acceptodds
Under review as a conference paper at ICLR 2027

Residual Transferability in Neural Image Watermarking

10 views

Abstract

Neural image watermarks can be forged by extracting watermark-bearing residuals from released images and transferring them to unrelated content. While prior work has demonstrated this vulnerability, what makes these residuals transferable remains poorly understood. We formalize this vulnerability with **residual transferability (RT)**, a metric that quantifies how well watermark evidence remains decodable after transfer across unrelated images. Through comparative analyses and controlled interventions, we find that common training-side variations do not account for the large RT differences across watermarking systems; instead, architectural design plays a central role. By contrasting high- and low-RT systems and validating their architectural differences through controlled interventions, we identify two mechanisms that strengthen the dependence of watermark evidence on the cover image, thereby suppressing the residual transferability. These findings provide concrete design guidance for developing more forgery-resistant watermarking architectures. Complementarily, for existing watermarking systems where architectural redesign is impractical, we introduce **CoverLock**, a plug-and-play strategy for existing watermarking systems that strengthens such image dependence without architectural redesign. Across representative watermarking systems exhibiting high residual transferability, CoverLock achieves a more favorable security–robustness trade-off than both traditional handcrafted defenses and learned classifier-based defenses.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI Security, Invisible Watermarking, Watermark Forgery

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ps3bOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on the vulnerability of neural image watermarking systems to residual-based forgery attacks. The authors first introduce a metric called Residual Transferability (RT). Through empirical analysis, the paper reveals that models with high RT learn a cover-agnostic shortcut, while certain architectural designs (specifically Broadcast-GAP and Content-adaptive attention) enforce cover-dependent pathways that naturally resist residual transfer. To protect existing vulnerable systems, the authors propose CoverLock, a plug-and-play wrapper that XORs the watermark payload with a robust, image-conditioned binary hash generated by a frozen DINOv2 model and a lightweight projection head trained via contrastive learning.

Critical strengths
  1. The focused problem is interesting and important. I have seen several papers report the discussed phenomenon.
  2. The ablation studies in Section 3 and Appendix E are sufficient, providing a valuable guideline for future watermark model designs.
Critical weaknesses
  1. The core defensive mechanism of CoverLock is not innovative. As the authors acknowledge, traditional methods like MHDW (Lu et al., 2006) utilized this exact pipeline. While updating the hash extractor to a modern DINOv2 backbone and training it with contrastive learning improves robustness, the system-level workflow lacks conceptual novelty. It feels more like a modern engineering patch on an old idea.
  2. The authors motivate CoverLock as a deployment-friendly alternative that avoids the cost of redesigning and retraining the watermarking architecture. However, based on my training experience, in the context of deep image watermarking, training a separate lightweight projection head does not incur a significantly lower absolute cost than training a lightweight, semantic-bound watermark encoder.
  3. While Section 6 briefly acknowledges in-generation/latent diffusion watermarking methods, the paper uses overly broad terminology (such as Neural Image Watermarking) throughout the title, abstract, and introduction. The authors should either: (1) Explicitly scope their claims, title, and abstract to "Post-Processing Neural Image Watermarking" to avoid overclaiming; or (2) Provide a discussion (and empirical verification) on whether residual transferability poses a practical vulnerability to in-generation watermarking.
  4. The related work discussion overlooks recent studies on additive watermark residuals and their cross-image behavior, such as Removing Box-Free Watermarks for Image-to-Image Models via Query-Based Reverse Engineering (AAAI 2026). While the attack objectives differ, its analysis of additive watermark residuals is relevant to the motivation and interpretation of residual transferability in this paper.
  5. There is a slight structural disconnect in the paper. The first half brilliantly identifies that architectural flaws (lack of cover-dependency) cause high RT and proves that architectural interventions (Broadcast-GAP) fix it. However, the proposed solution abandons this architectural insight in favor of a post-hoc wrapper. It leads to the question: why not propose a new, robust watermark architecture based directly on the insights from Section 3? Furthermore, what is the vulnerability of CoverLock to adaptive attacks? (e.g., an attacker who knows CoverLock is deployed and attempts to find a collision in the lightweight hash head).
  6. Although the architectural ablations in Section 3 and Appendix E demonstrate substantial reductions in RT, these improvements come at a considerable cost to image fidelity. For instance, introducing Broadcast–GAP reduces the PSNR of MBRS from 42.32 dB to 27.07 dB and that of CIN from 41.55 dB to 21.71 dB. Moreover, the modified VINE fails to converge. These results raise concerns about the practical applicability of the identified architectural design. The authors should investigate whether the reduction in RT can be maintained under comparable image-fidelity and watermark-robustness constraints.
  7. The quality of Figure 3 is not good. For example, some lines are overlapped with text.
Questions for authors

See weaknesses. I may adjust my rating according to the authors' rebuttal.

