acceptodds
Under review as a conference paper at ICLR 2027

GuardMarkGS: Unified Ownership Tracing and Edit Deterrence for 3D Gaussian Splatting

Abstract

3D Gaussian Splatting (3DGS) is becoming a practical representation for novel view synthesis, but its growing adoption, together with rapid advances in instruction-driven 3DGS editing, also exposes a dual copyright risk: once released, a 3DGS-based asset can be used without permission and manipulated through 3D editing. Existing protection methods address only one side of this problem. Watermarking can trace ownership after unauthorized use, but it cannot prevent malicious editing. Adversarial edit-deterrence methods can disrupt editing, but they do not provide evidence of ownership. To the best of our knowledge, we present the first unified protection framework for 3DGS that jointly optimizes ownership tracing and unauthorized editing deterrence. Our framework combines a scene-wide watermarking objective over all Gaussians with an adversarial edit-deterrence objective. The adversarial branch combines latent-anchor separation, denoising-trajectory diversion, and cross-attention diversion to divert the editing trajectory, while an update-saliency-motivated Gaussian selection strategy assigns stronger adversarial updates to mask-selected Gaussians, improving the balance among watermark recovery, edit deterrence, and rendering fidelity. Experiments on scenes from Mip-NeRF 360 and Instruct-NeRF2NeRF demonstrate that the framework achieves a favorable balance among bit accuracy, edit deterrence, and rendering quality. These results suggest that practical copyright protection of 3DGS-based assets can be more effectively addressed by integrating ownership tracing and unauthorized editing deterrence into a single optimization framework.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
3D Gaussian Splatting, copyright protection, digital watermarking, adversarial perturbations, 3D scene editing

Reviews

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

This paper introduces GuardMarkGS, a unified copyright-protection framework for 3D Gaussian Splatting (3DGS) that jointly optimises two previously separate goals: (1) ownership tracing through watermarking and (2) deterrence of unauthorised text-driven 3D editing. The core contribution is a dual-branch optimisation framework combining scene-wide watermark embedding with adversarial edit-deterrence mechanisms, including update-saliency-guided Gaussian selection, parameter-role-aware modulation, latent-anchor separation, denoising-trajectory diversion, and cross-attention diversion.

The paper addresses a timely and practically relevant problem. The proposed formulation is novel relative to prior work that treats watermarking and edit protection independently. Experimental results consistently show that the method outperforms individual watermarking baselines and a naïve combination baseline on the authors' unified metric (sUCPS), while maintaining strong watermark recovery and edit-deterrence performance. Table 1 reports a 10.7% improvement over the strongest baseline according to the proposed unified score.

The major concern is not whether the method works, but whether the evidence is sufficient for a top-tier venue such as ICLR. The evaluation remains relatively limited in scale, relies heavily on a newly introduced composite metric (sUCPS), and lacks stronger analyses against adaptive attackers, larger datasets, and broader editing ecosystems.

Critical strengths

Strengths

  1. Clear and novel problem formulation The paper identifies a genuine gap in the literature: watermarking and editing defence have largely evolved as separate research directions. The unified formulation appears original and well motivated.

  2. Strong practical motivation The threat model is realistic. Text-guided editing systems increasingly enable repurposing of released 3D assets, creating copyright and ownership concerns that are not adequately addressed by passive watermarking alone.

  3. Well-designed optimisation strategy The paper recognises the conflict between: scene-wide watermark embedding localised adversarial perturbations and proposes mechanisms specifically intended to reconcile them. This is more sophisticated than simply summing losses.

  4. Strong empirical evidence against naïve composition The 3DGSW+DEGauss baseline is particularly valuable. It demonstrates that joint optimisation is non-trivial and that performance gains are not merely obtained by adding existing objectives together. Table 1 shows severe bit-accuracy degradation (62.79%) for the naïve combination compared with 97.23% for GuardMarkGS.

  5. Comprehensive ablation study The authors systematically analyse: Gaussian selection Role-aware modulation Diversion-term combinations Hard vs soft masking Perturbed parameter groups Tables 2–3 and appendices provide convincing support for the design decisions.

  6. Good paper organisation and clarity The manuscript is generally easy to follow. Figures effectively communicate the framework and motivation. Figure 1 clearly illustrates the application scenario. Figure 3 provides useful intuition for the saliency-guided Gaussian selection strategy.

  7. Transferability experiments The GaussianEditor and SDXL-backbone evaluations strengthen the claim that the defence generalises beyond the exact optimisation environment.

