acceptodds
Under review as a conference paper at ICLR 2027

ProxyEraseAgent: Blind Watermark Removal in the Wild

Abstract

Invisible image watermark removal has received growing attention. Despite sub- stantial progress, existing attacks still face a tension between practicality and at- tack specificity. Attacks that exploit detector outputs, decoder responses, or paired watermarked and clean images can be tailored to the watermark decision bound- ary, but they require information that is rarely available in realistic dissemination scenarios. In contrast, attacks based on compression, geometric distortion, or image reconstruction are easy to deploy from a single watermarked image, but they are largely open-loop: they apply generic transformations without knowing whether the current image is moving toward watermark failure. Thus, the key challenge in single-image blind watermark removal is not merely how to trans- form the image, but how to obtain a useful direction for removal without access to the hidden decoder. To bridge this gap, we propose ProxyEraseAgent, an agent-driven framework that recovers attack specificity through proxy decoder responses. Publicly available watermarking schemes provide a natural external knowledge base of candidate decoders, where some are informative for a given unknown watermarked image. Our insight is that a decoder producing a stronger calibrated response to the query image is more likely to share a nearby decoding boundary with the hidden target mechanism. ProxyEraseAgent ranks knowledge- base decoders by calibrated response strength and uses the top ones as proxy boundary estimators. Their responses then guide a progressive search over hetero- geneous removal operations, including geometric distortion, JPEG compression, image reconstruction, and gradient-based perturbation, under perceptual-quality constraints. Experiments across 11 image watermarking systems show that Proxy- EraseAgent achieves an attack success rate of 94.8%, demonstrating the effective- ness of response-guided proxy retrieval and feedback-driven sequential planning for blind watermark removal

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Watermark Removal, Agent Attack, Proxy Decoders

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_gkMpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ProxyEraseAgent, a blind watermark-removal framework that removes invisible image watermarks without accessing the target decoder or knowing the underlying watermarking algorithm. The framework retrieves informative proxy decoders from a knowledge base of publicly available watermarking models and uses their responses to guide adaptive watermark-removal attacks through sequential search. Experiments across 11 watermarking systems report an overall constrained attack success rate (cASR) of 94.8%, outperforming existing baselines.

Critical strengths
  • Interesting and practical attack setting. The paper investigates blind watermark removal, where attackers have access only to watermarked images without querying the target decoder. This is a relevant security problem for real-world watermarking systems.

  • Interesting proxy-guided attack framework. ProxyEraseAgent combines response-guided proxy decoder retrieval with sequential attack planning, enabling adaptive composition of different watermark-removal operations without direct feedback from the target decoder.

Critical weaknesses
  • Insufficient validation of the proxy decoder selection assumption. The proposed response strength is derived from the prediction certainty of watermark bits. However, a decoder may produce highly confident bit predictions even for images watermarked by unrelated algorithms. Such confidence only reflects the decoder's own prediction certainty and does not necessarily indicate similar decision boundaries or transferable responses to perturbations. Although entropy-based selection outperforms random selection, the paper lacks a direct analysis of the correlation between calibrated response strength and actual attack transferability.

  • Insufficient comparison with strong watermark-removal baselines. The paper compares ProxyEraseAgent with several transformation-based and dedicated watermark-removal attacks, but omits relevant methods such as A Transfer Attack to Image Watermarks[1] and RAVEN[2], which are discussed in the related work. Comparisons with these methods under consistent threat models would help establish the effectiveness of the proposed approach.

  • Questionable attack success criterion. The paper defines a successful attack as achieving BER ≥ 0.1 under LPIPS ≤ 0.05. However, a 10% bit error rate does not necessarily indicate successful watermark removal, particularly for watermarking systems with error-correction capabilities or different detection thresholds. The authors should report actual watermark detection failure rates and verify that the original watermarks are correctly recoverable before applying attacks.

  • Potentially unfair optimization and search budgets. ProxyEraseAgent benefits from multi-step Pareto beam search, while the evaluated baselines do not necessarily receive comparable optimization budgets. The ablation study shows that increasing the search depth from one to four improves cASR from 83.91% to 94.82%. Comparisons under matched GPU time or optimization budgets are needed to determine whether the improvements arise from the proposed proxy guidance or additional computation.

  • Limited evaluation scale and image resolution. The experiments use only 100 MS-COCO images, with all images resized to 128×128. Some watermarking systems also require higher internal resolutions, with their outputs subsequently resized to 128×128, potentially affecting watermark integrity. Evaluations on larger datasets and higher-resolution images would strengthen the practical applicability claims.

[1] Hu, Yuepeng, et al. "A transfer attack to image watermarks." International Conference on Learning Representations. Vol. 2025. 2025.

[2] Shamshad, Fahad, Nils Lukas, and Karthik Nandakumar. "RAVEN: Erasing Invisible Watermarks via Novel View Synthesis." CVPR 2026.

Questions for authors
  • Proxy decoder selection. Could the authors evaluate the correlation between calibrated response strength and actual attack transferability to validate the assumption behind proxy decoder selection?

  • Baseline comparison. Could the authors include stronger watermark-removal baselines, particularly A Transfer Attack to Image Watermarks[1] and RAVEN[2], under comparable threat models?

