- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
Overall this paper introduces a benchmark for 3DGS posioning detection. This work collects 100 base scenes and then positions them into 1253 or 1353 scenes via existing security methods on 3DGS like StealthAttack, GaussTrap, GuardSplat, etc. Results on detector performance across S1, S2, S3, and S4 are reported along with some findings. Based on the current version and reported results, I think it's difficult to accept for ICLR.
Critical strengths
(+) The idea of detecting the posioned 3DGS on multi-stage is interesting.
Critical weaknesses
(-) This paper mainly conducts experiments with previous methods, and no novel designation is made, although with some findings. I think the current findings may not be enough to reflect the introduced idea ``representation-centric design enables an attack-agnostic, stage-wise evaluation." The idea of detecting the posioned 3DGS on multi-stage is interesting, but the solution and findings is not novel enough. I think introducing methods that could detect these attacks at different stages persepctive could greatly strengthen this manuscript.
(-) Also, the current 100-base scenes are not enough for a benchmark. Current results may not be able to effectively indicate the detection method, as there are many scenarios (outdoor, indoor, open terrain...). Also, the results could be categorized on different scenes. It's better to include more scenes and position them.
(-) Correct me if I'm wrong: In Tab. 1, S2, different 3D reconstruction methods are used to calculate multi-view consistency probes. It seems that previous poison methods (GaussTrap, 3D-GSW, and GuardSplat) can also be calculated via the models and get one consistency score to verify detectability. I wonder why the results are not included.
(-) Overall, this paper is not well written and presented. For example, the citation format could be ~\citep{} rather than ~cite{} sometime. Mismatch results in Figure 1(B), 1,253 or 1,353?
Questions for authors
Please see above.
6 comments
Comment for Reviewer 744E (2/3)
(Continued from part 1/3)
I think introducing methods that could detect these attacks at different stages persepctive could greatly strengthen this manuscript.
As the first step, our previous manuscript has already explored a method for detection at different stages (Appendix C.3, Table 9). We studied a multi-stage ensemble that averages standardized anomaly scores from a fixed detector panel across S1 to S4 (SparseViT at S1, MET3R at S2, GradNorm-style at S3, and SplatFormer at S4):
| Attack | Best single-stage member of panel | AUROC ↑ | FPR@95 ↓ | ensemble AUROC ↑ | ensemble FPR@95 ↓ | Δ AUROC |
|---|---|---|---|---|---|---|
| StealthAttack | SparseViT (S1) | 0.773 ± 0.028 | 0.780 ± 0.042 | 0.891 ± 0.018 | 0.400 ± 0.024 | +0.118 ± 0.035 |
| Poison-Splat | SplatFormer (S4) | 0.707 ± 0.033 | 0.740 ± 0.016 | 0.665 ± 0.032 | 0.840 ± 0.015 | −0.042 ± 0.024 |
Interestingly, this ensemble detection method improves StealthAttack detection, where evidence is spread across stages, and lowers Poison-Splat detection, where evidence is concentrated at S4. Poison-3DGS enables future work on stage-aware detectors and defenses.
[1] Li, Y., Liu, Z., Li, Z., Lin, Z., & Zhang, J. (2026). RemedyGS: Defend 3D Gaussian Splatting against Computation Cost Attacks. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).
[2] Qiao, Y., Lu, Y., Zhou, Y., Yang, R., Hou, L., Yin, Y., & Ma, J. (2026). DefenseSplat: Enhancing the Robustness of 3D Gaussian Splatting via Frequency-Aware Filtering. European Conference on Computer Vision (ECCV).
(W2) Also, the current 100-base scenes are not enough for a benchmark. Current results may not be able to effectively indicate the detection method, as there are many scenarios (outdoor, indoor, open terrain...). Also, the results could be categorized on different scenes. It's better to include more scenes and position them.
A (W2): We thank the reviewer for the comment.
Scale. We believe 100 base scenes is very substantial for this area (as illustrated in Fig. 1(b)). Standard 3DGS benchmarks have 5–21 scenes each (e.g., Mip-NeRF 360: 9; Tanks and Temples: 21), so our scene pool is about 5 to 20 times larger. Our benchmark contains 1,253 poisoned variants and 1,353 evaluation scenes in total. Every scene comes with a trained 3DGS model and the artifacts of all four stages.
Diversity. As illustrated in Fig. 1(b), the scenes are split evenly between indoor (50) and outdoor (50). They cover architecture (61), interior objects (43), open terrain (39), standalone objects (34), and vegetation (30). A scene can belong to more than one category.
Results by scene type. Appendix C.2 (Table 8) reports indoor and outdoor results. As the reviewer suggests, we now also report results for the five categories: (a) AUROC and (b) FPR@95 (added as Appendix C.5, Table 11, marked in blue).
