acceptodds
Under review as a conference paper at ICLR 2027

EndoRDH: Constrained Reversible Information Transport over Non-Stationary Endoscopic 3D Gaussian Representations

Abstract

Dynamic 3D Gaussian Splatting (3DGS) models are increasingly reused as persistent 3D assets, raising the need to embed ownership information without compromising the underlying representation. This is particularly challenging in endoscopic scenes, where clinically sensitive regions restrict admissible modifications and camera-coupled active illumination induces spatially non-stationary spherical harmonic (SH) carriers whose reliability varies across the scene. To address these issues, we introduce EndoRDH, a geometry-aware reversible transport framework that formulates reversible watermarking as constrained information transport over a non-stationary Gaussian representation, allocating information to perceptually low-cost and locally stable carriers while preserving an explicit inverse path to the original host. First, clinically admissible carrier projection restricts watermark updates to a prescribed editable SH subspace while freezing structural Gaussian parameters and excluding protected regions. Second, Riemannian photometric decoupling converts local photometric variation into a carrier-dependent transport geometry without assuming recovery of the underlying illumination physics. Within this geometry, an optimal transport inspired reversible flow adaptively redistributes the watermark payload toward low cost, locally stable carriers while retaining a key conditioned inverse for host recovery. Finally, dual-branch redundant coding distributes error corrected evidence across complementary carrier groups, improving robustness without increasing the per-carrier perturbation budget. Experiments on dynamic endoscopic 3DGS scenes show that EndoRDH achieves a favorable trade-off among representation fidelity, copyright verification, and authorized reversibility, while remaining robust to both rendering-domain and model-domain perturbations. These results suggest that modeling watermark embedding as constrained reversible transport offers a principled alternative to uniform or independently optimized carrier perturbations in non-stationary neural scene representations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, reversible information transport, optimal transport, Riemannian geometry, robust watermarking, medical imaging

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DwtROpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces EndoRDH, a watermarking framework for dynamic endoscopic 3DGS. The authors introduce four modules: Clinically Admissible Carrier Projection (CACP), Riemannian Photometric Decoupling (RPD), Optimal-transport-inspired Reversible Flow (ORF), and Dual-Branch Redundant Coding (DBRC).

Critical strengths
  1. Protecting reusable endoscopic 3DGS assets is clinically relevant, the motivation is reasonable.
  2. The writing is clear and easy to follow, and the content and figures are easy to read and understand.
  3. The methods are introduced in detail, with released code in the supplement.
Critical weaknesses
  1. The limitation is also quite clear: the experiments are mainly done on only two EndoNeRF sequences. There is a lack of sufficient experiments on multiple datasets and also sequence number to support the proposed method. e.g. StereoMIS, SCARED, Hamlyn, D4D, EndoSLAM, C3VD dataset.
  2. The novelty is also limited. Since the proposed RPD does not recover the real illumination physics and the ORF is “OT-inspired” rather than solving the OT problem.
  3. The evaluation is also not complete. The recovery evidence has only qualitative results, and no quantitative results are provided in the paper.
  4. While the paper compares with general domain GS watermarking method, there are only in-domain experiments. It seems a bit unfair since the compared methods are not used in their domain. The comparison can only claim that this paper introduces a method that is better for the Endoscopic-specific domain. However, the authors overclaim their contributions in their introduction and method. I think the correct way is also to compare with in-domain methods or compare general domain methods in the general domain datasets.

Hayoz M, Hahne C, Gallardo M, et al. Learning how to robustly estimate camera pose in endoscopic videos[J]. International journal of computer assisted radiology and surgery, 2023, 18(7): 1185-1192.

Allan M, Mcleod J, Wang C, et al. Stereo correspondence and reconstruction of endoscopic data challenge[J]. arXiv preprint arXiv:2101.01133, 2021.

