[ICLR 2025] "Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances" (Official Implementation)
[ICLR 2025] "Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances" (Official Implementation)
Official implementation of Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances
Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances
Shilin Lu, Zihan Zhou, Jiayou Lu, Yuanzhi Zhu, and Adams Wai-Kin Kong
ICLR 2025
Abstract:
Current image watermarking methods are vulnerable to advanced image editing techniques enabled by large-scale text-to-image models. These models can distort embedded watermarks during editing, posing significant challenges to copyright protection. In this work, we introduce W-Bench, the first comprehensive benchmark designed to evaluate the robustness of watermarking methods against a wide range of image editing techniques, including image regeneration, global editing, local editing, and image-to-video generation. Through extensive evaluations of eleven representative watermarking methods against prevalent editing techniques, we demonstrate that most methods fail to detect watermarks after such edits. To address this limitation, we propose VINE, a watermarking method that significantly enhances robustness against various image editing techniques while maintaining high image quality. Our approach involves two key innovations: (1) we analyze the frequency characteristics of image editing and identify that blurring distortions exhibit similar frequency properties, which allows us to use them as surrogate attacks during training to bolster watermark robustness; (2) we leverage a large-scale pretrained diffusion model SDXL-Turbo, adapting it for the watermarking task to achieve more imperceptible and robust watermark embedding. Experimental results show that our method achieves outstanding watermarking performance under various image editing techniques, outperforming existing methods in both image quality and robustness.
git clone https://github.com/Shilin-LU/VINE.git
cd VINE
conda create -n vine python=3.10
conda activate vine
pip install git+https://github.com/Shilin-LU/VINE
# To train the model, you also need to install the following package:
conda activate vine
cd vine/src/training_src/imagenet_c
pip install -e .Tip: Note that when editing images using MagicBrush and SVD, the environment should use the specific environments listed in their respective sections below. In all other cases, the vine environment is sufficient to run all code, including watermark encoding, decoding, regeneration, local editing, and other global editing tasks.
Our models, VINE-B and VINE-R, have been released on HuggingFace (VINE-B-Enc, VINE-B-Dec, VINE-R-Enc, VINE-R-Dec).
An interactive Colab demo is available. It walks you through watermark encoding, image editing, watermark decoding, and quality metrics calculation. This demo is a starting point to familiarize yourself with the workflow.
export TORCH_HOME=$(pwd) && export PYTHONPATH=$(pwd)
accelerate launch --num_processes=8 --main_process_port 17736 vine/src/train.py --enable_xformers_memory_efficient_attention --train_batch_size 14 --secret_loss_scale 1.5 --G_loss_scale 0.5 --l2_loss_scale 2.0 --lpips_loss_scale 1.5 --tracker_project_name pretraining --key_change pretraining --learning_rate 1e-4 --output_dir output/pretraining --fixed_inputaccelerate launch --num_processes=8 --main_process_port 17358 vine/src/finetune.py --enable_xformers_memory_efficient_attention --train_batch_size 20 --secret_loss_scale 1.5 --G_loss_scale 0.5 --l2_loss_scale 2.0 --lpips_loss_scale 1.5 --tracker_project_name finetuning --key_change finetuning --learning_rate 1e-4 --output_dir output/finetune --no_im_loss_steps 0 --fixed_input --imagenetc_step 500000000 --l2_loss_ramp 1 --l2_edge_ramp 1 --G_loss_ramp 1Embed a message into a single image:
python vine/src/watermark_encoding.py \
--pretrained_model_name Shilin-LU/VINE-R-Enc \
--input_path ./example/input/2.png \
--output_dir ./example/watermarked_img \
--message 'Hello World!'Watermark a list of images:
python vine/src/watermark_encoding_list.py \
--pretrained_model_name Shilin-LU/VINE-R-Enc \
--input_dir ./example/input \
--output_dir ./example/watermarked_img \
--message 'Your Secret'Watermark the entire W-Bench:
python vine/src/watermark_encoding_wbench.py \
--pretrained_model_name Shilin-LU/VINE-R-Enc \
--input_dir ./W-Bench \
--output_dir ./vine_encoded_wbench \
--message 'Your Secret'Apply image editing techniques to watermarked images. For example, using InstructPix2Pix, UltraEdit, or Image Inversion:
python vine/src/image_editing.py \
--model ultraedit \
--input_path ./example/watermarked_img/2_wm.png \
--output_dir ./example/edited_watermarked_img For batch processing and additional editing methods (e.g., InstructPix2Pix, MagicBrush), see the W-Bench section.
