acceptodds
Withdrawn from ICLR 2027

ValidGS: Validation-Guided Defense Against View-Dependent Poisoning in 3D Gaussian Splatting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The rapid progress of 3D Gaussian Splatting (3DGS), characterized by its explicit representation and highly efficient rendering, has facilitated the broad adoption of realistic 3D scene reconstruction. However, current 3DGS training pipelines lack a robust validation process to separate trustworthy Gaussians, which enable clean and accurate scene modeling, from questionable Gaussians that generate target-view-specific illusions, leaving 3DGS systems exposed to view-dependent poisoning attacks. A recent work, StealthAttack, shows that attackers can subtly corrupt only a small portion of the training views, causing the resulting 3DGS models to render illusions from certain target viewpoints while largely preserving rendering quality for innocent views. To address this problem, we introduce ValidGS, a validation-guided defense method against view-dependent poisoning. Specifically, ValidGS first constructs clean reference supervision by enforcing cross-view consistency, then applies a candidate target-view validation criterion to determine when these references are trustworthy. With this validation, our approach aggregates Gaussian-level statistics from validated clean regions, excludes Gaussians that also influence suspicious areas, and safeguards only the validated clean Gaussians, while periodically pruning Gaussians that remain suspicious. Comprehensive experiments on the well-known Mip-NeRF360, Tanks & Temples, and Free datasets show that ValidGS effectively mitigates target-view illusions while preserving high reconstruction fidelity for innocent views.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer Vision, 3D Scene Reconstruction, 3D Gaussian Splatting, Poisoning 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_8ocrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a defense ValidGS against the StealthAttack (which injects fake Gaussians to render a target appearance in the poisoned view). The idea is to utilize multi-view consistency: find image regions/patches with low neighboring consistency support, and replace this area with neighboring colors, and then map the suspicious and trusted image patches back to Gaussians. Trusted Gaussians are protected and suspicious Gaussians are deleted from time to time.

Critical strengths

The paper addresses view-dependent poisoning of 3DGS, a recent and under-explored security risk. The design of methodology is sensible and the evaluated settings are broad.

Critical weaknesses
  1. Several rendering-quality numbers are implausible. On Mip-NeRF360, ValidGS trained on clean data reaches 26.86 dB PSNR (Table 14), whereas the same method trained on poisoned data reaches 30.31 dB V-TEST PSNR on innocent views (Table 2). It is unclear how a defense can render innocent views 3.5 dB better when trained on poisoned data than on clean data. The same pattern appears elsewhere. The undefended Poisoned 3DGS (27.76 dB) already exceeds clean Vanilla 3DGS (27.21 dB). On Free, ValidGS raises V-TEST PSNR from 25.25 to 30.47 dB, even though the attack objective (Eq. 3) explicitly keeps innocent views close to the clean model, so a defense should at best recover clean-model quality. A V-TEST PSNR of 30.31 dB is also well above what state-of-the-art methods report on Mip-NeRF360 under clean training (around 28.5 dB, e.g., Zip-NeRF). Please clarify exactly how V-TEST is computed.

  2. The compared baselines are not designed for defending such attack: MaskGaussian targets model compaction; Deng et al. (2024) targets efficient density control; RemedyGS defends against computation-cost attacks (Poison-splat). Their failure in this setting is expected. Meanwhile, robust reconstruction methods for transient distractors, such as RobustNeRF (Sabour et al., 2023), SpotLessSplats, and WildGaussians seem to be more proper as baseline, since they rest on the same principle as Stage 1 of ValidGS: identifying image content that is not supported by the other views and down-weighting it. An object that appears in only one training image is, from their perspective, exactly a transient distractor. It will be helpful for the readers to understand the difficulty of this problem by reporting their performance.

  3. The reliability of the clean reference in the poisoned region is not analyzed. Stage 1 builds the clean reference by unprojecting pixels with the current rendered depth and reprojecting them into neighboring views. In the injected region, however, the rendered depth is largely determined by the poisoned Gaussians. StealthAttack places these at low-density locations along the target ray, in front of the true surface (Appendix A.2). Reprojecting with this depth samples the neighboring images at parallax-shifted locations, so the fused color comes from a different part of the scene rather than from the occluded background the reference is meant to recover. The reliability weight w(x) may down-weight some of these pixels, and the depth may become correct once the illusion is suppressed. However, references archived early in training could lock in incorrect colors. The paper does not discuss this.

Questions for authors
  1. How exactly is V-TEST computed? How does ValidGS trained on poisoned data outperform the same method trained on clean data by about 3.5 dB on Mip-NeRF360?

