Abstract

Dataset is essential for advancing deep learning in healthcare, yet its unauthorized use raises serious concerns regarding patient privacy leakage and intellectual property infringement. Existing copyright protection methods typically rely on backdoor-triggered verification mechanisms, but these methods inherently introduce security vulnerabilities by embedding malicious backdoor behaviors that can lead to abnormal outputs. To address this issue, we propose a novel harmless copyright verification threat model. Unlike traditional methods that depend on trigger-activated abnormal outputs as verification signals, our approach instead uses prediction accuracy on watermarked samples as the verification criterion. The core idea is to enforce improved performance on trigger-injected samples, enabling reliable ownership verification without inducing malicious behaviors. Specifically, we propose a Learnable Measurement-Domain Harmless Backdoor Watermark (LMHBW) method. Unlike traditional image-domain watermarking techniques, which operate on reconstructed images that are easily accessible and thus vulnerable to tampering, we shift the protection to the measurement domain prior to image reconstruction. Motivated by the fact that raw measurement data are typically confined within imaging devices (e.g., scanners) and are rarely exposed externally, this design enhances the stealthiness and robustness of the embedded watermark against detection and removal. Experimental results demonstrate that the proposed threat model enables effective and reliable watermark verification while remaining behaviorally harmless, and further validate the feasibility and advantages of measurement-domain watermarking for medical imaging copyright protection.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2875_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)

N/A

BibTex

@InProceedings{XuFen_Harmless_MICCAI2026,
        author = { Xu, Fengzhi AND Yang, Ziyuan AND Liang, Yan AND Chen, Yingyu AND Xia, Wenjun AND Zhang, Yi},
        title = { { Harmless Copyright Protection for CT Dataset } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16896},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The main contribution of this paper is proposing a novel “harmless” copyright protection method for medical CT datasets. Traditional backdoor watermark methods prove dataset ownership by forcing the AI model to make incorrect or abnormal predictions when a trigger is present, which creates severe security risks in healthcare applications. To fix this, the authors introduce a new verification approach that checks if a suspicious model has a higher prediction accuracy on watermarked samples, rather than making it fail. Furthermore, to keep the watermark robust and prevent it from ruining the visual quality of the medical images, their method (LMHBW) embeds a learnable watermark directly into the raw, pre-imaging “measurement data”. This ensures the final reconstructed images remain completely safe for clinical diagnosis while successfully proving data ownership.

  • 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.It improves security by using prediction accuracy for verification instead of causing the model to make dangerous mistakes 2.It embeds the watermark in the raw measurement data, making the protection much harder to detect and remove compared to traditional image-level methods. 3.It uses a learnable mask that successfully proves data ownership without ruining the visual quality of the reconstructed medical images.

  • 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 experiments are conducted on a very small subset of the COVID-19 Low-Dose CT dataset, using only 15 COVID and 15 normal cases for training, and 5 of each for testing. Furthermore, it appears to treat the data as 2D slices rather than continuous 3D volumes. This limited scale fails to demonstrate clinical feasibility or prove that the learnable mask can maintain both visual fidelity and verification accuracy across massive, highly diverse, real-world 3D medical datasets. 2.The methodology relies solely on standard Filtered Back-Projection (FBP) to reconstruct the images from the watermarked measurement data. In modern clinical practice, advanced Iterative Reconstruction (IR) or Deep Learning-based Reconstruction (DLR) techniques are heavily used for noise reduction. These non-linear, aggressive denoising algorithms could easily filter out or distort the subtle measurement-domain watermark (LMHBW), potentially breaking the downstream prediction-accuracy verification entirely.

  • 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 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.

    (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?

    The authors introduce a compelling approach to medical dataset copyright protection by shifting the verification signal from harmful, abnormal predictions to benign prediction accuracy gains, thereby mitigating critical security risks associated with traditional backdoor watermarks. Furthermore, embedding the learnable watermark (LMHBW) within the pre-imaging measurement domain is a clever strategy that inherently enhances stealthiness against post-imaging tampering while preserving visual fidelity. However, my enthusiasm is dampened by the limited scope of the experiments. The validation relies on a very small-scale dataset evaluated as 2D slices rather than complete 3D volumes, which fails to fully demonstrate clinical feasibility. More importantly, the current framework solely utilizes standard Filtered Back-Projection (FBP); it remains unclear whether the subtle measurement-domain watermark would survive the non-linear, aggressive denoising of modern Iterative Reconstruction (IR) or Deep Learning-based Reconstruction (DLR) algorithms standard in current clinical practice.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    The rebuttal has adequately addressed my main concerns, so I am satisfied with the authors’ response and recommend acceptance.



