acceptodds
Withdrawn from ICLR 2027

4DGuard: Towards Proactive 4D Gaussian Assets Protection for Dynamic Novel-View Synthesis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Dynamic objects reconstructed as 4D Gaussian Splatting (4DGS) assets are valuable for simulation and content creation, yet they are shared as rendered videos from which anyone can segment, reconstruct, and reuse the object. We study proactive protection of 4D Gaussian assets: imperceptibly modifying the asset so that the target object cannot be reliably segmented in any video rendered from it. Beyond protecting images or static scenes, this raises two challenges: the protection must hold at every renderable time step, including those not seen during optimization, and it must survive video post-processing such as compression and frame averaging. We propose 4DGuard, which embeds adversarial perturbations into the asset itself rather than into individual frames. 4DGuard perturbs the appearance and opacity of the target Gaussians before deformation and lets the perturbation vary smoothly over time, so that every rendered frame carries it. To survive post-processing, it optimizes the perturbation against a promptable segmenter under a spatio-temporal expectation over transformations that covers compression and frame averaging, and a global calibration bounds the change of every released frame. Experiments on real-world dynamic scenes show that 4DGuard keeps the released video visually unchanged while disrupting target segmentation, generalizes to unoptimized time steps, transfers to unseen segmenters, and retains its effect under compression.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
4D Gaussian Splatting, Adversarial Perturbation, Dynamic Novel View Synthesis, Proactive Protection

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

This paper proposes 4DGuard, a proactive protection method for dynamic 4D Gaussian Splatting (4DGS) assets. It aims to reduce the risk of unauthorized asset extraction and reuse by introducing imperceptible adversarial perturbations that disrupt target object segmentation in rendered videos. Specifically, the method embeds perturbations into the color and opacity parameters of target Gaussians and employs Bernstein temporal basis functions to ensure smooth temporal variation. Furthermore, a spatio-temporal Expectation over Transformations (EOT) strategy is introduced to enhance robustness against video post-processing operations, such as JPEG compression and frame averaging, while global calibration maintains the visual fidelity of the released videos. Experimental results demonstrate that 4DGuard can effectively disrupt SAM-based segmentation in certain scenes, exhibiting a degree of temporal generalization and transferability across different segmentation models.

Critical strengths
  1. The research problem is novel and practically significant. This paper focuses on proactive protection of dynamic 4DGS assets and explores challenging issues such as temporal generalization and robustness against video post-processing, offering a new perspective on dynamic 3D asset protection.
  2. The robustness evaluation is comprehensive. The paper considers various downstream video post-processing operations encountered in real-world applications and systematically evaluates the proposed method under different conditions, including JPEG compression, temporal frame averaging, and their combinations, enhancing the completeness and credibility of the experimental evaluation.
Critical weaknesses

(1) Gap between the claimed contribution and the actual protection effectiveness. Although 4DGuard can significantly disrupt SAM segmentation in certain scenes, as acknowledged by the authors in Section 4.4, it has almost no impact on RGB-only reconstruction quality and only a limited effect on mask-guided object extraction, showing no clear advantage over ordinary perturbations (Table 6). This suggests that segmentation degradation does not necessarily translate into reduced reconstruction or extraction quality. Therefore, the current experimental evidence is insufficient to support the broader claim of proactive 4DGS asset protection. I suggest that the authors narrow their claims and more clearly position the contribution as adversarial disruption of target segmentation in videos rendered from dynamic 4DGS assets, rather than effective end-to-end asset protection.

(2) Limited temporal generalization and scene-dependent effectiveness. The effectiveness of 4DGuard varies substantially across different timestamps and scenes. Regarding temporal generalization, on the Split-cookie scene, the average IoU reduction for SAM ViT-H drops from 60.78 percentage points on optimized frames to only 1.97 percentage points on unoptimized timestamps, indicating a substantial degradation in attack effectiveness at unseen timestamps. Regarding performance across scenes, the method achieves strong segmentation disruption on Split-cookie but shows very limited effectiveness on Americano. These results suggest that the method is highly dependent on the specific timestamps and scenes, and its generalization and robustness require further validation (Section 4.3, Table 9).

Questions for authors

Q1. Effectiveness of End-to-End Asset Protection.

Given that 4DGuard shows limited degradation in mask-guided object extraction compared to ordinary perturbations and almost no impact on RGB-only reconstruction (Section 4.4, Table 6), could the authors clarify what evidence supports the broader claim of proactive 4DGS asset protection beyond segmentation disruption?

