acceptodds
Under review as a conference paper at ICLR 2027

PhantomSplat: Stealthy Post-Reconstruction Attacks on 3D Gaussian Splatting

Abstract

3D Gaussian Splatting (3DGS) has become a widely adopted representation for 3D scene reconstruction, real-time rendering, and downstream perception. As optimized 3DGS representations are increasingly distributed and reused as scene assets, their integrity after reconstruction becomes an important security concern: a compromised asset may be directly consumed by downstream systems even when the original data and reconstruction process are trusted. Existing 3DGS attacks primarily target the construction stage through poisoned training views or perturbed inputs, leaving this post-reconstruction failure mode largely unexplored. We study whether an already optimized 3DGS asset can be transformed into a self-contained, view-dependent malicious representation while remaining visually plausible under benign inspection. We propose PhantomSplat, which injects a small set of low-opacity adversarial Gaussians while keeping the original representation fixed. The design targets three requirements for asset-level security evaluation: localized adversarial behavior, benign-view consistency, and structural inconspicuousness. Surface-constrained placement reduces detectable geometric artifacts, while bounded opacity optimization limits degradation under normal views. A first-order rendering surrogate further makes this constrained optimization practical without repeated full-scene rendering. Across diverse scenes and downstream perception models, PhantomSplat induces strong target-view effects while largely preserving the original asset under benign views. On Mip-NeRF360, PhantomSplat achieves 33.96 dB attack-view PSNR while retaining 29.62 dB benign-view PSNR, compared with 30.36 dB for the clean asset, with an average optimization time of 4.25 minutes. These findings reveal a gap in current 3DGS security assumptions and motivate post-reconstruction integrity verification before scene assets are deployed in downstream systems.

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

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_8XnCOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a new 3DGS poisoning attack which targets post-reconstruction tampering. The attacker leaves the original benign Gaussian scene unchanged, and injects a small set of new Gaussians. The goal is to render a target pattern at a chosen victim viewpoint, while keeping the benign viewpoints stay close to the original clean appearance, without introducing geometric artifacts. Experiments cover 3 datasets, and the attack exhibit a strong attack-view PSNR and faster optimization time compared with existing reconstruction phase attacks.

Critical strengths
  1. The paper identifies a research gap of post-reconstruction tampering. This new setting introduces less optimization time compared with previous reconstruction stage attacks.

  2. The proposed method is elegant and mathematically sound.

  3. The evaluation is broad on three datasets and 24 scenes. The authors introduces semantic metrics (CLIP and YOLO) which go beyond PSNR.

Critical weaknesses
  1. The first-order approximation is not well-motivated: The paper's narrative is that low-opacity Gaussians make the problem "first-order". This is supposed to enable decoupled and cached rendering. However, caching does not require the approximation.
Questions for authors
  1. Can you compare exact cached compositing (prefix product) against the linearized surrogate, with both evaluated by the true renderer? Does the linearization improve anything?

  2. How robust is the attack to opacity-threshold pruning (e.g., removing o<0.05o<0.05) and to detectors based on scale/opacity statistics?

5 comments

We thank the reviewer for the constructive comments and positive feedback on our paper. Regarding the concerns of the reviewer 8XnC, we provide the following responses.

Response to W1. We thank the reviewer for this observation. We would like to clarify the role of first-order approximation in our decoupled rendering formulation. Our caching mechanism does not involve any approximation. Specifically, we precompute and cache the exact contributions of the frozen original Gaussians, including the original-scene color CbaseC_{\mathrm{base}}, transmittance TibaseT_i^{\mathrm{base}}, and suffix color contribution Γi\Gamma_i.

The first-order approximation is introduced during subsequent adversarial compositing, where we approximate the transmittance induced by the injected low-opacity Gaussians:

