acceptodds
Under review as a conference paper at ICLR 2027

Forging Image Watermarks by Reversing Watermark Removal Attacks

Abstract

Image generative models have accelerated the need for robust image watermarking to track and verify AI-generated images. While watermark removal attacks have been extensively studied, the threat of watermark forgery, where benign images are maliciously modified to appear watermarked, remains underexplored, especially in the no-box setting. In this work, we introduce WForge, a no-box and query-free forgery attack that reframes forgery as the inverse of removal. Our key insight is that residual perturbations from removal attacks approximate watermark signals and can be repurposed to forge watermarks. Concretely, we train a forger network to learn the pattern of residuals and apply it to unwatermarked images, making them falsely detected as watermarked. We evaluate WForge across three datasets and four state-of-the-art watermarking methods, demonstrating that it consistently outperforms existing forgery baselines. Our results further reveal a critical vulnerability: the existence of a successful removal attack implies the feasibility of forgery for the same watermarking method.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Image Watermark, Watermark Forgery, Watermark Removal, Imge Generative Model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_HFdCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces WForge, a no-box and query-free image watermark forgery attack. WForge treats forgery as the inverse of watermark removal, using removal residuals as supervision to train a network that predicts watermark-like perturbations for clean images. Experiments across multiple watermarking schemes and datasets show that WForge significantly outperforms existing forgery baselines. The paper also studies the relationship between watermark removal and forgery from a residual-based perspective.

Critical strengths
  1. This paper studies an important and relatively under-explored attack scenario. Most existing studies focus on watermark removal, while watermark forgery can also cause serious problems for traceability and the source system. WForge considers a strict non-black-box environment where the attacker cannot access the watermark detector or its API, making this attack scenario practically significant.
  2. The proposed method is simple and has good experimental results. The experimental results show that this method is indeed effective under the evaluated conditions, and it strongly supports the idea presented in the paper of extracting supervision from the residual information of the watermark removal. Extensive ablation studies have also verified the effectiveness of the proposed method.
Critical weaknesses

1.The assumptions underlying the WForge method have not been fully proven. This method assumes that the residuals generated after watermark removal can serve as an alternative indicator of the actual watermark signal. However, experiments have shown that different removal methods produce completely different residuals, leading to significantly different forgery performance. Therefore, it is unclear under what circumstances the removal residuals mainly capture information related to the watermark rather than the reconstruction artifacts caused by a specific removal algorithm. 2.Compared with the recent watermark forgery methods, the research scope is limited. This paper mainly compares WForge with Watermark Faker and Watermark Steganalysis. A closely related recent work, "WMCopier: Forging Invisible Watermarks on Arbitrary Images" (Dong et al., NeurIPS 2025), studies similar out-of-box forgery scenarios and evaluates overlapping watermark methods such as RivaGAN and Stable Signature. A comparison with WMCopier would help to more fully demonstrate the experimental advantages of WForge.

Questions for authors

The proposed method relies on the assumption that the residual produced by a watermark remover can serve as a good surrogate for the actual watermark residual. Could the authors provide a more principled justification for this assumption? In particular, why should R'(Iw)-Iw) mainly capture watermark-related information rather than remover-specific reconstruction artifacts, and how can one determine whether a given removal method will produce residuals suitable for forgery?

Reviewer_eb5cOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes WForge, a no-box and query-free attack that forges image watermarks. The attacker has only a set of watermarked images, without the detector and without the clean versions of these images. The method applies a known removal attack (by default, VAE regeneration) to each watermarked image, takes the difference to the watermarked image as a residual, trains a forger network (by default, a Stable Diffusion VAE with a fine-tuned decoder) to predict this residual with an L1 loss, and then adds the predicted residual to clean images. The paper also concludes that a successful removal attack implies that forgery is possible.

Critical strengths

A simple, low-cost attack that works on four current learned watermarks. Forgery of the localized WAM watermark, also in selected regions, This is interesting, but it is shown only with example images, not with average results, so I give it less weight.

Critical weaknesses

The relation to close existing work is not clear, and the closest methods are not baselines. WARFARE (Li et al., 2023) is cited in the related work, and from its abstract it also seems to be one framework for both watermark removal and forgery, but the paper does not describe it or compare with it. WMCopier(Dong et al., arXiv 2503.22330), which is not cited, also seems to forge watermarks without access to the watermarking algorithm. Your work seems close to these two, so a discussion and, if possible, a direct comparison may help to show what is new. The baselines in the paper could be weak: the adapted watermark faker, which trains a U-Net, gets .0 success everywhere in Table 1, while a UNet forger in WForge gets 0.964 on WAM(table 6).