  2. How do distractor-robust methods such as SpotLessSplats and WildGaussians perform under the same attack setting and protocol as Tables 1 and 2? If they fail, why?

  3. Within the injected region, how accurate is the fused reference I_ref compared with the clean target image, especially early in training, when the poisoned Gaussians still dominate the rendered depth and references may already be archived?

Reviewer_KzMzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ValidGS, a validation-guided defense framework against view-dependent poisoning attacks in 3D Gaussian Splatting (3DGS). The method aims to eliminate target-view-specific illusions while preserving reconstruction quality in innocent views. ValidGS consists of three main components: consistency-based clean reference construction, validation-guided Gaussian filtering, and selective Gaussian purification. Specifically, the method constructs reliable supervision through cross-view photometric consistency, identifies trustworthy Gaussian primitives using accumulated visibility and region statistics, and subsequently protects validated Gaussians while pruning persistently suspicious ones. Experiments on Mip-NeRF360, Tanks & Temples, and Free demonstrate significant improvements over several defense baselines. The authors further evaluate the method under multi-view poisoning, adaptive attacks, different injected objects, and sparse training views. Ablation studies and supplementary experiments are provided to assess the contributions of individual components and the robustness of the proposed framework.

Critical strengths
  1. The paper shows that, on the Mip-NeRF360, Tanks & Temples, and Free datasets, this method significantly outperforms existing defensive baselines (Section 5.2, Tables 1, 2). ValidGS achieves a defense success rate of 9/9, 8/8, and 7/7 on these respective datasets, while maintaining high-quality reconstruction for "normal views" (i.e., views unaffected by poisoning). Furthermore, evaluations conducted under varying attack difficulty levels and multi-view poisoning scenarios (Tables 3, 4) further validate the framework's effectiveness across diverse experimental settings.

  2. The framework combines consistency-based reference construction, verification-guided Gaussian filtering, and selective Gaussian purification. By identifying reliable "clean" regions and selectively preserving the associated Gaussian primitives, ValidGS successfully addresses the challenge of mitigating "hallucinations" in targeted views without significantly compromising the reconstruction quality of normal views. This "verification-guided" design offers a practical and effective approach to enhancing the robustness of 3D Gaussian Splatting (3DGS) against view-specific poisoning attacks.

  3. The paper presents extensive supplementary experiments, including component ablation studies, statistical evaluations based on five random seeds, assessments against adaptive attacks, and an analysis of training-view sparsity. Additionally, the study explores the framework's compatibility, its performance under standard training conditions, and its computational overhead.

Critical weaknesses
  1. The proposed framework primarily combines cross-view photometric consistency, statistical Gaussian filtering, feature regularization, and periodic pruning (Eqs. (5)–(8)). Although their integration is meaningful, these components are closely related to existing consistency-based reconstruction and Gaussian management techniques. The paper does not sufficiently clarify which specific methodological innovations distinguish ValidGS from conventional robust reconstruction and Gaussian filtering strategies.

  2. The Related Work does not discuss recent research such as PoInit-of-View (CVPR 2026), which investigates transferable poisoning attacks targeting the initialization stage of 3D reconstruction systems. Although the two works address different threat models, discussing this distinction would help clarify the scope and novelty of ValidGS.

  3. Although the authors use matched poisoned inputs, 3DGS backbones, and training schedules (Lines 673–701), most competing methods are adapted from general Gaussian optimization techniques rather than specifically designed for view-dependent poisoning defense. The paper does not sufficiently establish whether these adapted baselines have been adequately optimized for the target threat model.

  4. The ablation study (Table 5) shows that adding Gaussian Filtering (B) to Reference Construction (A) improves the defense success rate from 0/8 to 5/8. However, Appendix A.4 (Lines 736–742) states that the (A)+(B) variant only constructs validated Gaussian sets without applying protection, regularization, or pruning. It is therefore unclear how this configuration produces such a substantial performance improvement. These issues raise concerns about the interpretation and reproducibility of the experimental comparisons.

