Abstract

Watermarking surgical scene representations learned with 4D Gaussian Splatting (4DGS) is challenging due to continuous deformation of anatomical structures and articulated motion of surgical instruments. Although watermarking has been widely studied for images, videos, and 3D neural scene representations, these approaches do not directly generalize to Watermarking surgical scene representations learned with 4D Gaussian Splatting (4DGS) is challenging due to continuous deformation of anatomical structures and articulated motion of surgical instruments. Although watermarking has been widely studied for images, videos, and static 3D scene representations, these approaches do not directly generalize to dynamic 4D surgical scenes, where naive watermark embedding can disrupt temporal stability or interfere with the analysis of surgical scenes. To address this problem, we propose SurgMark, a watermarking framework tailored to 4DGS in surgical scenes. SurgMark consists of: 1) a temporal coherence regularization that preserves the original motion trajectories of Gaussians after watermark embedding; 2) a motion-aware gradient weighting strategy that adaptively scales watermark strength according to the motion magnitude of individual Gaussians; and 3) a segmentation feature consistency loss based on foundation segmentation models to prevent degradation of segmentation performance on anatomical regions and surgical instruments. Extensive experiments on two public surgical datasets demonstrate that SurgMark enables robust and imperceptible watermark embedding while maintaining rendering quality and segmentation performance.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2259_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

CholecSeg8K dataset: https://www.kaggle.com/datasets/newslab/cholecseg8k EndoVis2018 dataset: https://github.com/mli0603/EndoVis2018

BibTex

@InProceedings{JeoMin_SurgMark_MICCAI2026,
        author = { Jeong, Minjae AND Lee, Jaejin AND Hwang, Yechan AND Kim, Won Hwa},
        title = { { SurgMark: Structure-Aware Watermarking for 4D Gaussian Splatting in Surgical Scenes } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16892},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The authors introduce SurgMask, a watermarking approach for 4D Gaussian splatting scenes acquired from surgical scenes. It is argued, that 3D surgery scene data will become a major asset for hospitals, and therefore ownership needs to be attributable. To do so, the authors present a system targeted towards dynamic 4D Gaussian splatting scenes. By combining a temporal coherence regularization with a motion-aware watermark strength, the authors are able to preserve semantic features important for segmentation tasks.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The major strength of the paper is the application of watermarking to dynamic surgical scenes, while keeping coherence and segmentation outcome in mind. The authors further evaluate their approach against several baselines, such as for instance SOTA 3DGS watermarking, namely GuardSplat. Furthermore, methods like 4DGS+HiDDeN and 4DGS+SSL are compared against. All these and more baselines are compared against on two data sets, across different segmentation models, which makes the analysis quite thorough.

    As, to my knowledge this paper is the first 4DGS watermarking paper, which takes into account coherence and segmentation loss, I believe that the authors make a creative technical contribution, which is adequately evaluated.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    While the approach seems to work and is also robust to certain attacks, I am unsure about it s relevance for the MICCAI crowd. The authors claim, that security and ownership are important aspects of 3D surgical videos. While I can imagine that 3D surgical videos will have even higher value in the future, e. g. , for training surgical robots, I am unsure that watermarking approaches are essential here. And if watermarking of 4DGS is relevant, then segmentation would also be relevant in other domains. So why not proposing/evaluating the approach on general real-world data?

    Furthermore, the studied attacks are rather simple, how about more advanced adversarial attacks, for instance driven by GANs?

    Finally, a deeper discussion on the computational requirements and the acceptance of these in the medical domain would be nice.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    As stated above, the idea is creative and to my knowledge the approach seems new, as it takes into account segmentation as an important target to be considered when watermarking. However, my main concern is the relevance of this work for the MICCAI community, as I am unsure that the authors address a relevant problem. But I am happy to be convinced otherwise during the rebuttal phase, by having more data-based insights into the needs alongside the other suggestions above.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    I feel that the rebuttal addresses my main concern. The authors should also include the more sophisticated attacks in the paper.