4 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer ps3b (Part 1)

We greatly appreciate the reviewers' recognition of our work and their thoughtful suggestions, which have helped us further improve the manuscript. We provide detailed responses to each comment below.

W1: Clarification of CoverLock's Novelty

A1:

We thank the reviewer for raising this important point. We acknowledge that content-dependent watermarking has been explored in earlier work, including MHDW (Lu et al., 2006). However, the novelty of our work does not rest solely on introducing a new content-binding mechanism. Our primary contribution is a systematic investigation of residual transferability in modern neural image watermarking, where the encoder–decoder architectures, learned representations, and training objectives introduce complexities that are not captured by traditional watermarking methods. Through controlled experiments and architectural interventions, we identify key factors governing residual transferability and derive practical guidance for designing more secure watermarking systems.

CoverLock builds on these findings as a practical, post-hoc defense against residual-transfer attacks. Its contribution lies in adapting content binding to existing neural watermarking systems without modifying or retraining their original architectures, while maintaining robustness under common image distortions. We will further clarify the distinction between the established principle of content binding and the specific contributions of CoverLock in the revised paper.

W2: Why Not Propose a New Watermarking Architecture?

A2:

We thank the reviewer for this question. We would like to clarify the motivation and contributions of our work. Our architectural analysis aims to answer our original research question: what are the key factors governing residual transferability in neural image watermarking? Through systematic experiments, we examine and rule out several potential factors, ultimately identifying network architecture as the dominant factor affecting RT. We further pinpoint two architectural designs, Broadcast-GAP (BG) and RivaGAN-style attention, that contribute to low RT. However, our experiments also show that these designs cannot be directly transferred to other watermarking architectures without substantial modifications. For example, introducing BG significantly reduces the PSNR of MBRS and CIN, while the modified VINE fails to converge. These findings indicate that achieving low RT through architectural changes requires case-by-case redesign and optimization of the watermarking system, which is far from straightforward. This is precisely why we propose CoverLock as a complementary post-hoc defense: it provides protection against residual-transfer attacks without requiring the redesign or retraining of existing watermarking architectures. We will clarify this connection between our systematic findings and the motivation for CoverLock in the revised paper.

W3: Scope of Neural Image Watermarking

A3:

We thank the reviewer for pointing this out. We acknowledge that our previous terminology may have led to overly broad claims. Following your suggestion, we have revised the title, abstract, and introduction to explicitly scope our study to post-processing neural image watermarking. These changes will be incorporated into the revised manuscript.

W4: Additional Related Work on Watermark Residuals

A4:

We thank the reviewer for highlighting this valuable work. An et al. (AAAI 2026) provide an insightful study of watermark removal by exploiting the additive properties of watermark residuals through query-based reverse engineering. While their work focuses on watermark removal, ours investigates cross-image residual transferability for watermark forgery and its underlying factors. We have cited this work and added the following sentence to the Related Work section of the revised manuscript:

"Relatedly, An et al. (2026) exploit the additive properties of watermark residuals through query-based reverse engineering to remove watermarks from image-to-image models, whereas our work investigates residual transferability for cross-image watermark forgery."

AuthorsOpenReview ↗

Rebuttal to Reviewer ps3b(Part 2)

W5: Robustness Against Adaptive Attacks

A5:

We sincerely thank the reviewer for raising this important concern. To further evaluate CoverLock against adaptive attacks, we conducted additional experiments using white-box PGD and black-box NES. The results show that both the no-box residual-transfer attack and the black-box adaptive attack achieve 0% forgery ASR. While the white-box attack reaches 100% ASR, it requires full access to the private hash model and its gradients, which is generally unrealistic in practical watermarking deployments, and substantially degrades image quality (PSNR: 27.68 dB). These findings support CoverLock's effectiveness under practical no-box and black-box threat models while acknowledging its limitations under full white-box access. For detailed experimental settings and results, please refer to W3 in our rebuttal to Reviewer wSNK (Part 2).

