acceptodds
Withdrawn from ICLR 2027

HiddenSplat: Injecting View-Specific Content into 3D Gaussian Splatting via “Invisible” Gaussians

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

3D Gaussian Splatting (3DGS) has rapidly emerged as a leading representation for photorealistic novel-view synthesis and is increasingly adopted in AR/VR, robotics, and digital content creation. As 3DGS scenes become widely distributed and reused as shareable assets, they introduce an underexplored security risk: an adversary can inject a set of structurally optimized Gaussians into a pretrained 3DGS scene such that a chosen target content (e.g., a sensitive image, logo, watermark, or hidden message) appears clearly when the scene is rendered from a specific target viewpoint while remaining visually imperceptible from other viewpoints. We term this attack HiddenSplat. Unlike prior approaches, HiddenSplat efficiently operates in both white-box and gray-box settings, enables the attacker to arbitrarily select the target viewpoint, and requires no auxiliary encoders or decoders. Extensive experiments across diverse scenes and target patterns demonstrate that HiddenSplat achieves high-fidelity target reconstruction at the attack viewpoint while keeping the injected content effectively “invisible” from non-target views. Our findings reveal a new, practical attack surface to 3DGS assets and highlight the need for efficient defenses against such vulnerabilities.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
3D Gaussian Splatting, view-dependent attack, content injection, adversarial gaussians

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_X6EoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies view-specific content injection into pretrained 3D Gaussian Splatting assets. HiddenSplat freezes the original scene and adds low-opacity Gaussians along target-camera rays. Target reconstruction, non-target preservation, spatial repulsion, and color dispersion objectives are combined to reveal designated content at a selected viewpoint while limiting its visibility elsewhere. The paper presents alpha-compositing analyses and experiments on three datasets, supplemented by nearby-view, pruning, overhead, multi-view, and preliminary defense studies.

Critical strengths
  1. The attack construction addresses a concrete asset-integrity problem. Adding standard Gaussians to a pretrained scene allows the payload to be displayed through the conventional rendering pipeline, without modifying the original scene parameters or requiring a separate decoder. This makes the proposed threat relevant to the distribution and reuse of 3DGS assets. (§§3.2–4)

  2. The submission provides substantial experimental coverage and inspectable source code. The paper investigates several important aspects beyond target reconstruction, including neighboring viewpoints, computational overhead, pruning, and preliminary defenses. The supplementary source makes the implementation accessible for inspection, which is valuable for assessing the method and its evaluation.

Critical weaknesses
  1. The released pipeline does not support interpreting the test-view results as held-out generalization. The training script places both training views and all non-target validation views into the non-target optimization pool and samples that pool for the stealth loss. The evaluation script subsequently evaluates those same non-target validation views. Thus, under the supplied pipeline, these views are available during attack optimization rather than held out from it. This limits what the reported test-view results establish about stealth at unseen viewpoints. The concern is the interpretation of generalization, rather than whether access to these cameras is allowed by the threat model.

  2. The implemented stealth objective differs materially from Equation (8). The manuscript defines a masked discrepancy between poisoned and clean renderings. For dataset cameras, however, the supplied implementation minimizes full-frame L1 against the original photographs; only interpolated cameras use clean-rendered reference images. These objectives can produce different gradients when the clean reconstruction contains errors, and the full-frame reduction also changes the relative weighting of the loss. Consequently, the supplementary implementation does not provide a consistent account of the objective underlying the reported results.

  3. The evaluation implementation and reported metric definitions are not fully consistent. Stealth PSNR is weighted by accumulated adversarial opacity, while SSIM and LPIPS use a bounding-box crop that can include unaffected pixels. These differ from uniformly evaluating the affected-pixel mask described in the manuscript. Target PSNR also averages over the entire patch after setting pixels outside the target mask to zero, which can inflate it relative to averaging only over valid target pixels. Separately, Table 2 reports αmean=0.156>αmax⁡=0.122\alpha_{\mathrm{mean}}=0.156>\alpha_{\max}=0.122 for StealthAttack-Adapt. These discrepancies weaken confidence in the interpretation and comparability of the reported metrics.

  4. Appendix C.2 does not establish the claimed uniform perturbation bound. The occupancy analysis counts Gaussians only within a Mahalanobis radius of 0.5 and averages the count across pixels and scenes. Substituting this average into the pointwise bound of Theorem 2 cannot establish an “at most 3%” deviation, and contributions outside the counted region are omitted. The observed occupancy decay is informative, but interpreting it as a worst-case guarantee is stronger than the evidence supports. This concern addresses the analysis already provided in the appendix, rather than its absence.

Questions for authors

See weaknesses.

