acceptodds
Under review as a conference paper at ICLR 2027

SwiftRegen: Efficient Image Watermark Evasion Beyond Output Imitation and Evidence Removal

Abstract

Robust invisible image watermarking relies on embedded evidence that remains detectable under post-processing and adversarial manipulation. Regeneration attacks can suppress this evidence, but iterative denoising incurs substantial computational cost. We present SwiftRegen, an efficient watermark evasion attack that replaces iterative regeneration with a single-pass operator learned offline from clean images and their regenerated counterparts, without watermark supervision, detector queries, or keys. Our study reveals two key insights. First, faithfully reproducing the regeneration output does not guarantee effective removal transfer: a full-image predictor achieves higher output fidelity yet weaker removal than a bounded residual predictor, so learning the regeneration-induced update transfers removal more effectively. Second, watermark evidence can survive while detection fails. For schemes that depend on correspondence between surviving evidence and the detector-side reference, disrupting this correspondence can defeat detection without removing the evidence. This motivates Watermark Carrier Desynchronization (WCD), which uses mild cropping and resizing to disrupt synchronization while largely preserving visual content. SwiftRegen combines the learned operator with WCD, and across eleven watermarking schemes achieves a mean TPR@1% FPR of 0.080 and a 72× speedup over CtrlRegen. Our results show that robust watermark detection depends on both evidence persistence and synchronization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
image watermarking, watermark removal, watermark evasion, regeneration attacks, invisible watermarks, Watermark Carrier Desynchronization, residual learning

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_zZzoOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents SwiftRegen, an efficient watermark evasion attack that replaces the original iterative regeneration process with a single-pass operator. Based on two key insights, it proposes a bounded residual predictor and Watermark Carrier Desynchronization (WCD), which employs mild cropping and resizing to disrupt synchronization while largely preserving visual content. By combining these two techniques, SwiftRegen achieves strong performance across eleven watermarking schemes.

Critical strengths

1. Effective, Efficient, and Generalizable Watermark Evasion. The paper proposes an effective and efficient watermark evasion method that demonstrates strong generalization across different watermarking schemes. The authors also conduct extensive experiments and provide comprehensive evaluations to validate the effectiveness of the proposed approach.

2. Insights into Watermark Robustness. The value of this work lies not only in proposing an effective attack method but also in revealing potential vulnerabilities in existing watermark detection systems, providing useful insights into their robustness and limitations.

Critical weaknesses

1. Insufficiently Justified Motivation. The primary motivation of this work is to reduce the computational overhead of CtrlRegen's iterative regeneration process. However, the authors do not sufficiently explain why inference efficiency is a critical bottleneck in existing watermark removal attacks or what practical attack scenarios would require efficient, large-scale watermark removal. As a result, the connection between the need for improved efficiency and the underlying threat model remains underdeveloped.


2. Lack of Real Computational Cost Analysis. SwiftRegen requires running CtrlRegen in advance to generate a large number of training image pairs, followed by training an additional model. Although this approach improves inference efficiency, it essentially shifts the computational overhead to offline data generation and model training. When accounting for these additional costs, at what scale of watermark removal can the upfront computational overhead be amortized? Is this trade-off computationally worthwhile in practice?


3. Limited Novelty Beyond Combining Existing Attacks. SwiftRegen consists of two relatively independent components, and the strong results in Table 1 may largely stem from combining two existing attack mechanisms. Existing studies suggest that certain content-agnostic watermarks (e.g., Tree-Ring) are vulnerable to geometric perturbations, such as rotation [1][2] and translation [3], while content-dependent watermarks (e.g., Stable Signature) are susceptible to reconstruction-based attacks [4][5]. Therefore, it is not surprising that a method like SwiftRegen (small U-Net CtrlRegen + WCD), which combines two attack mechanisms, achieves better performance than methods targeting only a single type of watermark. It would be beneficial to include additional ablation results in Table 1, similar to those presented in Table 2, such as SwiftRegen(Aθ\mathcal{A}_θ only) and SwiftRegen(Aθ\mathcal{A}_θ + C).