W6: Practicality of BG and Comparisons at Matched PSNR

A6:

We thank the reviewer for raising this important concern. We would like to clarify that Broadcast-GAP (BG) is not a defense proposed by our work, but an existing architectural design investigated to understand the factors governing residual transferability. As shown in our experiments, directly introducing BG into MBRS and CIN substantially reduces their PSNR, while the modified VINE fails to train successfully. Despite our training efforts, the BG-modified models could not achieve PSNR comparable to their original counterparts, making a matched-PSNR comparison infeasible under these architectural modifications. This further highlights that low-RT architectural designs cannot be directly transferred across watermarking systems without case-by-case redesign and optimization.

Importantly, low RT is not necessarily associated with lower image fidelity. In our RivaGAN ablation, fixing the attention maps increases RT@100 from 0.077 to 0.944, while PSNR remains nearly unchanged (42.12 vs. 42.11 dB). Moreover, the original HiDDeN incorporates BG and achieves 28.23 dB PSNR. Together, these findings suggest that residual transferability is governed by architectural design rather than image fidelity alone.

W7: Presentation Quality of Figure 3

A7:

We thank the reviewer for pointing this out. This was an oversight on our part, and we will correct the overlapping text and improve the presentation quality of Figure 3 in the revised manuscript.

Dear Authors,

Thank you for the detailed rebuttal and the additional experiments.

I appreciate that the authors have addressed my concerns regarding the scope of the paper, the missing related work, and the adaptive attack evaluation against CoverLock.

However, I have decided to maintain my initial rating for the following reasons:

  1. No revised manuscript has been uploaded for now, so the promised revisions cannot yet be verified in the manuscript.

  2. Remaining conceptual and practical limitations: In A2 and A6, the authors clarify that the identified architectural mechanism (Broadcast–GAP) cannot be readily transferred across watermarking schemes without severe image quality degradation or optimization failure, and thus position CoverLock as a post-hoc workaround. While practically motivated, this leaves the architectural insight largely unexploited for designing high-fidelity, low-RT watermarking encoders, and CoverLock itself remains an engineering modernization of the established content-binding paradigm. Furthermore, as shown in A5, CoverLock's security under adaptive threats still relies on keeping the hash model itself confidential rather than relying purely on a secret key, which runs counter to Kerckhoffs's principle.

Overall, the systematic analysis of residual transferability in post-processing watermarking is interesting, and I remain slightly positive on the submission.

We sincerely thank you once again for your positive assessment of our work and your valuable feedback!

  1. Revised Manuscript. We have now uploaded an updated version of the revised manuscript, incorporating the additional experiments and clarifications discussed in our rebuttal. Please refer to the light purple highlights for revisions addressing your concerns and the light orange highlights for shared concerns.

  2. Remaining Limitations. We sincerely appreciate the reviewer's suggestion regarding broader architectural exploration. We believe this is a valuable direction for future research and have explicitly discussed it in the revised Limitations section.

    Regarding white-box security and Kerckhoffs's principle, CoverLock is intended as a content-binding mechanism, not as a cryptographic secret key. Its purpose is to strengthen the dependence between watermark evidence and image content, consistent with the motivation behind content-dependent watermarking designs. We therefore recommend treating CoverLock as an integral part of the watermarking system rather than a separate, publicly accessible component.

    This assumption is consistent with practical deployments such as Google SynthID and Amazon Titan, where watermark embedding is controlled by the provider, while users can access watermarked content or verification services without obtaining arbitrary embedding capabilities. Keeping security-critical model parameters confidential is a reasonable assumption in such provider-controlled settings. Furthermore, CoverLock and the underlying watermarking model can be viewed as a unified system that could also incorporate cryptographic secret-key mechanisms for stronger security guarantees.

