acceptodds
Withdrawn from ICLR 2027

Anisotropy-Gated Densification Against Poison-splat Attacks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

3D Gaussian Splatting (3DGS) achieves efficient novel-view synthesis through explicit Gaussian representations and Adaptive Density Control (ADC). However, recent Poison-splat attacks exploit the gradient-driven densification mechanism to trigger excessive clone and split operations, leading to rapid Gaussian proliferation and substantial computational overhead. We observe that poisoned scenes exhibit stronger spectral anisotropy than clean scenes, suggesting that frequency-domain characteristics provide useful cues for regulating densification. Based on this observation, we propose Anisotropy-Gated Densification (AGD), a defense method that integrates spectral anisotropy into ADC. AGD adaptively adjusts the densification threshold according to scene-level spectral statistics and guides gradient accumulation using local spectral orientation, suppressing unnecessary Gaussian generation while preserving legitimate scene structures. Extensive experiments on Mip-NeRF360, NeRF Synthetic, and Tanks and Temples demonstrate that AGD effectively mitigates Poison-splat attacks by reducing Gaussian proliferation, GPU memory consumption, and training overhead, while maintaining rendering quality comparable to the original 3DGS on clean scenes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer Vision, 3D Gaussian Splatting, Attacks and Defense

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4QaNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Anisotropy-Gated Densification (AGD) as a defense against Poison-splat, which is a computation-cost attack that perturbs training images to inflate 3DGS densification. AGD computes a tile-level spectral anisotropy score from an amplitude-weighted orientation tensor over mid-band FFT coefficients. Experiments are done on NeRF-Synthetic, Mip-NeRF360 and Tanks&Temples. Under poisoning, AGD reduces peak Gaussian count by 17–28%, peak memory by 14–51% and training time by 15–23% relative to undefended 3DGS, at a cost of ≤0.35 dB PSNR. Clean-scene quality is essentially unchanged. The paper also combines AGD with three existing defenses.

Critical strengths
  1. The research problem is interesting and important: Compute-cost attacks matter for 3DGS training services. Regulating the densification decision directly is a reasonable complement to input purification.

  2. Near-zero clean-scene cost. Clean quality is nearly unchanged with the proposed defense.

  3. Lightweight and plug-in. Spectral maps are precomputed and the whole pipeline does not require learning. The method can be integrated into four 3DGS variants.

Critical weaknesses
  1. The core mechanism of the proposed method lacks interpretation. STI aggregates tile-level anisotropy with a median over tiles (Eq. 5) and then over images (Eq. 6). Why use median but not other statistics? What is the advantage? If the design is known by the attacker, would it be easy to evade (an adaptive attack)? These methodology choice should be discussed and analysed more in-depth.

  2. The effectiveness of AGD is not strongly supported by the results. From Table 1 and Table 2, we can see that AGD alone is weaker than existing defenses (comparing result of AGD in Table 1 and result of all three other defenses in Table 2), e.g., #Gaussians (M) of AGD alone is 0.561 while the #Gaussians (M) of image smoothing, remedy GS and DefenseSplat are all under 0.3. Additionally, adding AGD to RemedyGS increases #G, memory and time under poisoning (see Table 2, the row of RemedyGS)

  3. Regarding removing needle-like Gaussians: how to define and distinguish needle-like, plate-like and sphere Gaussians? Are there only 3 categories in total? Removing needle-like Gaussians exhibit similar (or even worse) degradation in PSNR compared with random removal, it seems that such needle-like Gaussians removal can be further improved by finer-grained detection or specification. If the removal of Gaussians will cause a degradation less than random removal, the design will make sense more. SR's memory saving effect also seem to only appear in Nerf-Synthetic, but not significant on MIP and TT (Tables 10-11).

Questions for authors
  1. Choice of median. What is the rationale for using the median at both levels: over tiles (Eq. 5) and over images (Eq. 6)? Please consider comparing it against alternatives such as the mean, a trimmed mean, a high percentile (e.g., 90th), or the max.

  2. Adaptive attacks. How does AGD perform against an attacker who knows the design? In particular: (a) poisoning fewer than half of the training views, so the image-level median is unchanged; (b) confining the perturbation to fewer than half of the tiles in each image, so the tile-level median is unchanged; (c) adding a differentiable penalty to the Poison-splat objective that keeps AsceneA_{\text{scene}} below ATA_T.

  3. Why does adding AGD to RemedyGS increase cost under poisoning?

  4. How exactly are needle-like, plate-like and spherical Gaussians defined, e.g., which thresholds on the ratios between scale axes? Are these three categories exhaustive and mutually exclusive, or do some Gaussians fall outside all three? What fraction of Gaussians falls into each category in clean vs. poisoned scenes?

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

This paper proposes Anisotropy-Gated Densification (AGD), a defense against Poison-Splat attacks on 3D Gaussian Splatting. The method estimates scene-level spectral statistics, introduces direction-aware gradient accumulation, and regularizes Gaussian scales during later optimization. Experiments on NeRF-Synthetic, Tanks & Temples, and Mip-NeRF 360 show reductions in Gaussian count, peak memory, and training time under poisoned inputs while largely preserving rendering quality.

Critical strengths
  1. The paper targets the densification mechanism through which Poison-Splat increases computational cost, rather than relying exclusively on input purification.
  2. The method is lightweight and can be integrated into existing 3DGS optimization pipelines.
Critical weaknesses
  1. The evaluation considers only Poison-Splat-style attacks. An adaptive attacker could explicitly optimize perturbations to evade the anisotropy gate or manipulate the estimated spectral statistics.