4. Figure Quality and Readability. The figures in the paper could benefit from further refinement, particularly in terms of formatting and attention to detail. For example, in Figure 4, some numerical labels overlap with the bars, while in Figure 5, the legend overlaps with the plotted lines.

Questions for authors
  1. As discussed in Q3, SwiftRegen combines two attack mechanisms. It is therefore reasonable that SwiftRegen achieves attack performance similar to that of the better-performing variant between SwiftRegen w/o WCD and WCD only. However, the results for Tree-Ring suggest an additional synergistic effect. Could the authors provide further analysis, supporting experiments, or a more detailed explanation for this observation?

  1. SwiftRegen heavily relies on CtrlRegen as a core component, despite the existence of various regeneration-based watermark removal attacks. Could the authors clarify the rationale behind choosing CtrlRegen? Have they explored using alternative regeneration-based attacks or combining different watermark removal methods to generate more diverse training data?

[1] Dongjun Hwang, Sungwon Woo, Tom Gao, Raymond Luo, and Sunghwan Baek. Invisible watermarks: Attacks and robustness. arXiv preprint arXiv:2412.12511, 2024.

[2] Cheng-Han Yeh, Kuan-chun Yu, Cheng-Chang Tsai, and Chun-Shien Lu. On the intrinsic limited robustness of latent-based watermarking. arXiv preprint arXiv:2610.08178, 2026.

[3] Xiaoxi Kong, Jieyu Yuan, Pengdi Chen, Yuanlin Zhang, Chongyi Li, and Bin Li. MarkCleaner: High-fidelity watermark removal via imperceptible micro-geometric perturbation. arXiv preprint arXiv:2602.01513, 2026.

[4] Xuandong Zhao, Kexun Zhang, Zihao Su, Saastha Vasan, Ilya Grishchenko, Christopher Kruegel,Giovanni Vigna, Yu-Xiang Wang, and Lei Li. Invisible image watermarks are provably removable using generative ai. In Advances in neural information processing systems, 2024.

[5] Yepeng Liu, Yiren Song, Hai Ci, Yu Zhang, Haofan Wang, Mike Zheng Shou, and Yuheng Bu. Image watermarks are removable using controllable regeneration from clean noise. In International Conference on Learning Representations, 2025.

Reviewer_3kowOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SwiftRegen achieves watermark removal through a single-step operator, chroma restoration, and cropping and resizing. The goal of the research is to achieve faster watermark removal. It presents two findings: bounded residual prediction performs better than full-image prediction for watermark removal, and watermark detection can fail even though the watermark may still remain.

Critical strengths

The idea of combining a single-step operator with cropping and resizing to make them complementary is quite interesting. In terms of reducing the time cost of multiple diffusion iterations alone, the reported 72× speedup is indeed quite good.

Critical weaknesses
  1. The paper finds and claims that bounded residual prediction performs better than full-image prediction for watermark removal, but it does not adequately explain why it performs better, and the relevant ablation experiments are insufficient.

  2. It is fine to point out that detection failure does not mean that the watermark evidence has been eliminated, but the paper does not demonstrate the practical implications of this finding.

  3. WCD seems to be just cropping and resizing, which are very common operations, so I feel it does not offer much novelty.

Questions for authors
  1. The conclusion that bounded residual prediction is better than full image prediction for watermark removal is only based on your specific scenario and needs more rigorous experimental proof and analysis.

  2. The paper points out that watermark detection failure doesn’t mean the watermark evidence has been erased. Besides experiments showing that detection performance is restored after reversing known geometric transformations, what else is this discovery useful for?

  3. The WCD proposed in the paper mainly relies on cropping and scaling. How is it different from the geometric transformations commonly used in previous work?

Reviewer_bxTjOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper considers the problem of blind watermark removal problem: making a watermarked image's watermark non-detectable without the knowledge of watermark algorithm and its corresponding detector.