Review #2

  • Please describe the contribution of the paper

    This paper proposes SurgMark, a structure-aware watermarking framework designed to protect 4D Gaussian Splatting (4DGS) assets in dynamic surgical scenes. The method embeds invisible watermarks into the time-varying position and spherical harmonic parameters of the Gaussians. To preserve rendering quality and clinical utility, SurgMark introduces three key constraints: (1) a temporal coherence regularization to maintain original motion trajectories; (2) a Motion-Aware Gradient Scaling (MAGS) strategy that adaptively modulates the watermarking strength based on the motion magnitude of individual Gaussians ; and (3) a segmentation feature consistency loss to preserve semantic cues for downstream 2D anatomical segmentation tasks.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The method cleverly introduces a Motion-Aware Gradient Scaling (MAGS) strategy to adapt watermark strength based on Gaussian motion , and a segmentation consistency loss to preserve critical semantic features for downstream clinical tasks.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The paper lacks a comparison or discussion with existing 4DGS watermarking methods, most notably Hide-in-Motion (ICRA 2025). The authors must clarify the distinct contributions of SurgMark compared to this recent 4DGS watermarking work, rather than framing 4DGS watermarking as an entirely unexplored domain. 2.The authors claim that SurgMark preserves the performance of downstream analysis tasks critical to “safe robotic navigation,” yet the evaluation is strictly limited to 2D semantic segmentation. Furthermore, the proposed segmentation feature consistency loss (Lseg) is highly task-specific. The paper lacks a discussion and evaluation of whether optimizing the watermark specifically for 2D segmentation features might inadvertently disrupt or degrade the performance of other critical downstream tasks (e.g., depth estimation, 3D geometry extraction). The authors should clarify whether this task-specific design compromises the broader utility of the 4DGS asset. 3.The robustness evaluation (Table 1) relies entirely on simple 2D image-level perturbations (e.g., cropping, rotation, resizing). Given the recent advancements in Gaussian Splatting watermarking, there are already existing attack methods explicitly designed to target 3DGS/4DGS representations. To substantiate the robustness claims, the authors must test SurgMark against these existing, domain-specific 3DGS attack methods, rather than limiting the evaluation to conventional 2D noise. 4.SurgMark’s Motion-Aware Gradient Scaling (MAGS) shares a fundamentally similar design philosophy with GaussianMarker’s uncertainty-aware embedding. While GaussianMarker exploits spatial uncertainty to find perturbation-tolerant regions in 3DGS, SurgMark exploits temporal motion magnitude to achieve the exact same goal in 4DGS. Given this strong conceptual overlap in adaptive embedding strategies, the omission of GaussianMarker as a primary baseline in the quantitative evaluation (Table 1) is a critical flaw. 5.There are notable formatting inconsistencies in the references section.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper has three major weaknesses: First, it misses critical baselines, including recent 4DGS watermarking methods and GaussianMarker (which shares a similar adaptive embedding concept). Second, it fails to verify whether the watermark degrades other critical downstream tasks beyond 2D segmentation. Finally, the robustness evaluation is insufficient, as it only tests against 2D image perturbations while ignoring existing 3DGS-specific attacks.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    After reading the rebuttal, I still recommend rejection. My main concern is the complete omission of Hide-in-Motion, which I consider a highly relevant and necessary baseline for a 4DGS watermarking paper. Although the authors argue that Hide-in-Motion focuses on steganography rather than watermarking, it is still a recent 4DGS-based information embedding method and therefore a direct point of comparison for SurgMark. The paper frames 4DGS watermarking as largely unexplored while not discussing or comparing against this closely related work, which substantially weakens the novelty claim and positioning of the contribution. This omission cannot be resolved by a brief clarification in the rebuttal; at minimum, a careful discussion and empirical comparison/adaptation should be included. Other concerns also remain insufficiently addressed. The robustness evaluation is still limited to simple 2D image-level perturbations and does not test against representation-level 3DGS/4DGS attacks. The downstream-task claim is also too broad given that the evaluation only considers 2D semantic segmentation, while the proposed segmentation feature consistency loss may be task-specific and may not preserve other important 3D/temporal properties. Finally, the absence of GaussianMarker as a quantitative baseline remains a weakness; implementation difficulty alone does not provide a fair substitute for comparison or convincing ablation. Overall, the rebuttal does not sufficiently address the key concerns regarding novelty, robustness, and evaluation breadth.



Review #3

  • Please describe the contribution of the paper

    1.The paper proposes SurgMark, a structure-aware watermarking framework tailored specifically for 4D Gaussian Splatting (4DGS) in dynamic surgical scenes.

    1. Introduces a regularization technique that constrains watermark embedding to follow the original motion trajectories of Gaussians, preventing the distortion of learned deformation dynamics. 3.Develops an adaptive strategy that modulates the watermark strength according to the motion magnitude of individual Gaussians.
  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.Protecting the intellectual property of clinical 4DGS assets is an emerging and critical problem. This paper is among the first to directly tackle the complexities of watermarking in dynamic, non-rigid surgical environments.

    2.A standout strength is the explicit focus on maintaining segmentation performance. By using a segmentation feature consistency loss, SurgMark successfully maintains high mIoU across models like SAM, SAM2, and MedSAM. This ensures the models remain viable for robot-assisted surgery tasks.

    3.The method demonstrates strong resilience to realistic image-level degradations.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The evaluations are all conducted with a fixed message length of 48 bits ($L=48$). The paper lacks an ablation study on payload capacity to demonstrate at what bit-length the visual fidelity or segmentation performance begins to break down.

    2.While the segmentation feature consistency loss preserves 2D semantic representations, the paper does not evaluate whether the watermarking affects 3D-specific downstream tasks, such as depth estimation or 3D instrument tracking, which are equally vital in surgical navigation.

    3.While keeping opacity and covariance unperturbed effectively prevents visual artifacts, this restriction also limits the parameter space available for watermark embedding, potentially capping the maximum theoretical capacity compared to full parameter optimization.