Critical weaknesses

Major Weaknesses

  1. Evaluation scale is relatively small The entire quantitative evaluation uses only six scenes:
  • bicycle --garden -face
  • person
  • bear
  • fangzhou T his raises concerns regarding statistical robustness and generalisability. For an ICLR-level contribution, broader evaluation across more diverse scene types would substantially strengthen confidence.
  1. Heavy dependence on the proposed sUCPS metric The headline improvement is largely expressed through sUCPS. However: sUCPS is introduced by the authors. Weighting and aggregation choices are subjective. Harmonic-mean aggregation may favour particular trade-offs. Appendix H formally defines sUCPS, but there is limited justification that the metric corresponds to real-world copyright utility. A sensitivity study on metric design choices would improve credibility.

  2. Limited adaptive attacker analysis The robustness experiments mainly evaluate: noise cloning pruning simple distortions

Table 8 demonstrates resilience under these settings.

However, a knowledgeable attacker aware of GuardMarkGS might:

  • perform adversarial fine-tuning
  • optimise explicitly against watermark recovery
  • reverse engineer mask-selected regions
  • etrain edited scenes from rendered images
  • Such adaptive attacks are not thoroughly examined.
  1. Missing computational cost analysis The paper introduces multiple optimisation components: -watermark optimisation -latent diversion -trajectory diversion -cross-attention diversion -saliency masking

but provides little discussion of: -training time increase -memory overhead -optimisation stability

This omission is important for practical deployment.

  1. Threat model remains somewhat narrow The threat model focuses heavily on instruction-driven editing pipelines. Many realistic misuse scenarios involve: -mesh conversion -scene distillation -Gaussian-to-mesh reconstruction -neural-to-neural transfer

The paper acknowledges this limitation in the conclusion and future work.

  1. Minor Weaknesses

  2. Lack of statistical significance reporting The paper reports averages but no variance, confidence intervals, or repeated-run statistics.

  3. Hyperparameter sensitivity is not explored Important parameters include: λadv λlat λtraj λxattn τmask γmask The robustness of performance to these settings is unclear.

  4. Some qualitative comparisons are limited The paper would benefit from more failure-case analysis showing situations where edit deterrence does not work.

  5. Limited comparison against newer watermarking paradigms The comparison suite is reasonable, but watermarking research is evolving rapidly. Discussion of broader watermarking directions could strengthen related work positioning.

  6. Reproducibility still depends on appendices The reproducibility statement is encouraging, but full reproducibility requires release of: code pretrained models evaluation scripts These are not yet available in the submission

Questions for authors
  1. How robust is GuardMarkGS against an adaptive attacker who explicitly tries to remove both the watermark and the edit-deterrence signal? The paper evaluates random noise, pruning, cloning, and image/model distortions, but not a knowledgeable adversary that knows GuardMarkGS is being used and actively optimises against it. The central claim is practical copyright protection, so the strongest threat model is an adaptive attacker.

  2. How much of the reported gain depends on the newly proposed sUCPS metric? The headline result is that GuardMarkGS achieves the highest sUCPS (0.8622 versus 0.7791 for the strongest baseline), but sUCPS is designed by the authors and combines traceability, deterrence, and fidelity through several normalisation and aggregation choices.

  3. Does the method generalise beyond the six evaluated scenes and the specific editing pipelines studied? The main benchmark consists of six scenes from Mip-NeRF 360 and Instruct-NeRF2NeRF, with evaluation primarily on DGE and GaussianEditor.

3 comments

We sincerely thank the reviewer for the positive assessment of our work and for recognizing the novelty and practical significance of jointly optimizing ownership tracing and edit deterrence for 3DGS. We appreciate the constructive suggestions and address the main concerns below.

Q1 & W3. Robustness against adaptive attacks

We agree that evaluating stronger attacks beyond random noise, cloning, and pruning is important.

To address this concern, we additionally evaluated a re-render-and-reconstruct attack across all six evaluation scenes. Specifically, we rendered multi-view images from the protected GuardMarkGS assets, constructed fresh COLMAP sparse reconstructions from these renders, and trained new vanilla 3DGS. This procedure did not reuse the original clean training images, original sparse point clouds, protected Gaussian parameters, or optimizer states.

EvaluationMetricProtected GuardMarkGSFresh reconstructionChange after reconstruction
Ownership tracingBit accuracy (%) ↑97.2396.70−0.53 pp
Edit deterrenceCLIP Method ↓0.90930.9103+0.0010
CLIP Diff. ↑0.09070.0897−0.0010
CLIP-T Method ↓0.26760.2678+0.0002
CLIP-T Diff. ↑0.01490.0147−0.0002
CLIP-D Method ↓0.10860.1153+0.0067
CLIP-D Diff. ↑0.03900.0323−0.0067
Fidelity to clean 3DGSPSNR ↑30.3628.92−1.44 dB
SSIM ↑0.89350.8496−0.0439
LPIPS ↓0.14710.1799+0.0328
Reconstruction fidelityReconstructed vs. protected PSNR ↑N/A35.76N/A
Reconstructed vs. protected SSIM ↑N/A0.9479N/A
Reconstructed vs. protected LPIPS ↓N/A0.0787N/A

Importantly, the reconstructed models remain visually faithful to the protected assets (PSNR: 35.76 dB, SSIM: 0.9479, LPIPS: 0.0787), confirming that the reconstruction itself was successful.

