acceptodds
Under review as a conference paper at ICLR 2027

Characterizing Detectability in 3DGS Poisoning: A Stage-wise Benchmark

Abstract

3D Gaussian Splatting (3DGS) has rapidly emerged as a leading representation for real-time novel view synthesis, but recent work has shown that it is vulnerable to diverse poisoning attacks, including illusory object injection, computation cost amplification, and post hoc model watermarking. Despite this expanding threat surface, existing studies primarily focus on attack success, while defense and detection remain underexplored. From a detection perspective, a key challenge and opportunity arise from the multi-stage nature of the 3DGS reconstruction pipeline, which produces heterogeneous intermediate representations. Importantly, forensic signals for detecting poisoning are inherently stage dependent: an attack introduced at one stage may produce signals that emerge only at later stages or become more detectable as the pipeline progresses. This motivates a \em stage-wise view of detectability that goes beyond single-stage evaluation. In this work, we introduce Poison-3DGS, a benchmark designed for stage-wise characterization of poisoning detection in 3DGS, comprising 100 diverse base scenes and 1,253 poisoned variants. Moving beyond attack-specific evaluation, Poison-3DGS systematically exposes the stage-specific representations naturally produced throughout the standard 3DGS reconstruction pipeline, including multi-view images, SfM-derived geometry, training dynamics, and the final Gaussian representation. This representation-centric design enables an attack-agnostic, stage-wise evaluation protocol that naturally accommodates existing and future poisoning attacks. Using this benchmark, we conduct the first systematic study of poisoning detectability across the 3DGS reconstruction pipeline. Our study reveals that the stage where an attack is introduced is not necessarily the stage where it is most detectable. Instead, downstream reconstruction transforms attack effects into stage-specific forensic signals whose detectability varies substantially across stages and attack types. Overall, Poison-3DGS provides a principled benchmark and the first systematic characterization of stage-dependent detectability in 3DGS, establishing a foundation for future research on robust and reliable 3DGS systems. We include the code and benchmark in the submission.

Primary area:
datasets and benchmarks
Keywords:
3DGS, Stage-wise Benchmark, Poison-3DGS

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_744EOpenReview ↗
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
AuthorsOpenReview ↗

Comment for Reviewer 744E (1/3)

(W1) 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.

A (W1): We thank the reviewer for the comment.

This paper mainly conducts experiments with previous methods, and no novel designation is made, although with some findings.

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.

I think the current findings may not be enough to reflect the introduced idea representation-centric design enables an attack-agnostic, stage-wise evaluation.

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.

The idea of detecting the posioned 3DGS on multi-stage is interesting, but the solution and findings is not novel enough.

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)

AuthorsOpenReview ↗

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):

AttackBest single-stage member of panelAUROC ↑FPR@95 ↓ensemble AUROC ↑ensemble FPR@95 ↓Δ AUROC
StealthAttackSparseViT (S1)0.773 ± 0.0280.780 ± 0.0420.891 ± 0.0180.400 ± 0.024+0.118 ± 0.035
Poison-SplatSplatFormer (S4)0.707 ± 0.0330.740 ± 0.0160.665 ± 0.0320.840 ± 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 ↑

StageDetectorVegetationArchitectureInterior objectsStandalone objectOpen terrain
S1SoftPatch+0.633 ± 0.0340.647 ± 0.0210.653 ± 0.0270.673 ± 0.0280.680 ± 0.029
S1SparseViT0.652 ± 0.0390.678 ± 0.0300.641 ± 0.0370.655 ± 0.0370.712 ± 0.032
S1MVSS-Net0.548 ± 0.0300.502 ± 0.0280.502 ± 0.0360.531 ± 0.0270.530 ± 0.026
S1IML-ViT0.574 ± 0.0420.577 ± 0.0270.555 ± 0.0320.619 ± 0.0340.613 ± 0.037
S2MET3R0.617 ± 0.0350.625 ± 0.0250.623 ± 0.0300.686 ± 0.0330.656 ± 0.030
S2VGGT0.514 ± 0.0280.570 ± 0.0270.509 ± 0.0340.550 ± 0.0280.575 ± 0.025
S2MV-DUSt3R+0.501 ± 0.0240.508 ± 0.0200.500 ± 0.0230.537 ± 0.0200.532 ± 0.019
S2NoPoSplat0.455 ± 0.0280.394 ± 0.0170.396 ± 0.0210.455 ± 0.0180.437 ± 0.019
S2-PCSonata0.773 ± 0.0500.921 ± 0.0260.884 ± 0.0350.875 ± 0.0380.881 ± 0.037
S2-PCConcerto0.810 ± 0.0500.845 ± 0.0320.835 ± 0.0390.863 ± 0.0390.836 ± 0.041
S3GradNorm-style0.611 ± 0.0420.634 ± 0.0340.692 ± 0.0360.621 ± 0.0410.642 ± 0.044
S3Spectral-style0.662 ± 0.0450.519 ± 0.0380.490 ± 0.0440.635 ± 0.0360.638 ± 0.040
S4SplatFormer0.649 ± 0.0430.596 ± 0.0300.557 ± 0.0320.701 ± 0.0330.669 ± 0.037
S4Gaussian-MAE0.585 ± 0.0430.536 ± 0.0320.463 ± 0.0320.613 ± 0.0400.667 ± 0.034
S4SceneSplat0.495 ± 0.0460.545 ± 0.0320.511 ± 0.0380.496 ± 0.0350.477 ± 0.035

