- 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
- 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.
- 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)?