The central claim seems stronger than the evidence. The Abstract says that the existence of a successful removal attack implies the feasibility of forgery, and the Conclusion says the same. But Lemma 1 seems to follow directly from its two assumptions, because the second assumption already defines the forger, so it does not add support in practice. The paper's own data also seem to go against the general form of the claim. In Table 3, Ctrl-M removes the watermark fully, but forgery stays at chance. The paper explains this by image damage(A.3.1), but then the claim needs this condition. VAE-L hardly removes the watermark(0.958), but forgery partly works(0.831). The paper itself uses two forms of the claim: "forgery strength closely correlated with removal capability", and a "conditional positive correlation" in Section 5.3. The Abstract and Conclusion use only the general form. The premise is shown only with pictures(L323), and the ablation goes against it(L434–436). The abstract says "consistently outperforms", but Tables 1 and 5 show cases where a baseline is better.

Questions for authors
  1. Which sign of the residual do you use for training and at test time? Eq. 5 and L310 seem to use opposite signs.
  2. How would a no-box, query-free attacker choose the remover, its strength and the forger network without the detector?
  3. How is WForge different from WARFARE and WMCopier? A direct comparison on the same data would help.
  4. Do all training images carry the same watermark message? Does WForge still work if each image has a different message?
Reviewer_tBq2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WForge learns to forge watermarks without access to the target watermarking algorithm, detector, or paired clean originals. It uses removal residuals to train a network that predicts perturbations for new images. The authors evaluate four watermarking schemes on MS-COCO, ImageNet, and DiffusionDB, with ablations on removers, networks, and training set size, plus localized WAM examples. They report strong forgery performance in several settings and argue that removal and forgery are approximately inverse processes.

Critical strengths
  1. The attack addresses a meaningful security problem. The experiments support the feasibility of learning forgery from watermarked samples without detector queries or genuine clean-image pairs.

  2. Three test datasets, four watermarking schemes, and ablations on removers, backbones, and sample sizes provide evidence beyond isolated examples. The localized WAM experiments further help assess the method's scope.

Critical weaknesses
  1. Limited novelty and insufficient comparison with prior work.

Much of the motivation overlaps with existing work: WMCopier [1] learns no-box forgery from watermarked samples, DAPAO [2] reverses removal perturbations for forgery, and Warfare [3] learns a forgery mapping from denoised intermediate images. WMCopier and DAPAO are not cited, while the distinction from Warfare is underdeveloped. Comparisons with Watermark Steganalysis and Watermark Faker do not establish an advance over these closer methods. WAVES [4] and TrustMark [5] also share attack objectives or components, although their complete methods differ.

Explicit residual supervision may distinguish WForge from these methods, but its added value has not been isolated. With the correct sign and no projection, residual L1 regression can be rewritten as paired-image L1 reconstruction with a skip connection. This does not establish equivalence of the complete algorithms, but it calls for controlled comparisons of residual parameterization, supervision, backbone, and loss to establish the contribution.

  1. Unclear residual direction and an unsupported general reversibility claim.

Equation (5), page 5, defines Rg=R′(Iw)−IwR_g=R'(I_w)-I_w, approximately the negative watermark residual under ideal recovery. Equation (9), page 6, fits this target, but Equation (8) and Algorithm 1, page 18, add the prediction to the input. This conflicts with Equation (4) and the subtraction order in Section 4.4, leaving the implemented operation unclear.

Lemma 1 assumes exact recovery and an ideal inverse; it does not establish that practical removal yields learnable residuals that generalize to new images. The ablations also show dependence on the remover and image quality. Even if the signs are only writing errors, the general claim in the abstract and conclusion remains insufficiently supported by the reported performance.

References

[1] Dong, Ziping, et al. "WMCopier: forging invisible watermarks on arbitrary images." Advances in Neural Information Processing Systems 38 (2026): 138035-138065.

[2] Ba, Zhongjie, et al. "More Robust, More Vulnerable: The Robustness–Security Paradox of Post-Processing Watermarking." IEEE Transactions on Dependable and Secure Computing (2026).