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

This paper introduces HiddenSplat, a view-specific content injection attack designed for pretrained 3DGS scenes. The method injects multiple low-opacity Gaussians along the pixel rays of a target viewpoint. By leveraging the accumulation of their opacity from that specific perspective, it enables the reconstruction of desired content while minimizing visibility from non-target viewpoints through spatial dispersion. To achieve this, the authors propose four loss functions, including target reconstruction, stealth, spatial repulsion, and color dispersion. The effectiveness of the proposed method is validated through theoretical analysis and experiments across multiple datasets.

Critical strengths

(1)This paper addresses the security risks associated with pre-trained 3DGS assets. It proposes an incremental content-injection attack that requires no modifications to the original Gaussian parameters, offering a valuable perspective on the security of 3DGS scenes.

(2)The proposed attack leverages the ray-accumulation properties of low-opacity Gaussians and their spatial dispersion across different viewpoints. By incorporating various optimization constraints, the method achieves a balance between the accurate reconstruction of the target content and the concealment of that content from non-target viewpoints.

(3)The authors conducted comparative experiments across three public datasets and validated the effectiveness of the approach through ablation studies, gray-box attacks, and parameter sensitivity analysis. The results demonstrate high-quality reconstruction of the target content while maintaining effective concealment from non-target viewpoints.

Critical weaknesses

(1)Theorem 1 focuses on the accumulation capacity of Gaussians along a single pixel ray, yet it fails to account for projection overlap and mutual interference between different rays. The upper bound on perturbations in Theorem 2 depends on the number of overlapping Gaussians from non-target viewpoints; however, there is no rigorous proof that this number decreases as the viewpoint shifts. Consequently, the current theoretical analysis is insufficient to provide robust guarantees regarding overall reconstruction capability and stealth.

(2)The paper does not clarify whether the comparisons with other methods were conducted under equivalent conditions—specifically, using the same number of Gaussians and the same computational budget. Given that HiddenSplat requires the initialization of a large number of Gaussians along each target ray, its performance advantages might stem, at least in part, from a higher representation capacity rather than the proposed spatial dispersion mechanism.

(3)The experiments lack direct comparisons with simpler strategies, such as random depth sampling or uniform distribution, making it difficult to fully justify the necessity of the core design elements.

(4)Although the paper utilizes a 3D spatial repulsion loss to reduce projection overlap from non-target viewpoints, 3D spatial distance does not necessarily correlate directly with the degree of overlap on the 2D image plane. Additionally, the effectiveness of the color dispersion loss in mitigating visual anomalies remains inadequately explained. It is recommended to include comparisons with optimization strategies based directly on projection overlap or rendered color discrepancies to further validate the necessity of these design choices.

Questions for authors

Clarification Questions:

(1) Under what assumptions does the number of overlapping Gaussians decrease as the viewpoint moves away from the target view, as suggested by Theorem 2?

(2) Are the baseline methods compared under the same Gaussian budget and computational cost?

(3) Why should the proposed 3D repulsion and color dispersion losses effectively reduce visual artifacts in non-target views?

Suggestions for Improvement:

(1) Comparisons with simpler sampling strategies, such as random or uniform depth sampling, would help justify the proposed design.

(2) Comparisons with losses that directly penalize 2D projection overlap or rendering differences would further support the proposed regularization terms.

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

This work injects optimized Gaussians into pretrained 3DGS scenes so that chosen content appears from a target viewpoint while remaining visually imperceptible from other views.

Critical strengths

Experiments across three benchmarks, together with gray-box transfer, multi-view injection, and ablation studies, provide useful evidence of the method’s flexibility and effectiveness.

Critical weaknesses
  1. The paper demonstrates visual imperceptibility from non-target viewpoints, but does not establish whether the injected content remains difficult to detect when a defender uses AI-based analysis of the complete 3DGS representation.

  2. Please evaluate representation-level detection to determine whether visual imperceptibility translates into resistance to automated scene inspection.

  3. Section 3.2 requires invisibility from arbitrary non-target viewpoints, but rotating a camera around its fixed optical center does not create the parallax needed to disperse Gaussians stacked along the target rays. Their viewing directions for SH evaluation also remain unchanged, so the content should remain visible while it stays within the camera frustum. Please define the trigger region in terms of camera position and orientation, and evaluate rotations and translations separately.

  4. Appendix C.2 counts only Gaussians within a Mahalanobis radius of 0.5 and then averages this count over affected pixels, obtaining approximately five overlaps. Substituting this average into Theorem 2 cannot establish the claimed per-pixel upper bound of 0.03, and Gaussians outside the chosen radius can still contribute to rendering. Please measure the actual contributing Gaussians or accumulated effective opacity per pixel, including worst-case values.

  5. Table 2 reports a test-view maximum opacity of 0.122 and mean opacity of 0.156 for StealthAttack-Adapt. A mean cannot exceed the maximum when both are computed over the same samples. Please clarify the accumulation and aggregation definitions and correct this inconsistency before using these statistics to support the stealth comparison.

Questions for authors

Please see the weakness. Specifically, I am interested in Q1 and Q2.

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

Related papers

Discussion (0)

Sign in to comment.