(continue in part 3/3)

AuthorsOpenReview ↗

Comment for Reviewer 744E (3/3)

(Continued from part 2/3)

(b) FPR@95 ↓

StageDetectorVegetationArchitectureInterior objectsStandalone objectOpen terrain
S1SoftPatch+0.700 ± 0.0170.820 ± 0.0360.872 ± 0.0150.691 ± 0.0140.679 ± 0.014
S1SparseViT0.867 ± 0.0390.787 ± 0.0130.837 ± 0.0130.838 ± 0.0540.795 ± 0.032
S1MVSS-Net0.950 ± 0.0180.893 ± 0.0080.919 ± 0.0090.926 ± 0.0130.949 ± 0.026
S1IML-ViT0.900 ± 0.0150.934 ± 0.0380.965 ± 0.0180.897 ± 0.0100.897 ± 0.021
S2MET3R0.733 ± 0.0170.738 ± 0.0200.872 ± 0.0450.662 ± 0.0170.744 ± 0.015
S2VGGT0.817 ± 0.0150.902 ± 0.0120.919 ± 0.0130.985 ± 0.0080.987 ± 0.030
S2MV-DUSt3R+0.950 ± 0.0120.951 ± 0.0080.930 ± 0.0110.971 ± 0.0110.910 ± 0.014
S2NoPoSplat0.967 ± 0.0060.959 ± 0.0070.988 ± 0.0050.971 ± 0.0070.962 ± 0.006
S2-PCSonata0.733 ± 0.0350.393 ± 0.0500.581 ± 0.0670.559 ± 0.0450.615 ± 0.059
S2-PCConcerto0.800 ± 0.0460.508 ± 0.0360.442 ± 0.0240.441 ± 0.0260.615 ± 0.050
S3GradNorm-style0.900 ± 0.0310.828 ± 0.0130.791 ± 0.0220.868 ± 0.0150.846 ± 0.019
S3Spectral-style0.767 ± 0.0250.877 ± 0.0130.884 ± 0.0160.838 ± 0.0180.897 ± 0.022
S4SplatFormer0.847 ± 0.0520.830 ± 0.0290.884 ± 0.0260.682 ± 0.0600.815 ± 0.046
S4Gaussian-MAE0.873 ± 0.0490.905 ± 0.0240.953 ± 0.0280.894 ± 0.0480.800 ± 0.051
S4SceneSplat0.913 ± 0.0400.902 ± 0.0350.916 ± 0.0290.929 ± 0.0350.908 ± 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.

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

This paper introduces Poison-3DGS, a benchmark to detect poisoning attacks on 3D Gaussian Splatting (3DGS) with a stage-wise perspective. The authors decompose the 3DGS reconstruction pipeline in 4 stages. The five attack occurs at specific stage(s) of the pipeline and the author use "attack-agnostic" detectors at each stage. The benchmark has 100 clean base scenes gathered from public datasets and has 1253 poisoned variants (derived from 4 attack types). The paper contains an evaluation of the benchmark (Table 1) based on 2 metrics: AUROC and FPR@95. One interesting finding (in my opinion) is an attack injected at the first stage (e.g., Poison-Splat) is the most detectable in the last stage of the pipeline (fourth stage). Whereas, Stealth-Attack (also injected in the first stage) is highly detectable in the first stage. Overall, the paper provides a good overview of the field.