Despite rebuilding the 3D representations, watermark bit accuracy decreases by only 0.53 percentage points (97.23% → 96.70%), while the CLIP, CLIP-T, and CLIP-D deterrence gaps retain 98.9%, 98.7%, and 82.8% of their original values, respectively.

These findings demonstrate that both ownership tracing and substantial edit-deterrence effects persist even after the original Gaussian representation is discarded and reconstructed, providing additional robustness evidence beyond the model distortions evaluated in Appendix F.

Q2 & W2. Motivation and sensitivity of sUCPS

We appreciate the concern regarding the design choices of sUCPS. However, we would like to clarify that sUCPS is primarily intended as a summary metric designed to reflect the three protection requirements motivated by Fig. 1 and formally defined in Sec. 3 (Eqs. 1–3): ownership traceability, edit deterrence, and rendering fidelity.

Importantly, the advantages of GuardMarkGS are observable directly from the individual metrics in Table 1. Watermark-only methods, such as 3DGSW [5], provide strong ownership tracing but do not explicitly deter editing, while DEGauss [6] provides edit deterrence without ownership verification. Moreover, straightforward composition (3DGSW+DEGauss) substantially reduces bit accuracy to 62.79%. In contrast, GuardMarkGS maintains 97.23% bit accuracy while achieving stronger edit deterrence (CLIP Diff.: 0.0907) with competitive rendering fidelity.

Thus, sUCPS is intended to summarize this multi-objective trade-off, rather than to establish the contribution through a newly designed metric alone.

Regarding its mathematical construction, Appendix H employs geometric means within the edit-deterrence and fidelity dimensions, followed by a harmonic mean across the three protection dimensions. This final aggregation intentionally penalizes weak protection dimensions, consistent with our objective of satisfying all three requirements simultaneously.

To further examine sensitivity to the aggregation choice, we recomputed the unified scores using alternative final aggregation functions while retaining the same normalized component scores:

AggregationGuardMarkGSBest Baseline
Harmonic Mean (Original)0.86220.7791
Geometric Mean0.87590.7992
Arithmetic Mean0.88880.8328

GuardMarkGS remains the highest-ranked method under all three aggregation choices. These results indicate that its overall advantage is not solely dependent on the harmonic-mean formulation.

Q3 & W1. Evaluation scale and generalizability

We would like to clarify that our evaluation adopts a six-scene benchmark scale comparable to that of DEGauss [6], the most directly relevant prior work on adversarial protection against unauthorized 3DGS editing.

Importantly, our scene selection is not arbitrary. We include spatially extended, unbounded scenes from Mip-NeRF 360 and object- and subject-centric scenes from Instruct-NeRF2NeRF. These settings provide diverse scene structures, background complexity, and semantic editing targets.

Furthermore, our evaluation extends beyond the main DGE experiments through two complementary transfer settings:

  • Cross-editor transfer: GaussianEditor (Appendix D).
  • Cross-backbone transfer: An SDXL-based editing backbone within the DGE pipeline (Appendix E).

These experiments demonstrate that the edit-deterrence effect is not restricted to a single editing pipeline or diffusion backbone.

Additional Clarifications: Deployment, Scope, and Implementation Details

Computational cost. GuardMarkGS is a one-time post-training protection stage rather than a reconstruction-from-scratch method. We measured optimization times on a single RTX A6000:

MethodObject-centric: bear (min) ↓Unbounded: garden (min) ↓
DEGauss [6]218.74231.83
GuardMarkGS (Ours)215.52 (−1.47%)235.10 (+1.41%)

These results indicate that GuardMarkGS provides both protection functions with optimization time comparable to DEGauss [6], which provides edit deterrence alone.

Threat model. We would like to clarify that GuardMarkGS specifically targets unauthorized diffusion-guided, instruction-driven 3DGS editing, as formulated in Sec. 3 (Eq. 2) and Appendix B, with its scope explicitly clarified in Sec. 6 (Limitations and future work).