(a) AUROC ↑
| Stage | Detector | Vegetation | Architecture | Interior objects | Standalone object | Open terrain |
|---|---|---|---|---|---|---|
| S1 | SoftPatch+ | 0.633 ± 0.034 | 0.647 ± 0.021 | 0.653 ± 0.027 | 0.673 ± 0.028 | 0.680 ± 0.029 |
| S1 | SparseViT | 0.652 ± 0.039 | 0.678 ± 0.030 | 0.641 ± 0.037 | 0.655 ± 0.037 | 0.712 ± 0.032 |
| S1 | MVSS-Net | 0.548 ± 0.030 | 0.502 ± 0.028 | 0.502 ± 0.036 | 0.531 ± 0.027 | 0.530 ± 0.026 |
| S1 | IML-ViT | 0.574 ± 0.042 | 0.577 ± 0.027 | 0.555 ± 0.032 | 0.619 ± 0.034 | 0.613 ± 0.037 |
| S2 | MET3R | 0.617 ± 0.035 | 0.625 ± 0.025 | 0.623 ± 0.030 | 0.686 ± 0.033 | 0.656 ± 0.030 |
| S2 | VGGT | 0.514 ± 0.028 | 0.570 ± 0.027 | 0.509 ± 0.034 | 0.550 ± 0.028 | 0.575 ± 0.025 |
| S2 | MV-DUSt3R+ | 0.501 ± 0.024 | 0.508 ± 0.020 | 0.500 ± 0.023 | 0.537 ± 0.020 | 0.532 ± 0.019 |
| S2 | NoPoSplat | 0.455 ± 0.028 | 0.394 ± 0.017 | 0.396 ± 0.021 | 0.455 ± 0.018 | 0.437 ± 0.019 |
| S2-PC | Sonata | 0.773 ± 0.050 | 0.921 ± 0.026 | 0.884 ± 0.035 | 0.875 ± 0.038 | 0.881 ± 0.037 |
| S2-PC | Concerto | 0.810 ± 0.050 | 0.845 ± 0.032 | 0.835 ± 0.039 | 0.863 ± 0.039 | 0.836 ± 0.041 |
| S3 | GradNorm-style | 0.611 ± 0.042 | 0.634 ± 0.034 | 0.692 ± 0.036 | 0.621 ± 0.041 | 0.642 ± 0.044 |
| S3 | Spectral-style | 0.662 ± 0.045 | 0.519 ± 0.038 | 0.490 ± 0.044 | 0.635 ± 0.036 | 0.638 ± 0.040 |
| S4 | SplatFormer | 0.649 ± 0.043 | 0.596 ± 0.030 | 0.557 ± 0.032 | 0.701 ± 0.033 | 0.669 ± 0.037 |
| S4 | Gaussian-MAE | 0.585 ± 0.043 | 0.536 ± 0.032 | 0.463 ± 0.032 | 0.613 ± 0.040 | 0.667 ± 0.034 |
| S4 | SceneSplat | 0.495 ± 0.046 | 0.545 ± 0.032 | 0.511 ± 0.038 | 0.496 ± 0.035 | 0.477 ± 0.035 |
(continue in part 3/3)
Comment for Reviewer 744E (3/3)
(Continued from part 2/3)
(b) FPR@95 ↓
| Stage | Detector | Vegetation | Architecture | Interior objects | Standalone object | Open terrain |
|---|---|---|---|---|---|---|
| S1 | SoftPatch+ | 0.700 ± 0.017 | 0.820 ± 0.036 | 0.872 ± 0.015 | 0.691 ± 0.014 | 0.679 ± 0.014 |
| S1 | SparseViT | 0.867 ± 0.039 | 0.787 ± 0.013 | 0.837 ± 0.013 | 0.838 ± 0.054 | 0.795 ± 0.032 |
| S1 | MVSS-Net | 0.950 ± 0.018 | 0.893 ± 0.008 | 0.919 ± 0.009 | 0.926 ± 0.013 | 0.949 ± 0.026 |
| S1 | IML-ViT | 0.900 ± 0.015 | 0.934 ± 0.038 | 0.965 ± 0.018 | 0.897 ± 0.010 | 0.897 ± 0.021 |
| S2 | MET3R | 0.733 ± 0.017 | 0.738 ± 0.020 | 0.872 ± 0.045 | 0.662 ± 0.017 | 0.744 ± 0.015 |
| S2 | VGGT | 0.817 ± 0.015 | 0.902 ± 0.012 | 0.919 ± 0.013 | 0.985 ± 0.008 | 0.987 ± 0.030 |
| S2 | MV-DUSt3R+ | 0.950 ± 0.012 | 0.951 ± 0.008 | 0.930 ± 0.011 | 0.971 ± 0.011 | 0.910 ± 0.014 |
| S2 | NoPoSplat | 0.967 ± 0.006 | 0.959 ± 0.007 | 0.988 ± 0.005 | 0.971 ± 0.007 | 0.962 ± 0.006 |
| S2-PC | Sonata | 0.733 ± 0.035 | 0.393 ± 0.050 | 0.581 ± 0.067 | 0.559 ± 0.045 | 0.615 ± 0.059 |
| S2-PC | Concerto | 0.800 ± 0.046 | 0.508 ± 0.036 | 0.442 ± 0.024 | 0.441 ± 0.026 | 0.615 ± 0.050 |
| S3 | GradNorm-style | 0.900 ± 0.031 | 0.828 ± 0.013 | 0.791 ± 0.022 | 0.868 ± 0.015 | 0.846 ± 0.019 |
| S3 | Spectral-style | 0.767 ± 0.025 | 0.877 ± 0.013 | 0.884 ± 0.016 | 0.838 ± 0.018 | 0.897 ± 0.022 |
| S4 | SplatFormer | 0.847 ± 0.052 | 0.830 ± 0.029 | 0.884 ± 0.026 | 0.682 ± 0.060 | 0.815 ± 0.046 |
| S4 | Gaussian-MAE | 0.873 ± 0.049 | 0.905 ± 0.024 | 0.953 ± 0.028 | 0.894 ± 0.048 | 0.800 ± 0.051 |
| S4 | SceneSplat | 0.913 ± 0.040 | 0.902 ± 0.035 | 0.916 ± 0.029 | 0.929 ± 0.035 | 0.908 ± 0.029 |
The current benchmark provides diverse scenarios for cross-attack, stage-wise evaluation.