Q2. Temporal Generalization and Scene-Dependent Effectiveness.

Could the authors explain why the attack effectiveness drops substantially on unoptimized timestamps (e.g., from 60.78 pp to 1.97 pp on Split-cookie) and varies significantly across scenes? Are there any analyses identifying the factors responsible for these differences (Section 4.3, Table 9)?

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

This paper proposes 4DGuard, a proactive protection framework for dynamic 4D Gaussian Splatting assets released as rendered videos. Instead of perturbing individual frames, the method embeds adversarial perturbations into the appearance and opacity parameters of target Gaussians in the canonical asset, allowing the perturbation to propagate across viewpoints and time through the frozen deformation field. Experiments evaluate segmentation disruption on optimized and unoptimized time steps, robustness to compression and temporal averaging, transfer to unseen segmenters, and downstream recipient-side reconstruction and object extraction.

Critical strengths
  1. The work studies a timely and underexplored problem of protecting dynamic 4D Gaussian assets, extending proactive protection beyond images and static 3D scenes.
  2. The asset-level perturbation and spatio-temporal EOT are well motivated for improving temporal consistency and robustness.
Critical weaknesses
  1. This paper utilizes disrupted segmentation to protect 4D assets. In practice, this method only reduces the quality of specific intermediate processing steps, and does not completely prevent reconstruction and extraction.
  2. There is a lack of direct comparison with existing methods for segmentation attacks and 3D protection. While the paper mentions Attack-SAM, related segmentation attacks, and the static 3DGS protection method AdLift in the “Related Work” section, the experiments do not actually adapt these methods to video or 4D scenarios for comparison.
Questions for authors
  1. Is “Segmentation Destruction” Sufficient as a 4D Asset Protection Solution?
  2. As shown in Figure 7 and Table 6, the segmentation disruption does not translate into substantial degradation of downstream reconstruction or object extraction.
  3. The performance of the segmentation interference effect varies too widely across different frames; on some optimized frames, the SAM IoU shows almost no decrease (e.g., the IoU before and after protection in frame 399 is 0.957 in both cases). This raises concerns about the reliability of the proposed method, since even frames directly used during optimization are not consistently protected.
  4. Split-cookie exhibits very large segmentation degradation, whereas Americano shows almost no effect. Does this relationship have a quantitative description?
  5. How robust is it in video processing? For example, motion-compensated filtering, video denoising, or standard video codecs such as H.264/H.265.
Reviewer_pwotOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies proactive protection of dynamic 4D Gaussian Splatting assets released through rendered videos. It proposes 4DGuard, which stores adversarial perturbations in canonical Gaussian parameters and propagates them through the original deformation field. The method uses Bernstein basis functions for temporal variation and optimizes the perturbation against a frozen SAM surrogate under JPEG compression and temporal averaging. Experiments evaluate segmentation, transfer, temporal robustness, and recipient-side reconstruction and extraction. The method degrades segmentation in some settings, but the downstream results do not yet demonstrate reliable asset protection.

Critical strengths
  1. The paper studies a meaningful protection problem for dynamic 3D assets and proposes an appropriate asset-level solution. The method stores perturbations in canonical Gaussian parameters and propagates them through the original deformation field, which provides temporal persistence.

  2. The paper evaluates the method under several practical settings, including unseen segmenters, unoptimized time steps, JPEG compression, temporal averaging, reconstruction, and object extraction. The downstream experiments also reveal the gap between segmentation disruption and actual asset protection.

Critical weaknesses
  1. Segmentation degradation does not translate into effective downstream asset protection, and the core protection objective is not fully achieved. RGB-only reconstruction can bypass segmentation, and the reconstructed foreground quality remains similar between clean and protected releases. The mask-guided extraction results also show limited degradation, and the extracted objects remain recognizable.

  2. The empirical evidence for general protection is limited. The method shows strong scene dependence, does not clearly establish protection at novel viewpoints, and mainly evaluates SAM-style segmentation pipelines. The current experiments therefore do not fully support broad protection claims across scenes and recipient strategies.

Questions for authors
  1. What downstream criterion do the authors use to define a successfully protected asset?

  2. Can the method optimize against recipient-side reconstruction or extraction instead of only SAM?

  3. Can the authors evaluate held-out viewpoints that are not used during optimization or calibration?

  4. What causes the large performance differences across scenes?

  5. How robust is the method to non-SAM extraction pipelines, such as tracking, video segmentation, or manual mask correction?

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

Related papers

Discussion (0)

Sign in to comment.