Regarding mesh conversion and Gaussian-to-mesh reconstruction, SuGaR [1] primarily focuses on enabling mesh-based editability by constructing a mesh representation from 3DGS, rather than performing diffusion-guided semantic editing. Such representation-conversion approaches differ from DGE [2], which employs diffusion-guided multi-view editing followed by Gaussian optimization to realize semantic editing instructions.

Our goal is not to prevent every possible form of 3D asset manipulation, but to jointly optimize ownership tracing and edit deterrence within the defined threat model. Within this scope, we additionally evaluate transferability beyond DGE [2] using GaussianEditor [3] and an SDXL-based editing backbone [4] (Appendices D and E).

We acknowledge that other representation transformations beyond the re-render-and-reconstruct setting evaluated above remain outside the scope of our current evaluation.

Related work and implementation details. Recent generalizable watermarking methods, such as MarkSplatter [7], employ pretrained embedding networks rather than the scene-wise optimization considered in our joint framework. This distinction is now clarified in the revised Sec. 2 (Related Work). We have also expanded Appendix C with optimization settings and implementation details, supplemented by Appendix I (trajectory definitions) and Appendix N (prompt libraries), to facilitate reproducibility. We intend to release our code and evaluation scripts upon acceptance.

We sincerely appreciate the reviewer's constructive suggestions, which have helped us clarify the evaluation scope and strengthen the experimental evidence supporting GuardMarkGS.

[1] SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering, CVPR 2024.
[2] DGE: Direct Gaussian 3D Editing by Consistent Multi-view Editing, ECCV 2024.
[3] GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splatting, CVPR 2024.
[4] SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis, ICLR 2024.
[5] 3D-GSW: 3D Gaussian Splatting for Robust Watermarking, CVPR 2025.
[6] DEGauss: Defending Against Malicious 3D Editing for Gaussian Splatting, NeurIPS 2025.
[7] MarkSplatter: Generalizable Watermarking for 3D Gaussian Splatting Model via Splatter Image Structure, ACM MM 2025.

Reviewer_YpgyOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies joint watermarking and edit deterrence for released 3D Gaussian Splatting assets. It optimizes a scene-wide watermark objective alongside latent, denoising-trajectory, and cross-attention diversion objectives. A fixed soft semantic mask allocates the adversarial gradients across Gaussians, and parameter-role coefficients scale different parameter groups. Experiments on six scenes report a useful trade-off between pre-edit watermark recovery, rendering fidelity, and CLIP-based editing deterrence, with additional transfer, distortion, and component studies.

Critical strengths
  1. Jointly considering watermark recovery and edit deterrence is a useful problem formulation, and the paper demonstrates that allocating the adversarial updates spatially can alleviate the interference between these objectives. With parameter-role modulation fixed, adding Gaussian selection raises bit accuracy from 65.02% to 97.23% and PSNR from 28.74 to 30.36 dB (Table 2). The full model also improves the reported DGE deterrence metrics over DEGauss while retaining watermark decoding, which DEGauss does not provide (Table 1).

  2. The evaluation goes beyond the main DGE experiment: it includes transfer to GaussianEditor and an SDXL-based editor, image- and model-level distortion tests, diversion-term ablations, a hard/soft-mask comparison, and parameter-group ablations (Appendices D-F and I-K). These studies are relevant to the proposed design and provide useful evidence that the observed effect is not confined to a single editor configuration.

Critical weaknesses
  1. The central adversarial mechanism is not specified sufficiently to reproduce or assess the claimed improvement. In particular, Phi_traj in Eq. (10) is described as summarizing a prompt-induced edit direction and a short denoising trajectory, but the actual descriptor, trajectory length, timestep selection, and normalization are not defined. This is consequential because Table 3 identifies the trajectory term as the strongest individual diversion term. Appendix I does define the cross-attention descriptor, but it does not resolve this separate omission. Appendix C provides several loss weights and training settings, yet not lambda_adv, the parameter-role coefficients, the mask threshold/exponent, or the optimization prompt library and its overlap with the evaluation prompts. Algorithm 1 treats these as supplied inputs. The paper's central novelty is an allocation and balancing mechanism, so its exact coefficients, construction, and tuning protocol are necessary to determine what was evaluated and to reproduce it; these are not merely peripheral implementation details.

  2. The reported trade-off is promising, but the evidence for its advantage over a properly tuned composition is incomplete. Table 2 convincingly shows that the selected allocation helps at the reported setting, yet the comparison to 3DGSW+DEGauss is a single operating point without a reported balance-weight sweep or comparison at matched rendering fidelity. Thus, the conclusion that straightforward composition cannot obtain joint protection is stronger than the evidence establishes. The marginal role-modulation gains are also small (97.02% to 97.23% bit accuracy and 30.22 to 30.36 dB), with no repeated-run uncertainty. Finally, the practical framing should distinguish the formal pre-edit tracing task in Eq. (1) from tracing after attempted manipulation: Appendix L/Table 10 reports post-edit recovery for three baselines but omits GuardMarkGS. This omission does not contradict Eq. (1), but it leaves the broader claim of preserved ownership evidence after an edit-update process unverified. These limitations warrant more careful attribution and scoping of the claimed protection advantage.