Extract the embedded message from a watermarked image:
python vine/src/watermark_decoding.py \
--pretrained_model_name Shilin-LU/VINE-R-Dec \
--input_path ./example/edited_watermarked_img/2_wm_edit.png \
--groundtruth_message 'Hello World!' Extract the embedded message from a list of images:
python vine/src/watermark_decoding_list.py \
--pretrained_model_name Shilin-LU/VINE-R-Dec \
--input_path ./example/watermarked_img \
--groundtruth_message 'Hello World!' Decode messages from the edited W-Bench:
python vine/src/watermark_decoding_wbench.py \
--pretrained_model_name Shilin-LU/VINE-R-Dec \
--groundtruth_message 'Your Secret' \
--unwm_images_dir ./W-Bench \
--wm_images_dir ./output/edited_wmed_wbench \
--output_dir ./output/detection_results/vine_r \
--transformation edit \
--decode_wbench_raw_data nEvaluate watermark imperceptibility (PSNR, SSIM, LPIPS):
python vine/src/quality_metrics.py \
--input_path ./example/input/2.png \
--wmed_input_path ./example/watermarked_img/2_wm.png To compute statistical decoding metrics over a batch of images, including [email protected]% FPR, TPR@1% FPR, and AUROC, refer to the W-Bench section. These metrics can be calculated using vine/src/watermark_decoding_wbench.py.
W-Bench is our comprehensive benchmark designed to evaluate the robustness of watermarking across four image editing techniques:
The benchmark comprises 10,000 samples sourced from datasets like COCO, Flickr, and ShareGPT4V.
The images of W-Bench have been released on HuggingFace. Below is a detailed guide on how to use all the image editing techniques listed in W-Bench.
cd VINE
huggingface-cli download Shilin-LU/W-Bench --repo-type=dataset --local-dir W-BenchEmbed watermarks into every image in W-Bench:
python vine/src/watermark_encoding_wbench.py \
--pretrained_model_name Shilin-LU/VINE-R-Enc \
--input_dir ./W-Bench \
--output_dir ./vine_encoded_wbench \
--message 'Your Secret'Optionally, run quality assessments on the encoded dataset:
python vine/src/quality_metrics_wbench.py \
--input_dir ./vine_encoded_wbench \
--wbench_path ./W-Benchpython vine/w_bench_utils/regeneration/stochastic_regeneration.py \
--wm_images_folder ./vine_encoded_wbench/512/STO_REGENERATION_1K \
--edited_output_folder ./output/edited_wmed_wbench/STO_REGENERATION_1Kpython vine/w_bench_utils/regeneration/stochastic_regeneration.py \
--wm_images_folder ./vine_encoded_wbench/512/DET_INVERSION_1K \
--edited_output_folder ./output/edited_wmed_wbench/DET_INVERSION_1Kpython vine/w_bench_utils/global_editing/global_editing_ultraedit.py \
--wm_images_folder ./vine_encoded_wbench/512/INSTRUCT_1K \
--edited_output_folder ./output/edited_wmed_wbench \
--editing_prompt_path ./W-Bench/INSTRUCT_1K/prompts.csv python vine/w_bench_utils/global_editing/global_editing_instructpix2pix.py \
--wm_images_folder ./vine_encoded_wbench/512/INSTRUCT_1K \
--edited_output_folder ./output/edited_wmed_wbench \
--editing_prompt_path ./W-Bench/INSTRUCT_1K/prompts.csv cd vine/w_bench_utils/global_editing
git clone https://gNo open issues yet, or sync has not completed.