Recasens D, Lamarca J, Fácil J M, et al. Endo-depth-and-motion: Reconstruction and tracking in endoscopic videos using depth networks and photometric constraints[J]. IEEE Robotics and Automation Letters, 2021, 6(4): 7225-7232.

Docea R, Younis R, Long Y, et al. The Dresden Dataset for 4D Reconstruction of Non-Rigid Abdominal Surgical Scenes[J]. Scientific Data, 2026, 13(1): 1334.

Ozyoruk K B, Gokceler G I, Bobrow T L, et al. EndoSLAM dataset and an unsupervised monocular visual odometry and depth estimation approach for endoscopic videos[J]. Medical image analysis, 2021, 71: 102058.

Bobrow T L, Golhar M, Vijayan R, et al. Colonoscopy 3d video dataset with paired depth from 2d-3d registration[J]. Medical image analysis, 2023, 90: 102956.

Questions for authors

I have mentioned all concerns in the above weaknesses.

2 comments

We thank the reviewer for the careful reading, answered below one by one.

Comment 1: Experiments are limited to two EndoNeRF sequences; additional datasets are needed.

The revision adds multiple StereoMIS evaluation sequences, a qualitative comparison in the main text and panels in Appendix F. The StereoMIS sequences enter the shared-protocol quantitative evaluation; the figures are visual complements, while the quantitative claims rest on the aggregate metrics reported in the manuscript. Further datasets and acquisition systems are needed for broader generalization, which we acknowledge.

Revision location: Section 4.1; Appendix F; Section 5 and Appendix G.

Comment 2: Limited novelty; RPD does not recover illumination physics and ORF is OT-inspired but does not solve optimal transport.

RPD does not estimate physical illumination or intrinsic reflectance. It builds a carrier-dependent cost geometry meant to reflect local photometric variation for allocation. ORF uses an optimal-transport-inspired prior without solving a global transport problem. Invertibility comes from the additive-coupling maps; the prior only biases transport toward lower-cost, locally stable carriers. The description now separates feasible-set restriction (CACP), carrier geometry (RPD), allocation prior (ORF) and redundant evidence aggregation (DBRC), and no longer presents them as illumination recovery or an OT solver.

Revision location: Abstract; Sections 1 and 3; Appendices B and C.

Comment 3: Recovery results are qualitative and lack quantitative results.

Float32 SH recovery RMSE is now in Section 4.5 (Table 4, reproduced below as Table 2). EndoRDH gives 5.2×10−65.2\times10^{-6} in the stated no-post-modification setting. The embedding-mechanism comparison is manuscript Table 5, below as Table 3, and the factorized ablation is manuscript Table 3. The measured setting is now kept distinct from recovery after quantization or model-domain attacks.

Embedding mechanismPSNRSSIMLPIPSRecovery SH RMSE
EndoRDH34.290.96630.02045.2×10−65.2\times10^{-6}
No ORF prior33.860.95120.03185.4×10−65.4\times10^{-6}
Standard residual embedding31.720.89350.13624.8×10−64.8\times10^{-6}
Direct quantization33.180.92460.04174.7×10−24.7\times10^{-2}

Table 2: Fidelity and authorized recovery. Lower SH RMSE is better.

Embedding mechanismClean 2DClean 3D2D attack avg.3D attack avg.
EndoRDH98.4786.2094.6983.69
No ORF prior93.9880.8592.9478.83
Standard residual embedding93.3156.4790.1253.28
Direct quantization94.0579.6393.8476.91

Table 3: Verification accuracy (%).

Revision location: Section 4.3; Section 4.5, Tables 4–5; Appendix H.

Comment 4: Generic-domain comparisons may be unfair and contributions may be overclaimed.

The controlled-transfer protocol is now spelled out: external methods are migrated to a common endoscopic 3DGS host and evaluated with shared sequences, ROI definitions, rendering interface and metrics. That controls for reconstruction differences, but it does not reproduce each method's original benchmark or establish superiority on generic 3DGS data. The main results are in Table 4 and attack-averaged verification in Table 5, using the manuscript's reported values.