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

This paper looks at residual-based watermark forgery: someone pulls the watermark residual out of a released image and pastes it onto unrelated content. The authors define residual transferability, an oracle metric that uses ground-truth residuals, so it doesn't depend on any specific estimation attack. They compare five systems and ablate training-side factors. Their conclusion is that architecture is what mainly drives RT. They point to two mechanisms that tie the watermark to the cover: Broadcast–GAP and content-adaptive attention.

Critical strengths
  • The RT metric is simple and attack-agnostic. It separates a system-level property from how good the residual estimator is.
  • Dataset, message length, loss weights and noise are tested over three seeds, which makes a fair case that these alone don't explain the RT gap.
Critical weaknesses
  • CoverLock is never tested against an adaptive attack. The threat model assumes the attacker doesn't know about the defense. That's basically security through obscurity, and the authors themselves list adaptive attacks as open.
  • The interventions use a different training setup from the deployed models. We don't know whether BG still keeps RT low in robustness-trained models, and the BG variants' robustness isn't reported.
Questions for authors
  • Can you report RT for PSNR-matched variants, e.g., HiDDeN−BG and MBRS+BG trained to the same PSNR? Or plot RT against residual norm for each architecture?
  • What's the forgery ASR when targets are near-duplicates of, or semantically close to, the reference images?
  • Do BG models trained with distortion simulation still have low RT?
  • Please say which COCO split is used to evaluate CoverLock, since CoverLock is also trained on COCO.
3 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer SeHh

We sincerely thank the reviewers for their positive assessment of our work and their valuable and constructive feedback. We address your comments and concerns point by point below.

W1: Robustness Against Adaptive Attacks

A1:

We sincerely thank the reviewer for raising this important concern. To further evaluate CoverLock against adaptive attacks, we conducted additional experiments using white-box PGD and black-box NES. The results show that both the no-box residual-transfer attack and the black-box adaptive attack achieve 0% forgery ASR. While the white-box attack reaches 100% ASR, it requires full access to the private hash model and its gradients, which is generally unrealistic in practical watermarking deployments, and substantially degrades image quality (PSNR: 27.68 dB). These findings support CoverLock's effectiveness under practical no-box and black-box threat models while acknowledging its limitations under full white-box access. For detailed experimental settings and results, please refer to W3 in our rebuttal to Reviewer wSNK (Part 2).

W2&Q3: Do BG Models Trained with Distortion Simulation Still Exhibit Low RT?

A2:

We sincerely thank the reviewer for raising this important question. Yes, BG models continue to exhibit low RT when trained with distortion simulation. To verify this, we evaluate HiDDeN and MBRS+BG using their original training configurations, including distortion simulation. The results are reported below.

BG ModelRT@1RT@100Self BitAccPSNR
HiDDeN (distortion simulation)0.3240.27592.37%33.85 dB
MBRS+BG (distortion simulation)0.1720.11892.66%26.19 dB

As discussed in our BG ablation, BG itself increases the difficulty of jointly optimizing image fidelity and watermark recovery. Introducing distortion simulation may further complicate this optimization, potentially affecting bit accuracy and imperceptibility. Nevertheless, these results confirm that BG models maintain low residual transferability even when trained with distortion simulation.

Q1: Lack of Comparisons at Matched PSNR

A3:

We sincerely thank the reviewer for this valuable suggestion. We found that obtaining PSNR-matched variants is particularly challenging in practice, as introducing BG substantially increases the difficulty of jointly optimizing image fidelity and watermark recovery. Even the original HiDDeN model with BG achieves only approximately 27–30 dB PSNR, illustrating this optimization challenge. More importantly, we would like to clarify that low RT is not necessarily associated with low PSNR. As demonstrated by our RivaGAN ablation (Fig. 2, left), fixing its attention across images increases RT@100 from 0.08 to 0.94 while changing PSNR only slightly. This suggests that the reduction in RT is primarily related to architectural choices rather than being an inevitable consequence of reduced imperceptibility. We therefore interpret the lower PSNR observed with BG as evidence of increased optimization difficulty, rather than a necessary cost of achieving low RT.

Q2: Forgery ASR When Target Images Are Semantically Similar to Reference Images

A4:

We sincerely thank the reviewer for highlighting this important concern. To further evaluate CoverLock against forgery attempts involving similar target images, we consider two settings: semantic regeneration using SDXL-Turbo and local editing using SD1.5 inpainting. The results show that CoverLock maintains very low forgery ASRs of 0% and 2%. For detailed experimental settings and results, please refer to W2 in our rebuttal to Reviewer wSNK (Part 1).