\approx 1-\sum_{\substack{j<i\\\j\in\mathcal I_{\mathrm{inj}}}}\alpha_j =1-S_i.$$ This leads to our simplified decoupled rendering formulation: $$\widetilde C_p=C_{\mathrm{base}}+ \sum_{i\in\mathcal I_{\mathrm{inj}}} \left[ T_i^{\mathrm{base}}(1-S_i)\alpha_i c_i -\alpha_i\Gamma_i \right].$$ Therefore, **first-order approximation is part of our decoupled rendering formulation, but is not required for caching itself**. Its purpose is to simplify adversarial transmittance interactions and gradient computation during optimization. We also acknowledge that the phrase *"cached first-order approximation"* in our ablation description may have conflated these two operations. We will revise this wording to explicitly distinguish exact caching of the original scene from first-order approximation of adversarial compositing. **Response to Q1.** As clarified in our response to W1, **we do not use first-order approximation during caching**. The contributions of the frozen original Gaussians are computed exactly. The first-order approximation is introduced only when computing the transmittance effects of the injected adversarial Gaussians during optimization. We believe that this approximation can facilitate optimization by reducing nonlinear interactions among injected Gaussians. Specifically, exact transmittance involves multiplicative dependencies, $$U_i=\prod_{j<i}(1-\alpha_j),$$ whereas our first-order approximation simplifies them into an additive form, $$\widetilde U_i=1-\sum_{j<i}\alpha_j.$$ This simplification reduces higher-order interactions and may alleviate gradient attenuation, potentially improving optimization stability and convergence. To further investigate this effect, we extend our existing ablation by introducing Decoupled Opt (Exact), which uses exact adversarial compositing with prefix products, alongside the original Decoupled Opt, which uses our first-order approximation. Both settings are identical except for the compositing strategy, and all results are evaluated using the full 3DGS renderer. This comparison isolates the contribution of first-order approximation to optimization efficiency and attack performance. | Method | Attack PSNR ↑ | Benign PSNR ↑ | Time (min) ↓ | |:---|---:|---:|---:| | Decoupled Opt (Exact) | 32.12 | 28.78 | 34.89 | | Decoupled Opt (Ours) | 34.42 | 28.47 | 24.28 | With exact compositing, optimization becomes more challenging due to the multiplicative transmittance interactions among injected Gaussians. As a result, optimization may fail for some viewpoints selected by the outer loop, leading to additional viewpoint searches and longer overall optimization time. The increased optimization difficulty can still reduce attack-view quality or require more viewpoint selection attempts to achieve a successful attack. **Response to Q2.** We additionally evaluate opacity-based pruning ($\alpha < 0.05$) on the original Bicycle scene, with visual comparisons shown in **Appendix Fig. 12.** At this threshold, 27.69% (1,333,851/4,816,275) of the original Gaussians are removed. This introduces visible rendering differences, with SSIM dropping below 0.95, PSNR dropping below 30 dB. We also test a lower pruning threshold of $\alpha < 0.01$ and find that the attack remains largely unaffected, with attack-view PSNR still above 30 dB. These results highlight the trade-off between attack removal and scene fidelity. These results show that aggressive pruning can remove low-opacity adversarial Gaussians, but also affects legitimate scene content. Moreover, our injected Gaussians account for less than 1% of all Gaussians below this opacity threshold, making them difficult to distinguish from legitimate Gaussians using opacity statistics alone. Standard 3DGS implementations also use much lower pruning thresholds (e.g., 0.005), and downstream users do not know the attacker's opacity bound in advance.
Reviewer_8XnCOpenReview ↗

Following up questions

Thank you for your response. Regarding the response, I have two follow-up questions:

  1. Can you show the value distribution of exact UiU_i and the approximated U~i\tilde{U}_i ? For example, what is the mean ±\pm standard deviation of these two values, across both attacked views and benign views? What would the distribution of their difference U−U~U - \tilde{U} be? How would that change according to the number of injected adversarial Gaussians?

  2. I can understand that such first-order approximation can make optimization faster and thus reduce attack time. But why approximation will make the optimization results even better than the exact computation on the attacked views? Why is it not the case for benign views? Could you explain the details?

Response to Follow-up Q1. We appreciate your follow-up and the opportunity to address the remaining concerns. Below, we report the mean and standard deviation of UiU_i and its first-order approximation U~i\widetilde{U}_i, evaluated on our 500×500500 \times 500 Earth image attack presented in Appendix D.10, which includes one attack view and 145 benign views.

ViewExact UiU_iApprox. U~i\widetilde{U}_iUi−U~iU_i-\widetilde{U}_i
Attack0.861 ±\pm 0.1280.840 ±\pm 0.1700.021 ±\pm 0.047
Benign0.774 ±\pm 0.2100.694 ±\pm 0.3680.079 ±\pm 0.174

Overall, the first-order approximation underestimates the exact transmittance, with a larger discrepancy at benign viewpoints than at the attack viewpoint. To see how Ui−U~iU_i-\widetilde{U}_i changes, we average it across rays grouped by the number of injected Gaussians along each ray.

