Xiang and Iynkaran Natgunanathan and Uthayasanker       
          Thayasivam},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026},
  url={https://openreview.net/forum?id=st1hrLTP14}
}