Review #2

  • Please describe the contribution of the paper

    They address the important problem of copyright protection for CT datasets and propose a Learnable Measurement-Domain Harmless Backdoor Watermark (LMHBW). Compared with prior watermarking methods, LMHBW enables ownership verification while preserving prediction accuracy, whereas earlier approaches often introduced malicious output behavior and increased adversarial vulnerability. The paper further demonstrates superior verification performance over existing methods.

  • 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.A key strength of the paper is the novelty of its loss design: it combines multiple loss terms to preserve reconstruction quality, ensure that the watermark remains verifiable, and maintain high predictive accuracy on watermarked samples.

    2.Another strength is that the paper defines the problem very clearly. The introduction is well written and accessible even to readers outside the specific field, and the topic itself is timely and important, with broad relevance to researchers and practitioners concerned with AI privacy and data protection.

  • 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.A major weakness of the paper is the choice of baselines. The compared methods are from 2019 and 2022, which are relatively old for a rapidly evolving field. Although the paper cites the more recent work by Sun et al. (2025) on EntropyMark, it neither includes a direct comparison with that method nor explains why such a comparison was omitted. This makes it harder to assess the true novelty and competitiveness of the proposed approach against the current state of the art.

    2.There is also some ambiguity in the presentation of the results. For example, in Table 1, it is unclear what backbone model is being used and what watermarking ratio is assumed. In Table 2, only the proposed method is evaluated across different architectures, whereas it would be more informative to include the same architectural comparison for the baselines as well. That would allow for a fairer and more complete evaluation.

    3.In addition, the paper lacks clarity in several technical details and contains a few minor writing issues. For instance, in Equation 5, (m_c) is not defined, and the abbreviation “FBP” appears before its full form is introduced. Some model details are also missing, such as which version of CLIP was used, whether it was pretrained on medical data, and which ResNet variant was adopted. The absence of citations for these architectures further reduces clarity.

  • 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 submission does not provide sufficient information for 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.

    (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?

    My main concern regarding the paper is the lack of comparison with current state-of-the-art methods.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    One of my main concerns was the comparison with recent methods. The authors have addressed the concerns I raised, and I believe the research topic is timely and important for the field.



Review #3

  • Please describe the contribution of the paper

    The authors propose a novel backdoor watermarking method for CT datasets aimed at protecting the intellectual property of medical imaging data. The key innovation lies in embedding a learnable mask within the measurement domain rather than the image domain, which increases robustness against adversarial replication. To support this approach, the authors design three custom loss functions that jointly preserve reconstruction fidelity while ensuring reliable watermark verification. A pre-trained CLIP model is leveraged as part of the feature consistency objective, and experimental results are reported to demonstrate the effectiveness of the proposed method. Overall, the paper addresses a timely and relevant problem in the intersection of medical imaging and data copyright protection, offering a fresh perspective through its measurement-domain formulation.

  • 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 central idea is novel, with contributions that are twofold. First, the watermark is implemented as a learnable mask inserted in the measurement domain rather than the image domain, which makes it more difficult for an adversary to replicate through adversarial watermarking. Second, the authors introduce three custom loss functions that jointly ensure images are reconstructed without loss of fidelity while preserving the effectiveness of copyright protection.

  • 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.

    Major Comments

    The authors refer to CLIP as the pre-trained model but do not clarify whether they employ the original CLIP model or a custom variant pre-trained on domain-specific data. Since the features generated by CLIP are highly dependent on the pre-training data, this ambiguity raises concerns regarding the validity of the feature consistency loss. An ablation study is strongly needed, particularly given the use of three distinct loss functions. Although page constraints may limit the scope, a thorough ablation study would provide empirical evidence that each of the three losses contributes meaningfully to the overall performance. The introduction of the learnable mask is described in vague terms, and additional details would strengthen the presentation. A brief background on the measurement domain, along with some intuition regarding what the data represents in this domain, would provide valuable context for the reader. An important open question is the practical adoption of measurement-domain watermarking in physical devices such as CT scanners. While the proposed approach is novel, its feasibility for real-world deployment warrants further discussion.

    Minor Comments

    The abstract would benefit from including specific numerical results rather than the current general statement that “experimental results demonstrate that the proposed threat model enables effective watermark verification while remaining harmless.” There is a typo on Page 2, paragraph 2: “embeddedinto” should be corrected. The acronym “FBP” is used prior to its formal introduction (Page 5, following Equation 1). The organization of Section 2.2 could be improved. The authors introduce two objectives but, rather than proceeding directly to the loss functions, first describe a masking strategy before returning to the loss formulation. Restructuring this section would enhance readability. The authors should justify the selection of SSIM as the image-fidelity loss metric. A citation to prior work establishing SSIM as a standard choice for this purpose would be sufficient. Figure 3 would benefit from explicit annotations to clearly highlight the issues present in the images.

