- 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
-
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.
-
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.
-
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.
-
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.
-
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.
-
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.
-
Transferability experiments The GaussianEditor and SDXL-backbone evaluations strengthen the claim that the defence generalises beyond the exact optimisation environment.
Critical weaknesses
Major Weaknesses
- 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.
-
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.
-
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.
- 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.
- 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.
-
Minor Weaknesses
-
Lack of statistical significance reporting The paper reports averages but no variance, confidence intervals, or repeated-run statistics.
-
Hyperparameter sensitivity is not explored Important parameters include: λadv λlat λtraj λxattn τmask γmask The robustness of performance to these settings is unclear.
-
Some qualitative comparisons are limited The paper would benefit from more failure-case analysis showing situations where edit deterrence does not work.
-
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.
-
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
-
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.
-
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.
-
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
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:
| Aggregation | GuardMarkGS | Best Baseline |
|---|---|---|
| Harmonic Mean (Original) | 0.8622 | 0.7791 |
| Geometric Mean | 0.8759 | 0.7992 |
| Arithmetic Mean | 0.8888 | 0.8328 |
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:
| Method | Object-centric: bear (min) ↓ | Unbounded: garden (min) ↓ |
|---|---|---|
| DEGauss [6] | 218.74 | 231.83 |
| 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.
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.
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.
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.