# Injected Gaussians NumberAttack View AvgBenign Views Avg
1–200
50.00190.0018
100.0110.011
200.0700.057
300.1470.111

As more Gaussians overlap along a ray, the omitted higher-order opacity interactions accumulate, increasing the approximation error.

Response to Follow-up Q2. The adversarial optimization, just as 3DGS, is nonlinear and nonconvex, so the first-order approximation may outperform exact computation. Moreover, our objective maximizes attack-view fidelity while keeping benign-view degradation within a small bound. Thus, greater benign-view degradation may coexist with improved attack-view fidelity. To further investigate how approximation errors contribute to these performance differences, we compare the gradients obtained using the first-order approximation and exact computation on the 500×500500 \times 500 Earth image attack:

ViewGradient Cosine Similarity ↑Sign Disagreement ↓
Attack view0.9940.38%
Benign views (per-view)0.951 ±\pm 0.0743.20%
Benign views (aggregated)0.9930.77%

We observe that gradients obtained using the first-order approximation align more closely with exact gradients at the attack view. Despite strong alignment when benign-view gradient contributions are aggregated (0.993 cosine similarity and 0.77% sign disagreement), individual benign-view gradients show weaker alignment (0.951 cosine similarity and 3.2% sign disagreement). This suggests that the approximation is more likely to misrepresent gradient directions for individual benign views, potentially leading to greater fidelity degradation.

Reviewer_Be79OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PhantomSplat, an efficient and stealthy adversarial attack method targeting post-reconstruction 3D Gaussian Splatting (3DGS) assets. Unlike traditional training-stage data poisoning attacks, PhantomSplat does not require retraining the original 3DGS model. Instead, it injects a small number of low-opacity adversarial Gaussians near existing object surfaces, causing the scene to display attacker-specified malicious content from selected viewpoints while preserving visual and structural inconspicuousness from benign viewpoints. To improve efficiency, the authors introduce Decoupled Rendering and first-order Taylor approximation, which cache the original scene's rendering information and benign-view loss gradients, avoiding repeated full-scene rendering and backpropagation. Experiments on three datasets—Mip-NeRF360, Tanks & Temples, and Free—demonstrate that PhantomSplat can effectively influence downstream perception models such as CLIP and YOLO while maintaining high visual fidelity under benign views. On Mip-NeRF360, the attack requires an average optimization time of only 4.25 minutes. The main contribution of this work is to reveal the security risks associated with post-reconstruction tampering of 3DGS assets and introduce an efficient attack framework that balances adversarial effectiveness, stealthiness, and computational efficiency.

Critical strengths
  1. The paper introduces a novel post-reconstruction attack setting for 3D Gaussian Splatting, addressing a security vulnerability that has received limited attention in prior work. Unlike existing reconstruction-stage attacks, PhantomSplat directly manipulates optimized 3DGS assets without requiring access to the original training data or retraining the entire scene. This highlights an important security risk in the deployment and distribution of 3DGS assets.
  2. The proposed method combines low-opacity Gaussian injection, decoupled rendering, and Taylor-based stealth constraints to significantly reduce computational overhead while maintaining attack effectiveness and benign-view fidelity. Experimental results demonstrate substantial improvements in optimization efficiency, reducing the average attack time to only 4.25 minutes, while achieving higher attack-view PSNR and better benign-view preservation than the compared baselines.
Critical weaknesses
  1. Insufficient Evaluation and Limited Baseline Comparisons. The experimental evaluation is not sufficiently comprehensive. The paper primarily reports PSNR on training views rather than held-out test views, making it difficult to assess the attack's stealthiness and generalization to unseen viewpoints. Additionally, the baseline comparisons are limited, with only one existing attack method (StealthAttack) included, while other relevant methods, such as GaussTrap and ComplicitSplat, are not evaluated. These limitations weaken the experimental evidence supporting the effectiveness and generalizability of the proposed method.

  2. Limited Practical Validation of Attack Effectiveness and Robustness. The experiments primarily focus on a single adversarial target (a STOP sign) with a fixed size of 125 × 125 pixels, lacking systematic evaluation on larger and more complex adversarial objects. Furthermore, the evaluation mainly relies on image similarity and perception metrics, without demonstrating whether the attack can lead to meaningful errors in realistic downstream applications. The robustness evaluation is also limited, as the paper does not sufficiently investigate the impact of common post-processing operations, such as Gaussian pruning and incremental fine-tuning, on attack effectiveness.

