acceptodds
Under review as a conference paper at ICLR 2027

PhysMark: Population-Guided Carrier Learning for Generalizable Physical-Channel Watermarking in 3D Gaussian Splatting

Abstract

Medical volumes reconstructed with 3D Gaussian Splatting (3DGS) become digital assets for case discussion, teaching, and online sharing. To support provenance and copyright tracking during both digital and physical dissemination, a verifiable message can be embedded into the 3DGS representation. However, most existing 3DGS watermarking methods design carriers for a single asset and do not explicitly exploit cross-asset regularities in carrier reliability, which limits generalization to unseen assets. We observe that the relative reliability of candidate carriers transfers across 3DGS assets: carriers that are relatively reliable on a finite set of seen assets tend to remain relatively reliable on unseen ones. Based on this observation, we propose PhysMark, a generalizable physical-channel 3DGS watermarking framework. PhysMark converts this regularity into a learnable, structured carrier prior: it initializes structured sparse carriers from population-level responses and then refines the carrier representation through the watermark recovery objective, turning asset-specific carrier search into prior-guided cross-asset carrier learning. In addition, PhysMark integrates the print-and-capture process into training, allowing the carrier to adapt to distortions introduced by printing, halftoning, and photographing. Experiments on medical 3DGS assets across multiple message lengths, unseen assets, digital-domain transformations, and real print-and-capture conditions show that the carrier reliability ranking transfers well across assets, that the prior-guided carrier improves recovery stability on unseen assets compared with single-asset carrier design, and that practical detectability is achieved under real printing and photographing.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Watermarking, Physical-Channel Robustness, Population Prior, Cross-Instance Generalization

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_nqk8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PhysMark addresses watermark recovery from medical 3DGS assets after digital dissemination and print-and-capture. It constructs local candidate carriers in a registered canonical space, ranks their reliability using response statistics across multiple subjects under digital and simulated physical channels, and guides learnable carriers through structured support masks and initialization. Watermarks are embedded in Gaussian log-scales, with training incorporating a differentiable print-and-capture channel, EOT, and several regularization terms. At the receiver, decoding follows corner-based rectification and extraction of the low-frequency Haar subband.

Critical strengths
  1. The proposed combination of a population prior and learnable residuals is clearly motivated. Separating spatial support, initial carriers, and subsequent learning avoids repeating carrier search for every new asset.

  2. The method directly optimizes recovery after propagation and achieves a favorable fidelity–recovery trade-off on unseen subjects. The authors also use the same physical channel for scene-specific baselines and allow them to be trained separately on each test subject, making the comparison more convincing than one against unadapted digital-domain baselines.

Critical weaknesses
  1. Section 3 assumes stable Gaussian correspondences across instances, and Equation (4) depends on the row-alignment operator Πᵢ. However, the paper does not sufficiently explain how Conditional3DGS establishes these consistent correspondences from T2 MRI, how registration is performed, which subjects are used to construct the population prior, or which quantities remain fixed versus require optimization at test time.

  2. The evidence for real physical-channel performance and practical ownership verification is too limited. Section 4.4 explicitly reports only one unseen subject and a single real photograph. The method recovers 14 of 16 bits, while the best baseline recovers 13. This one-bit difference is insufficient to establish a statistically reliable advantage under real physical-channel conditions.

  3. The paper reports only average Bitacc, without complete-message recovery rates, effective capacity after error correction, false-positive rates on unwatermarked assets or incorrect messages, or detection rates at a fixed false-positive rate.

Questions for authors
  1. Are the prior statistics computed exclusively from training subjects? For test subjects, are B, ρ, the decoder, or any registration parameters updated, and what are the associated costs?
  2. Are results available for multiple unseen subjects and repeated printing/capture? Please report per-sample detection results or results with uncertainty intervals, and specify the devices, paper, capture conditions, and corner-based rectification requirements.
1 comment
AuthorsOpenReview ↗

Response to Reviewer nqk8

Thank you for your detailed review.

W1. Gaussian correspondence and application to new subjects. The pretrained Conditional3DGS model takes an MNI-aligned T2 volume and predicts subject-specific offsets from a shared, ordered Gaussian template. Each reconstructed Gaussian retains its template index, which identifies the corresponding row of the learned carrier in Eq. (4). Given an aligned volume, PhysMark reuses this carrier and completes reconstruction and watermark embedding through forward passes.

W2. Real print-and-capture evaluation. Figure 3 shows one real print-and-capture example from a subject held out from watermark training. PhysMark recovered 14 of 16 bits (87.50%), and the best-performing baseline in that example recovered 13 of 16 bits (81.25%). This example demonstrates partial payload recovery after physical print-and-capture. Table 1 quantifies recovery under simulated print-and-capture conditions. Assessing variability across subjects and capture conditions would require repeated physical captures.