MethodPSNRSSIMLPIPS2DACC3DACC
GS-Marker31.160.89810.127597.4655.39
3D-GSW32.090.90360.153297.9956.03
P-NODE31.940.88790.138095.2855.41
SFP-NODE33.960.90870.114997.1680.42
EndoRDH34.290.96630.020498.4786.20

Table 4: Equal-weight averages over the evaluated EndoNeRF and StereoMIS sequences under the shared protocol. Rendering measured against GT; accuracies in %.

MethodClean 2DAvg. 2D attacksClean 3DAvg. 3D attacksOverall
GS-Marker97.4689.7755.3951.6670.72
3D-GSW97.9989.7556.0352.0670.90
P-NODE95.2887.9955.4152.5970.29
SFP-NODE97.1693.2780.4275.7384.50
EndoRDH98.4794.6986.2083.6989.19

Table 5: Verification accuracy (%) clean and under averaged attacks; overall is the mean of the 2D and 3D averages.

The conclusion is narrowed to the stated endoscopic transfer protocol, with no cross-domain superiority claim.

Revision location: Sections 4.1–4.2; Appendix D.

Thank you for the response. However, my concern is not fully resolved. Since:

  1. The newly added StereoMIS experiment contains only two sequences, with no experiment setup details. This is only an incremental experiment that lacks sufficient evidence to support my previous comment, where I mentioned StereoMIS, SCARED, Hamlyn, D4D, EndoSLAM, C3VD a total of 6 datasets.

  2. This clarification for comment 4 does not resolve my original concern. The current comparisons remain in-domain only. The experiments can support a claim that the proposed method is better suited to the endoscopic setting under this shared host, but they still do not justify broader claims of superiority over general-domain GS watermarking methods. To address this fairly, the authors should either compare with in-domain methods or evaluate the general-domain baselines on general-domain datasets (or both).

Therefore, I retain my score.

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

This paper proposes EndoRDH, a reversible watermarking framework for dynamic endoscopic 3D Gaussian Splatting (3DGS) representations, aiming to provide copyright protection while preserving rendering quality and clinically sensitive information. The authors formulate watermark embedding as a constrained reversible information transport problem, addressing challenges associated with protected anatomical regions, spatially varying illumination, and the need for authorized recovery of the original representation.

The proposed framework consists of four main components: (1) Clinically Admissible Carrier Projection (CACP), which restricts watermark modifications to selected spherical harmonic coefficients while preserving structural Gaussian parameters and protected regions; (2) Riemannian Photometric Decoupling (RPD), which models carrier-dependent photometric variation to guide watermark allocation; (3) Optimal-Transport Reversible Flow (ORF), which employs invertible coupling transformations and key-conditioned message embedding to enable authorized host recovery; and (4) Dual-Branch Redundant Coding (DBRC), which distributes error-corrected watermark information across complementary carrier groups to improve robustness.

Experiments on two endoscopic sequences (cutting and pulling) using a pretrained EndoGS representation report improvements over the evaluated baselines in rendering quality and watermark verification. EndoRDH achieves 34.29 dB PSNR, 0.9663 SSIM, 0.0204 LPIPS, 98.47% rendering-domain verification accuracy, and 86.20% model-domain verification accuracy. The method also reports improved robustness under the evaluated rendering-domain and model-domain perturbations. Overall, the paper presents a geometry-aware reversible watermarking approach that integrates constrained embedding, photometric conditioning, invertible transformations, and redundant coding for copyright protection of dynamic endoscopic 3DGS representations.

Critical strengths
  1. Well-motivated problem formulation: The paper formulates reversible watermarking for dynamic endoscopic 3D Gaussian Splatting as constrained information transport. It explicitly separates the admissible parameter space, carrier-dependent modification costs, and invertible embedding operations. This provides a coherent framework for jointly considering watermark fidelity, verification, and host recovery.

