acceptodds
Withdrawn from ICLR 2027

Gaussian Carrier: Model-Level Watermarking in Feed-Forward 3D Gaussian Splatting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Feed-forward 3D Gaussian Splatting (3DGS) models are increasingly deployed as reusable pretrained reconstruction engines, yet their editable parameters introduce an underexplored model-level risk: an attacker can encode an attacker-specified invisible message into novel view renderings through lightweight fine-tuning of the Gaussian prediction head, without perturbing the input images or retraining the full model. We investigate this model-level attack surface through adversarial watermark injection. Naively fine-tuning the Gaussian prediction head, however, modifies predicted primitives at all spatial locations, resulting in unnecessary representation changes and degrading the rendering fidelity. To address this, we propose Gaussian Carrier, which identifies reliable watermark carriers according to opacity and cross-view observability. We develop a carrier-restricted mechanism that applies the adapted predictions only to selected primitives while retaining the original pretrained predictions. Combined with complementary watermark-success and fidelity-preserving objectives, the predictions are optimized. Experiments on RealEstate10K and ACID datasets with two feed-forward 3DGS models achieve over watermarking accuracy while limiting the PSNR difference to at most dB. The watermarking approach also maintains above accuracy under various distortions. These results expose Gaussian adaptation as a model-level attack surface in pretrained feed-forward 3DGS.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
3D Security, Watermarking, 3D Gaussian Splatting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_78koOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies watermark injection into feed-forward 3DGS models. It freezes the feature, depth, and opacity branches and fine-tunes only the Gaussian prediction head to embed a specified 48-bit message in the rendered outputs. The method selects carriers based on the original Gaussians' opacity and coverage across target views, restricts modifications through masked mixing of the original and adapted heads' outputs, and jointly uses BCE, a bit-level margin loss, and rendering-fidelity losses.

Critical strengths
  1. The paper introduces a simple and interpretable local adaptation mechanism for feed-forward Gaussian prediction. Equation (10) directly retains the original head's predictions at non-carrier locations, constraining modifications in output space despite shared model parameters.

  2. The experiments show consistent improvements in the recovery–fidelity trade-off across two backbones and two datasets, with limited additional overhead.

Critical weaknesses
  1. There is a substantial gap between the central framing as a “model-level attack” and the scope of the empirical validation. Appendix A.2 explicitly states that the adapted model is specific to the current scene and that the original head parameters are restored after each scene. Thus, the experiments do not establish a model-level attack that is implanted once and remains effective on future unseen scenes.

  2. Per-scene optimization is a valid setting in itself. However, if the evaluation views also participate in optimization, the current results establish only that the watermark can be embedded in the training views. This distinction determines the security implications of the work and the extent of its contribution relative to scene-level watermarking methods.

  3. Equation (11) adds the margin loss to BCE. After the margin gradient vanishes for correctly decoded bits, BCE continues to drive optimization. Therefore, the claim that this objective “prevents excessive optimization of correctly decoded bits” does not follow directly from the stated formulation.

Questions for authors
  1. How are the numbers of scenes and messages, as well as the actual evaluation views, determined for Table 1? Are the three target views used for carrier selection, optimization, and final evaluation? Please provide results with an explicit separation between optimization and test views.
  2. Please specify the reference images used for each image-quality metric in Tables 1 and 3, and report PSNR/SSIM/LPIPS directly comparing renders before and after watermarking.
  3. How do random masks and opacity-only masks perform under the same BCE-plus-margin objective? Please also explain how the margin term suppresses excessive optimization when BCE remains active.
2 comments
AuthorsOpenReview ↗

Reply to W1 & W2 & W3 & Q1

W1: Scene-specific adaptation versus persistent checkpoint modification.

Model-level attack identifies the modified variables as the Gaussian prediction-head parameters. We acknowledge that the original adaptation is scene-specific, and we additionally introduce extensions that retain one adapted head from the source scenes and are evaluated across unseen scenes on the same checkpoint without test-time parameter updates. We conduct the shared-head baseline, with a universal setting and a strengthened margin setting. These extensions use 64 source scenes, 8 validation scenes, and 25 unseen scenes for testing.

The shared-head baseline adapts a single shared Gaussian prediction head across the source scenes, using 150 updates per scene, and retains the final checkpoint. The Universal setting uses the same source scenes and the original watermark and fidelity objectives in Eq. (11) and Eq. (12), but trains the shared head by averaging gradients over four scenes per update. It sets every carrier-mask entry in Eq. (10) to one, so all Gaussians use the adapted head predictions without blending with the original head. The resulting checkpoint is reused unchanged on unseen scenes. Building on the Universal setting, we strengthen the existing margin term in Eq. (11). We gradually increase the confidence margin to 0.05 and give equal weight to the mean margin loss over all 48 bits and the mean over the eight largest margin violations in each rendering. This increases bit accuracy on unseen scenes from 89.17% to 93.06%, again without test-time parameter updates.

Figure 11 in Appendix B.8 shows the visual results.

SettingSource Scenes: Bit Acc. (%)Source Scenes: PSNR ↑Unseen Scenes: Bit Acc. (%)Unseen Scenes: PSNR ↑
Shared-head99.3226.2972.2525.82
+ Universal99.4925.7189.1725.55
+ Strengthened margin99.9725.6593.0625.03

W2: Independence of evaluation cameras.