Questions for authors

1. Evaluation on Unseen Views and Additional Baselines

Could the authors provide PSNR, SSIM, and LPIPS results on held-out test views to further validate the attack's generalization and stealthiness? Additionally, could the authors include comparisons with relevant 3DGS adversarial attack methods, such as GaussTrap and ComplicitSplat?

2. Diversity and Scale of Adversarial Targets

Could the authors evaluate the method on more diverse and complex adversarial targets, particularly larger attack regions, to demonstrate its generalizability and scalability?

3. Practical Impact on Downstream Applications

Beyond CLIP similarity and YOLO detection confidence, could the authors provide stronger evidence that the proposed attack can cause meaningful misperception or incorrect decisions in realistic downstream applications?

4. Robustness Against Post-processing Operations

Could the authors evaluate the robustness of PhantomSplat against common post-processing operations, such as opacity-based Gaussian pruning and incremental fine-tuning with clean observations? How do these operations affect attack effectiveness and stealthiness?

9 comments

We thank the reviewer for the constructive comments. Regarding the concerns of the reviewer Be79, we provide the following responses.

W1 & Q1 Held-out viewpoint evaluation and Baseline comparisons. We agree that additional quantitative evidence is helpful. Fig. 10 already provides visualizations from viewpoints excluded from attack optimization. We further evaluate six held-out viewpoints sampled around the target object, obtaining clean-versus-attacked rendering PSNR values of 33.04–36.23 dB(> 30dB[]) using the full 3DGS renderer. These results provide additional evidence that our attack preserves visual fidelity beyond the optimization viewpoints.

Angle4590135180225270
PSNR (dB)33.0436.2333.2434.1234.8935.32
SSIM0.9320.9180.9830.9580.9440.969
LPIPS0.0890.0180.0430.0490.0770.061

The results show consistently high PSNR values across different viewing angles, indicating limited visual deviations from the original clean rendering outside the attack region.

For more baseline, We would like to clarify that GaussTrap and most of the other attack focus on reconstruction-stage attacks, which fall outside our post-reconstruction threat model. In particular, they rely on manipulating training observations during reconstruction, whereas our method directly modifies a reconstructed 3DGS asset without accessing the original training data or reconstruction process. Nevertheless, we agree that expanding the baseline comparisons would provide a more comprehensive evaluation, we have additionally included Gausstrap as a baseline in our evaluation.

MethodMetricTime (min) ↓BIFLGASTTRROCOKIBOAvg.
GaussTrapAttack view PSNR ↑93.2433.0135.2336.5437.9832.8331.3234.5931.3836.2834.35
GaussTrapTraining view PSNR ↑--24.3223.5028.2926.5821.5428.9825.3827.4932.1226.46
PhantomSplatAttack view PSNR ↑4.2534.0938.6237.0636.8533.6531.6931.4031.6230.7033.96
PhantomSplatTraining view PSNR ↑--26.5324.1230.1730.7924.6532.8430.9032.6033.9729.62

Under the same attack setting, PhantomSplat achieves similar attack-view PSNR to GaussTrap (33.96 vs. 34.35 dB), but is 21.9× faster (4.25 vs. 93.24 min) without retraining the original 3DGS. Our training-view PSNR is also 3.16 dB higher (29.62 vs. 26.46 dB). Both GaussTrap and StealthAttack show more degradation in non-attack views.

[1] Tan et al., "CodedStereo: Learned Phase Masks for Large Depth-of-Field Stereo," CVPR, 2021.

Response to W2 & Q2. We thank the reviewer for this suggestion. To further demonstrate the generalizability and scalability of our method, we evaluate a larger and more complex Earth image, similar to the adversarial target used in GaussTrap. Specifically, we increase the target resolution from 125×125125 \times 125 to 500×500500 \times 500 pixels, representing a 16-fold increase in the number of pixels. The corresponding attack and stealthiness results are presented in Appendix Figs. 13 and 14, demonstrating that PhantomSplat can accommodate substantially larger and more complex adversarial targets while preserving viewpoint-dependent stealthiness.