    4.Lack of discussion with prior watermarking methods in NeRF/3DGS, like [1-3]. It would strengthen the paper for a broader audience.

    [1] Protecting nerfs’ copyright via plug-and-play watermarking base model. [2] MarkSplatter: Generalizable watermarking for 3D gaussian splatting model via splatter image structure. [3] GS-Hider: Hiding Messages into 3D Gaussian Splatting

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Please see above

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank the reviewers for valuable feedback. Due to policy, we can’t include additional ablation results in the rebuttal, but will add them in the revised version along with the suggested references. We address the major concerns below.

R1: Q) Relevance to the MICCAI community A) Unlike general 4DGS settings (e.g., human or outdoor scenes), where motion is typically distributed across large spatial regions, surgical 4DGS scenes contain confined endoscopic views, thin surgical tools, and complex tissue deformation with rapid local motion. In such settings, watermarking can easily disrupt temporal consistency and fine-grained anatomical/tool representations important for surgical scene understanding. We address these surgical-specific challenges through motion- and structure-aware watermarking, which we believe is particularly relevant to the MICCAI community.

R1 & R2: Q) Limited scope of robustness evaluation A) We agree that more advanced attacks would further strengthen the analysis and plan to expand the robustness evaluation with Gaussian parameter perturbation, Gaussian removal/pruning, deformation-network perturbation, and GAN-based adaptive attacks.

R1: Q) Computational requirements and feasibility A) SurgMark introduces only lightweight overhead beyond standard 4DGS optimization. The watermark decoder is a simple 3-layer MLP, while the remaining pipeline follows standard 4DGS rendering/training with learnable watermark offsets, which is computationally practical for surgical 4DGS settings.

R2: Q) Lack of comparison with Hide-in-Motion A) Hide-in-Motion and SurgMark address different problem settings: steganography vs. watermarking. While watermarking focuses on ownership verification and robust message recovery, steganography targets covert information transmission [1]. Accordingly, Hide-in-Motion ensures that data is “recoverable from specific viewpoints” while SurgMark recovers the watermark message “for an arbitrary viewpoint v and timestamp t”. In addition, Hide-in-Motion lacks robustness evaluation against attacks, while SurgMark explicitly evaluates robust recovery under such conditions. Additional discussion regarding Hide-in-Motion will be included to clarify such distinctions more explicitly. [1] Shih, F. Y. (2017). Digital watermarking and steganography: fundamentals and techniques. CRC press.

R2 & R3: Q) Limited evaluation on other downstream tasks A) We agree that evaluating additional downstream tasks (e.g., depth estimation, tracking) would further strengthen the analysis, and we will include additional discussion in the revised version. Importantly, opacity and covariance parameters remain fixed, and temporal coherence is explicitly regularized to help prevent inadvertent disruption of other downstream tasks.

R2: Q) Why was GaussianMarker not included as a baseline? A) Direct adaptation of GaussianMarker to dynamic 4DGS is non-trivial as its uncertainty-aware embedding requires repeated Hessian computation over temporally varying Gaussians across temporal states, resulting in excessive computation and memory consumption (OOM on 48GB NVIDIA RTX 6000 Ada GPU). Also, the two methods are fundamentally different; GaussianMarker selects high-uncertainty Gaussians and additionally injects message-embedded Gaussians, whereas SurgMark continuously scales message-loss gradients according to temporal motion magnitude.

R3: Q) Lack of payload-capacity analysis A) SurgMark and GuardSplat, which are based on CLIP embeddings, support larger payload capacities (e.g., up to 72 bits within the current setting) while maintaining stable performance. We will include additional payload-capacity analysis in the revised version.

R3: Q) Restricting optimization to selected parameters A) We observed that perturbing opacity and covariance parameters caused noticeably larger visual/temporal artifacts with limited gains in embedding capacity. Therefore, we did not find perturbing these parameters beneficial in practice.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    Reviewer 2 suggests comparing with SOTA techniques including 4DGS and GaussianMarker. Please provide more clarity to the methods and demonstrate and discuss robustness in depth.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    R1 suggested accept post-rebuttal while suggesting that the authors to include the more sophisticated attacks. R2 suggested reject based on the lack of comparison and discussion with closely related work. R3 suggested accept in the first round. However, all reviewers did agree on the creative design of this work. I suggest accept but at the same time would strongly suggest authors to include a discussion on relevant works and revise the final paper accordingly adhering to the scientific code of conduct.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    two accept and one reject.

    i read through and agree the paper can be accepted subject to minor revision.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    All the reviewers found the research problem interesting; however, they were not convinced by the evaluations. After going through the reviewers’ comments and the authors’ rebuttal, I found that some of the concerns raised by the reviewers (R1 and R2), particularly regarding the study of robustness against attacks specifically designed for 4DGS, were not properly addressed in the rebuttal. As the method focuses on 4D watermarking, it is important either to align the hypothesis with the experiments or to provide a clear justification for not including such a study.



back to top