Questions for authors
  1. Please specify the exact Phi_traj construction, including trajectory length, timesteps, descriptors, and normalization, together with lambda_adv, the role coefficients, mask parameters, and optimization prompt library. Are Table 4's evaluation prompts held out, and were the same protected assets used unchanged in the GaussianEditor and SDXL transfer tests?

  2. Can the authors report the 3DGSW+DEGauss trade-off when its objective balance is tuned at comparable rendering fidelity and optimization cost? Per-scene results or repeated runs would also help determine whether the small role-modulation improvements are consistent.

  3. What is GuardMarkGS's post-edit watermark recovery under the same DGE protocol as Table 10? If preservation after an attempted edit is outside the intended tracing requirement, please make that scope explicit in the practical ownership-protection claims.

2 comments

We sincerely thank the reviewer for the thoughtful assessment of our work and for recognizing the novelty and effectiveness of our approach. We appreciate the constructive suggestions and address each concern below.

Q1 & W1. Implementation details and reproducibility

We thank the reviewer for highlighting the importance of implementation details. We provide the following clarifications, which are now incorporated into the revised Appendices C, I, and N.

Denoising-trajectory diversion (Eq. 10). Our trajectory diversion combines two terms: (1) an edit-direction term, defined as the squared distance between ℓ2\ell_2-normalized prompt-induced edit directions (ϵtext−ϵimage\epsilon_{\mathrm{text}} - \epsilon_{\mathrm{image}}), and (2) a clean-latent term, defined as the mean squared distance between predicted clean latents.

To capture editing responses at different noise levels, we evaluate the two terms at different diffusion timesteps. For the latter, we obtain a single clean-latent prediction from each original image latent noised with the same sampled noise. Intuitively, the first term targets how the instruction steers the editor, while the second targets the clean-latent outcome the editor predicts.

We select one of the intervals [250, 399], [400, 899], and [900, 998] with probabilities 0.1, 0.3, and 0.6, respectively. For each term, protected and reference renders share the same prompt, timestep, and sampled noise. The adversarial objective is averaged across six sampled training views, as formulated in Eq. (8). The revised Appendix I (Eqs. 30–32) now provides a mathematical definition of Φtraj\Phi_{\mathrm{traj}}, specifying its edit-direction and clean-latent components, their normalization, and the timestep construction.

Optimization settings. We use λadv=0.2\lambda_{\mathrm{adv}} = 0.2 and parameter-role coefficients ρk=1.0\rho_k = 1.0 (scale/rotation), 0.550.55 (opacity), 0.700.70 (DC color), and 0.950.95 (remaining SH coefficients). Gaussian positions are excluded from adversarial updates. The soft-mask parameters are τmask=0.7\tau_{\mathrm{mask}} = 0.7 and γmask=0.5\gamma_{\mathrm{mask}} = 0.5.

Optimization prompts and transferability. Each scene is independently optimized using its own prompt library containing approximately 45 instructions, uniformly sampled during optimization. Importantly, the evaluation prompts in Table 4 are held out from the corresponding scene's optimization library. Thus, no evaluation prompt is used to optimize the protected 3DGS on which it is evaluated. The scene-wise optimization prompt libraries are now provided in Appendix N (Tables 11–16), while their sampling and held-out evaluation protocols are specified in Appendix C.

Furthermore, as clarified in the revised Appendix C, the GaussianEditor and SDXL transfer experiments use exactly the same protected 3DGS assets as the main DGE experiments, without additional protection optimization, fine-tuning, or hyperparameter adjustment. Only the editing pipeline or diffusion backbone is changed, ensuring a direct evaluation of transferability.

Q2 & W2. Tuned composition baseline and role-modulation consistency

We appreciate the reviewer's suggestion to further examine whether tuning the objective balance of 3DGSW+DEGauss could improve its joint protection performance.

Additional baseline tuning. We conducted an additional sweep of the adversarial objective weight (where larger weights place greater emphasis on optimizing adversarial perturbations) for 3DGSW+DEGauss, using scaling factors c∈0.5,1,2,4c \in \\{0.5, 1, 2, 4\\} relative to its original weights.