MetricAttack viewBenign views
PSNR (dB) ↑29.4031.32
SSIM ↑0.9430.942
LPIPS ↓0.0250.076

Response to W2 & Q3. Downstream impact. Motivated by the downstream object removal experiment in GaussTrap, we extend our evaluation to investigate the impact of PhantomSplat on 3D scene editing. Specifically, we apply single-view, mask-guided Gaussian segmentation to remove the adversarial stop sign. As shown in Appendix Fig. 11, the removal process also mistakenly removes parts of the original object, causing unintended damage to the scene. This occurs because the injected Gaussians are spatially intertwined with the original Gaussians, making them difficult to distinguish during segmentation. These results demonstrate that our attack can disrupt downstream 3D editing tasks beyond image-level perception.

Response to W2 & Q4. Robustness Against Post-processing Operations We further evaluate opacity-based Gaussian pruning (α<0.05\alpha < 0.05) on Bicycle. Notably, 27.69% (1,333,851/4,816,275) of the original Gaussians fall below this threshold and are removed.

Setting# GaussiansClean View PSNR (dB)SSIM
Original4,816,275——
Pruned3,482,424< 30<0.95

As shown in Appendix Fig. 12, we evaluate pruning on 10 sampled views randomly and observe visible color differences, with SSIM dropping below 0.95 and PSNR below 30 dB, indicating a trade-off between aggressive pruning and scene fidelity. We further test a lower pruning threshold of α<0.01\alpha < 0.01 and find that the attack remains largely unaffected, with attack-view PSNR decreasing only slightly from 34.09 to 33.17 dB. These results highlight the trade-off between attack removal and scene fidelity.

Regarding fine-tuning with clean observations, this falls outside our post-reconstruction threat model, as downstream recipients are assumed to have access to the distributed 3DGS asset but not the original training data.

Thank you to the authors for the detailed response. I still have one concern regarding scalability. The reported attack-view PSNR of 29.40 dB is noticeably lower than the previously reported results. Does this suggest that increasing the size and complexity of the adversarial target may substantially affect the performance of the proposed method?

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

This paper proposes PhantomSplat, a method for conducting stealthy post-reconstruction attacks on 3D Gaussian Splatting (3DGS) scenes. Instead of manipulating the reconstruction process, it inserts optimized adversarial Gaussians into an existing scene to produce a target appearance from selected viewpoints while preserving the original appearance from benign views. The method combines adversarial Gaussian initialization, a cached first-order rendering approximation to improve optimization efficiency, and a benign-view constraint to maintain visual consistency.

Experiments evaluate attack quality, semantic alignment, optimization efficiency, and stealthiness. The post-reconstruction threat model is relevant, and the proposed approach offers a promising direction for securing 3D scene representations. However, the evidence for consistent semantic attack success, stealthiness from unseen viewpoints, and robustness against potential defenses remains limited. Further evaluation across scenes and viewpoints is needed to establish the method's effectiveness in realistic settings.

Critical strengths

1.This paper investigates post-reconstruction attacks for 3DGS, a security threat separate from training-stage attacks. It reveals vulnerabilities of pre-built 3D assets and expands the scope of 3DGS security research.

2.The method leverages benign-view constraints and cached first-order rendering approximation to balance target visual manipulation, benign-view fidelity and optimization efficiency.

Critical weaknesses
  1. The evaluation does not fully establish the consistency of semantic attack success across scenes.

Although the paper reports strong image-space attack quality (33.96 dB average attack-view PSNR on Mip-NeRF360, Table 2), pixel-level similarity to the target image does not necessarily guarantee reliable semantic effects on downstream perception models. Table 3 reports an average normalized CLIP ratio of 0.982 and an average YOLO confidence of 0.925, compared with target-image reference values of 1.0 and 0.946, respectively. Some scene-level differences are more pronounced: Garden achieves a YOLO confidence of 0.823 (reference 0.936), while Bonsai achieves 0.892 (reference 0.961). These differences do not establish attack failure, but they motivate a more systematic assessment of semantic attack reliability across scenes. More importantly, the paper does not clearly define an explicit semantic attack success criterion or report scene-wise success rates under a predefined criterion. Consequently, the aggregate metrics do not reveal how consistently the method satisfies a specified success threshold across different scenes. Without such an analysis, it remains difficult to determine whether the reported results substantiate the claimed reliability of the semantic manipulation beyond favorable aggregate performance.