Critical strengths
  • I believe the proposal of a unified benchmark for characterizing poisoning detection across 3DGS pipeline stages has its value from an engineering perspective. However, it would have been more valuable to derive new attacks and defenses that are built based on Poison-3DGS. For instance, a defense that leverages each defense at a specific stage to detect an attack that remains undetected by baseline defenses at a stage level.
Critical weaknesses
  • It is unclear how the paper positions itself relative to existing work. The contribution feels relatively modest from a research perspective. The paper argues their approach provides a detection at a lifeycle/pipeline level instead of being limited to a single stage of thelifeycle/the pipeline (prior art). However, the paper does not provide a defense that requires a lifecycle/pipeline level approach. The current paper sticks to an aggregation of results (Table 1) derived from single stage detectors. Therefore, it is unclear to me what the actual benefit (from a research perspective) of the proposed approach is.

  • Table 1 lacks of statistical evidences such as confidence intervals. How many times was the experiment to get Table 1 numerical values run? The absence of statistical information for a benchmark paper is problematic. For instance, Table 1 reports FPR@95. However, section 5 focuses only about AUROC values. In the context of attack detection, it is important to clarify the trade-off between detection rate and false positive rate when it comes to evaluating a detection system. For instance, as a reader, I am curious to know what would be a satisfactory FPR@95 value for a real-world deployment scenario.

  • In Appendix B.2, authors claim that "We verify attack effectiveness". However, I do not see a table with a quantitative analysis of the attack success or the attack stealthiness per attack. The absence of quantitative results regarding the attack effectiveness makes it difficult to assess the validity of the attack.

Questions for authors
  • What is the computational to run the detection methods (e.g, inference time and memory usage)?
  • Can you provide the confidence interval for Table 1?
  • What is the novelty of the approach compared to prior work? (any new attacks or new defenses)
7 comments
AuthorsOpenReview ↗

Comment for Reviewer ojty (1/5)

(W1a) It is unclear how the paper positions itself relative to existing work. The contribution feels relatively modest from a research perspective.

(Q3) What is the novelty of the approach compared to prior work? (any new attacks or new defenses)

A (W1a, Q3): We thank the reviewer for the question.

Positioning and gap. Detection research for 3DGS poisoning remains limited and fragmented. A few existing detectors [1–2] were developed very recently. However, they are attack-specific, developed using known attack-specific characteristics [2] or trained with attack-specific samples [1].

Limitation of attack-specific detector: In a practical, open-world defense setting, one cannot assume a specific attack type because poisoning signals are diverse and new attacks will continue to emerge. Existing attack-specific detectors are either inapplicable to other attacks or perform very poorly when facing unseen poisoning. Our results show that attack-specific detectors perform very poorly on unseen poisoning: 0.463 - 0.519 AUROC (added as Appendix C.8, Table 14, marked in blue)

Furthermore, even among these attack-specific detectors, each study evaluates on different scene subsets and metrics, hindering direct comparison across studies.

The critical gap is that there is no unified benchmark for studying detection across diverse attacks and stages of the 3DGS lifecycle/pipeline. This leaves an important practical question unanswered: how does a defender assess a potentially compromised 3DGS scene in the open-world setting when the attack identity is not known in advance?

Contribution and novelty. As a benchmark submission, 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 lifecycle/pipeline-level detectability study: Poison-3DGS is structured based on the representations naturally exposed at four lifecycle/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 lifecycle/pipeline.

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 2/5)

AuthorsOpenReview ↗

Comment for Reviewer ojty (2/5)

(Continued from part 1/5)

(W1b) The paper argues their approach provides a detection at a lifeycle/pipeline level instead of being limited to a single stage of thelifeycle/the pipeline (prior art). However, the paper does not provide a defense that requires a lifecycle/pipeline level approach. The current paper sticks to an aggregation of results (Table 1) derived from single stage detectors. Therefore, it is unclear to me what the actual benefit (from a research perspective) of the proposed approach is.