MethodBit Acc. (%) ↑PSNR ↑CLIP Diff. ↑
3DGSW+DEGauss (c=0.5)66.6429.890.0393
3DGSW+DEGauss (c=1)62.7929.900.0404
3DGSW+DEGauss (c=2)59.7629.920.0412
3DGSW+DEGauss (c=4)57.8629.930.0417
GuardMarkGS97.2330.360.0907

Across the evaluated settings, increasing adversarial strength substantially reduces watermark accuracy while yielding only marginal changes in rendering fidelity and edit deterrence. Although an operating point with fidelity comparable to that of GuardMarkGS was not obtained, these results provide further evidence that adjusting the objective balance alone does not resolve the interference between watermarking and adversarial protection.

Role-modulation consistency. We additionally conducted three paired experiments with different random seeds, keeping Gaussian selection enabled and varying only parameter-role modulation. The resulting improvements (mean ± sample SD) were consistent across all three runs: +0.217 ± 0.012 pp in bit accuracy, +0.133 ± 0.012 dB in PSNR, +0.01367 ± 0.00122 in SSIM, and −0.00313 ± 0.00012 in LPIPS. The sUCPS improvement was +0.00550 ± 0.00026.

These results further support our intended design: Gaussian selection provides the principal improvement in balancing the conflicting objectives, while parameter-role modulation offers a smaller but consistent refinement.

Q3 & W2. Post-edit watermark recovery and the intended protection scope

We thank the reviewer for highlighting the distinction between pre-edit ownership tracing and post-edit watermark recovery. As the reviewer correctly observes, this distinction is intentional in our problem formulation, as illustrated in Fig. 1 and formally defined in Sec. 3 (Eqs. 1–2) and Appendix B.

We would like to emphasize that Table 10 in Appendix L was not intended to establish post-edit watermark robustness. Rather, as explicitly discussed in Appendix L, this experiment highlights an important limitation of existing watermark-only methods: instruction-driven editing can substantially degrade watermark recoverability. Moreover, even when a watermark survives editing, watermarking remains a passive protection mechanism that provides post-hoc ownership evidence without actively deterring unauthorized manipulation.

This observation directly reinforces our central motivation: ownership tracing and active edit deterrence address complementary copyright-protection objectives and should be jointly considered, rather than relying exclusively on either protection mechanism.

For completeness, we additionally report GuardMarkGS's post-edit watermark accuracy under the same DGE protocol as Table 10. GuardMarkGS achieves 86.61% after editing, compared with 97.23% before editing. This result is now included in Table 10 (Appendix L) of the revised manuscript. However, our framework does not claim guaranteed watermark preservation after an edit-update process. Instead, it jointly optimizes ownership tracing from the protected representation and deterrence of unauthorized instruction-driven editing.

We sincerely appreciate the reviewer's constructive feedback, which has helped us further refine and strengthen our work.

Reviewer_hdv9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GuardMarkGS. It is a unified 3DGS copyright protection framework that jointly embeds traceable watermarks and adversarial perturbations to block text-guided diffusion 3D editing. It contributes saliency-aware Gaussian selection, parameter-aware adversarial modulation and the sUCPS metric. Experiments verify its balanced performance in watermark recovery, editing deterrence and rendering fidelity.

Critical strengths

1.It jointly optimizes traceable watermarks and adversarial perturbations for 3DGS, and demonstrates that directly combining separate methods yields poor performance. 2.It proposes saliency-aware Gaussian selection, parameter-aware adversarial modulation and the sUCPS metric, supported by ablation and robustness experiments.

Critical weaknesses

1.The defense only targets text‑guided diffusion‑based 3D editing, with no evaluation of alternative modifications such as direct Gaussian parameter editing and mesh‑driven re‑baking. 2.Watermark extraction depends on rendered images; ownership verification cannot be performed if attackers distribute only raw Gaussian files. 3.Relative to the watermark‑only baseline 3DGSW, GuardMarkGS exhibits reduced bit accuracy (97.23% vs 99.00%) and decreased PSNR. Despite obtaining the highest sUCPS, its quality and accuracy overhead requires further discussion for high‑fidelity commercial scenarios.

Questions for authors

1.The paper lacks evaluation of the defense against non-diffusion modifications such as direct Gaussian parameter editing and mesh-based re-baking. 2.The paper does not present a watermark extraction method that operates directly on raw Gaussian files without rendered images. 3.The paper relies only on CLIP-based metrics and does not include human evaluation for the visual effect of editing suppression.

4 comments

We sincerely thank the reviewer for the positive assessment of our work and for recognizing the novelty of jointly optimizing ownership tracing and edit deterrence for 3DGS. We appreciate the constructive comments and address each concern below.

