acceptodds
Withdrawn from ICLR 2027

AdvSplat: Adversarial Attacks on Feed-Forward Gaussian Splatting Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

3D Gaussian Splatting (3DGS) is increasingly recognized as a powerful paradigm for real-time, high-fidelity 3D reconstruction. However, its per-scene optimization pipeline limits scalability and generalization, and prevents efficient inference. Recently emerged feed-forward 3DGS models address these limitations by enabling fast reconstruction from a few input views after large-scale pretraining, without scene-specific optimization. Despite their advantages and strong potential for commercial deployment, the use of neural networks as the backbone also amplifies the risk of adversarial manipulation. In this paper, we introduce AdvSplat, the first systematic study of adversarial attacks on feed-forward 3DGS. We first employ white-box attacks to reveal fundamental vulnerabilities of this model family. We then develop two improved, practically relevant, query-efficient black-box algorithms that optimize pixel-space perturbations via a frequency domain parameterization: one based on gradient estimation and the other gradient-free, without requiring any access to model internals. Extensive experiments across multiple datasets demonstrate that AdvSplat can significantly disrupt reconstruction results by injecting imperceptible perturbations into the input images. Our findings surface an overlooked yet urgent problem in this domain, and we hope to draw the community’s attention to this emerging security and robustness challenge.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Feed-forward 3DGS, Adversarival Attack

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_uLh5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the adversarial robustness of feed-forward 3D Gaussian Splatting (3DGS) models and proposes AdvSplat, a black-box attack framework. It uses low-frequency DCT perturbations to reduce the search dimensionality, combined with either NES-based gradient estimation or gradient-free CMA-ES optimization to generate adversarial examples. Experiments show that small input perturbations can substantially degrade novel-view reconstruction quality.

Critical strengths
  1. The work addresses adversarial security in feed-forward 3DGS, a problem with practical relevance.
  2. The method is straightforward, using low-frequency perturbations to reduce the search dimensionality while covering two black-box optimization approaches.
  3. The attack effects are substantial: reconstruction quality degrades markedly across three models and two datasets. DCT control experiments further validate the role of frequency-domain search.
Critical weaknesses
  1. The core design mainly combines low-frequency DCT perturbations with established NES and CMA-ES algorithms, without a clearly distinctive attack mechanism tailored to 3DGS. It is unclear whether the work offers a unique task-specific motivation or insight.
  2. The main experiments primarily compare performance before and after attacks, without comparisons against existing black-box attacks under the same query budget. This makes it difficult to assess the method's competitiveness.
  3. Although the paper emphasizes query efficiency, the experiments involve numerous iterations and candidate samples. The actual query cost and efficiency gains over existing methods are not explicitly reported.
  4. For pose-free models, drift in pose estimation may affect rendering metrics. The paper does not adequately distinguish geometric reconstruction errors from quality degradation caused by viewpoint shifts.
Questions for authors
  1. What new technical designs does this work introduce beyond directly combining DCT with standard NES/CMA-ES?
  2. Could the authors provide comparisons with existing black-box attacks under identical query budgets?
  3. For pose-free models, can the authors distinguish pose drift from actual geometric reconstruction errors?
Reviewer_37CmOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduced AdvSplat, which is claimed to be the first systematic study of adversarial attacks on feed-forward 3DGS. The paper developed two query-efficient black-box algorithms that optimize pixel-space perturbations via a frequency domain parameterization. The paper claims to be the first to investigate the security issues in feed-forward 3DGS models through designing adversarial attacks.

Critical strengths

This paper studied a important topic of 3D reconstruction model security. This paper implemented two black-box attacks for feed-forward 3DGS models in the methods section.

Critical weaknesses
  1. The contribution of this paper is significantly overclaimed. The adversarial robustness of 3DGS models have been studied and published before, actually both attacks and defenses. Claiming main contribution to be the first to explore feed-forward 3DGS is a bit ambitious. For example, PoisonSplat [1], RemedyGS [2], AdLift [3], Gaussian Splatting Under Attack [4], ShieldSplat [5], StealthAttack [6]. There are far more papers in this field that acknowledged in this paper. Just because this paper investigates feed-forward 3DGS only, doesn't mean that their contribution claims are valid.

  2. Considering that there are lots of works in the area of 3DGS security vulnerabilities, this paper failed to highlight why feed-forward 3DGS is different and why it is more vulnerable.

  3. A good portion of the claimed technical novelty is simply to fit PGD onto feed-forward 3DGS models, which in my opinion should not take a significant portion in the method section.

  4. No error bar or standard deviation is reported in the experimental tables, need to prove statistical significance.

  5. The experimental setup with detailed parameter setting information is missing in the experiments section. The implementation details only indicated the GPU settings.