(W3) Correct me if I'm wrong: In Tab. 1, S2, different 3D reconstruction methods are used to calculate multi-view consistency probes. It seems that previous poison methods (GaussTrap, 3D-GSW, and GuardSplat) can also be calculated via the models and get one consistency score to verify detectability. I wonder why the results are not included.
A (W3): We thank the reviewer for the question.
GaussTrap, 3D-GSW, and GuardSplat modify a model that has already been trained (Figure 1(c)). The training images and COLMAP output are exactly the same as in the clean scene, so an S2 probe returns the same score for both and cannot tell them apart. We therefore mark these entries “N/A”. Only S4 detectors can see these attacks.
StealthAttack and Poison-Splat change the training multiview images, so the S2 probes do apply to them.
(W4) Overall, this paper is not well written and presented. For example, the citation format could be ~\citep{} rather than ~cite{} sometime. Mismatch results in Figure 1(B), 1,253 (poisoned) or 1,353 (all)?
A (W4): We thank the reviewer for pointing these out.
Citations. We have fixed the citation format throughout the revision, using \citep{} for parenthetical citations.
Figure 1(b). The two numbers are consistent; they count different things. The benchmark has 100 clean base scenes and 1,253 poisoned variants, which gives 1,353 evaluation scenes in total. The bar chart shows the total (1,353), and the panel beside it shows the poisoned variants (1,253). We will label both numbers explicitly in the figure and its caption in the final version.
Presentation. We have also proofread the manuscript and will continue to improve its clarity in the final version.
We thank the reviewer for the constructive feedback.
Comment for Reviewer 744E (1/3)
A (W1): We thank the reviewer for the comment.
Research gap. Detection research for 3DGS poisoning remains limited and fragmented. The few existing detectors [1–2], developed only recently, target specific attack mechanisms (particularly Poison-Splat), or are trained using attack-specific samples. They were also evaluated on different scene datasets and with different metrics, hindering direct comparison across studies. Their attack-specific design could limit transfer to other attacks: as shown in Appendix C.8, Table 14 (blue), they achieve only 0.463–0.519 AUROC on attacks outside their target setting.
The critical gap is the absence of a unified benchmark for studying detection across diverse attacks and stages of the 3DGS pipeline. This leaves an important practical question unanswered: How can a defender assess a potentially compromised 3DGS scene when the attack identity is unknown?
Contribution and novelty. As a benchmark submission, our Poison-3DGS brings five attack families, 100 base scenes from eight datasets, and 1,253 poisoned variants into a common benchmark. Rather than proposing a single attack-specific defense, Poison-3DGS’s novelty lies in moving beyond isolated single-stage evaluations into a unified, pipeline-level detectability framework:
Unifies fragmented evaluations: It standardizes the evaluation of heterogeneous 3DGS poisoning attacks within one benchmark and common protocol.
Enables pipeline-level detectability study: Poison-3DGS is structured based on the representations naturally exposed at four pipeline stages (S1 to S4), with stages defined independently of the attacks. Poison-3DGS enables pipeline-level detectability study for 3DGS for the first time.
Moves beyond isolated attack-specific evaluation: It applies fixed, stage-appropriate detectors across attacks without attack labels, poisoned-sample training, or per-attack adaptation, allowing systematic comparison of where different attacks become detectable.
Together, these design choices enable, to our knowledge, the first systematic study of poisoning detectability across attack families and stages of the 3DGS pipeline.
Our representation-centric design maps each 3DGS pipeline stage to its naturally exposed representation and assigns detectors based on that representation, independently of attack identity. This enables the same fixed, stage-appropriate detectors to be applied across diverse attacks—without attack labels, poisoned-sample training, or attack-specific adaptation—and thus enables a common attack-agnostic, stage-wise evaluation.
Key insight and significance. Poison-3DGS enables unified, stage-wise evaluation. Our experiment results show that, for the first time, the attack injection stage need not be the stage where an attack is most detectable. Importantly, this suggests existing detection methods limited to attacks' entry points can miss downstream forensic evidence. Poison-3DGS enables future work on stage-aware detectors and defenses.
(continue in part 2/3)