Q1 & W1. Defense against non-diffusion 3D editing

We agree that direct Gaussian parameter manipulation and mesh-based editing are relevant 3D asset manipulation paradigms. However, we would like to clarify that GuardMarkGS specifically targets unauthorized diffusion-guided, instruction-driven 3DGS editing, as formulated in Sec. 3 (Eq. 2) and Appendix B (Threat Model and Application Scenario), with its scope explicitly clarified in Sec. 6 (Limitations and future work).

While Gaussian Grouping [1] supports direct Gaussian manipulation for relatively simple operations such as object removal, relocation, and recoloring, more complex semantic editing, such as instruction-driven style transfer, relies on diffusion-based guidance followed by Gaussian optimization. This is also consistent with DGE [3], which employs diffusion-guided multi-view editing to reflect the user's semantic intent in 3DGS. Meanwhile, SuGaR [2] primarily focuses on enabling mesh-based editability by constructing a mesh representation from 3DGS, rather than performing diffusion-guided semantic editing. These approaches therefore address different editing mechanisms from those targeted by our adversarial objectives.

Importantly, adversarial protection against unauthorized 3DGS editing is itself a relatively recent research direction [7]. Our central motivation is to address an important gap between two previously separate protection paradigms: passive ownership tracing through watermarking and active edit deterrence through adversarial optimization. Rather than attempting to prevent every possible form of 3D asset manipulation, GuardMarkGS introduces a unified framework that jointly optimizes these complementary protection objectives within a single 3DGS representation while balancing traceability, deterrence, and rendering fidelity.

Within our defined threat model, we have additionally evaluated transferability beyond DGE [3], using GaussianEditor [4] and an SDXL-based editing backbone [5] (Appendices D and E), demonstrating that the proposed protection is not restricted to a single diffusion-based editing pipeline.

We acknowledge that arbitrary direct parameter modifications and mesh-based reconstruction fall outside the current edit-deterrence scope. We have further clarified this distinction and explicitly discussed these scenarios in Sec. 6 (Limitations and future work) of the revised manuscript.

Q2 & W2. Ownership verification from raw Gaussian files

We would like to clarify that our ownership verification does not require attackers to distribute rendered images.

Even when only a raw 3DGS asset is redistributed, the rightful owner can obtain the suspected asset, render it from selected viewpoints, and recover the embedded watermark using the pretrained decoder.

This procedure is already illustrated in Fig. 1 and explicitly described in Appendix B (Ownership Verification). Moreover, our scene-wide watermarking objective (Sec. 4.2) is specifically designed to support watermark recovery across diverse rendered viewpoints.

Therefore, even when only raw Gaussian files are redistributed, ownership verification remains fully supported through rendering-based watermark extraction, consistent with our intended application scenario.

W3. Trade-off between watermark accuracy, rendering fidelity, and edit deterrence

We acknowledge the relative differences in bit accuracy and PSNR between GuardMarkGS and the watermark-only baseline 3DGSW [6].

However, we respectfully emphasize that our primary objective is not to outperform watermark-only methods on individual watermarking or fidelity metrics, but to jointly satisfy three protection requirements: traceability, deterrence, and fidelity, as formulated in Sec. 3 (Eqs. 1–3).

3DGSW primarily focuses on ownership tracking and rendering fidelity, offering little to no defense against unauthorized editing. In contrast, GuardMarkGS additionally addresses edit deterrence while maintaining 97.23% bit accuracy, achieving a larger edit-deterrence gap (CLIP difference: 0.0907 vs. 0.0371) and the highest unified protection score (0.8622 vs. 0.7791), as reported in Table 1.

Furthermore, Table 2 demonstrates that our Gaussian selection strategy substantially alleviates the interference between watermarking and adversarial protection, improving both bit accuracy and rendering fidelity under joint optimization.

We also note that PSNR measures rendering fidelity relative to the original pretrained 3DGS, rather than defining an absolute quality threshold (Appendix G). While our primary focus is on jointly optimizing the three protection objectives, we agree that further improving performance along each individual axis would be valuable. We have highlighted the opportunity to further improve watermark recovery and rendering fidelity in Sec. 6 (Limitations and future work) of the revised manuscript.

Q3. Human evaluation of editing deterrence

We thank the reviewer for suggesting human perceptual evaluation.

Following the established evaluation protocol adopted by DEGauss [7], we employ three complementary CLIP-based metrics that assess editing deterrence from different perspectives:

  • CLIP: Similarity between the protected and unprotected editing results.
  • CLIP-T: Semantic alignment between the edited result and the target editing instruction.
  • CLIP-D: Alignment between the image-editing direction and the corresponding text-editing direction.

These metrics are already formally defined in Appendix G. Nevertheless, we acknowledge that all three rely on CLIP representations and therefore share certain limitations.