A(W1b):

Our framework directly enables lifecycle/pipeline-level defenses that combine evidence across stages. We clarify that, as the first step, our previous manuscript has already explored a lifecycle-level detection method (Appendix C.3, Table 9). The method 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):

AttackBest single-stage member of panelAUROC ↑FPR@95 ↓Lifecycle ensemble AUROC ↑Lifecycle ensemble FPR@95 ↓Δ AUROC
StealthAttackSparseViT (S1)0.773 ± 0.0280.780 ± 0.0420.891 ± 0.0180.400 ± 0.024+0.118 ± 0.035
Poison-SplatSplatFormer (S4)0.707 ± 0.0330.740 ± 0.0160.665 ± 0.0320.840 ± 0.015−0.042 ± 0.024

The ensemble improves StealthAttack detection, where evidence is spread across stages, and lowers Poison-Splat detection, where evidence is concentrated at S4. This gives a concrete baseline for future stage-aware 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) Table 1 lacks of statistical evidences such as confidence intervals. How many times was the experiment to get Table 1 numerical values run? The absence of statistical information for a benchmark paper is problematic. For instance, Table 1 reports FPR@95. However, section 5 focuses only about AUROC values.

(Q2) Can you provide the confidence interval for Table 1?

A (W2 + Q2):

We thank the reviewer for the comment. Each value in Table 1 comes from a single evaluation on the fixed benchmark: all detectors use officially released pretrained weights and are applied at inference only, with no retraining or fine-tuning on our data, so there is no training randomness to average over. The remaining variability is across scenes. We therefore now report bootstrap standard errors for every AUROC and FPR@95 entry (added to Table 1, marked in blue). They are estimated from 2,000 resamples of the 100 base scenes drawn with replacement; each resampled scene keeps its clean score and all of its poisoned variants.

(a) AUROC ↑

StageDetectorStealthAttackPoison-SplatGaussTrap3D-GSWGuardSplatOverall
S1SoftPatch+0.837 ± 0.0220.459 ± 0.027N/AN/AN/A0.648 ± 0.017
S1SparseViT0.773 ± 0.0280.540 ± 0.029N/AN/AN/A0.657 ± 0.022
S1MVSS-Net0.707 ± 0.0290.318 ± 0.026N/AN/AN/A0.513 ± 0.019
S1IML-ViT0.701 ± 0.0300.427 ± 0.028N/AN/AN/A0.564 ± 0.021
S2MET3R0.744 ± 0.0250.521 ± 0.026N/AN/AN/A0.633 ± 0.020
S2VGGT0.635 ± 0.0230.464 ± 0.027N/AN/AN/A0.550 ± 0.018
S2MV-DUSt3R+0.486 ± 0.0120.526 ± 0.024N/AN/AN/A0.506 ± 0.013
S2NoPoSplat0.508 ± 0.0110.314 ± 0.025N/AN/AN/A0.411 ± 0.013
S2-PCSonata0.878 ± 0.023N/AN/AN/AN/A0.878 ± 0.023
S2-PCConcerto0.833 ± 0.026N/AN/AN/AN/A0.833 ± 0.026
S3GradNorm-style0.728 ± 0.0370.579 ± 0.025N/AN/AN/A0.654 ± 0.023
S3Spectral-style0.568 ± 0.0390.550 ± 0.027N/AN/AN/A0.559 ± 0.027
S4SplatFormer0.657 ± 0.0380.707 ± 0.0330.798 ± 0.0210.566 ± 0.0320.460 ± 0.0290.638 ± 0.023
S4Gaussian-MAE0.514 ± 0.0360.657 ± 0.0300.712 ± 0.0310.506 ± 0.0300.478 ± 0.0280.573 ± 0.024
S4SceneSplat0.582 ± 0.0330.520 ± 0.0260.489 ± 0.0290.497 ± 0.0300.483 ± 0.0290.514 ± 0.024

(continue in part 3/5)

AuthorsOpenReview ↗

Comment for Reviewer ojty (3/5)

(Continued from part 2/5)

