- 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
-
It addresses a real and recent threat, and the research question is interesting and important.
-
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.
-
The paper is well-written; the motivation is well explained and the method clearly described.
Critical weaknesses
- 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:
-
The authors should use \citep{} and \citet{} more properly. Currently they use \citet{} all the time and it affects reading.
-
Some presentation could be improved:
- Line 081: "become substantially more clean-like"?
Questions for authors
-
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).
-
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?
-
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?
-
What is the training overhead of autoencoder?
8 comments
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.
| Method | PSNR ↑ | SSIM ↑ | LPIPS ↓ |
|---|---|---|---|
| Standard 3DGS | 33.917 | 0.96946 | 0.03046 |
| Gaussian Blur (r=1) | 31.043 | 0.95250 | 0.06140 |
| DefenseSplat (wavelet-zero) | 27.132 | 0.90616 | 0.10026 |
| RemedyGS | 27.356 | 0.51454 | 0.05762 |
| CAGE-GS | 33.852 | 0.96937 | 0.03089 |
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.
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.
| Dataset | Configuration | Gaussians (M) ↓ | Memory (GB) ↓ | Time (min) ↓ | PSNR ↑ | SSIM ↑ | LPIPS ↓ |
|---|---|---|---|---|---|---|---|
| NS | 3DGS (Clean) | 0.326 | 4.00 | 7.37 | 33.645 | 0.9854 | 0.0175 |
| 3DGS (Poisoned) | 0.482 | 4.13 | 10.05 | — | — | — | |
| Resolution only | 0.402 | 5.01 | 9.54 | 33.603 | 0.9853 | 0.0175 | |
| + Mild Wavelet | 0.332 | 3.59 | 8.95 | 33.565 | 0.9852 | 0.0176 | |
| + AE Gate | 0.240 | 3.18 | 8.22 | 33.100 | 0.9830 | 0.0180 | |
| + Congestion Control (Full) | 0.252 | 3.22 | 9.06 | 33.540 | 0.9940 | 0.0120 | |
| Wavelet-only (full-res.) | 0.404 | 2.80 | 9.66 | 33.602 | 0.9852 | 0.0175 | |
| 3D Priors w/o Wavelet | 0.282 | 2.54 | 8.66 | 33.568 | 0.9849 | 0.0178 | |
| TT | 3DGS (Clean) | 2.860 | 8.67 | 23.21 | 25.379 | 0.8728 | 0.1327 |
| 3DGS (Poisoned) | 3.727 | 10.82 | 29.19 | — | — | — | |
| Resolution only | 3.389 | 9.27 | 24.71 | 25.391 | 0.8732 | 0.1338 | |
| + Mild Wavelet | 2.134 | 6.93 | 19.46 | 25.146 | 0.8493 | 0.1799 | |
| + AE Gate | 1.708 | 5.86 | 17.57 | 24.810 | 0.8430 | 0.1850 | |
| + Congestion Control (Full) | 1.788 | 6.31 | 18.59 | 25.030 | 0.8580 | 0.1650 | |
| Wavelet-only (full-res.) | 2.305 | 7.32 | 22.08 | 25.121 | 0.8432 | 0.1797 | |
| 3D Priors w/o Wavelet | 2.499 | 7.77 | 21.37 | 25.176 | 0.8685 | 0.1483 |
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.
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):
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.
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.