W3. Bit recovery and provenance verification. Bit accuracy measures the fraction of embedded bits recovered correctly, and payload length refers to the number of raw embedded bits. A provenance verifier could compare the decoded sequence with a registered payload using a calibrated Hamming-distance threshold. Evaluating this decision rule would require measuring false accepts on unwatermarked samples and on watermarked samples checked against an incorrect registered payload, then reporting the detection rate at a specified false-positive rate. Exact-message recovery and effective payload capacity after error correction are separate metrics.

Q1. Prior construction and test-time procedure. We compute the population response statistics using subjects from the watermark training split. Final carrier-group selection also uses a brain mask. The source volume for that mask remains unverified. At test time, we load the reconstruction model, carrier BB, and decoder from their checkpoints and set the embedding strength ρ\rho from the evaluation configuration. MNI alignment is a separate preprocessing step. Once the volume is aligned, reconstruction and embedding are forward-pass operations that use these fixed parameters.

Q2. Printing, capture, and rectification. The physical setup uses a standard printer, A4 paper, and artificial lighting. We define the capture angle as the angle between the camera’s optical axis and the paper plane. The tested angles range from 45∘45^\circ (oblique) to 90∘90^\circ (front-on). Before decoding, we use the four printed corner markers to estimate a perspective transform and rectify the image. Figure 3 reports the bit accuracy for one capture from this setup.

Reviewer_GQ8YOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a generalizable 3DGS watermarking framework intended to support carrier transfer across assets and robustness to print-and-capture degradation. Its central idea is to construct a population-informed carrier prior in canonical space and transfer the learned carriers to unseen assets through registration.

Critical strengths
  • Cross-asset carrier learning is a potentially interesting contribution
Critical weaknesses
  • Unclear motivation: The motivation for learning reusable carrier patterns is insufficiently explained, particularly in Lines 53–55. The paper should clarify the limitations of per-asset carrier selection and the practical benefits of a shared carrier prior.

  • Insufficiently established novelty: The novelty of Sections 3.1 and 3.2 is unclear. The potential contribution appears to lie in population-informed carrier selection and cross-asset transfer, rather than in the embedding mechanism itself. The distinction from existing approaches needs to be articulated more clearly.

  • Poor methodological clarity: The presentation relies on lengthy sentences and substantial mathematical notation, while leaving essential definitions and implementation choices unspecified. Notation overloading further obscures the method’s main contributions and limits reproducibility.

  • Insufficient experimental support for carrier learning: The reported ablations do not clearly isolate the benefits of population-informed selection or cross-asset carrier learning.

Questions for authors

Major Questions

  • How exactly is the population response set constructed in Lines 193–196? Please clarify which subjects, rendering views, and channel conditions are used. How are the normalization functions and fusion weights chosen? It is difficult to reproduce the ranking process from Section 3.1 as written.

  • What is the actual benefit of cross-asset carrier learning? The ablation study does not seem to isolate this component. Could you compare your method with a directly learned shared carrier and per-asset carrier selection, particularly on unseen assets?

  • How does the scale modulation Δ𝑠_i connect to 𝐿_{rob} in Section 3.2? I can infer that the connection is through 𝐺_i, but this should be stated explicitly.

  • What print-and-capture conditions are used in Figure 2? Please clarify the settings and how they are sampled.

  • Is the bit accuracy in Table 1 measured before or after print-and-capture? Please make the evaluation condition clear for each reported result.

Notation and Clarity:

  • Could you use different symbols for the region index, response, and rank? Using 𝑟 for all three in Line 152, Eq. (1), and around Line 210 makes the notation difficult to follow.

  • What does 𝑤 represent in Eq. (4)? It appears to be the watermark message, but it is not explicitly defined in this passage.

1 comment
AuthorsOpenReview ↗

Response to Reviewer GQ8Y

Thank you for your detailed review.

W1. Why learn a reusable carrier? Per-asset carrier selection requires rendering and scoring candidate perturbations for each new asset. PhysMark aggregates responses across training subjects to construct a carrier prior, then refines the shared carrier using the watermark recovery objective. For a new aligned asset, the Gaussian template indices identify where to apply the learned carrier, eliminating the need to repeat carrier selection.

W2. Methodological contribution. Sections 3.1–3.2 introduce a population-guided procedure for selecting and initializing a transferable carrier, followed by refinement during training. Log-scale modulation implements embedding by using the learned carrier to map the watermark message to bounded perturbations of the Gaussian scales.

W3, Q3, Q6, Q7. Equations and notation. In Eq. (4), w∈{−1,+1}Kw\in\{-1,+1\}^{K} is the KK-bit watermark message. The perturbation Δsi\Delta s_i updates the Gaussian scales as si′=exp⁡(log⁡si+Δsi)s'_i=\exp(\log s_i+\Delta s_i), producing the watermarked asset Gi′G'_i. Equation (5) renders Gi′G'_i, applies the simulated physical channel, and decodes the watermark from the low-frequency Haar component. The recovery loss supervises the carrier and embedding strengths through these operations. We will define ww next to Eq. (4) and use distinct symbols for the region index, response, and rank.