  • Attack success criterion. Why is BER ≥ 0.1 considered sufficient for successful watermark removal? Could the authors report actual watermark detection failure rates and initial watermark recovery accuracy?

  • Optimization fairness. How does ProxyEraseAgent compare with existing baselines under identical GPU time or optimization budgets? Would additional optimization or multi-step search improve the performance of the baselines?

  • Evaluation scale and resolution. How does ProxyEraseAgent perform on larger datasets and higher-resolution images? Does resizing watermarked images to 128×128 affect their initial watermark integrity and the reported attack success rates?

[1] Hu, Yuepeng, et al. "A transfer attack to image watermarks." International Conference on Learning Representations. Vol. 2025. 2025.

[2] Shamshad, Fahad, Nils Lukas, and Karthik Nandakumar. "RAVEN: Erasing Invisible Watermarks via Novel View Synthesis." CVPR 2026.

Reviewer_vnbjOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents ProxyEraseAgent, a framework that combines existing watermarking removal attacks in a no-box setting using public decoders from various watermarking techniques as surrogates. Using these surrogates, ProxyEraseAgent searches over possible combinations until it finds one that meets the image quality and surrogate-removal signal requirements.

Critical strengths

The core idea is interesting and has merit.

Critical weaknesses
  1. The paper mentions another paper on watermark removal that seems to be a direct competitor to their work [1]. Yet, you do not evaluate it or compare it against their method. Also, something seems to have gone wrong with the citations in that section, as they are missing their wrapping brackets. Additionally, you compare mainly against attacks whose threat model is not originally the no-box model, which makes the comparison unfair.
  2. The self-watermarked reference method seems fundamentally flawed as it is presented. Using only a watermarked sample, without calibrating the metric with unwatermarked images or watermarked images from other watermarking methods, severely limits its ability to discard false positives.
  3. Furthermore, for the self-watermarked reference, the mathematical definition of SmS_m does not align with the text description. You claim to use a reference image, yet the math uses xwx^w rather than a separate reference image.
  4. SmS_m is never used after it is defined. It serves no purpose in your method.
  5. The equation for bmb_m adds no information beyond the preceding sentence; you should use it to define BER instead, mathematically. bmb_m is also never used again and seems to serve no purpose in your method.
  6. B(x)B(x) is not defined anywhere, yet it is integral to your method. I assume, following the flow of the paper, that it is based on bmb_m and SmS_m, but it is not specified anywhere. This is a critical flaw.
  7. The dataset is far too small (100 images) to produce meaningful results. This needs to be significantly increased. The experimental results have little to no significance with such a small sample size. Also, varying the provenance of the images used could help establish the generality of your results.
  8. You report no seed-variation experiments and provide no standard deviation or confidence intervals.
  9. A bit error rate of only 10% seems way too small to call an attack successful. Most watermarking methods could still decode a message while limiting false positives with 90% correct bits.
  10. Table 1 has one too many significant digits; since you use only 100 images, the “.0” across the entire table is useless.
  11. For Table 2, how are you getting four significant digits if you only use 100 images? Even if you average over all defenses, that would still only mean three significant digits.
  12. Figure 4 lacks clarity and description of what is being measured, how, and on what data/inputs.
  13. Appendix 6.2 and 6.4 could be merged.
  14. The text in Figures 2 and 4 is too small to read easily.
  15. Please add equation numbers.

[1] Hu, Yuepeng, et al. "A transfer attack to image watermarks." International Conference on Learning Representations. Vol. 2025. 2025.

Questions for authors
  1. For the self-watermarked reference, wouldn’t you want SmS_m for non-watermarked images so you can use that as a stopping signal that the watermark is gone?
  2. Have you tried using K=11? Also, an experiment plotting K vs. cASR or BER would provide an interesting insight into the hyperparameter choice.
Reviewer_GY4DOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes ProxyEraseAgent, a framework for removing invisible image watermarks when the attacker has only a watermarked image and cannot access the target decoder. The problem is practically relevant, and the method achieves promising results on 11 image watermarking systems.

Critical strengths
  1. Removing watermarks without querying the target decoder addresses a meaningful practical scenario.
  2. Proxy retrieval and sequential attack composition provide a clear mechanism for adapting attacks to individual images.
Critical weaknesses
  1. The main potential contribution is retrieving public proxy decoders through calibrated responses and using their feedback to guide attack composition. The framework is primarily a proxy-guided and constrained attack sequence selection system, with relatively limited methodology novelty. The “Agent” terminology does not itself correspond to an additional learning or reasoning mechanism.

  2. Initial watermark validity is unclear. All images are resized to 128x128, including downsampling after embedding for some systems. Pre-attack decoding accuracy and detection rates are needed to distinguish attack effects from preprocessing damage.

  3. Evaluation uses only 100 images from one dataset, without sufficient evidence across resolutions, scalability, random seeds, or runtime.

  4. Assertion such as “in the wild” and “prior-free” extend beyond the demonstrated evaluation setting, which needs more detailed qualification.

  5. Fig. 2 (a) contains two HiDDeN, typo?

Questions for authors

Can ProxyEraseAgent erase invisible watermark embedded in NeRF/3DGS?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.