The paper proposed a two-stage approach: a trained network (SwiftRegen) to purge the watermark pattern and Watermark Carrier Desynchronization (WCD) to disturb the localization of the watermark pattern in the image.

The authors claimed three contributions:

  1. This paper shows that the watermark can be effectively removed using a single feed-forward-pass method trained only from clean images.
  2. This paper shows that the watermark evasion methods do not necessarily require eliminating the embedded watermark evidence.
  3. The proposed method is superior than existing methods in watermark removal.
Critical strengths
  1. The paper addresses a practically relevant problem.
  2. The proposed method is technically reasonable.
Critical weaknesses

Writing and framing

I have concerns about several statements in the introduction and their consistency with existing methods:

  • The discussion of "iterative regeneration" is too broad. For example, the VAE-based methods included in the paper's comparisons do not require iterative regeneration. This limits the validity of the introduction's subsequent arguments to characterize regeneration methods to require iterations.

  • The statement that "successful evasion does not necessarily require eliminating the embedded evidence" is presented as challenging an existing assumption, but I do not see that assumption as generally established. For example, sufficiently strong Gaussian noise can cause a watermark decoder to fail without necessarily eliminating the underlying embedded evidence. Reference [1] discusses the distinction between "corruption methods" and "purification methods." The authors' framing may apply to purification methods, but it does not account for corruption methods.

Contributions and evaluation

  1. Novelty of single-pass regeneration without watermark information. The first claimed contribution states: "We show that the removal behavior of iterative regeneration can be learned from clean images and their regenerated counterparts and transferred to a single-pass operator without watermark supervision, detector queries, or keys." Existing VAE-based regeneration methods already achieve single-pass watermark evasion without access to watermark information. As currently framed, I therefore do not consider the general feasibility of single-pass evasion under these conditions to be a novel contribution. The novelty should instead be articulated in terms of the specific proposed method and how it differs from VAE-based approaches.

  2. Novelty of "Evidence Survives, Detection Fails." I do not consider this general observation to establish a novel contribution, given the distinction between corruption and purification methods discussed above. The authors should clarify what their findings add beyond the existing understanding that detection can fail even when watermark evidence remains.

  3. Fairness of the evaluation supporting "Effective and Efficient Watermark Evasion." I have concerns about whether the reported evaluation supports a fair comparison across evasion methods. As discussed in [1] and [2], the decoder's detection threshold, the visual quality of the attacked images, and evasion effectiveness are three dimensions that must be considered together when evaluating watermark robustness or attack performance. Two reasonable evaluation approaches are:

  • to fix two dimensions and compare the third, as in Table 2 of [1],

  • or to fix one dimension and plot the trade-off between the other two, as in Figure 3 of [2].

In this paper, the authors fix the detector setting but report evasion effectiveness (Table 1) and image quality (Table 2) separately. In my view, this presentation is insufficient to establish a fair comparison across methods because it does not show their relative evasion effectiveness at comparable image quality.

  1. Missing comparison with VAE + WCD. I strongly recommend including a comparison between the proposed SwiftRegen + WCD pipeline and VAE + WCD, using the VAE-based regeneration method from [3]. This comparison would help clarify whether both SwiftRegen and WCD make substantial contributions to watermark evasion, or whether WCD is the main source of the gains and SwiftRegen offers little additional benefit over VAE.

References

[1] Liang, Hengyue, Taihui Li, and Ju Sun. "A baseline method for removing invisible image watermarks using deep image prior." Transactions on machine learning research 2025 (2025): 4248.

[2] An, Bang, et al. "Waves: Benchmarking the robustness of image watermarks." arXiv preprint arXiv:2401.08573 (2024).

[3] Zhao, Xuandong, et al. "Invisible image watermarks are provably removable using generative ai." Advances in neural information processing systems 37 (2024): 8643-8672.

Questions for authors

Is there any misunderstanding / lack of knowledge you can see that can result to my assessment of the weakness above?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.