Q4: Evaluation of CoverLock on COCO Despite Training on the Same Dataset

A5:

Thank you for pointing out this ambiguity. CoverLock is trained on images from the COCO 2017 training split, whereas all reported COCO evaluations use images from the official validation split. We will explicitly state these splits in the experimental setup.

Thank you for your detailed responses, which have answered many of the shared concerns raised by the reviewers. Please update the manuscript accordingly and highlight all changes in a different color. I will provide my final score once Q1-Q4 are properly resolved in the revised version.

Thanks for your reply! We have uploaded an updated version of our revised manuscript and will continue proofreading and refining it. For clarity, revisions are color-coded: light blue for reviewer wSNK, light green for SeHh, light purple for ps3b, and light orange for concerns shared by multiple reviewers. To avoid disrupting readability, we highlight only the section headings and table captions for newly added sections and substantially revised tables, respectively.

Reviewer_wSNKOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies why neural image watermarks can be forged by transferring residuals. It proposes the residual transferability (RT) metric and identifies two low-RT designs: Broadcast–GAP (BG) and content-adaptive attention. It also proposes CoverLock, a plug-in that binds the message to image content.

Critical strengths
  • The problem is important and well motivated.
  • The RT metric measures transferability independently of any specific attack.
  • CoverLock achieves low attack success across watermarks, datasets and attacks.
Critical weaknesses
  • Limited novelty of the analysis. Residual transfer is a known issue, first shown by the copy attack (Kutter et al., 2000). The natural defense is to bind the watermark to image content. The main insight, BG, can be seen as an architectural form of this idea. However, Table 5 shows BG greatly lowers PSNR (e.g., MBRS 42.3 → 27.1 dB; CIN 41.6 → 21.7 dB). In watermarking, improving one property such as security usually costs another, such as imperceptibility. So the lower RT may partly reflect this trade-off. Without comparing RT at matched PSNR, the analysis feels incremental.
  • CoverLock is trained on 256-bit codes, but short codes (30 bits for CIN) raise the risk of collisions. The paper does not report collision rates or results on semantically similar/ locally edited images.
  • The threat model assumes the attacker does not know the defense. No adaptive black-box, gray-box or white-box attacks are tested. Replay attacks, where the attacker extracts and re-embeds the watermark, are also not evaluated. This limits the security claims.
  • Only pixel-level distortions and crop are tested. Please evaluate rotation and flips. Please also report Hamming distance between codes of original and transformed images.
  • Lbin and Lcode appear in Eq. 15 but are only defined in Appendix B.
  • HiDDeN is analyzed for RT but not used in the main experiments. Does CoverLock also help low-RT systems?
Questions for authors

Please refer to the critical weaknesses

6 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer wSNK (Part 1)

Thank you for your comments and constructive suggestions! We address your concerns point by point below. We will incorporate these changes, together with those requested by the other reviewers, into the revised manuscript shortly.

W1: Previously studied problem & lack of a matched-PSNR comparison

A1:

  1. Previously studied problem

    We agree that Kutter et al. (2000) introduced an early watermark copy attack based on residual transfer. Our work differs in two important respects.

    • Ground-truth residuals as the object of analysis. Kutter et al. used an estimated watermark signal to carry out the copy attack. We study the ground-truth embedding residual itself and formalize residual transferability (RT) as a system property that measures the residual's cross-image decodability, independent of residual estimation error.

    • What determines RT in neural watermarking. In hand-designed watermarking, the embedding rule explicitly specifies which image features or coefficients are used and how they are modified. In modern neural watermarking, the encoder and decoder learn this behavior jointly. The learned embedding and decoding rules are therefore harder to inspect, and RT may depend on several factors, including message length, loss weighting, training distortions, and network architecture. Our analysis identifies a training failure mode that leads to high RT. It also rules out message length, robustness distortion layers, dataset, and other factors, and traces low RT to architectural design: BG in HiDDeN and content-adaptive attention in RivaGAN, the two low-RT watermarks.

    We also clarify that BG is an existing component of HiDDeN's original architecture. Our contribution in this section is an analysis of which design choices govern residual transferability in neural image watermarking; we use BG as an ablation to answer that question.