W4, Q2. Carrier ablation. In the 8-bit analysis in Sec. 4.3, pooling responses across subjects reduced the cross-subject standard deviation from 0.033 to 0.028. In the carrier-refinement comparison, the standard deviation decreased from 0.099 to 0.055, and PSNR increased from 32.86 to 42.62 dB. The full-model aggregate and the two component variants in Table 4 summarize different sets of observations. We will specify the observation set for each row. To isolate the effect of the population prior, the proposed comparisons with a directly learned shared carrier and with per-asset carrier selection would use the same unseen assets and evaluation conditions.

Q1. Population response computation and ranking. We compute population response statistics using subjects from the watermark training split. For each candidate and selected axial view, we apply positive and negative log-scale perturbations and measure the resulting rendering differences in the digital and simulated print-and-capture channels. Within each channel, Eq. (2) combines the population mean, the 10th percentile, and the weakest-subject response, then normalizes the combined statistic. The final score combines the normalized ranks for population stability and response energy. We will specify the subject and view selection rules, the normalization function, and the rank fusion weights.

Q4, Q5. Figure 2 and Table 1. Figure 2(a) shows the watermarked digital rendering, and Figure 2(b) shows the result after simulated print-and-capture. The PhysMark examples in Figure 2 use a fixed evaluation preset. The random sampling described in Sec. 3.2 applies during training. We will identify the evaluation preset in the figure caption. Table 1 reports bit accuracy after simulated print-and-capture. The fidelity metrics compare the original and watermarked digital renderings before the simulated channel is applied.

Reviewer_2pscOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PhysMark, a watermarking framework for medical 3D Gaussian Splatting assets that targets generalization to unseen subjects and recovery after printing and photographing rendered images. Its main contribution is to learn reusable watermark carriers from reliability patterns across subjects and refine them through physical-channel-aware training. Experiments on reconstructed brain MRI assets report improved visual fidelity and watermark recovery over scene-specific baselines, with component ablations and a real print-and-capture demonstration.

Critical strengths
  1. Learning carrier reliability across subjects is a meaningful approach to reducing asset-specific watermark optimization. The combination of shared carrier structure and learned adaptation provides a coherent way to exploit population information, and the carrier ablation supports its usefulness for recovery.

  2. The method considers physical dissemination during training and reports improvements in both fidelity and recovery across multiple payload lengths. The use of physical-channel training for the compared baselines makes these results more informative than a comparison against methods trained only for clean renderings.

Critical weaknesses
  1. The evidence for generalization remains limited. The method relies on stable Gaussian correspondences across independently reconstructed subjects, but their construction and the adaptation required for a new asset are insufficiently explained. Evaluation on one brain MRI dataset, without comparisons to the cited generalizable methods GS-Marker and MarkSplatter, leaves unclear whether the benefit extends beyond this particular aligned setting or advances existing generalizable watermarking.

  2. The physical evaluation does not establish reliable provenance verification. The real experiment uses one unseen subject and a single photograph, with errors remaining in the decoded message. Most results concern simulated distortions and average bit accuracy, without establishing successful identification or false-positive behavior. The gap between recovering a watermark signal and making a dependable provenance decision is therefore unresolved.

Questions for authors
  1. How are Gaussian correspondences established for a new subject, and does embedding require any optimization on that subject? Were the template and population prior constructed exclusively from the training subjects?

  2. How does the system turn a message containing bit errors into a provenance decision, and what evidence establishes the reliability of that decision on unwatermarked assets or incorrect identities?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 2psc

Thank you for your detailed review.

W1. Cross-subject generalization. Table 1 evaluates whether a single learned carrier can be reused across 100 held-out IXI subjects with MNI-aligned T2 MRI. PhysMark reuses the same carrier across subjects. GaussianMarker, 3D-GSW, and GuardSplat are trained separately for each test subject using the same simulated print-and-capture channel.

W2. Real print-and-capture evaluation. Figure 3 shows one photograph of a printed rendering from a held-out subject. PhysMark correctly recovered 14 of 16 embedded bits (87.50%), and the best-performing baseline for that capture recovered 13 of 16 bits (81.25%). These per-capture results measure partial payload recovery through the physical pipeline.

Q1. Gaussian correspondence, test-time embedding, and training data. The pretrained Conditional3DGS model reconstructs each MNI-aligned T2 volume by predicting subject-specific offsets from a shared, ordered Gaussian template. Each reconstructed Gaussian retains its template index, which identifies the corresponding carrier row in Eq. (4). Once a new volume is aligned, PhysMark uses the fixed carrier and performs reconstruction and watermark embedding through forward passes. The population response statistics are computed exclusively from subjects in the watermark training split.

Q2. Interpreting a decoded message with bit errors. In Figure 3, the decoded sequence differs from the embedded message by two bits, giving a Hamming distance of 2. For a 16-bit payload, this corresponds to 1−2/16=87.50%1-2/16=87.50\% bit accuracy. The reported value quantifies agreement with the known embedded payload for that capture.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.