Ref: [1] PoInit-of-View: Poisoning Initialization of Views Transfers Across Multiple 3D Reconstruction Systems. [CVPR2026]

Questions for authors
  1. What are the key methodological innovations that distinguish ValidGS from existing cross-view consistency-based reconstruction and Gaussian filtering approaches? In addition, could the authors clarify the relationship between ValidGS and recent poisoning research such as PoInit-of-View (CVPR 2026), particularly regarding their different threat models and stages of 3D reconstruction?
  2. How were the adapted defense baselines optimized to ensure a fair comparison? More importantly, why does the (A)+(B) configuration achieve a substantial performance improvement when Appendix A.4 indicates that Gaussian Filtering only constructs validated Gaussian sets without applying protection or pruning? Are there additional optimization operations that are not explicitly described? Q3. Robustness against adaptive attacks (Appendix B.4, Table 9).
  3. The defense success rate decreases to 14/24 under stronger consistency-aware poisoning. What are the primary causes of these failures? How would ValidGS perform against adaptive attacks that jointly optimize poisoned content to maintain geometric and photometric consistency across multiple neighboring views?
Reviewer_86h1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ValidGS, a validation-guided defense method against view-dependent poisoning attacks in 3D Gaussian Splatting (3DGS), aiming to eliminate malicious illusions from specific viewpoints while preserving reconstruction quality in benign views. To address the limitations of existing defenses, which struggle to accurately distinguish benign from malicious Gaussians and often suffer from unintended removal of legitimate scene content or incomplete attack mitigation, the authors introduce a three-stage framework: (1) Consistency-Based Reference Construction, which detects suspicious regions through cross-view consistency, constructs clean reference images by fusing information from neighboring views, and validates their reliability to obtain trustworthy supervision signals; (2) Validation-Guided Gaussian Filtering, which maps validated clean regions in 2D image space to 3D Gaussians and identifies Gaussians to protect by analyzing their contributions to clean and suspicious regions; and (3) Selective Gaussian Purification, which protects benign Gaussians through feature regularization and parameter update attenuation while periodically pruning persistently suspicious Gaussians. Experiments on 24 scenes across three datasets, Mip-NeRF360, Tanks & Temples, and Free, demonstrate that ValidGS achieves a 100% Defense Success Rate (DSR) under the main single-view and multi-view poisoning settings, while outperforming existing baselines in terms of PSNR, SSIM, and LPIPS.

Critical strengths
  1. Well-Motivated and Innovative Method Design. The paper proposes a cross-view consistency-based validation framework specifically designed to address view-dependent poisoning attacks. By constructing reliable clean references, identifying trustworthy Gaussians, and combining selective protection with pruning, the proposed method effectively mitigates malicious content while minimizing degradation in benign-view reconstruction quality. The overall framework is well-designed, with logically connected components.

  2. Comprehensive Experimental Evaluation and Strong Defense Performance. The paper conducts extensive experiments on three representative 3DGS datasets, covering single-view and multi-view poisoning, different attack difficulty levels, adaptive attacks, and ablation studies. The results demonstrate that ValidGS achieves high defense success rates under the main attack settings and outperforms the selected baselines in terms of PSNR, SSIM, and LPIPS, supporting the effectiveness of the proposed method.

  3. Clarification on the Ablation Study Results. Could the authors further clarify the actual contribution of the Gaussian Filtering module? According to Appendix A.4, the A+B variant only constructs the Gaussian protection set without applying feature regularization, parameter update attenuation, or pruning. However, Table 5 shows that A+B achieves a significant 4.54 dB improvement in V-ILLUSORY PSNR over A alone. Could the authors explain how such a substantial performance gain is achieved without any actual Gaussian protection or pruning operations?

Critical weaknesses
  1. Insufficient Baseline Comparisons. The paper primarily adopts density-based pruning as its pruning baseline, without comparing against importance-based Gaussian pruning methods such as LightGaussian. Since malicious Gaussians typically have significant effects only from specific viewpoints, their overall contributions across other benign views may be relatively limited. Therefore, importance-based pruning strategies could also offer potential defensive capabilities. Without such comparisons, it is difficult to fully establish the advantages of the proposed validation-guided mechanism over existing advanced pruning strategies.

  2. Limited Diversity of Attack Scenarios. The experimental evaluation mainly focuses on StealthAttack and its variants, without considering other 3DGS attack methods with different mechanisms, such as GaussTrap and ComplicitSplat. The current results primarily demonstrate the effectiveness of ValidGS against StealthAttack, but provide insufficient evidence of its generalizability and defensive effectiveness against broader categories of 3DGS poisoning attacks.

Questions for authors
  1. Could the authors provide comparisons with importance-based Gaussian pruning methods, such as LightGaussian, to demonstrate the advantages of ValidGS over more advanced pruning strategies? In particular, how does the proposed validation-guided mechanism outperform methods that identify less important Gaussians based on their contributions across multiple views?

  2. Could the authors extend the evaluation to other 3DGS poisoning attacks, such as GaussTrap and ComplicitSplat, or discuss whether ValidGS can effectively defend against their different attack mechanisms? Such experiments or analysis would help clarify the generalizability of the proposed defense beyond StealthAttack.

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

Related papers

Discussion (0)

Sign in to comment.