  1. The evaluation provides limited evidence for stealthiness on held-out viewpoints and sustained attack effectiveness after defensive filtering.

(a) Benign-view fidelity is evaluated primarily on training camera poses, with an average training-view PSNR of 29.62 dB on Mip-NeRF360. Although the paper defines preserved fidelity as a PSNR above 25 dB with a decrease of no more than 3 dB relative to the clean asset, satisfying these criteria on the evaluated training views does not establish that the modified asset remains inconspicuous at camera poses excluded from optimization. Appendix D.7 provides evidence of viewpoint-localized attack behavior by examining camera offsets around the optimized attack viewpoint: the reported ground-truth PSNR increases from 13.28 dB at the attack viewpoint to approximately 23.6–23.8 dB at ±10° and 24.7–26.4 dB at ±20°. However, this controlled locality analysis does not fully characterize visual fidelity and semantic behavior across broader, held-out camera trajectories.

(b) The DBSCAN defense experiment reports that 96.6% of the adversarial Gaussians survive filtering (3,989 to 3,853). While this demonstrates that most injected primitives survive the tested filter, primitive survival is only an indirect measure of attack robustness. It does not establish whether the filtered scene retains the intended target-view appearance or preserves its downstream semantic effects, as the paper does not report corresponding post-filtering target-view rendering quality or CLIP/YOLO results. Since stealthiness under benign inspection and resilience to defensive filtering are central to the proposed threat model, quantitative evaluation on held-out viewpoints and direct measurements of attack effectiveness before and after filtering are needed to substantiate the method's practical robustness.

Questions for authors
  1. Could you explicitly define the semantic attack success criterion used in the evaluation and report scene-wise success rates under this criterion? In particular, how should the relatively low YOLO confidence values for Garden (0.823) and Bonsai (0.892) be interpreted relative to their corresponding target-image reference values?

  2. Figure 5 reports that 96.6% of the adversarial Gaussians survive DBSCAN filtering, but this statistic alone does not establish that the semantic attack remains effective after filtering. Could you provide quantitative post-defense results, such as target-view rendering quality and CLIP/YOLO scores, to assess how much attack effectiveness is retained? These results would help determine whether the reported evidence supports the claimed robustness against structural filtering.

  3. The paper evaluates benign-view fidelity mainly using training viewpoints and also discusses the localization of the attack around the optimized victim viewpoint. How well does the method preserve visual fidelity and maintain the intended attack behavior on held-out viewpoints excluded from benign-view optimization?

  4. The paper includes an approximation-error analysis in Appendix D.6. Could you clarify whether this analysis verifies that the cached first-order approximation and Taylor-based constraint preserve fidelity under the full renderer on held-out benign views? Please provide additional quantitative experimental evidence to determine whether the efficiency gains come at the cost of meaningful degradation in actual rendering quality.

6 comments

We thank the reviewer for the constructive comments. Regarding the concerns of the reviewer quzu, we provide the following responses.

W1 & Q1. Semantic attack success and cross-scene consistency. Continuous CLIP and YOLO scores alone do not explicitly quantify semantic attack success. Following prior evaluations of targeted object-detection attacks [1, 2], we introduce a binary semantic ASR with an additional fidelity requirement criterion based on the target-class detection confidence defined in Eq below.