(b) FPR@95 ↓

StageDetectorStealthAttackPoison-SplatGaussTrap3D-GSWGuardSplatOverall
S1SoftPatch+0.570 ± 0.0160.970 ± 0.006N/AN/AN/A0.770 ± 0.009
S1SparseViT0.780 ± 0.0420.900 ± 0.011N/AN/AN/A0.840 ± 0.022
S1MVSS-Net0.940 ± 0.0170.950 ± 0.007N/AN/AN/A0.945 ± 0.008
S1IML-ViT0.940 ± 0.0130.970 ± 0.018N/AN/AN/A0.955 ± 0.011
S2MET3R0.540 ± 0.0270.930 ± 0.006N/AN/AN/A0.735 ± 0.014
S2VGGT0.830 ± 0.0191.000 ± 0.000N/AN/AN/A0.915 ± 0.010
S2MV-DUSt3R+0.960 ± 0.0060.920 ± 0.032N/AN/AN/A0.940 ± 0.017
S2NoPoSplat0.950 ± 0.0220.980 ± 0.016N/AN/AN/A0.965 ± 0.012
S2-PCSonata0.550 ± 0.041N/AN/AN/AN/A0.550 ± 0.041
S2-PCConcerto0.530 ± 0.018N/AN/AN/AN/A0.530 ± 0.018
S3GradNorm-style0.800 ± 0.0250.880 ± 0.010N/AN/AN/A0.840 ± 0.014
S3Spectral-style0.820 ± 0.0130.930 ± 0.009N/AN/AN/A0.875 ± 0.008
S4SplatFormer0.910 ± 0.0360.740 ± 0.0160.730 ± 0.0190.860 ± 0.0200.960 ± 0.0070.840 ± 0.010
S4Gaussian-MAE0.940 ± 0.0170.780 ± 0.0150.790 ± 0.0140.990 ± 0.0130.990 ± 0.0050.898 ± 0.008
S4SceneSplat0.840 ± 0.0200.920 ± 0.0100.900 ± 0.0090.970 ± 0.0090.970 ± 0.0100.920 ± 0.006

The uncertainty estimates remain consistent with the broad trends underlying our findings, and our findings remain valid. In particular, Poison-Splat achieves higher AUROC at S4 with SplatFormer (0.707 ± 0.033) than with the best evaluated S1 detector (0.540 ± 0.029), together with lower FPR@95 (0.740 ± 0.016 versus 0.900 ± 0.011). This supports our main finding (Finding 2):

Ln 030: “the stage where an attack is introduced is not necessarily the stage where it is most detectable. Instead, downstream reconstruction transforms attack effects into stage-specific forensic signals whose detectability varies substantially across stages and attack types.”

The results also retain the variation across stages and detectors underlying Findings 1 and 3.

(continue in part 4/5)

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

This manuscript studies poisoning detection for 3D Gaussian Splatting, noting that prior work only evaluates whether attacks succeed and lacks a systematic defender-side characterization of at which stage poisoning can be detected. The authors propose the Poison-3DGS benchmark, comprising 100 indoor/outdoor-balanced clean scenes and 1,253 poisoned variants covering five attack families, and exposing the artifacts of four reconstruction stages: multi-view images, SfM geometry, training dynamics, and the final Gaussian model. The evaluation protocol is attack-agnostic: detectors are assigned by stage representation rather than attack type, all use public pretrained weights, see no poisoned samples and no attack identity, and output only scene-level anomaly scores measured by AUROC and FPR@95.

Critical strengths
  1. The problem formulation is novel. Existing work has mainly focused on whether an attack succeeds. This paper takes the defender’s perspective and systematically examines at which stage of the reconstruction pipeline poisoning can be detected.

  2. Decoupling the stage structure from the attack space is a clear and extensible design principle. Detectors are organized according to the representation at each stage rather than a specific attack type.

  3. The dataset is reasonably large and diverse for this area.

Critical weaknesses
  1. The detector suite is mismatched with poisoning signals, so the core findings may be an artifact of detector choice. All S1–S3 detectors were designed for 2D image manipulation localization or multi-view consistency; none of them was designed for poisoning. As a result, it is hard to tell whether the findings "no single stage dominates" and "poisoning is undetectable at S1" reflect properties of poisoning itself, or simply that the authors did not include detectors suited to the task.

  2. The evaluation protocol suffers from cross-scene confounding and lacks a paired evaluation. The question that really matters is whether a poisoned variant scores higher than its own clean base scene.

  3. The main table has no uncertainty estimates at all. There are only 100 negative scenes, and none of the AUROC values in Table 1 come with confidence intervals or bootstrap standard deviations. Many of the comparisons in Section 5.3 rest on gaps of around 0.05, and there is no way to tell whether these are significant.

  4. Strength of conclusions and nature of the contribution. The paper does not propose a new attack, a new detector, or a new mitigation; it is a combination of five existing attacks, eight existing datasets, and thirteen existing detectors, plus empirical analysis. For a benchmark paper this is acceptable in itself, but among the three findings, Finding 1 and Finding 3 are fairly generic, and only Finding 2 carries real information.

Questions for authors

Please refer to Critical Weaknesses. Q1, 2, 3 are more important.

6 comments
AuthorsOpenReview ↗

Comment for Reviewer hsjC (1/4)

(W1) The detector suite is mismatched with poisoning signals, so the core findings may be an artifact of detector choice. All S1–S3 detectors were designed for 2D image manipulation localization or multi-view consistency; none of them was designed for poisoning. As a result, it is hard to tell whether the findings "no single stage dominates" and "poisoning is undetectable at S1" reflect properties of poisoning itself, or simply that the authors did not include detectors suited to the task.