  2. Explicitly invertible watermark embedding: The proposed ORF module employs additive coupling transformations with analytically defined inverse operations. Together with the CACP module, which restricts updates to selected spherical harmonic coefficients while keeping structural Gaussian parameters and protected carriers unchanged, the framework provides a well-defined authorized recovery procedure.

  3. Photometrically conditioned carrier representation: The RPD module incorporates carrier-dependent photometric information and distinguishes between direction-independent and higher-order spherical harmonic coefficients. This design is motivated by the spatially varying illumination and appearance characteristics of endoscopic scenes. The controlled comparisons with P-NODE and SFP-NODE provide experimental evidence supporting the usefulness of the proposed photometric conditioning.

Critical weaknesses
  1. Limited experimental evaluation and generalizability: The experimental evaluation is restricted to only two endoscopic sequences (cutting and pulling) using the pretrained EndoGS representation. Although the reported results are promising, this limited evaluation makes it difficult to assess the generalizability of the proposed framework across different surgical scenes, lighting conditions, endoscope systems, and 3DGS reconstruction architectures. Additional experiments on diverse datasets and reconstruction frameworks would strengthen the empirical evidence.

  2. Limited practical evaluation of reversibility: One of the central contributions of the paper is the ability to recover the original host representation through an explicitly invertible transformation. However, the exact recovery guarantee is derived under ideal assumptions, including exact arithmetic, correct key/codeword, deterministic rendering, and the absence of irreversible modifications. The paper does not provide a systematic quantitative analysis of practical recovery errors under finite-precision computation, quantization, or post-embedding model processing. Such experiments would strengthen the practical significance of the proposed reversible framework.

  3. Lack of downstream clinical task evaluation: The proposed CACP module restricts watermark modifications to selected spherical harmonic coefficients while preserving structural Gaussian parameters and protected regions. However, preserving these parameters does not necessarily guarantee that clinically relevant information remains unaffected. The experimental evaluation primarily focuses on rendering quality and watermark verification rather than downstream clinical tasks. Evaluating the impact of watermarking on tasks such as lesion analysis, surgical navigation, or anatomical measurement would provide stronger evidence for its suitability in clinically sensitive applications.

  4. Limited comparison with zero-watermarking approaches: The paper emphasizes copyright protection while preserving clinically sensitive information, but it does not include a comparison with zero-watermarking approaches, which can establish ownership evidence without modifying the original content. Although the authors cite prior work on reversible-zero watermarking for medical images (Taj et al., 2024), the relative benefits and limitations of their approach compared with zero-watermarking are not sufficiently discussed. Since representation preservation is a central motivation of this work, a comparison or more detailed discussion would help justify the need for reversible modification-based watermarking. In particular, it would be useful to understand the trade-offs in robustness, ownership verification, representation preservation, and applicability to dynamic 3DGS.

  5. Limitations of external baseline comparisons: The paper compares EndoRDH with GS-Marker and 3D-GSW, but their watermark embedding and extraction procedures are adapted to the pretrained EndoGS host rather than evaluated using their original reconstruction pipelines. While the shared evaluation protocol helps control reconstruction-related differences, the reported improvements are specific to these adapted implementations. Additional clarification of how the adaptation affects baseline performance, or comparisons under broader evaluation settings, would strengthen the conclusions about performance relative to existing methods.

Questions for authors
  1. The evaluation is limited to two endoscopic sequences using the pretrained EndoGS representation. Could the authors provide additional experiments on different endoscopic datasets or 3DGS reconstruction architectures to demonstrate the generalizability of EndoRDH? How would variations in illumination, tissue appearance, and camera configurations affect the proposed framework?