We provide results in Q1-(2).

W3: The margin term and the BCE term.

Thank you for pointing out this ambiguity. The per-bit watermark loss and its derivative are:

=-\frac{1}{1+e^{s}}-\lambda\mathbf{1}[s<\kappa].$$ Only the additional hinge contribution disappears once the margin is satisfied; BCE continues to contribute. At the original default $\kappa=0$, the hinge directly penalizes negative signed logits. Correctly decoded outputs can also change through shared-head parameter updates, even when their own hinge gradient is zero. The precise interpretation is: “The margin term adds an extra penalty to bits below the target margin and stops contributing once that margin is reached, while BCE continues to promote message recovery.” Table 3 supports the empirical benefit of this combination. **Q1-(1): Which scenes and messages were evaluated?** The original evaluation uses one fixed 48-bit target message across scenes, two context views, and three target views per scene. These three target views are used for carrier construction, adaptation, and the original reported evaluation. For the two datasets, RealEstate10K consists of 7289 testing scenes, and ACID comprises 1972 testing scenes, which are the actual evaluated scene counts by following default settings. **Q1-(2): Separation between optimization and test views.** To assess camera-pose generalization, we select ten additional camera views per scene, excluded from carrier selection and training objectives. Excluding context and optimization views, each candidate is compared with the three optimization cameras using a combined score: translation normalized by the context-camera baseline plus rotation normalized by 30 degrees. We retain the lowest score and its corresponding optimization camera. After excluding scores below 0.005, we rank the remaining candidates within each scene, divide them into three equal-sized distance groups, and deterministically sample three near, three intermediate, and four far views. The quantitative results show comparable rendering quality across these groups, supporting watermark transfer to cameras excluded from adaptation and indicating relatively stable rendering quality over the evaluated held-out views. **Figure 10 in Appendix B.7** shows visual results in the revised paper. | Views | Held-out | Normalized translation | Rotation angle | Bit Acc. (%) ↑ | PSNR ↑ | SSIM ↑ | LPIPS ↓ | |---|---|---:|---:|---:|---:|---:|---:| | Source | -- | -- | -- | 99.88 | 26.20 | 0.864 | 0.137 | | Unseen | Near | 0.0403 [0.0116, 0.1251] | 0.301 [0.000, 2.037] | 94.14 | 25.95 | 0.863 | 0.140 | | Unseen | Intermediate | 0.1254 [0.0599, 0.3202] | 1.016 [0.007, 4.904] | 90.22 | 26.28 | 0.864 | 0.136 | | Unseen | Far | 0.2805 [0.1157, 0.6490] | 2.083 [0.007, 8.661] | 86.10 | 26.31 | 0.865 | 0.136 |
AuthorsOpenReview ↗

Reply to Q1 & Q2 & Q3

Q1-(3): Does recovery depend on the original message? The original evaluations use one fixed 48-bit target message shared across all scenes. We additionally sample five independent target messages and repeat adaptation separately for each sampled message. Exact recovery is the percentage of evaluated renderings with all 48 bits correctly decoded. Mean bit error is the average Hamming distance between the decoded and target messages per rendering. The final row reports the mean and sample standard deviation across the five messages. Exact recovery ranges from 86.02% to 95.00%, with a mean of 91.12%, supporting adaptation to multiple target messages under the evaluated protocol. Figure 12 in Appendix B.10 shows renderings and pixel-wise residual maps for the default message and the five sampled messages in the revised paper.

Default: 111010110101000001010111010011010100010000100111

Message 1: 100100001111101001010001100110101010110001111111

Message 2: 011000111000111001000101011011001001100111101101

Message 3: 111111100001111111101110011000000011100100111010

Message 4: 000101111010111110111100001001001101100101100111

Message 5: 000010000101100100001101101001000100100001110010

Target messageBit Acc. (%)Exact recovery (%)Mean bit errors
Message 199.7991.670.1000
Message 299.9095.000.0500
Message 399.6286.020.1833
Message 499.8693.330.0667
Message 599.6989.580.1500
Mean ± std99.77 ± 0.1291.12 ± 3.490.1100 ± 0.0560

Q2: References for image-quality metrics and direct before/after comparison.

Thank you for requesting clarification. All PSNR/SSIM/LPIPS values reported in the paper, including Tables 1 and 3, are computed against the corresponding ground-truth images (GT). In each model block of Table 1, the first row reports rendering quality before watermarking, while the second and third rows report quality after watermarking using Unrestricted Fine-tuning and Carrier-Restricted Fine-tuning, respectively. Carrier-Restricted Fine-tuning achieves better rendering quality than Unrestricted Fine-tuning across both datasets and models.

We additionally quantify watermark-induced changes by directly comparing paired watermarked and original-model renderings at identical camera poses. This new comparison uses the original rendering, rather than GT, as its reference. The results below report PSNR / SSIM / LPIPS, and the after-watermarking results use carrier-restricted fine-tuning.

DatasetModelBefore watermarking vs. GTAfter watermarking vs. GTAfter watermarking vs. Before watermarking
RealEstate10KMVSplat26.36 / 0.868 / 0.12926.20 / 0.864 / 0.13741.79 / 0.994 / 0.016
ACIDMVSplat28.21 / 0.842 / 0.14528.00 / 0.837 / 0.15842.69 / 0.990 / 0.034
RealEstate10KIDESplat27.84 / 0.893 / 0.10827.66 / 0.888 / 0.11644.41 / 0.995 / 0.010
ACIDIDESplat29.09 / 0.859 / 0.13028.78 / 0.851 / 0.14642.03 / 0.988 / 0.036

Q3: Equal-budget carrier baselines and the BCE-plus-margin objective.

Q3-(1): How do random and opacity-only masks perform? Our original ablation evaluates the overall carrier mechanism. We add comparisons among random selection, opacity-only ranking, observability-only ranking, and joint opacity–observability ranking. The four ranking methods have the same number of carriers within each scene: mean 90,970.72, range 80,819–91,750. Joint ranking achieves the highest observed bit accuracy and exact recovery. Compared with observability-only ranking, it increases bit accuracy from 99.84% to 99.88% and exact recovery from 92.89% to 94.67%, corresponding to gains of 0.04 and 1.78 percentage points, respectively.

Carrier selectionMean number of carriersBit Acc. (%)Exact recovery (%)
Random eligible Gaussians90,970.7291.7385.21
Opacity-only ranking90,970.7298.0590.66
Observability-only ranking90,970.7299.8492.89
Joint opacity–observability ranking90,970.7299.8894.67

Q3-(2): Does the margin cancel continued BCE optimization?

No. As explained under W3, the hinge adds selective pressure below the margin without cancelling BCE. This interpretation applies to the original κ=0\kappa=0 setting. The positive-margin cross-scene extension additionally supervises correctly signed bits that remain below its confidence target.

We sincerely thank you for your insightful and constructive suggestions. Following your suggestion, we have modified the paper, with additional visualization results in the revised paper on pages 19-22.

Reviewer_FFuEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates model-integrity risks in pretrained feed-forward 3D Gaussian Splatting (3DGS) models and proposes Gaussian Carrier, a method for embedding an attacker-specified invisible watermark into novel-view renderings by fine-tuning the Gaussian prediction head while attempting to preserve rendering fidelity.

The method scores Gaussian primitives according to opacity and cross-view observability, selects a subset as watermark carriers, and uses masked mixing to apply adapted predictions only to the selected primitives. Its optimization combines a watermark-decoding objective with rendering-fidelity constraints.

The authors evaluate MVSplat and IDESplat on RealEstate10K and ACID. The protocol uses two context views, three target views, and 150 fine-tuning steps. The reported bit accuracy is approximately 99.63%–99.91%, with additional experiments on robustness to common image distortions.

The principal contribution is the identification of a potentially important attack surface: lightweight adaptation of a pretrained reconstruction model can alter its outputs without directly modifying the input images. However, the experiments have not yet convincingly established that the attack generalizes through a single deployable model modification to previously unseen scenes and camera views. This limitation directly affects the strength of the paper's central claim.

Critical strengths
  1. The research question is relevant and meaningful. The paper shifts attention from conventional owner-side 3DGS watermarking to unauthorized model adaptation and invisible watermark injection. This perspective helps expose a potentially important model-integrity risk in pretrained reconstruction systems.

  2. The experiments have reasonable breadth. The authors evaluate two datasets and two feed-forward 3DGS architectures, with ablations on carrier ratio, opacity threshold, observability threshold, and optimization objectives. The reported bit accuracy is high, and carrier-restricted adaptation yields better rendering fidelity than unrestricted fine-tuning.

Critical weaknesses

Weakness 1: The model-level threat claim is not adequately validated. The most important concern is the potential mismatch between the experimental protocol and the claim of a model-level attack against reusable pretrained models. First, the appendix states that the original Gaussian prediction head is restored after the attacked renderings of each scene are recorded. This suggests that each scene may be fine-tuned independently, rather than demonstrating that a single watermarked model checkpoint can directly process multiple unseen scenes. If per-scene optimization is required, the experiments establish the feasibility of scene-specific adaptation, not the cross-scene generalization of a single compromised model. Second, the protocol uses three target views, and both the watermark and fidelity losses are evaluated on target views. The paper does not clearly establish whether the reported bit accuracy is measured on independent camera views that were excluded from optimization. If the same views are used for optimization and evaluation, near-perfect bit accuracy does not demonstrate reliable watermark recovery from arbitrary novel views. Finally, Equation (10) mixes adapted predictions with original predictions. This suggests that inference may require the original prediction head and a carrier mask, rather than simply replacing the original head. This requirement affects the actual deployment model, storage overhead, and computational cost. Together, these issues limit the paper's conclusions regarding attacks on reusable feed-forward 3DGS models.

Weakness 2: Novelty and robustness claims are insufficiently established. The carrier-selection mechanism is intuitive and appears effective, but its independent contribution remains insufficiently established. Existing 3DGS watermarking research has explored adaptive allocation of watermark modifications. The authors should more clearly distinguish the substantive novelty of their threat model, algorithmic design, and empirical findings. Furthermore, the paper primarily reports bit accuracy rather than exact recovery of the complete 48-bit message or reliable watermark-detection metrics. Bit accuracy is not equivalent to successful recovery of an entire message. In particular, bit accuracy of approximately 80% under some distortions does not establish reliable full-message recovery. The robustness evaluation is largely limited to image noise, scaling, blur, and brightness changes. The paper does not adequately investigate model-level modifications, parameter pruning, subsequent fine-tuning, or dedicated watermark-removal attacks. In addition, the lack of comparisons against random carriers, opacity-only selection, and observability-only selection makes it difficult to isolate the contribution of the complete carrier-scoring mechanism.

Questions for authors

Q1. Scene-specific adaptation and generalization The appendix states that the original Gaussian prediction head is restored after each scene. Does this mean that every scene requires independent fine-tuning? If so, how does this support the claim of a model-level attack against reusable pretrained models? Please provide results using a fixed adapted checkpoint on scenes excluded from optimization. Q2. Independence of evaluation views Are the three target views used to compute the watermark loss also used to report the final bit accuracy? Please evaluate watermark recovery on camera views excluded from optimization and report performance across these held-out views. Q3. Actual inference and deployment requirements Equation (10) requires both the original and adapted prediction heads. Must both heads be retained and executed at deployment time? How is the carrier mask obtained for arbitrary target camera views? Please report the actual checkpoint size, inference latency, and computational overhead, and clarify the composition of the deployed attack model. Q4. Full-message recovery and carrier-selection baselines Beyond bit accuracy, can the authors report exact recovery of the complete 48-bit message, false-positive rates on clean models, and full-message detection performance under different distortions? Please also compare random carriers, opacity-only carriers, and observability-only carriers to validate the independent contribution of the proposed scoring mechanism.

3 comments
AuthorsOpenReview ↗

Reply to W1 & W2

Thank you. We address the concerns about generalization, deployment requirements, novelty, and watermark reliability below.

W1-(1). Scene-specific adaptation versus persistent modification.

The original experiments establish scene-specific parameter-based injection, but alone do not establish persistence on future scenes. Appendix B.8 adds fixed-checkpoint evaluation using 64 source, 8 validation, and 25 unseen test scenes. Unseen-scene bit accuracy reaches 72.25% with the shared-head baseline, 89.17% with Universal, and 93.06% with strengthened margin supervision, without test-time updates. Q1 explains these extensions.

W1-(2). Independence of evaluation cameras.

The original three target views participate in optimization. Appendix B.7 adds ten held-out cameras per scene, excluded from carrier construction and both losses. Bit accuracy is 94.14%, 90.22%, and 86.10% for near, intermediate, and far groups. This establishes transfer to excluded cameras while revealing degradation with pose distance, rather than reliable recovery at arbitrary viewpoints. Q2 gives the protocol.

W1-(3). Deployment composition.

Eq. (10) requires both original and adapted predictions when the carrier mask is nontrivial. We clarify the two-head reconstruction path, reusable scene representation, and all-ones-mask extension in Q3.

W2-(1). Contribution beyond adaptive allocation. We agree that adaptive watermark allocation has prior work. Our contribution is carrier-restricted adaptation of a pretrained feed-forward Gaussian prediction head. Because its parameters are spatially shared, restricting a loss to selected primitives does not itself ensure that other predictions remain unchanged. Eq. (10) explicitly controls which adapted predictions enter the reconstructed scene: at non-carrier locations, it uses the original predictions. The contribution therefore concerns the interaction between shared model parameters and selective changes to the predicted Gaussian representation, in addition to the carrier-ranking rule. The attacker-side setting motivates the model-integrity question. Q4 separately evaluates the proposed ranking components.

W2-(2). Bit accuracy versus complete-message recovery. We agree that bit accuracy cannot substitute for exact payload recovery, particularly under distortions. Appendix B.10 evaluates five independently sampled 48-bit messages and reports exact recovery and Hamming errors. Mean exact recovery is 91.12%, with a sample standard deviation of 3.49 percentage points, under the evaluated undistorted protocol. Q4 reports the message results and separately addresses distorted-image detection. Detection of a watermark and exact recovery of its entire payload are distinct outcomes.

W2-(3). Model modifications and removal attacks. The original robustness experiments cover image distortions. We additionally evaluate 10% Gaussian-head parameter pruning, 200 steps of clean-task fine-tuning, and 200 steps of targeted watermark-removal optimization. The latter two use a learning rate of 10−510^{-5}. The table reports recovery and image quality under these operations.

Model modificationStrength / Update BudgetBit Acc. (%)Exact recovery (%)PSNR ↑SSIM ↑LPIPS ↓
No modificationNone99.8894.6726.200.8640.137
Gaussian-head parameter pruning10% pruning99.1791.3326.200.8620.138
Subsequent clean-task fine-tuning200 steps, lr=10−510^{-5}93.0089.0526.290.8650.134
Dedicated watermark removal200 steps, lr=10−510^{-5}92.9487.8226.340.8680.131

W2-(4). Carrier-selection baselines. Appendix B.9 and Table 10 add random, opacity-only, and observability-only rankings with exactly matched carrier counts within each scene. Joint ranking obtains the highest observed bit accuracy and exact recovery. Q4 provides the numerical comparison and distinguishes the small bit-accuracy gain over observability-only ranking from the larger gain in exact recovery.

AuthorsOpenReview ↗

Reply to Q1 & Q2

Q1: Original protocol, scene-specific adaptation, and fixed-checkpoint transfer.

We clarify that the original adaptation is scene-specific, and we additionally introduce extensions that retain one adapted head from the source scenes and are evaluated across unseen scenes on the same checkpoint without test-time parameter updates. The training is conducted on MVSplat with gradient accumulation on the RealEstate10K dataset. We conduct the shared-head baseline, with a universal setting and a strengthened margin setting. These extensions use 64 source scenes, 8 validation scenes, and 25 unseen scenes for testing.

The shared-head baseline adapts a single shared Gaussian prediction head across the source scenes, using 150 updates per scene, and retains the final checkpoint. For each test scene, we construct a carrier mask from its reference cameras using the original opacity–observability rule and retain the original-head predictions outside the selected carriers.

The Universal setting uses the same source scenes and the original watermark and fidelity objectives in Eq. (11) and Eq. (12), but trains the shared head by averaging gradients over four scenes per update. It sets every carrier-mask entry in Eq. (10) to one, so all Gaussians use the adapted head predictions without blending with the original head. The resulting checkpoint is reused unchanged on unseen scenes. Bit accuracy increases from 72.25% to 89.17% on the same unseen scenes.

Building on the Universal setting, we strengthen the existing margin term in Eq. (11). We gradually increase the confidence margin to 0.05 and give equal weight to the mean margin loss over all 48 bits and the mean over the eight largest margin violations in each rendering. This adds margin supervision to correctly decoded bits that have not reached the confidence margin while emphasizing the largest violations. With the same training scenes, update budget, and all-ones carrier mask, this modification increases bit accuracy on unseen scenes from 89.17% to 93.06%, again without test-time parameter updates.

Figure 11 in Appendix B.8 shows the visual results.

SettingSource Scenes: Bit Acc. (%)Source Scenes: PSNR ↑Unseen Scenes: Bit Acc. (%)Unseen Scenes: PSNR ↑
Shared-head99.3226.2972.2525.82
+ Universal99.4925.7189.1725.55
+ Strengthened margin99.9725.6593.0625.03

Q2: Held-out camera evaluation.

To assess camera-pose generalization, we evaluate RealEstate10K scenes while retaining each scene's original two context images and three optimization cameras. We select ten additional camera views per scene. The additional cameras are excluded from carrier selection and from both the watermark and fidelity objectives. After adaptation, we render all evaluation cameras from the same fixed adapted Gaussian field, without further parameter updates or view-specific carrier selection.

We select real-frame cameras, excluding context and optimization views. Each candidate is compared with the three optimization cameras using a combined score: translation normalized by the context-camera baseline plus rotation normalized by 30 degrees. We retain the lowest score and its corresponding optimization camera. After excluding scores below 0.005, we rank the remaining candidates within each scene, divide them into three equal-sized distance groups, and deterministically sample three near, three intermediate, and four far views. Translation and rotation statistics are reported separately, both relative to the same optimization camera selected by the combined score. The quantitative results show comparable rendering quality across these groups: Bit accuracy decreases from 94.14% to 90.22% and 86.10% across the near, intermediate, and far held-out groups. These results support watermark transfer to cameras excluded from adaptation, while showing reduced decoding accuracy at larger pose distances. PSNR ranges from 25.95 to 26.31, SSIM from 0.863 to 0.865, and LPIPS from 0.136 to 0.140. These results indicate relatively stable rendering quality over the evaluated held-out views. Figure 10 in Appendix B.7 shows visual results in the revised paper.

ViewsHeld-outNormalized translationRotation angle (°)Bit Acc. (%) ↑PSNR ↑SSIM ↑LPIPS ↓
Source------99.8826.200.8640.137
UnseenNear0.0403 [0.0116, 0.1251]0.301 [0.000, 2.037]94.1425.950.8630.140
UnseenIntermediate0.1254 [0.0599, 0.3202]1.016 [0.007, 4.904]90.2226.280.8640.136
UnseenFar0.2805 [0.1157, 0.6490]2.083 [0.007, 8.661]86.1026.310.8650.136
AuthorsOpenReview ↗

Reply to Q3 & Q4

Q3: Actual inference and deployment requirements.

Under the carrier-restricted formulation in Eq. (10), deployment retains the frozen backbone and both the original and adapted prediction heads. For each new scene, both heads generate Gaussian predictions, with backbone features that can be shared between them. A carrier mask is computed from the scene’s reference cameras to select adapted predictions at carrier locations and original predictions elsewhere. The resulting mixed Gaussian representation can then be cached for subsequent rendering. Additional camera views of the same reconstructed scene can be rendered without rerunning either head or recomputing the mask. Table 6 (page 15) has already reported the costs on adaptation time, GPU memory, and FPS. We provide additional measurements on checkpoint size and latency. The deployment comparison below is measured on matched hardware and inputs: We measure deployment cost for MVSplat on RealEstate10K scenes using batch size one, two 256×256 context images, and three matched reference cameras. Each configuration ran in a fresh process in each of three rounds with rotated execution order, using five warm-up and twenty measured repetitions per scene. Latencies are equally averaged over scenes, rounds, and repetitions. Checkpoint sizes include consistently serialized inference weights and configuration, excluding training state and watermark decoding.

ConfigurationCheckpoint size (MB)Context-to-Gaussians latency (ms)Mask construction latency (ms)Rendering latency (ms/view)
Original model47.9949.34-2.28
Carrier-restricted model49.4981.431.202.34

Q4-(1). Exact recovery and multiple messages. We agree that mean bit accuracy does not indicate how often the complete 48-bit message is recovered. The original evaluations use one fixed 48-bit target message shared across all scenes. We additionally sample five independent 48-bit target messages and repeat adaptation separately for each sampled message. Exact recovery is the percentage of evaluated renderings with all 48 bits correctly decoded. Mean bit error is the average Hamming distance between the decoded and target messages per rendering. The final row reports the mean and sample standard deviation across the five messages. Exact recovery ranges from 86.02% to 95.00%, with a mean of 91.12%, supporting adaptation to multiple target messages under the evaluated protocol. Figure 12 in Appendix B.10 shows renderings and pixel-wise residual maps for the default message and the five sampled messages in the revised paper. Detailed bits of the 5 messages are defined in Appendix B.10 (page 22).

Target messageBit Acc. (%)Exact recovery (%)Mean bit errors
Message 199.7991.670.1000
Message 299.9095.000.0500
Message 399.6286.020.1833
Message 499.8693.330.0667
Message 599.6989.580.1500
Mean ± std99.77 ± 0.1291.12 ± 3.490.1100 ± 0.0560

Q4-(2). False positives and distorted-message detection. The bit accuracies below are the existing MVSplat/RealEstate10K results from Table 2 under different distortions, and we report additional metrics.

DistortionBit Acc. (%) ↑Exact recovery (%) ↑Detection TPR (%) ↑
None99.8894.67100.00
Noise93.2389.20100.00
Scaling82.7075.6698.67
Blur,99.1490.35100.00
Brightness89.8583.1698.67

Q4-(3). Equal-size carrier comparisons. Our original ablation evaluates the overall carrier mechanism. We add comparisons among random selection, opacity-only ranking, observability-only ranking, and joint opacity–observability ranking. The four ranking methods have the same number of carriers within each scene: mean 90,970.72, range 80,819–91,750. Joint ranking achieves the highest observed bit accuracy and exact recovery. Compared with observability-only ranking, it increases bit accuracy from 99.84% to 99.88% and exact recovery from 92.89% to 94.67%, corresponding to gains of 0.04 and 1.78 percentage points, respectively.

Carrier selectionMean number of carriersBit Acc. (%)Exact recovery (%)
Random eligible Gaussians90,970.7291.7385.21
Opacity-only ranking90,970.7298.0590.66
Observability-only ranking90,970.7299.8492.89
Joint opacity–observability ranking90,970.7299.8894.67

We sincerely thank you for the constructive feedback. Following your suggestion, we have modified the paper, with additional visualization results in the revised paper on pages 19-22.

Reviewer_Tns7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies adversarial watermark injection in feed-forward 3D Gaussian Splatting models. The proposed attack modifies only the Gaussian prediction head of a pretrained reconstruction model to make the rendered novel views carry an attacker-specified 48-bit message while keeping their appearance close to the original model's outputs.

The main technical idea, Gaussian Carrier, avoids applying the adapted head predictions to all Gaussian primitives. Instead, the method ranks Gaussians using opacity and cross-view observability, selects a subset of them as watermark carriers, and mixes the adapted predictions with the original frozen predictions according to this carrier mask. The optimization combines a watermark recovery objective, a bit-level margin term, and a rendering-fidelity objective.

The experiments use MVSplat and IDESplat on RealEstate10K and ACID. Under the reported protocol, carrier-restricted fine-tuning reaches close to 100% bit accuracy. It causes relatively small degradation in PSNR/SSIM/LPIPS compared with the original reconstruction. The paper also studies robustness to several image distortions. It provides ablations on carrier selection and optimization hyperparameters.

Critical strengths
  1. The paper identifies a useful and relatively unexplored attack surface in feed-forward 3DGS. Most 3DGS watermarking work is naturally framed from the asset owner's side. Looking instead at whether a small editable part of a reusable reconstruction model can be abused to inject an invisible message is a worthwhile security question. Restricting the attack to the Gaussian prediction head also makes the setting more interesting than simply retraining the whole reconstruction model.

  2. The carrier-restricted design gives a clear empirical improvement over unrestricted head fine-tuning. The main results are quite strong under the paper's protocol. Across two backbones and two datasets, the proposed method obtains roughly 99.6–99.9% bit accuracy while keeping the rendered images substantially closer to the original model than unrestricted fine-tuning. The component ablation is also useful: both carrier selection and the bit-level margin objective improve the trade-off, rather than the final result coming from an unexplained combination of components.

Critical weaknesses
  1. The experimental procedure does not yet demonstrate a persistent model-level attack across scenes. This is my main concern. The motivation is built around pretrained feed-forward 3DGS models as reusable reconstruction engines whose parameters can be compromised. However, the appendix states that the adapted Gaussian head defines the attacked model "for the current scene", and that the original head parameters are restored after evaluating each scene. In other words, the reported numbers come from performing a new optimization for each input scene.

This is materially different from the threat suggested by the paper's framing, where one would expect an attacker to modify a pretrained model once and have the malicious behavior persist when that model is subsequently used on new scenes. At present, the experiments show that model parameters can be used as the optimization variables for scene-specific watermark injection; they do not show that a compromised feed-forward model carries the attack across unseen inputs. I think this distinction needs to be resolved experimentally or the claims should be narrowed substantially.

  1. Watermark recovery is evaluated on views that appear to participate directly in the adaptation. Each scene uses two context views and three target views, and the optimization losses are averaged over the target views. Carrier selection itself also depends on cross-view observability with respect to these cameras. Bit accuracy is then reported over rendered target views. As written, I do not see a held-out-view experiment in which the adapted model is evaluated at camera poses that were not used for either carrier construction or optimization.

This matters because the paper repeatedly describes watermarking of "novel view renderings". High recovery on the optimization views does not by itself establish that the watermark is a stable property of the resulting 3D representation. A held-out camera evaluation is important here, particularly because the proposed carrier criterion is explicitly view dependent.

Questions for authors
  1. Are the three target views used to compute the adaptation losses also the views used to report the Bit Acc. and fidelity numbers in Tables 1–2? If so, could you report performance on additional held-out camera poses that are not used for carrier selection, watermark optimization, or fidelity optimization? Ideally, I would like to see Bit Acc. and rendering quality as the evaluation camera moves farther away from the optimization views. This would substantially affect my assessment of whether the watermark is genuinely encoded into the reconstructed representation rather than fitted to a small set of renderings.

  2. What is the intended meaning of "model-level" with respect to scene generalization? The appendix says that the optimized head is the attacked model for the current scene and that the original parameters are restored between scenes. What happens if one (\theta_h^{adv}), optimized once, is retained and the feed-forward model is subsequently given different context images from unseen scenes? Does the target message remain recoverable? If not, I think the threat model should be described more explicitly as scene-conditioned model adaptation rather than a persistent compromise of the reusable pretrained model.

  3. How many target messages were used in the experiments? Is one fixed 48-bit message used for all scenes, or are messages sampled independently? Since the paper claims attacker-specified message injection, results over multiple randomly sampled messages would be helpful. I would also prefer to see message-level exact recovery (or at least the distribution of per-message bit errors) in addition to mean bit accuracy, since an average Bit Acc. near 100% does not directly tell us how often the complete 48-bit payload is recovered correctly.

  4. Can the authors disentangle the two ingredients of the carrier score more directly? In particular, I would like to see equal-size carrier sets selected using (i) opacity only, (ii) observability only, and (iii) random eligible Gaussians. The current ablations show that carrier masking helps, but they do not fully establish that the specific opacity-plus-cross-view-observability criterion is responsible for the improvement.

2 comments
AuthorsOpenReview ↗

Reply to W1 & W2 & Q1

Thank you for the comments. We agree that the original protocol requires a clearer distinction between scene-specific adaptation and persistent cross-scene modification, and between optimization views and held-out cameras. We address these issues with the additional experiments below.

W1: Scene-specific adaptation versus persistent model compromise.

Thanks for pointing out this limitation. The original adaptation is scene-specific: the Gaussian prediction head is optimized independently for each scene, and the pre-trained parameters are restored between scenes. Here, “model-level” identifies the parameters being modified. To assess fixed-checkpoint transfer to unseen scenes, we provide additional experiments in Q2, evaluating retained heads and shared-head training protocols. We distinguish these cross-scene evaluations from the original scene-specific results.

W2: Independence of evaluation views.

The original three target views are novel relative to the context images but participate in adaptation and final evaluation. We agree that their recovery accuracy does not establish camera-pose generalization. In Q1, additional cameras were excluded from carrier construction and both optimization objectives. The adapted head, mask, and Gaussian field remained fixed during evaluation.

Q1: Watermark recovery on held-out camera poses.

To assess camera-pose generalization, we evaluate RealEstate10K scenes while retaining each scene's original two context images and three optimization cameras. We select ten additional camera views per scene. The additional cameras are excluded from carrier selection and from both the watermark and fidelity objectives. After adaptation, we render all evaluation cameras from the same fixed adapted Gaussian field, without further parameter updates or view-specific carrier selection.

We select real-frame cameras, excluding context and optimization views. Each candidate is compared with the three optimization cameras using a combined score: translation normalized by the context-camera baseline plus rotation normalized by 30 degrees. We retain the lowest score and its corresponding optimization camera. After excluding scores below 0.005, we rank the remaining candidates within each scene, divide them into three equal-sized distance groups, and deterministically sample three near, three intermediate, and four far views. Translation and rotation statistics are reported separately, both relative to the same optimization camera selected by the combined score. The quantitative results show comparable rendering quality across these groups: Bit accuracy decreases from 94.14% to 90.22% and 86.10% across the near, intermediate, and far held-out groups. These results support watermark transfer to cameras excluded from adaptation, while showing reduced decoding accuracy at larger pose distances. PSNR ranges from 25.95 to 26.31, SSIM from 0.863 to 0.865, and LPIPS from 0.136 to 0.140. These results indicate relatively stable rendering quality over the evaluated held-out views. Figure 10 in Appendix B.7 shows visual results in the revised paper.

ViewsHeld-outNormalized translationRotation angleBit Acc. (%) ↑PSNR ↑SSIM ↑LPIPS ↓
Source------99.8826.200.8640.137
UnseenNear0.0403 [0.0116, 0.1251]0.301 [0.000, 2.037]94.1425.950.8630.140
UnseenIntermediate0.1254 [0.0599, 0.3202]1.016 [0.007, 4.904]90.2226.280.8640.136
UnseenFar0.2805 [0.1157, 0.6490]2.083 [0.007, 8.661]86.1026.310.8650.136
AuthorsOpenReview ↗

Reply to Q2 & Q3 & Q4

Q2: Meaning of “model-level” and cross-scene generalization.

In our formulation, “model-level” identifies the modified variables as the Gaussian prediction-head parameters. We acknowledge that the original adaptation is scene-specific, and we additionally introduce extensions that retain one adapted head from the source scenes and are evaluated across unseen scenes on the same checkpoint without test-time parameter updates. The training is conducted on MVSplat with gradient accumulation on the RealEstate10K dataset. We conduct the shared-head baseline, with a universal setting and a strengthened margin setting. These extensions use 64 source scenes, 8 validation scenes, and 25 unseen scenes for testing.

The shared-head baseline adapts a single shared Gaussian prediction head across the source scenes, using 150 updates per scene, and retains the final checkpoint. For each test scene, we construct a carrier mask from its reference cameras using the original opacity–observability rule and retain the original-head predictions outside the selected carriers.

The Universal setting uses the same source scenes and the original watermark and fidelity objectives in Eq. (11) and Eq. (12), but trains the shared head by averaging gradients over four scenes per update. It sets every carrier-mask entry in Eq. (10) to one, so all Gaussians use the adapted head predictions without blending with the original head. The resulting checkpoint is reused unchanged on unseen scenes. Bit accuracy increases from 72.25% to 89.17% on the same unseen scenes.

Building on the Universal setting, we strengthen the existing margin term in Eq. (11). We gradually increase the confidence margin to 0.05 and give equal weight to the mean margin loss over all 48 bits and the mean over the eight largest margin violations in each rendering. This adds margin supervision to correctly decoded bits that have not reached the confidence margin while emphasizing the largest violations. With the same training scenes, update budget, and all-ones carrier mask, this modification increases bit accuracy on unseen scenes from 89.17% to 93.06%, again without test-time parameter updates.

Figure 11 in Appendix B.8 shows the visual results.

SettingSource Scenes: Bit Acc. (%)Source Scenes: PSNR ↑Unseen Scenes: Bit Acc. (%)Unseen Scenes: PSNR ↑
Shared-head99.3226.2972.2525.82
+ Universal99.4925.7189.1725.55
+ Strengthened margin99.9725.6593.0625.03

Q3: Message diversity and complete payload recovery.

We agree that mean bit accuracy does not indicate how often the complete 48-bit message is recovered. The original evaluations use one fixed 48-bit target message shared across all scenes. We additionally sample five independent 48-bit target messages and repeat adaptation separately for each sampled message. Exact recovery is the percentage of evaluated renderings with all 48 bits correctly decoded. Mean bit error is the average Hamming distance between the decoded and target messages per rendering. The final row reports the mean and sample standard deviation across the five messages. Exact recovery ranges from 86.02% to 95.00%, with a mean of 91.12%, supporting adaptation to multiple target messages under the evaluated protocol. Figure 12 in Appendix B.10 shows renderings and pixel-wise residual maps for the default message and the five sampled messages in the revised paper. Detailed bits of the 5 messages are defined in Appendix B.10 (page 22).

Target messageBit Acc. (%)Exact recovery (%)Mean bit errors
Message 199.7991.670.1000
Message 299.9095.000.0500
Message 399.6286.020.1833
Message 499.8693.330.0667
Message 599.6989.580.1500
Mean ± std99.77 ± 0.1291.12 ± 3.490.1100 ± 0.0560

Q4: Independent contribution of the carrier-score components.

Our original ablation evaluates the overall carrier mechanism. We add comparisons among random selection, opacity-only ranking, observability-only ranking, and joint opacity–observability ranking. The four ranking methods have the same number of carriers within each scene: mean 90,970.72, range 80,819–91,750. Joint ranking achieves the highest observed bit accuracy and exact recovery. Compared with observability-only ranking, it increases bit accuracy from 99.84% to 99.88% and exact recovery from 92.89% to 94.67%, corresponding to gains of 0.04 and 1.78 percentage points, respectively.

Carrier selectionMean number of carriersBit Acc. (%)Exact recovery (%)
Random eligible Gaussians90,970.7291.7385.21
Opacity-only ranking90,970.7298.0590.66
Observability-only ranking90,970.7299.8492.89
Joint opacity–observability ranking90,970.7299.8894.67

We sincerely thank you for your insightful suggestions.

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

Related papers

Discussion (0)

Sign in to comment.