A (W1):

We thank the reviewer for the comment. We clarify that existing detectors [1, 2] designed for 3DGS poisoning are attack-specific. They do not meet the eligibility requirements of our common attack-agnostic evaluation protocol.

Attack-agnostic evaluation protocol. Using Poison-3DGS, we evaluate detectors across attacks without attack labels, poisoned-sample training, or attack-specific adaptation.

Reason for attack-agnostic protocol. This protocol models an open-world setting in which the defender does not know the attack identity in advance and must handle potentially unseen, diverse and evolving poisoning.

Existing detectors for 3DGS poisoning. Detection research for 3DGS poisoning remains limited and fragmented. A few existing detectors were developed very recently[1, 2]. However, they are attack specific: they are trained with poisoned samples [1], or developed using attack specific characteristics [2]. As shown in the experiment below, attack-specific detectors are either inapplicable to other attacks or perform very poorly when facing unseen poisoning.

Stage-artifact detectors. In an open-world defense setting, one cannot assume a specific attack type because poisoning signals are diverse and new attacks will continue to emerge. As the first step to understand open-world defense, we select state-of-the-art, stage-appropriate anomaly detectors according to the artifacts naturally exposed at each stage of the 3DGS pipeline.

Experiment. In the following experiment, we evaluate existing detectors designed for 3DGS poisoning [1,2]. Both were proposed very recently and are Poison-Splat specific, using Poison-Splat samples in training or Poison-Splat footprints. We compare them alongside our attack-agnostic S4 detector (SplatFormer) across all five attack families (reporting AUROC ↑, with FPR@95 ↓ in parentheses; bootstrap standard errors over 2,000 resamples of the base scenes):

DetectorTypeStealthAttackPoison-SplatGaussTrap3D-GSWGuardSplat
DefenseSplat [2]Poison-Splat-specific0.519 ± 0.012 (0.960 ± 0.022)0.813 ± 0.021 (0.530 ± 0.067)N/AN/AN/A
RemedyGS [1]supervised, Poison-Splat-specific0.463 ± 0.013 (0.920 ± 0.011)0.989 ± 0.004 (0.040 ± 0.023)N/AN/AN/A
SplatFormerattack-agnostic0.657 ± 0.038 (0.910 ± 0.036)0.707 ± 0.033 (0.740 ± 0.016)0.798 ± 0.021 (0.730 ± 0.019)0.566 ± 0.032 (0.860 ± 0.020)0.460 ± 0.029 (0.960 ± 0.007)

(Note: "N/A" indicates that these Poison-Splat specific detectors are inapplicable to model-level attacks )

Note that a concurrent attack-specific detector, GaussGuard [3], will be published on 1 December 2026; the code has not been released, and it is not included in this comparison.

As shown, existing Poison-Splat specific detectors cannot handle unseen StealthAttack (0.463 to 0.519 AUROC). Furthermore, they cannot be applied to model-level attacks. Thus, attack-specific detectors are not suitable for open-world defense.

Finally, as a benchmark submission, Poison-3DGS brings five attack families, 100 base scenes, and 1,253 poisoned variants into a common benchmark. Its stage-wise, representation-centric design organizes the artifacts available at four pipeline stages independently of attack identity, enabling a fixed, attack-agnostic evaluation protocol that can accommodate additional and future attacks. The detector suite utilizes the proposed benchmark, and supports our empirical findings.

Revision: we added this comparison as Appendix C.8 (Table 14), marked in blue.

[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).

[3] Kwon, H., & Baek, J.-W. (2026). GaussGuard: Leave-One-Out Detection and Mitigation of View-Conditional Poisoning Attacks on 3D Gaussian Splatting. Neurocomputing, 704, 134830.

(continue in part 2/4)

AuthorsOpenReview ↗

Comment for Reviewer hsjC (2/4)

(Continued from part 1/4)

(W2) The evaluation protocol suffers from cross-scene confounding and lacks a paired evaluation. The question that really matters is whether a poisoned variant scores higher than its own clean base scene.

A (W2):

We thank the reviewer for the suggestion. Our evaluation protocol follows detectability study in other problems [4, 5]. Nevertheless we agree that the reviewer's suggestion on paired evaluation is a good addition.