  6. Only two dataset benchmarks are considered, which the authors also did not specify how many scenes total, and other relevant information.

  7. No robustness tests for the proposed attack results. How about simply smoothing the image? Will that change attack performance? How about some other defensive methods to detect and remove artifacts?

References:

  1. Lu, Jiahao, et al. "Poison-splat: Computation cost attack on 3d gaussian splatting." International Conference on Learning Representations. Vol. 2025. 2025.
  2. Li, Yanping, et al. "Remedygs: Defend 3d gaussian splatting against computation cost attacks." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.
  3. Hong, Ziming, et al. "Adlift: Lifting adversarial perturbations to safeguard 3d gaussian splatting assets against instruction-driven editing." (2025).
  4. Zeybey, Abdurrahman, Mehmet Ergezer, and Tommy Nguyen. "Gaussian splatting under attack: Investigating adversarial noise in 3d objects." arXiv preprint arXiv:2412.02803 (2024).
  5. Chen, Minshi, Hui Wei, and Xin Xu. "Shield-Splat: Mitigating Dos Computation-Cost Attacks on 3d Gaussian Splatting Service Via Input Sanitization." Available at SSRN 5392996.
  6. Ke, Bo-Hsu, et al. "Stealthattack: Robust 3d gaussian splatting poisoning via density-guided illusions." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.
Questions for authors
  1. On line 97 of the main paper, the referencing did not work, it showed up as "as shown in ??". Could you fix that?
  2. For PGD, why was L_inf norm selected as the main constraint? How are the results for L2 and L1?
Reviewer_jDjUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates the vulnerability of feed-forward 3DGS models (DepthSplat, NoPoSplat, AnySplat) to adversarial perturbations on input views. It shows that white-box PGD with ℓ∞=8/255\ell_\infty = 8/255 severely degrades reconstruction quality. To target black-box scenarios, the authors propose searching within a block-DCT low-frequency subspace using NES and CMA-ES. Experiments on RE10K and DL3DV show substantial drops in novel-view rendering quality.

Critical strengths
  1. The vulnerability itself is convincingly demonstrated: across all three feed-forward architectures (both pose-known and pose-free), a small bounded perturbation collapses novel-view quality (Tables 4–5).
  2. The attack loss relies only on input views and their re-renderings, without requiring ground-truth novel views. This is an appropriate and realistic formulation for black-box threat models.
Critical weaknesses
  1. Missing query costs and unsubstantiated efficiency claims. Query efficiency is presented as the primary merit of the proposed black-box attacks, yet the paper never reports actual query counts. Based on Appendix A (10,000 iterations, M=B=40M = B = 40), the search requires roughly 800K queries per scene for NES and 400K for CMA-ES. If this calculation holds, it is unclear in what context these attacks can be considered query-efficient—standard black-box image attacks typically operate within a budget of ≤10,000\le 10{,}000 queries (e.g., Bandits, SimBA, Square Attack). Furthermore, the paper lacks standard external baselines (e.g., random noise at the same ϵ\epsilon, Square Attack, SimBA-DCT, or a tuned pixel-space NES). Dismissing Square Attack and SimBA without empirical comparison is unjustified, given that both are score-based and require only scalar feedback. Consequently, the reported efficiency is demonstrated only against the authors' own unoptimized non-DCT ablation, rather than establishing any meaningful advantage over prior art.

  2. Uncredited prior work and limited technical novelty. Searching in a low-frequency DCT subspace with NES is the core technique of Guo et al. (UAI 2019, "Low Frequency Adversarial Perturbation"), and diagonal CMA-ES attacks in reduced search spaces date back to Meunier et al. (2019). Additionally, attacking source views in novel-view synthesis is not new: NeRFool (ICML 2023) perturbs inputs to generalizable NeRFs, and Geometry Cloak (NeurIPS 2024) attacks feed-forward Gaussians. The authors' claims of being "the first" need significant calibration, as the methodology largely amounts to applying known optimization techniques to a newer family of feed-forward models.

Questions for authors
  1. What are the exact query counts and wall-clock times per scene for each variant? Is the estimate of ~800K (NES) and ~400K (CMA-ES) accurate, and if so, against which existing black-box methods is this considered efficient?

  2. How many scenes are evaluated per dataset, how are the evaluation views selected, and at what resolution? The reported clean numbers differ significantly from the original papers (e.g., DepthSplat on RE10K is reported at 21.09 dB vs. 27.47 dB in the original work).

  3. Under matched query budgets (e.g., 1K, 10K, and 100K queries), how do random noise at equivalent ϵ\epsilon, Square Attack, and SimBA-DCT compare against the proposed NES and CMA-ES variants?

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

Related papers

Discussion (0)

Sign in to comment.