  2. While the paper establishes an analytical inverse under ideal conditions, how accurately can the original host representation be recovered in practical settings? Could the authors provide quantitative recovery results (e.g., MSE, PSNR, or SSIM) under finite-precision computation, quantization, and post-embedding model perturbations? It would be particularly useful to understand the trade-off between watermark robustness and host recoverability.

  3. Although CACP restricts modifications to selected spherical harmonic coefficients and preserves protected parameters, how can the authors ensure that clinically relevant information remains unaffected? Could the authors evaluate or discuss the impact of watermark embedding on downstream tasks such as lesion analysis, surgical navigation, or anatomical measurement?

  4. Since preserving sensitive medical information is a primary motivation, could the authors explain why reversible modification-based watermarking is preferable to zero-watermarking approaches that do not modify the original content? Given that the paper cites Taj et al. (2024), could the authors provide a comparative experiment or a more detailed discussion of the trade-offs between these approaches, particularly regarding robustness, ownership verification, representation preservation, and applicability to dynamic 3DGS?

  5. The paper adapts GS-Marker and 3D-GSW to the pretrained EndoGS representation rather than evaluating their original reconstruction pipelines. Could the authors clarify how these adaptations may influence baseline performance and whether comparable embedding budgets, payload capacities, and optimization settings were maintained across methods? How confident can we be that the reported improvements reflect the proposed watermarking framework rather than differences introduced during baseline adaptation?

1 comment

We thank the reviewer for the detailed report. The five weaknesses and five questions are answered in order.

W1. Limited experimental evaluation and generalizability.

The revision adds multiple StereoMIS sequences alongside the two EndoNeRF ones, and they enter the shared-protocol quantitative evaluation, so the main results are aggregate averages over the evaluated sequences. Further endoscope systems, datasets and backbones are future work. EndoRDH reads statistics of a pretrained 3DGS host, but that is not by itself validation on other architectures, so the claim stops at the evaluated EndoNeRF and StereoMIS hosts and scenes. Revision location: §3.1, §4.1; App. F; §5, App. G.

W2. Limited practical evaluation of reversibility.

Section 4.5 (Table 4) reports SH recovery RMSE for the tested mechanisms, with the matched setting in Appendix H. EndoRDH gives 5.2×10−65.2\times10^{-6} under float32, correct key, no post-embedding modification. That is not a guarantee under quantization, model attacks or arbitrary post-processing. The algebraic inverse and practical recovery are kept apart: quantization, finite precision, displacement error and irreversible model-domain changes all shift the residual. Revision location: §4.3; App. B.2; §4.5, Table 4.

VariantPSNRSSIMLPIPSRecovery SH RMSE
EndoRDH34.290.96630.02045.2×10−65.2\times10^{-6}
No ORF prior33.860.95120.03185.4×10−65.4\times10^{-6}
Standard residual embedding31.720.89350.13624.8×10−64.8\times10^{-6}
Direct quantization33.180.92460.04174.7×10−24.7\times10^{-2}

Table 1: Fidelity and authorized SH recovery under the matched setting of Appendix H. Lower RMSE is better.

W3. Lack of downstream clinical task evaluation.

CACP freezes structural Gaussian parameters and excludes the protected region from watermark updates, which is a parameter-level constraint, not proof that segmentation, navigation or measurement outputs are unchanged. The discussion states the limitation and leaves clinical safety to task-specific experiments. Revision location: §3.2; §5, App. G.

W4. Limited comparison with zero-watermarking approaches.

Section 4.6 sets the two families side by side. Zero-watermarking registers a host-derived signature and leaves the host bit-exact; EndoRDH carries an externally supplied payload through reversible edits on admissible SH carriers, verified by a keyed codeword decoded from parameters or renderings. Under float re-saving, SH quantization and primitive removal the comparison is uneven: survival on the zero-watermarking side is untested, and we have tested re-saving but not quantization. Taj et al. (2024) is image-domain with no parameter-level inverse for float primitives, so no numerical comparison is possible; we claim no superiority and a shared-protocol comparison is beyond our evidence. Revision location: §4.6.