\frac{1}{N}\sum_{i=1}^N \mathbb{1}\left[ C_{\mathrm{YOLO}}^{(i)} \geq \tau \land \Delta \mathrm{PSNR}_{\mathrm{benign}}^{(i)} < \delta \right]$$ where $N$ denotes the number of evaluated attack instances, $C_{\mathrm{YOLO}}^{(i)}$ is the YOLO detection confidence for the target class, $\Delta\mathrm{PSNR}_{\mathrm{benign}}^{(i)}$ represents the training-view PSNR degradation, $\tau$ is the fixed confidence threshold, and $\delta$ (in dB) is the maximum allowable PSNR degradation. We adopt a confidence threshold of $\tau=0.5$ following [1]. Since there is no established criterion for how to decide $\delta$, we pick a threshold\$\delta =3.0\$ , and report the raw values for the YOLO confidence and PSNR degradation. All 24 evaluated scenes across Mip-NeRF360, Tanks & Temples, and Free satisfy both criteria, achieving an ASR of 100% (24/24). We interpret the relatively low YOLO confidence scores for Garden (0.823) and Bonsai (0.892) as reflecting the optimization limits of our proposed method, which also provides a meaningful basis for comparing different methods. In the revised manuscript, we will present the results under the new evaluation criteria and report YOLO confidence scores and PSNR degradation in the appendix. [1] Zhao et al., "Seeing Isn't Believing: Towards More Robust Adversarial Attack Against Real World Object Detectors," ACM CCS, 2019. https://doi.org/10.1145/3319535.3354259 [2] Chen et al., "ShapeShifter: Robust Physical Adversarial Attack on Faster R-CNN Object Detector," ECML PKDD, 2018. https://doi.org/10.1007/978-3-030-10925-7_4 **W2(b) & Q2 Post-filtering robustness.** To address this concern, we select several representative scenes and report their post-defense YOLO confidence scores. | Scene | Bicycle | Flowers | Bonsai | Garden | Kitchen | Treehill | |:---|---:|---:|---:|---:|---:|---:| | YOLO Confidence ↑ | 0.921 | 0.917 | 0.877 | 0.826 | 0.926 | 0.914 | | Change | -1.3% | -0.3% | -1.7% | +0.4% | -0.2% | -3.8% | Our evaluation shows that most scenes exhibit little to no degradation in yolo confidence after defensive filtering. We attribute this to the fact that the removed outlier Gaussians contribute only a small fraction of the rendered image and are not the primary contributors to the attack effect. **W2 (a) & Q3Held-out viewpoint stealthiness**. Our Fig. 10 presents visualizations from benign viewpoints distinct from those used for training camera poses, qualitatively demonstrating rendering fidelity. **To provide a more comprehensive quantitative evaluation**, we further sample held-out viewpoints around the object from attack point and measure PSNR between clean and attacked 3DGS renderings, using the clean renderings as references. The additional results demonstrate that visual fidelity is largely preserved outside the attack region. | Angle (°) | 45 | 90 | 135 | 180 | 225 | 270 | |:---|---:|---:|---:|---:|---:|---:| | PSNR (dB) | 33.04 | 36.23 | 33.24 | 34.12 | 34.89 | 35.32 | All six held-out viewpoints achieve PSNR values above 33 dB (33.04–36.23 dB) when compared with the clean 3DGS renderings, exceeding our reference threshold of 30 dB [3]. These results indicate relatively small pixel-level deviations from the original scene across the evaluated viewpoints. [3] Tan et al., "CodedStereo: Learned Phase Masks for Large Depth-of-Field Stereo," CVPR, 2021.

Q4. Taylor-based constraint error analysis.

Appendix D.6 evaluates the discrepancy between approximate and full rendering during optimization, rather than final benign-view fidelity. We clarify that our caching is exact, while the first-order approximation is applied to adversarial compositing, and the Taylor constraint estimates benign-view loss changes. To further evaluate the Taylor approximation, we compare its predicted benign-view loss changes with the actual changes computed using the full renderer at different optimization iterations.

IterationTaylor ΔLossFull-render ΔLossAbsolute Error ↓
200−5.50×10−6-5.50 \times 10^{-6}−6.10×10−6-6.10 \times 10^{-6}6.00×10−76.00 \times 10^{-7}
400−1.32×10−3-1.32 \times 10^{-3}−1.87×10−3-1.87 \times 10^{-3}5.50×10−45.50 \times 10^{-4}
1400−1.54×10−3-1.54 \times 10^{-3}−2.28×10−3-2.28 \times 10^{-3}7.40×10−47.40 \times 10^{-4}

Importantly, all results in Tables 2 and 4 are evaluated using the full 3DGS renderer. As shown in Table 4(a), our method reduces optimization time from 83.64 to 4.25 minutes, while improving attack-view PSNR from 29.68 to 33.96 dB and benign-view PSNR from 27.89 to 29.62 dB. Compared with the clean 3DGS, benign-view PSNR decreases by only 0.74 dB.

We also evaluate six held-out viewpoints excluded from optimization using the full renderer. The PSNR between clean and attacked renderings ranges from 33.04 to 36.23 dB, as reported in W2. These results provide additional evidence that our efficiency gains do not come at the cost of substantial degradation in the evaluated benign views.

Thank you again for your thoughtful review. We would appreciate your feedback on whether these additions address your concerns about the problem you mentioned. We would be happy to clarify any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.