  2. Lack of a matched-PSNR comparison

    We sincerely appreciate the reviewer's thoughtful observation that improving watermark security may come at the cost of imperceptibility. While we acknowledge this potential trade-off, our experimental evidence leads us to favor an interpretation centered on increased optimization difficulty introduced by BG, rather than solely an inherent security–imperceptibility trade-off. HiDDeN originally uses BG and achieves approximately 28–30 dB PSNR. Removing BG improves both PSNR and bit accuracy. Moreover, Fig. 2 (right) shows that HiDDeN, MBRS, and CIN require substantially more training images to reach comparable bit accuracy when BG is applied. Together, these observations suggest that BG makes the joint optimization of image fidelity and message recovery more challenging. RivaGAN further supports this distinction. Fixing its attention across images changes PSNR only slightly, yet increases RT@100 from 0.08 to 0.94 (Fig. 2, left). These results suggest that architectural choices that bind watermarks to image content can substantially reduce RT at comparable PSNR, although certain designs, such as BG, may introduce additional challenges in multi-objective optimization.

W2: Risk of collisions and Forgery ASR When Target Images Are Semantically Similar to Reference Images

A2:

  1. Risk of collisions.

    We evaluate 50,000 ImageNet images and report the fraction of distinct image pairs that produce identical codes. The observed collision rates suggest that the risk of a collision between two images may be acceptable for watermarking verification.

    Code LengthCollision Rate
    30 bits3.944 × 10⁻⁷
    64 bits0
    100 bits0
    256 bits0
  2. Forgery ASR on semantically similar and local edited target images

    To further examine the robustness of CoverLock under this scenarios, we evaluate two settings using 100 randomly sampled COCO images:

    • Semantically similar targets: We derive a prompt from each image’s human-written COCO caption and use SDXL-Turbo to generate a new image without providing the original image as input.

    • Local edited targets: We modify the caption to describe a replacement subject and use SD1.5 inpainting to locally edit the main subject while preserving the remaining image content.

    We report the forgery ASR and hash similarity between target and reference images under CIN + CoverLock-G in the following table. Exact hash collision denotes identical 30-bit hashes (Hamming distance = 0), while mean bit agreement measures the average fraction of matching bits across image pairs.

    Target TypeSource ImagesForgery ASR ↓Exact Hash Collision ↓Mean Bit Agreement ↓
    Regeneration1000.00%0.00% (0/100)62.87%
    Local Editing1002.00%1.00% (1/100)67.40%

    Importantly, CoverLock maintains very low forgery ASRs in both settings, achieving 0% for regeneration and only 2% for local editing.

AuthorsOpenReview ↗

Rebuttal to Reviewer wSNK (Part 2)

W3: Robustness Against Adaptive Attacks

A3:

We sincerely appreciate you bringing this important aspect to our attention, as it has helped us strengthen our security evaluation. To address this concern, we conducted additional experiments with two established adaptive attacks: white-box PGD [1] and black-box NES [2]. For replay attack, if it is extracting the watermark residual from a watermarked image and re-embedding it into another image, we would like to clarify that this is precisely the attack evaluated in our GT Residual (GT1) setting.

We randomly select 50 source–target image pairs from the COCO dataset and use CIN as the underlying watermarking method with CoverLock. For each pair, the attacker's goal is to make the target image pass verification with the source image's watermark, thereby bypassing the content-binding mechanism. Both adaptive attacks proceed in two stages. The attacker first optimizes a perturbation on the target image so that its CoverLock hash matches the hash of the source image. The attacker then transfers the watermark residual from the watermarked source image onto this hash-matched image. As a reference, we also report the plain residual-transfer attack studied in our paper, which involves no optimization.

All attacks are evaluated under a total perturbation budget of ϵ=16/255\epsilon = 16/255, which covers both the hash-matching perturbation and the transferred residual. For white-box PGD, we assume full access to the CoverLock hash model and its gradients and minimize the binary cross-entropy between the predicted and target hash codes, using two initializations with 100 optimization steps each. For black-box NES, we assume access only to hash outputs through queries and estimate gradients from Hamming-distance feedback using 20 pairs of antithetic Gaussian directions per iteration, under a maximum budget of 1,000 queries. A forgery is counted as successful if at least 28 of the 30 payload bits are verified.