W5. Limitations of external baseline comparisons.

3D-GSW and GS-Marker were not run through their original reconstruction codebases. Shared: reconstruction, host, sequences, ROI definition, rendering interface, attack suite and metric pipeline. Not shared: perturbation budget and iteration count, fixed only for matched ablations; payload and ECC are matched throughout. These are controlled transfers, not reproductions; adaptation can only hurt a baseline that depends on geometry or opacity the host does not expose, so the advantage is bounded to this setting. Revision location: §4.1, App. D.

Q1. Only StereoMIS was added, and the claim stays inside it and EndoNeRF. Illumination, tissue appearance and camera configuration were not varied under control, so no invariance is claimed; cross-architecture tests need the host refitted.

Q2. Parameter-domain: 5.2×10−65.2\times10^{-6} SH RMSE under float32 without post-embedding modification, against 4.7×10−24.7\times10^{-2} for quantization, both in Table 1. Rendered-domain recovery and the perturbation curve are not in the evidence, and we do not extrapolate past the measured setting.

Q3. It is not guaranteed by CACP, which is a parameter-level constraint only. Validation on lesion analysis, navigation and anatomical measurement has not been done, and the text says so.

Q4. The two answer different requirements: zero-watermarking gives ownership evidence without modifying the host, while EndoRDH carries a message that an authorized party must later remove. Section 4.6 says where each is preferable, and no superiority is claimed.

Q5. Reconstruction, sequences, ROI, rendering interface and metrics are shared, so reconstruction differences are controlled. Payload, perturbation budget and iteration counts are not asserted equal, only for the matched ablations, so the gains hold for this setting.

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

This paper introduces EndoRDH, a reversible watermarking framework designed for dynamic 3D Gaussian Splatting (3DGS) representations of endoscopic surgical scenes. Protecting the ownership of persistent surgical 3D assets requires embedding metadata without inducing diagnostic artifacts, violating clinical constraints, or permanently altering the underlying tissue representation.

Critical strengths

(+) Rather than adopting standard additive noise or heuristic quantization index modulation (QIM) schemes that often break bijectivity, the paper constructs a mathematically rigorous framework where information transport is realized via bijective additive coupling blocks.

(+) The introduction of RPD explicitly addresses the non-stationarity of endoscopic lighting (inverse-square distance drop-off, tissue specularities, and light-source coupling

Critical weaknesses

(-) The empirical validation is restricted entirely to two sequences (cutting and pulling) derived from the standard EndoNeRF/EndoGS dataset. Endoscopic interventions exhibit high variability across anatomical domains (e.g., laparoscopy, colonoscopy, bronchoscopy), tissue reflectance, camera speeds, and illumination types (narrow-band imaging vs. standard white light). Evaluating the framework on a broader suite of surgical datasets would demonstrate generalizability across varied clinical visual conditions.

(-) While standard visual quality metrics (PSNR, SSIM, LPIPS) confirm perceptual preservation, surgical digital twins are utilized for quantitative downstream tasks such as tissue resection planning, depth estimation, tool tracking, and automated anatomical segmentation. The paper does not provide empirical evidence demonstrating that the SH perturbations leave downstream robotic or clinical measurement models unaffected.

(-) In Appendix G (Table 3), the ablation study groups CACP, RPD, and ORF into a single composite baseline (w/o CACP/RPD/ORF). While the comparison between P-NODE, SFP-NODE, and EndoRDH in Table 1 isolates aspects of RPD, the paper lacks an isolated ablation testing the standalone impact of the reversible flow (ORF) versus standard residual embedding or direct quantization on the same carrier set.

Questions for authors

Please see above.

1 comment

We thank the reviewer for the careful reading, answered below one by one.

Comment 1: Limited empirical validation on only two sequences; broader surgical datasets are needed.