To address the reviewer's concern, we additionally conducted a human evaluation covering all six scene settings. Five raters with ML/DL backgrounds evaluated anonymized pre- and post-edit results from GuardMarkGS, DEGauss, and 3DGSW. Method identities were hidden, and presentation order was randomized. The corresponding unprotected editing results were provided as references.

Raters assessed how successfully each result reflected the intended editing instruction on a five-point scale (1: not reflected at all, 5: clearly and successfully reflected). Lower scores indicate stronger edit deterrence.

MethodMean edit-success ↓Median [IQR] ↓
GuardMarkGS1.962 [1–3]
DEGauss2.362 [2–3]
3DGSW4.164 [4–5]

The overall ordinal Krippendorff's α\alpha was 0.8043, indicating consistent judgments among raters. GuardMarkGS achieved a mean edit-success score 0.40 points lower than DEGauss and 2.20 points lower than 3DGSW.

These human judgments are consistent with our CLIP-based results and provide complementary perceptual evidence that GuardMarkGS more effectively suppresses intended semantic edits while preserving recoverable ownership information.

We sincerely appreciate the reviewer's constructive suggestions, which have helped us clarify the scope of our contributions and strengthen the experimental evaluation.

[1] Gaussian Grouping: Segment and Edit Anything in 3D Scenes, ECCV 2024.
[2] SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering, CVPR 2024.
[3] DGE: Direct Gaussian 3D Editing by Consistent Multi-view Editing, ECCV 2024.
[4] GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splatting, CVPR 2024.
[5] SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis, ICLR 2024.
[6] 3D-GSW: 3D Gaussian Splatting for Robust Watermarking, CVPR 2025.
[7] DEGauss: Defending Against Malicious 3D Editing for Gaussian Splatting, NeurIPS 2025.

I have read your rebuttal and revisions. I appreciate the new experiments and improved exposition. However, my main concerns persist. I will keep my original score.

Dear Reviewer hdv9,

We sincerely thank you for your thoughtful review and constructive feedback. Your valuable comments provided us with an important opportunity to clarify our contributions and further strengthen our work.

We are also grateful that you continue to lean toward acceptance while maintaining your original score.

Thank you again for your time and careful consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Remarks

Dear Reviewers, Area Chairs, Senior Area Chairs, and Program Chairs,

Thank you for your careful reviews and constructive feedback. The reviews recognized the value of jointly addressing ownership tracing and edit deterrence for 3D Gaussian Splatting assets and highlighted how selective Gaussian updates help balance these objectives with rendering fidelity.

Two of the three reviewers (hdv9 and Ufqx) gave a score of 3 (Weak Acceptance). After reading the rebuttal and revised manuscript, Reviewer hdv9 appreciated the new experiments and clearer exposition and confirmed that they would retain their original recommendation, which leaned toward acceptance.

In response to the reviews, we conducted additional experiments and documented the method and evaluation protocol in more detail.

  • A substantial gap remains after tuning the composition baseline. We swept the adversarial loss weight for 3DGSW+DEGauss from 0.5× to 4× its original value. The baseline's best bit accuracy (66.64%) and best CLIP Diff. (0.0417) occurred at different weights. At a single operating point, GuardMarkGS achieved 97.23% bit accuracy and 0.0907 CLIP Diff., with higher rendering fidelity than any tested baseline setting in the sweep.
  • Implementation details. The revised manuscript defines the denoising-trajectory diversion objective, including its descriptor, timestep sampling, and normalization (Appendix I). It also documents the optimization settings, scene-specific prompt libraries, and held-out prompt protocol (Appendices C and N).
  • Re-render-and-reconstruct attack scenario. To address the question of adaptive attacks, we trained fresh 3DGS models from rendered views of the protected assets across the evaluation scenes. After reconstruction, watermark bit accuracy was 96.70% (97.23% before reconstruction), and CLIP Diff. was 0.0897 (0.0907 before reconstruction). These results demonstrate GuardMarkGS's robustness to this reconstruction-based adaptive attack, with both ownership tracing and edit deterrence largely preserved.
  • Human evaluation of edit deterrence. Raters assessed how closely the edited results followed the requested instructions, with method identities hidden. GuardMarkGS had the lowest mean edit-success score (1.96/5, compared with 2.36 for DEGauss and 4.16 for 3DGSW).

For ease of reference, newly added and revised text is shown in light blue\color{#00B0F0}{\textbf{light blue}} in the revised manuscript.

We appreciate the feedback that prompted these revisions and experiments. We would be grateful if you would take the revised manuscript and the additional evidence into account in your final assessment.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.