Can Simple Averaging Defeat Modern Watermarks?

Authors: Pei Yang, Hai Ci, Yiren Song, Mike Zheng Shou

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our quantitative and qualitative evaluations across twelve watermarking methods highlight the threat posed by steganalysis to content-agnostic watermarks and the importance of designing watermarking techniques resilient to such analytical attacks.
Researcher Affiliation Academia Pei Yang Hai Ci Yiren Song Mike Zheng Shou Show Lab, National University of Singapore yangpei@u.nus.edu, cihai03@gmail.com yiren@nus.edu.sg, mike.zheng.shou@gmail.com
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Github page: https://github.com/showlab/watermark-steganalysis.
Open Datasets Yes For the greybox setting, we use the COCO2017 [39] validation set for Wm Adapter [27] and Stable Signature [22], Stable Diffusion Prompts [40] for Tree-Ring [14] and Ring ID [16] prompts, and Diffusion DB [41] for the remaining methods as the non-watermarked images (x ). In the blackbox setting with no access to paired images, we substitute x with Image Net [42] test set.
Dataset Splits Yes For the greybox setting, we use the COCO2017 [39] validation set for Wm Adapter [27] and Stable Signature [22]... The selection of images within the datasets is random.
Hardware Specification Yes The experiments were conducted on an AMD EPYC 7413 24-Core Processor and an Nvidia RTX 3090 GPU, requiring around 200GB of disk space.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies used in the experiments.
Experiment Setup Yes The datasets are resized to 256 256 for Ro Ste ALS, SSL, and Hi DDe N, and 512 512 for other methods. We assess watermark removal under different n (number of images averaged) during watermark pattern extraction, and test on 100 images3 during watermark removal.