[3] Li, Guanlin, et al. "Warfare: Breaking the watermark protection of AI-Generated Content." arXiv preprint arXiv:2310.07726 (2023).

[4] An, Bang, et al. "Waves: Benchmarking the robustness of image watermarks." Proceedings of the 41st International Conference on Machine Learning (2024): 1456–1492.

[5] Bui, Tu, Shruti Agarwal, and John Collomosse. "Trustmark: Robust watermarking and watermark removal for arbitrary resolution images." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

Questions for authors
  1. Which residual target and inference sign were used in code? Please reconcile Equations (4) through (9), Section 4.4, and Algorithm 1, and identify configurations and results that verify the implemented convention.

  2. Please compare with WMCopier [1] on at least one shared target, such as RivaGAN or Stable Signature, under matched data size, resolution, message/key settings, detection criterion or false-positive rate, and image-quality constraints. Report forgery success and quality; if efficiency is the claimed advantage, include preprocessing, training, and inference costs on comparable hardware. If the comparison is inapplicable, explain why.

  3. Must training images share a message, key, model version, or identity, and how are suitable samples collected? The common default across four targets is useful evidence; how was it chosen in advance without target feedback, and how does it differ from the enhanced LaWa configuration selected using target detection performance?

  4. Section 5.1 (p. 7) defines forgery success as bit accuracy above 0.8. For the 32-, 48-, and 100-bit settings in Table 2, how does this criterion map to each method's native acceptance rule, target-identity acceptance, and region-level verification where applicable? Please report the corresponding theoretical and empirical false-positive rates. The distinctions between detection, identification, and provenance verification in WAVES [3] and TrustMark [4] are relevant here.

  5. Why is WAM/MS-COCO success 0.994 in Table 1 but 0.947 in the 5,000-sample row of Table 4 and the VAE FT-Decoder row of Table 6, with the other three metrics identical? Clarify any differences in thresholds, runs, or test settings, and which results apply.

Reviewer_G94yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes WForge, a detector-query-free approach that learns image-dependent watermark perturbations from watermarked images and their outputs under a watermark-removal procedure. A residual predictor is subsequently applied to unseen clean images.

Critical strengths

The problem is important, and the reported evaluation spans three image datasets and several watermarking schemes. The ablations on removers and predictor architectures are informative, and the appendix includes unsuccessful settings rather than reporting only favorable cases.

Critical weaknesses
  1. Sections 4.1–4.3 and Algorithm 1 define the supervised target as the removed image minus the watermarked image, but inference adds the predicted residual to a clean image. Under the paper's ideal-removal assumption, this learns the negative of the embedding residual. Section 4.4 uses the opposite difference. This is a confirmed inconsistency in the manuscript, not evidence that the undisclosed implementation necessarily uses the wrong sign. The authors should establish the convention used in the experiments and align the equations, pseudocode, captions, and implementation.

  2. Lemma 1 assumes exact recovery of the clean image and an ideal inverse that already returns the appropriate embedding residual. An ordinary successful removal attack only changes the detector decision; it need not recover the clean image or produce a transferable inverse perturbation. The empirical result should be stated conditionally, with analysis of reconstruction contamination and learnability of the residuals. Table 3 itself demonstrates substantial dependence on the remover.

  3. Warfare already uses denoised surrogate images to support watermark forgery without paired genuine clean images. This related workflow should be distinguished explicitly; its status as a verified preprint means that its missing numerical comparison alone is not a rejection reason under the current ICLR guidance. Independently, the modified Watermark Faker comparison does not isolate residual parameterization from backbone, initialization, optimization, and output constraints. Under a matched residual output parameterization, residual L1 supervision and image-reconstruction L1 supervision can be algebraically equivalent. A controlled comparison is important.

  4. The paper should state whether training images share a fixed unknown message and key, how checkpoints and configurations are selected without detector feedback, and how complete verification is performed for WAM. A fixed bit-accuracy threshold of 0.8 across different payload lengths is not a common false-positive operating point. Clean-image scores and calibrated attack success rates are needed.

  5. The same WAM default setting reports success rates of 0.994 in Table 1 and 0.947 in Tables 4 and 6, while the other reported metrics match. Furthermore, Watermark Steganalysis reaches 100% success on StegaStamp across all three datasets, exceeding WForge. The manuscript should correct the universal-superiority language and report the success–quality trade-off accurately.

Questions for authors

see weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.