We agree that broader evaluation matters. The original submission used only the two EndoNeRF sequences; the revision adds multiple StereoMIS sequences alongside them. Under the shared host and evaluation protocol, the main quantitative results aggregate the evaluated EndoNeRF and StereoMIS sequences. The new main-text Figure 3 shows two selected StereoMIS scenes, including an instrument view, with further panels in Appendix F. Those figures are qualitative complements to the numbers, not substitutes for them, and residual appearance is not read as an additional metric.

The generalization claim is now bounded to the two evaluated datasets and the controlled protocol. The results give evidence across more scenes and one more endoscopic dataset, but not across every anatomy, endoscope system, illumination regime or reconstruction architecture. SCARED, Hamlyn and C3VD remain future work.

Revision location: Section 4.1 and Appendix F; Section 5 and Appendix G.

Comment 2: Lack of evaluation on downstream clinical tasks.

We agree that downstream evaluation is needed before any clinical utility claim. Quantitative before/after experiments for segmentation, tracking, depth estimation and clinical measurements were not completed in this revision, and we do not offer rendering metrics in their place.

CACP freezes structural Gaussian parameters and excludes the supplied protected carriers from watermark updates. That limits where the watermark can act, but it does not make downstream outputs invariant. The revised text states that task-specific validation is required and lists it as future work.

Revision location: Section 3.2; Section 5 and Appendix G.

Comment 3: Missing isolated ablation of ORF and comparisons against standard residual embedding or direct quantization.

We added a factorized component ablation and an embedding-mechanism comparison, both in Section 4.5 (Tables 3–5), so the numbers sit next to the discussion they support. Removing RPD gives the largest fidelity drop among the component interventions; removing DBRC gives the largest drop in verification accuracy. Removing the OT-inspired prior while keeping the reversible coupling lowers fidelity and verification but keeps recovery error near zero, which separates the prior's allocation role from the coupling's inverse. Appendix H gives the matched-comparison settings.

ConfigurationPSNRSSIMLPIPSClean 2DClean 3DAvg. 2D attackAvg. 3D attack
EndoRDH (full)34.290.96630.020498.4786.2094.6983.69
Without CACP34.460.95720.030195.2883.3792.7680.31
Without RPD31.970.89050.047395.6372.7588.9365.70
Without ORF prior33.860.95120.031893.9880.8592.9478.83
Without DBRC33.610.93270.033678.7262.9370.7859.67

Table 1: Component ablation. Each row removes only the named module or prior. Accuracy values are percentages.

In the revised manuscript, the embedding-mechanism experiment is reported in Section 4.5 (Tables 4–5; reproduced below as Tables 2–3). Under the matched settings described in Appendix H, standard residual embedding has low float32 recovery RMSE but lower fidelity and model-domain verification than EndoRDH. Direct quantization has a much larger recovery error, consistent with its non-bijective mapping. These are results for the reported controlled ablation setting.

Embedding mechanismPSNRSSIMLPIPSRecovery SH RMSE
EndoRDH34.290.96630.02045.2×10−65.2\times10^{-6}
No ORF prior33.860.95120.03185.4×10−65.4\times10^{-6}
Standard residual embedding31.720.89350.13624.8×10−64.8\times10^{-6}
Direct quantization33.180.92460.04174.7×10−24.7\times10^{-2}

Table 2: Rendering fidelity and authorized recovery for the embedding-mechanism ablation. Lower recovery SH RMSE is better.

Embedding mechanismClean 2DClean 3D2D attack avg.3D attack avg.
EndoRDH98.4786.2094.6983.69
No ORF prior93.9880.8592.9478.83
Standard residual embedding93.3156.4790.1253.28
Direct quantization94.0579.6393.8476.91

Table 3: Verification accuracy (%) for the embedding-mechanism ablation.

Revision location: Section 4.5, Tables 3–5; Appendix H.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.