  • 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 submission does not provide sufficient information for 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.

    (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?

    The proposed approach of embedding a learnable watermark in the measurement domain is novel and makes this a strong contribution to MICCAI.

  • 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 sincerely thank reviewers for their valuable comments and recognition of our work as “compelling”, “novel” and “a strong contribution to MICCAI”. All typos and unclear descriptions have been revised and will be reflected in the next version.

@Meta Reviewer @R3Q1Q2 Core Idea and Loss Design. Re: The core attribution of this paper is to propose a novel harmless verification threat model. Unlike traditional methods that verify ownership through malicious behaviors, we prevent harmful behaviors by keeping watermarked labels identical to the originals, using prediction accuracy for verification. To implement this verification feasibility, we design a loss function driven by two objectives to optimize the watermark. 1)To ensure the watermark remains imperceptible, the visual loss combines MSE and SSIM to keep watermarked and original reconstructed images visually consistent. 2)To ensure effectively verification, the CLIP loss uses cosine similarity to distinguish features between clean and watermarked samples, while the CE loss optimizes the watermark for easier classification by the surrogate model, leading to higher prediction accuracy. Removing visual loss reduces PSNR (between the reconstructed images with and without watermark) by 27.28 dB. Removing CLIP loss reduces Verification Success Rate (VSR) by 15%. Removing CE loss reduces watermarked dataset prediction accuracy by 15.8% and lowers VSR by 10%.

@R1Q1 @R3Q4 Real-World Feasibility. Re: Our method is evaluated on a large-scale dataset comprising 104 COVID and 56 normal cases, achieving 61.42/100(mean Benign Accuracy/Verification Success Rate). Although class imbalance reduces mBA compared to small-scale datasets, the 100% VSR confirms the watermark’s effectiveness. We expand the watermark into multiple channels to match consecutive slices, yielding 68.11/100(mBA/VSR). In practical deployment, our watermark has low storage and simple embedding logic, enabling seamless integration through lightweight software or firmware updates. We are collaborating with hospitals and equipment vendors for future deployment.

@R1Q2 Generalization Under Aggressive Denoising Algorithms. Re: To show that our feature-level watermark verification is independent of pixel-level denoising algorithms, we apply the watermark optimized via FBP to the protected dataset and use a pre-trained FBPConvNet during verification, yielding 69.35/100(mBA/VSR).

@R2Q1 Comparison with SOTA. Re: Under the same setting as Table 1, EntropyMark(EM) achieves 57.55/50/91.77/91.38(mean BA/VSR/mC_b/mC_w). EM relies on a low-entropy patched set for verification, but patch-based methods like Badnet in our scenarios still produce high-entropy outputs, reducing verification performance.

@R2Q2Q3 @R3Q1 Experimental Setting. Re: In Table 1, ResNet18 is our backbone model and the watermarking ratio is 0.2.We use the original CLIP model with ViT-B/32 architecture. The original CLIP model extracts general visual features to facilitate the feature-level separation between clean and watermarked samples, as confirmed by our ablation studies. Furthermore, to verify the generalizability of our method across different CLIP models, we evaluated an architecture-identical medical pre-trained model, yielding 69.76/100 (mean BA/VSR).

@R2Q2 Cross-Architecture Comparison. Re: BadNet achieves 70.59/85 and 78.92/45 (mBA/VSR) on EfficientNet and MobileNet, respectively. FIBA\dagger achieves 68.26/65 and 77.65/70, while FIBA\ddagger achieves 70.61/85 and 76.99/65. @R2Q3 @R3Q3 Clarifying Vague Terms. Re: The learnable watermark is class-specific, where m_c denotes the watermark of the c-th class. Measurement-domain data represent the intensity attenuation of X-rays passing through the object from different projection views. We select SSIM as the fidelity metric because it emphasizes structural integrity critical for medical diagnosis [1].

[1] Structurally-sensitive multi-scale deep neural network for low-dose CT denoising.




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.

    Please address the concerns about core formulation, harmless verification mechanism, and loss design.

  • 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.

    three accept. i agree.



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.

    The rebuttal contains several additional experimental results, which violate the rebuttal guidelines. Hence, these results have been ignored while making this decision. The paper is recommended for acceptance solely based on the novel idea of inserting a watermark in the measurement domain that preserves clean accuracy and achieves good verification success rate. While the robustness of the proposed approach is questionable, this is still a good starting point.



Meta-review #3

  • 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.

    All reviewers recognized the significance of the work and recommended acceptance of the paper.



back to top