AttackForgery ASRPSNR (dB)
No-box residual0%41.64
Black-box0%38.64
White-box100%27.68

The white-box attack succeeds, but it requires full access to the hash model and its gradients and reduces the PSNR of the forged image to 27.68 dB, compared with 41.64 dB for plain residual transfer. In practical watermark deployments, the hash model is kept private by the provider, so attackers typically operate in the no-box setting, or at most the black-box setting when a verification API is exposed. CoverLock resists forgery in both settings, which supports its security under realistic threat models.

[1] Towards Deep Learning Models Resistant to Adversarial Attacks. ICLR 2018.

[2] Black-box Adversarial Attacks with Limited Queries and Information. ICML 2018.

W4: Robustness Evaluation under Flips and Rotations

A4:

Thank you for raising this point. We did not include rotation and flipping in our original training because the three underlying watermarking methods (CIN, MBRS, and VINE) have no resistance to these distortions: their bit accuracy drops to around 50%. Under such distortions the payload itself can no longer be decoded, regardless of the content-binding mechanism.

To answer your question, we retrained CoverLock with rotation and flipping added to the training distortions and evaluated the robustness of CoverLock itself, independent of the underlying watermark decoder. The table below reports the hash bit accuracy (%) under rotation by 30°, horizontal flip, and vertical flip.

Watermark / message lengthMethodRot. 30°HFlipVFlip
CIN / 30 bitMHDW58.1761.0559.67
CIN / 30 bitCoverLock92.9398.2191.91
MBRS / 256 bitMHDW59.9662.6659.80
MBRS / 256 bitCoverLock94.9998.5394.58
VINE / 100 bitMHDW59.7462.1059.33
VINE / 100 bitCoverLock94.9198.4294.41

CoverLock maintains 91.9–98.5% accuracy across all three watermarks and all three distortions, whereas MHDW stays at 58–63%.

W5: Lbin and Lcode appear in Eq. 15 but are only defined in Appendix B.

A5:

Thank you for pointing this out. We will move the definition to the revised manuscript's main text.

W6: Does CoverLock also help low-RT systems?

A6:

Thank you for your comments. We report the ASR of HiDDeN with and without CoverLock across three datasets below. With CoverLock, the ASR is 0% in all evaluated settings. We will include these results in Table 2 of the revised manuscript.

AttackDatasetOriginal HiDDeN ASRHiDDeN + CoverLock ASR
GT@1COCO2.52%0%
GT@1ImageNet4.27%0%
GT@1DIV2K2.11%0%
YangCOCO0%0%
YangImageNet0%0%
YangDIV2K0%0%
WmForgerCOCO2.77%0%
WmForgerImageNet3.54%0%
WmForgerDIV2K4.22%0%

I thank the authors for the detailed rebuttal. The new experiments strengthen the empirical evaluation. The rebuttal resolves W4–W6. However, several concerns remain

First, it is still unclear how the analysis leads to CoverLock. The analysis traces high RT to architecture, but adding BG lowers PSNR. So it gives no concrete way to improve existing encoders. CoverLock does not build on these findings. CoverLock appears to be an independent component, a learned robust image hash that wraps the payload, rather than a design that follows from the analysis.

Second, the collision results are reassuring for this dataset. However, 30-bit codes must collide beyond about 1.07B images. I suggest recommending CoverLock only for longer codes. Collisions across users also need an analysis. Suppose user 1 embeds message m₁ in image x₁, so the payload is m₁ ⊕ b(x₁). If an attacker copies this residual onto image x₂, the decoder returns m₁ ⊕ b(x₁) ⊕ b(x₂). If this equals another user’s message m₂, the image is attributed to the wrong user. Please report how often this happens.

Third, the no-box and black-box results are convincing, but the white-box attack succeeds. The hash relies on a frozen, public DINOv2 backbone, so it cannot be assumed to stay private. An attacker could also use DINOv2 as a partial surrogate. Since security is the main claim, this needs discussion.

Finally, my replay comment referred to re-encoding, not residual copying. GT@1 copies the residual. It does not cover an attacker who decodes the payload, recomputes the hash, and re-embeds a valid watermark.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.