GaussMarker: Robust Dual-Domain Watermarks for Diffusion Models

1. Contents

• GaussMarker: Robust Dual-Domain Watermarks for Diffusion Models
  • 1. Contents
  • 2. Introduction
  • 3. Get Start

2. Introduction

This is the official implementation of paper GaussMarker: Robust Dual-Domain Watermarks for Diffusion Models.

As Diffusion Models (DM) generate increasingly realistic images, related issues such as copyright and misuse have become a growing concern. Watermarking is one of the promising solutions. Existing methods inject the watermark into the single-domain of initial Gaussian noise for generation, which suffers from unsatisfactory robustness. This paper presents the first dual-domain DM watermarking approach using a pipelined injector to consistently embed watermarks in both the spatial and frequency domains. To further boost robustness against certain image manipulations and advanced attacks, we introduce a model-independent learnable Gaussian Noise Restorer (GNR) to refine Gaussian noise extracted from manipulated images and enhance detection robustness by integrating the detection scores of both watermarks. GaussMarker efficiently achieves state-of-the-art performance under eight image distortions and four advanced attacks across three versions of Stable Diffusion with better recall and lower false positive rates, as preferred in real applications.

3. Get Start

3.1 Installation

To setup the environment of GaussMarker, we use conda to manage our dependencies.

Run the following commands to install GaussMarker:

conda create -n gaussmarker python=3.9 -y && conda activate gaussmarker
pip install -r requirements.txt

3.2 Generating Watermarked Images

Run:

python gaussmarker_gen.py --chacha --num 10 --output_path './gen_10'  --w1_path w1_256.pth

The multi-bit and zero-bit watermarks will be saved into ./w1_256.pth and ./w2.pth, respectively. The watermarked and un-watermarked images will be saved into ./gen_10/w_img/ and ./gen_10/no_w_img/, respectively.

3.3 Training GNR

Run:

python train_GNR.py --train_steps 50000 --r 180 --s_min 1.0 --s_max 1.2 --fp 0.35 --neg_p 0.5 --model_nf 128 --batch_size 32 --num_workers 16 -ed 256bits --w_info_path w1_256.pth

After training, the model weight will be saved as ./GNR_256bits/model_final.pth.

3.3 Evaluation

Extract the watermark without any distortion:

python gaussmarker_det.py --chacha --model_nf 128 --GNR_path './GNR_bits256/model_final.pth' --input_path './gen_10'  --fpr 0.01 --num 10  --output_path './results/Clean' --reference_model ViT-g-14 --reference_model_pretrain laion2b_s12b_b42k --w1_path './w1_256.pth'

Extract the watermark with Rotation 75:

python gaussmarker_det.py --chacha --model_nf 128 --GNR_path './GNR_bits256/model_final.pth' --input_path './gen_10'  --fpr 0.01 --num 10  --output_path './results/Rotate75' --affine 1 --affine_params 75,0,1,0 --reference_model ViT-g-14 --reference_model_pretrain laion2b_s12b_b42k --w1_path './w1_256.pth'