acceptodds
Withdrawn from ICLR 2027

SHield: Watermarking 4D Gaussian Splatting via Spherical Harmonic Coefficient Perturbation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

4D Gaussian Splatting (4DGS) models are computationally expensive to train but distributed as easily copied parameter files. Existing watermarks protect these assets by fine-tuning the scene against a neural decoder to extract bit strings from rendered video, a heavy process vulnerable to the temporal dynamics of dynamic scenes. However, if the threat model targets the redistributed file itself, verification can be vastly simplified. Because the canonical Spherical-Harmonic (SH) coefficients of 4DGS are time- and view-invariant, parameter-space perturbations can be recovered via deterministic arithmetic. We propose **SHield**, the first non-blind, decoder-free watermarking approach that natively embeds full-resolution RGB images into 4DGS files. SHield tiles a zero-mean payload across the SH coefficients of a key-selected Gaussian subset, applying a brief anchored optimization to preserve visual fidelity. To verify ownership, it geometrically re-associates the key using canonical Gaussian centres to survive file tampering and recovers the payload in closed form. Across 30 scenes (D-NeRF, HyperNeRF, DyNeRF), SHield guarantees near-perfect payload recovery (Pearson ) on all 150 tested models with a mean rendering penalty of dB. Held-out evaluation yields perfect detection on 400 true trials and zero false positives across wrong-subset null trials. We also strictly bound the method's robustness: SHield survives INT8 quantization and opacity pruning for the image payload, and on aggressive SH-degree-1 truncation bit payloads are still decodable.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
4D Gaussian Splatting, Gaussian Splatting, Digital Watermarking, Model Watermarking, Spherical Harmonics, Ownership Verification, Dynamic Scene Representation, Neural Rendering

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_NG78OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

SHield targets redistributed 4DGS parameter files and uses non-blind verification. It tiles a zero-mean RGB image payload across the SH coefficients of a key-selected subset of Gaussians, applies 500 steps of anchored optimization to reduce rendering loss, and retains the original SH coefficients, positions, attributes, and payload metadata as a verification record. During extraction, it re-associates Gaussians through their canonical centers, subtracts the stored reference coefficients, and reconstructs the image by folding and averaging. For bit payloads, it uses a cyclic stride coprime to both the message length and the number of SH coefficients to reduce bit loss under truncation.

Critical strengths
  1. By retaining the original coefficients to cancel host interference, the method can recover full RGB payloads without a learned decoder.
  2. The paper includes extensive ablations and analyses of its operating limits.
Critical weaknesses
  1. The practical scope of protection is severely limited by the fragility of position-based matching, and the proposed fixes have not been incorporated into a complete final pipeline. Although the authors propose ways to resolve duplicate ambiguity and perform alignment, the main results still use the original matcher, and rotational alignment succeeds on only 25 out of 30 scenes. Even with correct matching, the SH-only fine-tuning attack in Table 3 defeats verification on both scenes evaluated at 10,000 steps, with an average GT-referenced PSNR cost of approximately 0.63 dB.

  2. The necessity and contribution of the design relative to simpler methods with the same information access have not been sufficiently established. It remains unclear how much of the near-perfect clean extraction and high capacity follows naturally from the non-blind reference information, and how much comes from the proposed design. The evaluation covers a broad range of 4D scenes, but file-level extraction from static canonical coefficients does not itself require a new temporal mechanism. The methodological novelty therefore needs to be defined more clearly.

Questions for authors
  1. Why have some of the improvements presented in the appendices not been incorporated into the final method? After integrating them, could the authors report results under a unified protocol covering clean files, quantization, pruning, Gaussian duplication, rigid transformations, and changes in position precision?
  2. How do simple non-blind SH-domain methods perform with the same original-coefficient side information, payload size, rendering quality, and verification-record storage budget? These results would help establish the independent value of the proposed payload allocation, optimization, and association mechanisms.
Reviewer_Mhi5OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

SHieLD focuses on embedding full-resolution RGB images into 4DGS files. It extracts the embedded image through closed-form arithmetic on canonical SH coefficients, using a key that stores the selected Gaussian indices, original SH coefficients, canonical centres, static tie-breaking attributes, and payload metadata.

Critical strengths

SHieLD tiles a zero-mean payload across the SH coefficients of a key-selected subset of Gaussians and applies a brief anchored optimization to preserve visual fidelity. For ownership verification, it uses canonical Gaussian centres to geometrically re-associate the key with the corresponding Gaussians after file tampering, then recovers the payload in closed form.

Critical weaknesses
  • Unclear novelty: If the contribution is full-resolution image watermarking, the paper does not clearly establish its capacity advantage over methods such as Splats in Splats++, which hide an entire scene. Moreover, comparisons with bit-level watermarking methods do not directly support the claimed image-watermarking contribution. If the contribution is decoder-free extraction, its benefits remain unclear given the original SH coefficients and Gaussian attributes explicitly retained in the key.

  • Key-security concerns: The stored Gaussian attributes may expose information about the underlying 4DGS representation. The paper should clarify how such leakage is assessed and prevented.

  • Insufficient comparisons: Relevant methods such as Splats in Splats++ and Guard4D, which also evaluate on D-NeRF and HyperNeRF, are missing from the comparison. Ablations alone do not establish an advantage in full-resolution image watermarking. Suitable baselines adapted from existing methods and evaluated under fair conditions would strengthen the claims.

Questions for authors

#Major:

  • Is the one-to-one matching required for key re-association computationally expensive? Please report its runtime.
  • Why does the baseline comparison (Mark4D) use binary payloads rather than full-resolution RGB images?
  • How does the pseudorandom candidate count K affect performance? An ablation is needed.
  • Given the original SH coefficients and Gaussian attributes stored in the key, how do the authors assess and prevent scene-information leakage?
  • What distinguishes wm1–wm5? Please provide clarification in the main paper.

#Miner:

  • Clearly indicate subfigures (a) and (b) in Figure 2.
  • Explicitly define S_kand cin Section 3.3 and Algorithm 1.
Reviewer_og3BOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work claims that their method is "the first non-blind, decoder-free watermarking approach that natively embeds full-resolution RGB images into 4DGS files".

Critical strengths

The authors conducted experiments on 150 models across 30 scenes, gaining a better understanding of scientific research through the process.

Critical weaknesses

First, I don't see enough differences between this work and the previous ones. