acceptodds
Under review as a conference paper at ICLR 2027

CAGE-GS: Low-Resolution Prior-Guided Gaussian Expansion Control Against Computation Cost Attacks on 3D Gaussian Splatting

Abstract

Computational cost attacks against 3D Gaussian Splatting (3DGS) can substantially increase the memory and training overhead of 3DGS optimization, potentially causing denial-of-service (DoS). Existing defenses mainly rely on image purification or frequency filtering. However, because adversarial perturbations overlap with legitimate high-frequency scene details, aggressive input processing can suppress valid textures while leaving abnormal Gaussian geometry and spatial congestion insufficiently constrained. Motivated by the reduced effectiveness of such attacks at low resolution, we propose CAGE-GS, a coarse-to-fine defense that combines mild wavelet denoising with low-resolution reconstruction to derive structural and density priors. During full-resolution optimization, these priors reject structurally implausible Gaussian candidates and suppress excessive local expansion. Extensive experiments across three benchmark datasets show that CAGE-GS restores computational overhead toward clean-training levels while preserving reconstruction fidelity, and achieves better clean-input reconstruction quality than existing defense methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Cost Attacks, Gaussian Expansion Control

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_7Ya3OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CAGE-GS as a defense for Poison-Splat. The motivation is an observation: at half resolution, the poisoned reconstructions are much closer to clean reconstructions compared with full resolution. Therefore, GS reconstruction can use its half-resolution reconstruction as a course, clean reference and carefully add high-frequency textures and gaussians based on structural and density priors learnt from the half-resolution reconstruction.

The CAGE-GS pipeline goes: (i) applies wavelet soft-thresholding to inputs; (ii) trains a half-resolution 3DGS for first 7k iters; (iii) derives a structural prior (an autoencoder for 12-D Gaussian geometry features) and voxel-wise occupancy budgets from this course model; and (iv) uses them to reject clone/split candidates and to prune congested voxels during full-resolution training.

On Nerf-Synthetic, Mip-Nerf 360 and TT datasets, resources under poisoning attack return to close to clean inputs, and the reconstruction quality is better than baselines including blur, DefenseSplat and RemedyGS.

Critical strengths
  1. It addresses a real and recent threat, and the research question is interesting and important.

  2. Compute cost and resource savings are consistent and significant across three datasets, different GS implementations (including Mip-Splatting), three perturabation budgets and a range of hyperparameter settings.

  3. The paper is well-written; the motivation is well explained and the method clearly described.

Critical weaknesses
  1. No adaptive attack. Although acknowledged in the limitations section, it would be important for the readers to understand the limitations of such proposed defense - since it does not has a theoretical guarantee and the defense effectiveness now seems to rest on the capability of the attackers. What if the attacker is aware of such half-resolution defense, and inject noise which is downsampling-resilient? Is such adaptive attack possible and how will the defense perform under such adaptive attack?

Minor weakness:

  1. The authors should use \citep{} and \citet{} more properly. Currently they use \citet{} all the time and it affects reading.

  2. Some presentation could be improved:

  • Line 081: "become substantially more clean-like"?
Questions for authors
  1. Evaluation protocol. Are PSNR/SSIM/LPIPS computed on held-out test views or on training views, and at what resolution? The vanilla 3DGS numbers in Table 3 (36.98 dB on NeRF-Synthetic, 29.51 dB on Mip-NeRF 360) are notably higher than commonly reported test-view results (~33.3 dB and ~27.2 dB in Kerbl et al., 2023).

  2. Source of the gains. In Table 4, most of the Gaussian reduction on the real scenes comes from the wavelet step. What do (a) wavelet denoising alone with vanilla 3DGS at full resolution and (b) the 3D priors without wavelet denoising achieve, in both resource usage and clean-input quality?

  3. Design constants. How were ρ, the voxel size v (not reported), κ, and the constants in Eqs. 17 and 19 chosen? How sensitive are the results to ρ and v?

  4. What is the training overhead of autoencoder?

8 comments
AuthorsOpenReview ↗

Response to Reviewer 7Ya3 (Part 1 / 5)

We deeply appreciate Reviewer 7Ya3 for the positive and insightful feedback. We are especially encouraged by your recognition of the timeliness and importance of the computation-cost attack problem, the consistency of our resource reductions across three datasets, multiple 3DGS implementations, and all evaluated perturbation budgets, and the clarity of our motivation and method description. Your overall evaluation means a lot to us and motivates us to further refine and advance this line of defense.

We have carefully prepared this response to address all of your concerns, and we kindly ask you to consider maintaining your positive assessment of our work. Below, we provide detailed responses to all the comments:

W1. Robustness Against Adaptive Attacks

Thank you for highlighting the importance of adaptive evaluation. We conducted three adaptive Poison-Splat variants under the original perturbation budget (ϵ=16/255\epsilon=16/255):

  1. Resolution-aware: Optimizes the attack objective and proxy reconstruction at half resolution (400×400), explicitly targeting downsampling-resilient perturbations.
  2. Mixed-resolution: Jointly optimizes half- and full-resolution objectives.
  3. Wavelet-aware: Incorporates our two-level db4 wavelet denoising into attack optimization, using identity BPDA for gradient approximation.

All poisoned images are generated at the original 800×800 resolution. The resolution-based variants are evaluated on Ficus/Materials/Mic/Drums, and the wavelet-aware variant on Chair/Ficus/Lego.

R3-Table1: Resource consumption under adaptive Poison-Splat attacks, averaged over four resolution-attack scenes or three wavelet-attack scenes.

AttackVictim ConfigurationGaussians (M) ↓Memory (MB) ↓Time (min) ↓
Resolution-basedClean 3DGS (half-res.)0.1750985.55.61
Clean 3DGS (full-res.)0.25342100.06.69
Resolution-aware3DGS (half-res.)0.24021122.56.50
3DGS (full-res.)0.33782931.08.61
CAGE-GS0.20702334.08.30
Mixed-resolution3DGS (half-res.)0.23301105.06.36
3DGS (full-res.)0.36462904.58.82
CAGE-GS0.22462357.08.47
Wavelet-basedClean 3DGS (full-res.)0.36892281.77.81
Wavelet-aware3DGS (full-res.)0.64785870.311.64
Wavelet-only (full-res.)0.53153605.710.66
CAGE-GS0.31082883.79.43

Importantly, the resolution-aware and mixed-resolution attacks increase Gaussian counts in half-resolution 3DGS by 37.3% and 33.1%, showing that adversarial perturbations remain effective after downsampling. CAGE-GS is thus evaluated under perturbations that survive its own preprocessing.

Nevertheless, compared with attacked full-resolution 3DGS, CAGE-GS reduces Gaussian counts by 38.7%/38.4% and peak memory by 20.4%/18.9% under these two attacks. Against the wavelet-aware attack, CAGE-GS further reduces Gaussian counts by 52.0% and peak memory by 50.9% relative to attacked vanilla 3DGS.

Overall, preprocessing-aware attacks do not bypass CAGE-GS: even the strongest variants still leave 38.4%–52.0% Gaussian-count reductions and 18.9%–50.9% peak-memory reductions relative to their attacked baselines. This supports the benefit of combining mild image purification with scene-aware 3D Gaussian expansion control, since a perturbation tuned to one preprocessing stage does not remove the expansion control imposed by the structural and density priors. We will add these adaptive-attack experiments and a brief discussion of end-to-end attacks targeting the priors to the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 7Ya3 (Part 2 / 5)

Q1. Clarification and Correction of the Evaluation Protocol

Thank you for identifying this important issue. We apologize for not clearly distinguishing the evaluation protocols in the original manuscript.

Our experiments use the poisoned datasets released by Poison-Splat, as reproducing its attack optimization at scale was computationally prohibitive. Specifically, the released NeRF-Synthetic (NS) data contain only training views, whereas the Mip-NeRF 360 images are resized to a width of 1600 pixels. To ensure consistency, we applied the same resizing to clean Mip-NeRF 360 inputs. Mip-NeRF 360 and Tanks and Temples (TT) use llffhold=8 for held-out evaluation, while the original NS quality results were evaluated on training views. This distinction, together with the evaluation resolution, was insufficiently documented and explains why our reported results are not directly comparable to the standard test-view numbers.

To address this issue, we additionally conducted a held-out camera evaluation on all eight NS scenes. We partitioned the 100 available views per scene using llffhold=8 (87 training / 13 testing), trained each method for 30,000 iterations, and evaluated against the original clean 800×800 ground-truth images of the held-out views.

R3-Table2: NS clean-input reconstruction quality under the new holdout-8 evaluation, averaged over eight scenes.

MethodPSNR ↑SSIM ↑LPIPS ↓
Standard 3DGS33.9170.969460.03046
Gaussian Blur (r=1)31.0430.952500.06140
DefenseSplat (wavelet-zero)27.1320.906160.10026
RemedyGS27.3560.514540.05762
CAGE-GS33.8520.969370.03089

Under this held-out camera split, CAGE-GS achieves reconstruction quality close to standard 3DGS, with only a 0.065 dB PSNR difference, while substantially outperforming the evaluated purification baselines. The corresponding average Gaussian counts are 0.2053M for CAGE-GS and 0.2917M for standard 3DGS.

The reruns reuse preprocessed images from the original pipeline; before finalizing these numbers, we will verify their provenance so that both the held-out split and the preprocessing are fully documented.

In the revised manuscript, we will correct the NS evaluation reporting, clearly distinguish training-view and held-out-view metrics, specify the resolution and split used for each dataset, and verify the preprocessing provenance before finalizing the held-out results. We appreciate the reviewer for helping us identify this reporting issue.

AuthorsOpenReview ↗

Response to Reviewer 7Ya3 (Part 3 / 5)

Q2. Source of the Gains

Thank you for this question. We conducted two additional ablations on Drums/Ficus (NS) and Ballroom/Truck (TT): (a) wavelet denoising alone with full-resolution vanilla 3DGS, and (b) 3D priors without wavelet denoising. We report these results alongside the existing component-wise ablations for a comprehensive comparison.

R3-Table3: Component-wise ablation of poisoned-input resource usage and clean-input reconstruction quality, averaged over two scenes per dataset.

DatasetConfigurationGaussians (M) ↓Memory (GB) ↓Time (min) ↓PSNR ↑SSIM ↑LPIPS ↓
NS3DGS (Clean)0.3264.007.3733.6450.98540.0175
3DGS (Poisoned)0.4824.1310.05———
Resolution only0.4025.019.5433.6030.98530.0175
+ Mild Wavelet0.3323.598.9533.5650.98520.0176
+ AE Gate0.2403.188.2233.1000.98300.0180
+ Congestion Control (Full)0.2523.229.0633.5400.99400.0120
Wavelet-only (full-res.)0.4042.809.6633.6020.98520.0175
3D Priors w/o Wavelet0.2822.548.6633.5680.98490.0178
TT3DGS (Clean)2.8608.6723.2125.3790.87280.1327
3DGS (Poisoned)3.72710.8229.19———
Resolution only3.3899.2724.7125.3910.87320.1338
+ Mild Wavelet2.1346.9319.4625.1460.84930.1799
+ AE Gate1.7085.8617.5724.8100.84300.1850
+ Congestion Control (Full)1.7886.3118.5925.0300.85800.1650
Wavelet-only (full-res.)2.3057.3222.0825.1210.84320.1797
3D Priors w/o Wavelet2.4997.7721.3725.1760.86850.1483

Resource metrics are measured under poisoned inputs (except the clean 3DGS reference), while quality metrics are evaluated on clean inputs. The two bold configurations are the newly added ablations requested by the reviewer.

As shown above, both defense mechanisms independently suppress attack-induced Gaussian expansion. Wavelet-only reduces Gaussian counts by 16.2%/38.2% on NS/TT, while 3D priors without wavelet achieve 41.5%/32.9% reductions, demonstrating the independent effectiveness of representation-level defense. Combining both further improves these reductions to 47.8%/52.0%, respectively.

These results support our complementary design: wavelet denoising provides initial input-level purification with a moderate, scene-dependent fidelity cost, while 3D priors further suppress abnormal Gaussian expansion. Within the 3D stage, the AE gate and congestion control also complement each other: the former provides stronger expansion suppression, while the latter recovers 0.44/0.22 dB in clean PSNR by preserving legitimate refinement. Together, these components achieve a favorable balance between computational-cost mitigation and reconstruction fidelity. We will add these two ablations to the revised manuscript.

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

This paper proposes CAGE-GS, a defense against computation-cost attacks on 3D Gaussian Splatting. It derives structural and density priors from a low-resolution reconstruction and uses them to regulate Gaussian expansion during full-resolution optimization. Experiments show reduced Gaussian count, memory usage, and training time under Poison-Splat attacks.

Critical strengths
  1. The method directly regulates attack-induced Gaussian expansion instead of relying only on image purification.

  2. The experiments show substantial resource reductions while preserving better clean-input quality than the compared defenses.

Critical weaknesses
  1. The novelty is limited. The method mainly combines wavelet denoising, coarse-to-fine reconstruction, autoencoder-based anomaly detection, and voxel occupancy control. Comparisons with simple coarse-to-fine training, Gaussian-count limits, or spatial occupancy budgets are missing.

2 . The security evaluation covers only Poison-Splat under a non-adaptive setting. An adaptive attacker could optimize through downsampling and wavelet preprocessing or directly corrupt the low-resolution priors.

Questions for authors
  1. How does CAGE-GS compare with the same coarse-to-fine schedule without the structural and density priors?

  2. How does it compare with a Gaussian-count cap or spatial occupancy budget matched to the same final Gaussian count?

3.Can the method resist an adaptive attacker that explicitly targets the low-resolution reconstruction and prior construction?

8 comments
AuthorsOpenReview ↗

Response to Reviewer koLt (Part 1 / 3)

We really appreciate the reviewer for the detailed and constructive feedback. We are pleased that you found our direct regulation of attack-induced Gaussian expansion, rather than image purification alone, to be a meaningful design, and that you valued the substantial resource savings achieved while preserving better clean-input quality than the compared defenses.

We believe our response has addressed most of your concerns and we respectfully ask if you're comfortable to re-consider the rating in light of our efforts and dedication. Below, we provide detailed responses to all the comments:

Q1. Novelty and Contribution Beyond Coarse-to-Fine Training

Thank you for the comment. We respectfully clarify that the key innovation of CAGE-GS goes beyond composing existing techniques: it extends computation-cost defense from image-level purification to scene-adaptive 3D Gaussian expansion control.

Existing defenses primarily suppress adversarial perturbations in the image domain. While aggressive high-frequency filtering can mitigate attacks, it also removes legitimate textures and details, leading to substantial clean-input quality degradation (Table 3). To address this trade-off, CAGE-GS introduces a complementary two-stage defense: mild wavelet denoising attenuates attack-related perturbations while preserving useful details, and low-resolution reconstruction provides scene-specific structural and density priors to regulate subsequent full-resolution Gaussian densification. Crucially, the coarse reconstruction is not merely used for initialization; its geometric and spatial statistics explicitly guide the acceptance and pruning of Gaussians.

The coarse-to-fine baseline without priors is already reported in Table 4 of the paper; below we summarize it together with additional independent variants.

R2-Table1: Gaussian counts under poisoned inputs and PSNR under clean inputs, averaged over Drums/Ficus (NS) and Ballroom/Truck (TT).

DatasetConfigurationGaussians (M) ↓Clean PSNR ↑
NSVanilla 3DGS (Poisoned)0.482—
Vanilla 3DGS (Clean)0.32633.645
Resolution only0.40233.603
+ Mild Wavelet0.33233.565
+ AE Gate0.24033.100
Full CAGE-GS0.25233.540
Wavelet-only (full-res.)0.40433.602
3D Priors w/o Wavelet0.28233.568
TTVanilla 3DGS (Poisoned)3.727—
Vanilla 3DGS (Clean)2.86025.379
Resolution only3.38925.391
+ Mild Wavelet2.13425.146
+ AE Gate1.70824.810
Full CAGE-GS1.78825.030
Wavelet-only (full-res.)2.30525.121
3D Priors w/o Wavelet2.49925.176

The first six configurations include the progressive ablations in the original paper. The two bold configurations are additional independent ablations.

As shown above, coarse-to-fine training alone reduces Gaussian counts by only 16.5%/9.1% on NS/TT. Under the same resolution schedule and wavelet preprocessing, introducing our 3D priors further reduces Gaussian counts by 24.1%/16.2%, with only 0.025/0.116 dB additional clean-input PSNR loss. Moreover, removing wavelet denoising while retaining the 3D priors still achieves 41.5%/32.9% Gaussian reductions, demonstrating that the 3D priors provide protection independently of image-level filtering.

These results highlight our core contribution: transforming the observed stability of low-resolution reconstruction into actionable 3D structural and spatial-density constraints. The structural prior suppresses anomalous Gaussian expansion, while density-aware congestion control preserves legitimate refinement. This complementary design enables CAGE-GS to mitigate computation-cost attacks without the severe fidelity degradation associated with aggressive image purification. We will add the two independent ablations (wavelet-only and 3D priors without wavelet) and this discussion of our contribution to the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer koLt (Part 2 / 3)

Q2. Comparison with Simple Gaussian Limits

Thank you for the suggestion. We conducted additional experiments comparing CAGE-GS with both a global Gaussian-count cap and a uniform voxel occupancy budget.

(1) Global Gaussian Count Cap. The original Poison-Splat paper (Lu et al., 2025, Sec. 4.4 and Appendix I) already investigated Gaussian-count limiting as a naive defense and highlighted the difficulty of selecting a threshold that balances resource protection and reconstruction fidelity across scenes.

To further examine this limitation, we trained vanilla full-resolution 3DGS for 30k iterations on four clean scenes, with and without a shared 266k Gaussian cap. This cap approximately matches Ficus, which has the smallest Gaussian count among the four unconstrained reconstructions.

R2-Table2: Clean reconstruction quality under a shared global Gaussian-count cap.

SceneConfigurationGaussians (M)PSNR ↑SSIM ↑LPIPS ↓
DrumsNo cap0.383529.4620.97640.0279
Global cap0.265829.3630.97590.0283
FicusNo cap0.265537.8390.99440.00688
Global cap0.265637.8400.99440.00686
BallroomNo cap3.113225.2630.86240.1189
Global cap0.266024.4100.83350.1659
TruckNo cap2.525925.4060.88230.1463
Global cap0.265722.5120.81730.2368

The same cap has negligible impact on Ficus but reduces PSNR by 0.853 dB on Ballroom and 2.894 dB on Truck, with LPIPS increasing by approximately 39.5% and 61.9%, respectively. These results illustrate the scene-dependent quality cost of a shared threshold. Raising the cap could alleviate quality degradation but would also weaken its resource constraint. (2) Uniform Voxel Occupancy Budget. We further implemented a spatially uniform budget using the same wavelet preprocessing, voxel grid, and 7k/23k coarse-to-fine training schedule as CAGE-GS, while disabling the structural AE gate and scene-adaptive density control. Each voxel, including newly occupied voxels, receives the same capacity.

For a controlled comparison, we independently calibrated the uniform voxel capacity for each scene and input condition (clean/poisoned), matching the corresponding CAGE-GS final Gaussian count within ±5%. Calibration used Gaussian counts only, not reconstruction quality.

R2-Table3: Clean-input reconstruction quality under approximately matched Gaussian counts.

SceneMethodGaussians (M)PSNR ↑SSIM ↑LPIPS ↓
DrumsUniform Voxel0.264129.3060.974940.02845
CAGE-GS0.270029.3730.975970.02819
FicusUniform Voxel0.156637.0820.984160.00823
CAGE-GS0.159337.6640.994140.00726
BallroomUniform Voxel2.038825.0100.850410.15244
CAGE-GS2.108225.0630.850900.14676
TruckUniform Voxel1.848624.7680.853020.19850
CAGE-GS1.769425.0080.866760.18352

R2-Table4: Poisoned-input resource consumption under approximately matched Gaussian counts.

SceneMethodGaussians (M)Peak Memory (MB) ↓Time (min) ↓
DrumsUniform Voxel0.3223272110.25
CAGE-GS0.331224199.45
FicusUniform Voxel0.161520317.92
CAGE-GS0.164820578.00
BallroomUniform Voxel1.9370669337.45
CAGE-GS2.0166753921.43
TruckUniform Voxel1.5582582326.62
CAGE-GS1.4874568115.92
MeanUniform Voxel0.9947431720.56
CAGE-GS1.0000442413.70

Despite comparable final Gaussian counts, CAGE-GS achieves higher PSNR and SSIM and lower LPIPS across all four clean scenes. Under poisoned inputs, it also reduces average training time by 33.4%, although average peak memory is slightly higher (2.5%).

Importantly, this is an oracle count-matched diagnostic, not a deployable uniform-budget defense: the voxel capacities were separately tuned for each scene and input condition. In practice, selecting such capacities without knowing the target Gaussian count is itself challenging.

Together, these experiments highlight the distinction between limiting Gaussian quantity and regulating Gaussian expansion. Global caps cannot readily accommodate varying scene complexity, while uniform voxel budgets ignore local structural and density differences. CAGE-GS instead derives scene-specific priors from coarse reconstruction to guide which Gaussian expansions should be retained or suppressed, providing a more favorable reconstruction-quality and training-efficiency trade-off in our evaluated settings. We will add these comparisons with the global Gaussian-count cap and the uniform voxel budget, together with a note on the diagnostic nature of the oracle-matched setting, to the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer koLt (Part 3 / 3)

Q3. Robustness Against Adaptive Attacks

Thank you for raising this concern. We evaluated three adaptive Poison-Splat variants under the same perturbation budget (ϵ=16/255\epsilon=16/255):

  1. Resolution-aware: Optimizes both the attack objective and proxy reconstruction at half resolution (400×400).
  2. Mixed-resolution: Jointly optimizes half- and full-resolution objectives.
  3. Wavelet-aware: Incorporates our two-level db4 wavelet denoising into attack optimization, using identity BPDA for gradient approximation.

All attacks generate 800×800 poisoned images. The first two explicitly target low-resolution reconstruction, while the third tests whether perturbations can survive wavelet preprocessing.

R2-Table5: Resource consumption under adaptive Poison-Splat attacks. Resolution-based attacks are averaged over Ficus/Materials/Mic/Drums; wavelet-aware results are averaged over Chair/Ficus/Lego.

AttackVictim ConfigurationGaussians (M) ↓Memory (MB) ↓Time (min) ↓
Resolution-basedClean 3DGS (half-res.)0.1750985.55.61
Clean 3DGS (full-res.)0.25342100.06.69
Resolution-aware3DGS (half-res.)0.24021122.56.50
3DGS (full-res.)0.33782931.08.61
CAGE-GS0.20702334.08.30
Mixed-resolution3DGS (half-res.)0.23301105.06.36
3DGS (full-res.)0.36462904.58.82
CAGE-GS0.22462357.08.47
Wavelet-basedClean 3DGS (full-res.)0.36892281.77.81
Wavelet-aware3DGS (full-res.)0.64785870.311.64
Wavelet-only (full-res.)0.53153605.710.66
CAGE-GS0.31082883.79.43

The two resolution-adaptive attacks increase Gaussian counts in half-resolution 3DGS by 37.3%/33.1%, showing that targeted perturbations remain effective at low resolution. Even so, CAGE-GS reduces Gaussian counts by 38.7%/38.4% and peak memory by 20.4%/18.9%, respectively, compared with attacked full-resolution 3DGS.

The wavelet-aware attack also remains effective against vanilla 3DGS, increasing Gaussian counts by 75.6% and peak memory by 157.3% over clean training. However, CAGE-GS reduces these costs by 52.0% and 50.9%, respectively, demonstrating substantial additional protection beyond wavelet-only filtering.

Overall, adapting Poison-Splat to low-resolution reconstruction or wavelet preprocessing does not bypass CAGE-GS: even the strongest adaptive variants still incur 38.4%–52.0% Gaussian reductions and 18.9%–50.9% memory reductions relative to their attacked baselines. This is consistent with our complementary design, in which input-level perturbation attenuation and scene-aware 3D Gaussian expansion control reinforce each other. We will add these adaptive-attack experiments and a brief discussion of end-to-end attacks that target the priors to the revised manuscript.

Reviewer_19pUOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies defenses for 3D Gaussian Splatting (3DGS) against computation-cost attacks. An attacker can add almost invisible noise to the training images, causing 3DGS to create far more Gaussians than usual during training, which consumes large amounts of GPU memory and time and can eventually crash a service.

The proposed defense, CAGE-GS, first applies a mild denoising step to the input images and downsamples them by half to train a coarse version of the scene. From this coarse version it learns two references: one judges whether the position, size, and shape of each Gaussian are "normal", and the other estimates how many Gaussians a spatial region should roughly contain. When training returns to full resolution, newly added Gaussians that look abnormal are rejected, and overly crowded regions are pruned further.

Experiments on three common datasets report that this brings the number of Gaussians, peak memory, and training time under attack back close to clean-training levels, while preserving good image quality on clean inputs.

Critical strengths
  1. The observation that a poisoned scene looks much closer to a clean one at low resolution is valuable. It may be useful to others building similar defenses.

  2. On clean images, the reconstruction quality is clearly better than that of the compared defenses. In Table 3, CAGE-GS reaches 36.80 dB on NeRF-Synthetic, almost the same as vanilla 3DGS at 36.98 dB, whereas DefenseSplat and RemedyGS reach only 27.64 and 28.12 dB. For the goal of stopping the attack without damaging the images, this is a convincing result.

Critical weaknesses
  1. No test of an attack designed specifically against this defense. The whole method rests on the premise that the attack becomes weaker at low resolution. But if the attacker knows that the defense downsamples the images by half, they can simply craft the noise for the downsampled images from the start. In that case, the coarse scene used as the reference may itself be poisoned.

  2. Image quality on clean inputs still drops, and the paper does not explain where the drop comes from. Table 3 shows that CAGE-GS is 0.18–0.68 dB below vanilla 3DGS. Because the defense also denoises clean images and also runs the half-resolution stage, this loss is unavoidable. The paper ablates only one component on clean inputs (congestion control) and does not say whether the loss comes from denoising, from the half-resolution stage, or from the new filtering. For a method whose selling point is preserving image quality, this cost should be explained.

Questions for authors
  1. If the attacker crafts the noise directly for the half-resolution images (or for the denoised images), does CAGE-GS still work? If so, please provide a quantitative experiment.
  2. Which step causes most of the 0.18–0.68 dB drop on clean images? Can you report results with denoising, the half-resolution stage, and the new filtering each turned off?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 19pU (Part 1 / 2)

We really appreciate the reviewer for the detailed and constructive feedback. We are pleased that you found our low-resolution stability observation valuable and acknowledged that CAGE-GS preserves clean-input reconstruction quality better than existing defenses.

We believe our response has addressed most of your concerns and we respectfully ask if you're comfortable to re-consider the rating in light of our efforts and dedication. Below, we provide detailed responses to all the comments:

Q1. Adaptive Attack Robustness

Thank you for raising this important concern. We constructed three adaptive Poison-Splat variants that explicitly target our defense preprocessing. All variants use the same optimization framework and perturbation budget (ϵ=16/255\epsilon=16/255) as the original attack, so any increase in destructiveness stems from adaptivity rather than from a larger budget:

  1. Resolution-aware: Optimizes the attack objective on half-resolution (400×400) images.
  2. Mixed-resolution: Jointly optimizes the attack objectives at half and full resolutions.
  3. Wavelet-aware: Incorporates our two-level db4 wavelet denoising into attack optimization, using identity BPDA to approximate gradients.

All perturbations are generated in the original 800×800 image space. We evaluate the first two attacks on Ficus/Materials/Mic/Drums and the third on Chair/Ficus/Lego.

R1-Table1: Resource consumption under adaptive Poison-Splat attacks, averaged over four scenes for resolution-based attacks and three scenes for the wavelet-aware attack.

AttackVictim ConfigurationGaussians (M) ↓Memory (MB) ↓Time (min) ↓
Resolution-basedClean 3DGS (half-res.)0.1750985.55.61
Clean 3DGS (full-res.)0.25342100.06.69
Resolution-aware3DGS (half-res.)0.24021122.56.50
3DGS (full-res.)0.33782931.08.61
CAGE-GS0.20702334.08.30
Mixed-resolution3DGS (half-res.)0.23301105.06.36
3DGS (full-res.)0.36462904.58.82
CAGE-GS0.22462357.08.47
Wavelet-basedClean 3DGS (full-res.)0.36892281.77.81
Wavelet-aware3DGS (full-res.)0.64785870.311.64
Wavelet-only (full-res.)0.53153605.710.66
CAGE-GS0.31082883.79.43

The results show that resolution-aware attacks are effective at half resolution, increasing Gaussian counts by 37.3%/33.1% over clean half-resolution training. Even so, CAGE-GS reduces Gaussian counts by 38.7%/38.4% and peak memory by 20.4%/18.9% compared with attacked full-resolution 3DGS.

Importantly, the wavelet-aware attack is also effective, increasing the Gaussian count and peak memory of vanilla 3DGS by 75.6% and 157.3% over clean training. Yet CAGE-GS reduces these costs by 52.0% and 50.9%, respectively, and further reduces Gaussian counts by 41.5% compared with wavelet-only defense.

Overall, adapting Poison-Splat to individual preprocessing stages does not bypass CAGE-GS: even the strongest adaptive variants remain 38.4%–52.0% below their attacked baselines in Gaussian count and 18.9%–50.9% below in peak memory. This is consistent with our complementary design: a perturbation tuned to one preprocessing stage does not remove the expansion control imposed by the structural and density priors in the full pipeline. We will add these adaptive-attack experiments and a brief discussion of more general end-to-end attacks to the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 19pU (Part 2 / 2)

Q2. Clean Reconstruction Quality Degradation

Thank you for the suggestion. We conducted additional component-wise ablations on Drums/Ficus (NS) and Ballroom/Truck (TT) to identify the sources of clean-input quality degradation. We also report the corresponding poisoned-input resource usage to clarify the quality–defense trade-off.

R1-Table2: Component-wise ablation of poisoned-input resource usage and clean-input reconstruction quality, averaged over two scenes per dataset.

DatasetConfigurationGaussians (M) ↓Memory (GB) ↓Time (min) ↓PSNR ↑SSIM ↑LPIPS ↓
NS3DGS (Clean)0.3264.007.3733.6450.98540.0175
3DGS (Poisoned)0.4824.1310.05———
Resolution only0.4025.019.5433.6030.98530.0175
+ Mild Wavelet0.3323.598.9533.5650.98520.0176
+ AE Gate0.2403.188.2233.1000.98300.0180
+ Congestion Control (Full)0.2523.229.0633.5400.99400.0120
Wavelet-only (full-res.)0.4042.809.6633.6020.98520.0175
3D Priors w/o Wavelet0.2822.548.6633.5680.98490.0178
TT3DGS (Clean)2.8608.6723.2125.3790.87280.1327
3DGS (Poisoned)3.72710.8229.19———
Resolution only3.3899.2724.7125.3910.87320.1338
+ Mild Wavelet2.1346.9319.4625.1460.84930.1799
+ AE Gate1.7085.8617.5724.8100.84300.1850
+ Congestion Control (Full)1.7886.3118.5925.0300.85800.1650
Wavelet-only (full-res.)2.3057.3222.0825.1210.84320.1797
3D Priors w/o Wavelet2.4997.7721.3725.1760.86850.1483

Resource metrics are measured under poisoned inputs (except the clean 3DGS reference), while quality metrics are evaluated on clean inputs. The first six configurations show progressive component ablations, and the two bold configurations are additional independent ablations.

As shown above, low-to-high resolution training has negligible quality impact (−0.042 dB on NS; +0.012 dB on TT). Wavelet denoising introduces a scene-dependent PSNR reduction (0.038/0.245 dB) while reducing attack-induced Gaussian counts, reflecting the trade-off of input-level purification. The AE gate further reduces PSNR by 0.465/0.336 dB while providing stronger expansion suppression, whereas congestion control recovers 0.440/0.220 dB by allowing legitimate Gaussian refinement. The drop is therefore dominated by the AE gate, while the half-resolution stage and wavelet denoising contribute only marginally.

These results align with our overall design: mild wavelet denoising provides initial attack mitigation with a manageable fidelity cost, while the structural AE gate and density-aware congestion control complement each other to balance expansion suppression and detail preservation. Consequently, Full CAGE-GS maintains near-vanilla reconstruction quality on these scenes (−0.105/−0.349 dB), demonstrating the intended balance between computational-cost defense and reconstruction fidelity. We will include this component-wise ablation and the accompanying quality-degradation analysis in the revised manuscript.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we’re eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper!

Comments on the paper

2 comments
Area_Chair_a2vMOpenReview ↗

Follow-Up on the Authors’ Rebuttal

Dear reviewers,

The authors have posted their rebuttal. Please take a moment to read their response and leave a brief follow-up comment on whether it addresses the concerns you raised in your review. If any points remain unclear or you have follow-up questions, please raise them during the rebuttal period so the authors have a chance to respond. Please also indicate whether the rebuttal changes your overall assessment of the paper.

Thank you !

AC

AuthorsOpenReview ↗

Summary Remarks to Area Chair and Reviewers

Dear Area Chair and Reviewers,

Greetings from the authors of submission #14047. We sincerely thank you for your time, effort, and constructive feedback throughout the review and rebuttal process.

CAGE-GS defends 3D Gaussian Splatting against computation-cost attacks. Its design is not a simple combination of existing techniques, but is derived from a specific limitation of current defenses and a specific observation. Existing defenses operate mainly in the image domain: aggressive purification suppresses perturbations but also removes legitimate high-frequency details, while leaving abnormal Gaussian geometry and spatial congestion insufficiently constrained. Simple counters such as global Gaussian-count caps or uniform occupancy budgets cannot adapt to scene complexity, as our new experiments confirm. Our observation is that poisoned reconstructions become much closer to clean ones at half resolution. CAGE-GS therefore pairs mild wavelet denoising with a half-resolution coarse stage to learn scene-specific structural and density priors, and uses these priors to decide which Gaussian expansions are plausible during full-resolution optimization. The contribution is this translation of a robustness observation into scene-adaptive expansion control, not the individual components.

During the rebuttal, we addressed the main concerns with new experiments: (i) three adaptive Poison-Splat variants targeting downsampling and wavelet preprocessing, under which CAGE-GS still reduces Gaussian counts by 38.4%–52.0% and peak memory by 18.9%–50.9% relative to attacked baselines; (ii) component-wise attribution of the clean-input quality drop, showing that it is dominated by the structural AE gate and largely recovered by congestion control; (iii) comparisons against coarse-to-fine training without priors, a global Gaussian-count cap, and a count-matched uniform voxel budget, where CAGE-GS achieves higher PSNR/SSIM and lower LPIPS, and 33.4% less training time than the voxel baseline; (iv) a corrected held-out evaluation protocol, on which CAGE-GS is only 0.065 dB below standard 3DGS; (v) sensitivity analyses for wavelet strength and voxel size; (vi) the AE training overhead, which accounts for only 0.19%–0.96% of total training time; and (vii) a 56-run study of the AE role and capacity, showing that it provides a better fidelity–resource balance than attribute-wise quantile and regularized Mahalanobis filters, and that an enlarged bottleneck weakens the reconstruction-error separation, as the reviewer anticipated.

We are grateful for the reviewers' engagement. Reviewer 19pU found that the adaptive-attack experiments and component-wise ablations substantially addressed the original concerns, and has updated the assessment accordingly. Reviewer 7Ya3, who has substantial direct research experience on this topic (confidence 4), recognized the threat as real and recent, the resource savings as consistent and significant across datasets, implementations, and perturbation budgets, and the presentation as clear. We regret that Reviewer koLt did not respond during the discussion period despite two follow-ups; we nevertheless believe our responses fully address the three questions raised (comparison with coarse-to-fine training, Gaussian-count caps and occupancy budgets, and adaptive attackers), and we hope the AC will take these responses into account. We respectfully ask the AC to weigh these results and the positive assessments of the reviewers familiar with this area, and we thank you again for your careful consideration.

Best regards, The Authors of Submission #14047

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.