We now compare each poisoned variant with its own clean base scene (added as Appendix C.6, Table 12, marked in blue). A variant counts as a win if it scores higher than its clean base, and a tie counts as half a win. We average within each base scene, then across scenes. Entries are paired ranking accuracy ± bootstrap standard error (2,000 resamples of base scenes with replacement; each scene keeps its clean reference and all its variants).

StageDetectorStealthAttackPoison-SplatGaussTrap3D-GSWGuardSplat
S1SoftPatch+0.942 ± 0.0250.445 ± 0.047N/AN/AN/A
S1SparseViT0.849 ± 0.0240.566 ± 0.047N/AN/AN/A
S1MVSS-Net0.773 ± 0.0280.106 ± 0.022N/AN/AN/A
S1IML-ViT0.826 ± 0.0260.413 ± 0.052N/AN/AN/A
S2MET3R0.860 ± 0.0250.558 ± 0.049N/AN/AN/A
S2VGGT0.657 ± 0.0490.547 ± 0.047N/AN/AN/A
S2MV-DUSt3R+0.523 ± 0.0270.511 ± 0.052N/AN/AN/A
S2NoPoSplat0.558 ± 0.0190.087 ± 0.025N/AN/AN/A
S2-PCSonata0.919 ± 0.029N/AN/AN/AN/A
S2-PCConcerto0.849 ± 0.038N/AN/AN/AN/A
S3GradNorm-style0.659 ± 0.0310.734 ± 0.040N/AN/AN/A
S3Spectral-style0.727 ± 0.0440.749 ± 0.034N/AN/AN/A
S4SplatFormer0.755 ± 0.0440.735 ± 0.0450.837 ± 0.0310.709 ± 0.0590.483 ± 0.059
S4Gaussian-MAE0.623 ± 0.0540.697 ± 0.0470.762 ± 0.0440.543 ± 0.0560.471 ± 0.045
S4SceneSplat0.714 ± 0.0520.527 ± 0.0500.528 ± 0.0530.543 ± 0.0370.475 ± 0.040

Importantly, all our findings remain valid under the paired evaluation:

  • Finding 1: No single stage consistently dominates. StealthAttack achieves its highest paired ranking accuracy at S1 with SoftPatch+ (0.942 ± 0.025), while Poison-Splat achieves its highest value at S3 with Spectral-style (0.749 ± 0.034).

  • Finding 2: Attack injection and strongest detectability are not always aligned. Although Poison-Splat is injected at S1, its best paired ranking accuracy is 0.566 ± 0.047 at S1, compared with 0.749 ± 0.034 at S3 and 0.735 ± 0.045 at S4. The downstream advantage therefore persists.

  • Finding 3: Detector performance remains uneven across attacks and representations. SoftPatch+ achieves 0.942 ± 0.025 on StealthAttack but 0.445 ± 0.047 on Poison-Splat. At S4, SplatFormer achieves 0.837 ± 0.031 on GaussTrap but 0.483 ± 0.059 on GuardSplat.

[4] Liu, W., Wang, X., Owens, J., & Li, Y. (2020). Energy-based Out-of-distribution Detection. NeurIPS.

[5] Liang, S., Li, Y., & Srikant, R. (2018). Enhancing the Reliability of Out-of-Distribution Image Detection in Neural Networks. ICLR.

(continue in part 3/4)

AuthorsOpenReview ↗

Comment for Reviewer hsjC (3/4)

(Continued from part 2/4)

(W3) The main table has no uncertainty estimates at all. There are only 100 negative scenes, and none of the AUROC values in Table 1 come with confidence intervals or bootstrap standard deviations. Many of the comparisons in Section 5.3 rest on gaps of around 0.05, and there is no way to tell whether these are significant.

A (W3)

We thank the reviewer for the comment. We now report bootstrap standard errors for every AUROC and FPR@95 entry (added to Table 1, marked in blue). They are estimated from 2,000 resamples of the 100 base scenes drawn with replacement; each resampled scene keeps its clean score and all of its poisoned variants.