  2. The paper does not compare against simple capacity-matched controls, such as increasing the standard densification threshold, imposing a Gaussian-count budget, or applying stronger scale regularization. It is therefore unclear how much benefit specifically comes from spectral anisotropy.

  3. Spectral anisotropy is not uniquely associated with poisoning. Legitimate thin structures and directional textures may exhibit similar statistics, but the paper provides limited analysis of such failure cases.

  4. The reported compatibility is inconsistent. AGD increases resource consumption for some combinations, such as RemedyGS and GDAGS, and generally causes small reconstruction-quality reductions.

Questions for authors
  1. How does AGD compare with simple controls, such as increasing the standard densification threshold or imposing a Gaussian-count cap, when all methods are matched for final Gaussian count or reconstruction quality?

  2. Have the authors evaluated, or can they discuss quantitatively, an adaptive Poison-Splat attacker that jointly maximizes resource consumption while minimizing the spectral-anisotropy signal used by AGD?

  3. Are the same hyperparameters used across all datasets and scenes, and how sensitive are the results to the scene-level threshold settings?

Reviewer_Svy3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Anisotropy-Gated Densification (AGD), a defense method against Poison-splat attacks on 3D Gaussian Splatting (3DGS). The method aims to mitigate excessive Gaussian proliferation caused by adversarially amplified gradients during Adaptive Density Control (ADC). AGD consists of three components: Scene-aware Threshold Initialization (STI), which adjusts the densification threshold based on scene-level spectral anisotropy; Direction-aware Gradient Accumulation (DGA), which incorporates local spectral orientation into gradient accumulation; and Scale Regularization (SR), which penalizes excessively elongated Gaussians. Experiments on three datasets demonstrate that AGD reduces Gaussian proliferation, GPU memory consumption, and training time under Poison-splat attacks while maintaining comparable rendering quality.

Critical strengths
  1. Well-Motivated Defense Perspective and Simple Modular Design. The paper directly targets the gradient-driven densification mechanism exploited by Poison-splat attacks, rather than focusing solely on input purification. This provides a reasonable approach to mitigating computation-cost attacks against 3DGS. Furthermore, the proposed method incorporates scene-level spectral statistics and local directional information into the existing ADC mechanism without modifying the underlying Gaussian cloning and splitting operations. The overall design is relatively simple and can be integrated into different 3DGS pipelines.

  2. Comprehensive Experimental Evaluation. The paper evaluates AGD on three datasets and reports multiple metrics, including rendering quality, Gaussian count, GPU memory consumption, and training time. The results demonstrate that AGD effectively reduces computational overhead with only a minor degradation in rendering quality. Additionally, the authors provide ablation studies and compatibility evaluations with existing defenses and representative 3DGS methods, offering further empirical evidence for assessing the effectiveness of the proposed approach.

Critical weaknesses
  1. Insufficient Direct Comparisons with Existing Defenses and Inconsistent Compatibility Gains. The paper lacks systematic head-to-head comparisons with existing defense methods, such as RemedyGS, DefenseSplat, and Spectral Defense. Instead, these methods are mainly evaluated by combining them with AGD. More importantly, adding AGD does not consistently improve existing defenses and can even worsen their performance. For example, combining AGD with RemedyGS on NeRF Synthetic increases Gaussian count, GPU memory consumption, and training time under poisoned training. These results raise concerns about AGD's practical advantages over existing defenses and the claimed complementarity with other methods.

  2. Missing Simple Fixed-Threshold Baselines. A major component of AGD, Scene-aware Threshold Initialization (STI), suppresses Gaussian proliferation by increasing the ADC densification threshold. The ablation results show that STI alone already accounts for a substantial portion of the defense effectiveness. However, the paper does not compare STI against simple fixed-threshold scaling strategies, such as multiplying the original threshold by a constant factor. Without such comparisons, especially under comparable reconstruction quality, it remains unclear whether the proposed spectral anisotropy-based adaptation provides meaningful benefits beyond simply increasing the densification threshold. This weakens the justification for the necessity and effectiveness of the proposed mechanism.

Questions for authors

Q1. Comparison with Existing Defenses and Compatibility

Could the authors provide direct, head-to-head comparisons between AGD and existing defense methods, such as RemedyGS, DefenseSplat, and Spectral Defense, under identical experimental settings? Additionally, could the authors explain why incorporating AGD sometimes leads to worse performance? For example, combining AGD with RemedyGS on NeRF Synthetic increases the Gaussian count, GPU memory consumption, and training time under poisoned training. These results raise questions about the practical advantages of AGD and its claimed complementarity with existing defenses.

Q2. Effectiveness of Spectral Anisotropy-Based Threshold Adaptation

Since STI achieves substantial defense improvements primarily by increasing the ADC densification threshold, could the authors compare it against simpler fixed-threshold scaling strategies (e.g., multiplying the original threshold by a constant factor)? In particular, could the authors demonstrate whether the proposed spectral anisotropy-based adaptive thresholding provides meaningful advantages over fixed-threshold baselines under comparable reconstruction quality? Such comparisons would help clarify the necessity and effectiveness of the proposed spectral anisotropy mechanism.

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

Related papers

Discussion (0)

Sign in to comment.