(a) AUROC ↑

StageDetectorStealthAttackPoison-SplatGaussTrap3D-GSWGuardSplatOverall
S1SoftPatch+0.837 ± 0.0220.459 ± 0.027N/AN/AN/A0.648 ± 0.017
S1SparseViT0.773 ± 0.0280.540 ± 0.029N/AN/AN/A0.657 ± 0.022
S1MVSS-Net0.707 ± 0.0290.318 ± 0.026N/AN/AN/A0.513 ± 0.019
S1IML-ViT0.701 ± 0.0300.427 ± 0.028N/AN/AN/A0.564 ± 0.021
S2MET3R0.744 ± 0.0250.521 ± 0.026N/AN/AN/A0.633 ± 0.020
S2VGGT0.635 ± 0.0230.464 ± 0.027N/AN/AN/A0.550 ± 0.018
S2MV-DUSt3R+0.486 ± 0.0120.526 ± 0.024N/AN/AN/A0.506 ± 0.013
S2NoPoSplat0.508 ± 0.0110.314 ± 0.025N/AN/AN/A0.411 ± 0.013
S2-PCSonata0.878 ± 0.023N/AN/AN/AN/A0.878 ± 0.023
S2-PCConcerto0.833 ± 0.026N/AN/AN/AN/A0.833 ± 0.026
S3GradNorm-style0.728 ± 0.0370.579 ± 0.025N/AN/AN/A0.654 ± 0.023
S3Spectral-style0.568 ± 0.0390.550 ± 0.027N/AN/AN/A0.559 ± 0.027
S4SplatFormer0.657 ± 0.0380.707 ± 0.0330.798 ± 0.0210.566 ± 0.0320.460 ± 0.0290.638 ± 0.023
S4Gaussian-MAE0.514 ± 0.0360.657 ± 0.0300.712 ± 0.0310.506 ± 0.0300.478 ± 0.0280.573 ± 0.024
S4SceneSplat0.582 ± 0.0330.520 ± 0.0260.489 ± 0.0290.497 ± 0.0300.483 ± 0.0290.514 ± 0.024

(continue in part 4/4)

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thank you Area Chair and Reviewers for the constructive feedback and discussion.

Dear Area Chair and Reviewers,

We sincerely thank the AC and reviewers for their thoughtful and constructive feedback. We are encouraged that reviewers recognized the novelty of our problem formulation (hsjC), the clear and extensible stage-wise design principle (hsjC), the value of a unified benchmark and overview of the field (ojty), and the interest of multi-stage poisoning detection for 3DGS (744E). Existing evaluations of 3DGS poisoning detection are fragmented, single-stage, and attack-specific. To our knowledge, ours is the first systematic study of poisoning detectability across attack families and stages of the 3DGS pipeline, showing for the first time that an attack’s injection stage need not be the stage where it is most detectable.

The discussion period has substantially strengthened the paper. During the rebuttal phase, two of the three reviewers responded positively and upgraded their ratings to Weak acceptance:

  • Reviewer hsjC (Weak acceptance): Concerns regarding detector selection, cross-scene confounding, missing uncertainty estimates, and the strength of our conclusions were addressed through an attack-specific detector comparison, a matched paired evaluation, and bootstrap standard errors. Reviewer hsjC subsequently stated, “Most of my concerns have been addressed, and I upgrade my recommendation to weak accept.”

  • Reviewer ojty (Weak acceptance): Questions on positioning, lifecycle-level defense, statistical uncertainty, detection-rate trade-offs, attack verification, and computational cost were resolved through clarifications, our existing cross-stage ensemble baseline, and new experimental tables. Reviewer ojty subsequently replied, “Thank you for your explanation - I have raised my rating.”

  • Reviewer 744E (Weak rejection): Concerns regarding novelty, benchmark scale and scene diversity, Stage-2 applicability for post-training attacks, and minor formatting/scene-count clarity were addressed through detailed clarifications and an added five-category scene breakdown. Reviewer 744E maintained the original rating, and no further questions were raised after we invited clarification on any remaining concerns.

We have revised the manuscript according to the reviewers' suggestions (changes highlighted in blue). Updates include: bootstrap standard errors for all main results in Table 1 (per hsjC, ojty); a five-category scene-content breakdown in Appendix C.5, Table 11 (per 744E); a matched paired evaluation in Appendix C.6, Table 12 (per hsjC); an expanded TPR/FPR operating-point analysis in Appendix C.7, Table 13 (per ojty); a comparison with attack-specific detectors in Appendix C.8, Table 14 (per hsjC); quantitative attack effectiveness and stealthiness verification in Appendix C.9, Table 15 (per ojty); per-scene runtime and GPU memory profiles in Appendix C.10, Table 16 (per ojty); and fixed parenthetical citation formatting throughout (per 744E).

We hope the positive discussions, score progressions, and manuscript improvements can be taken into account in the final decision. Thank you very much for your time and consideration.

Sincerely,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.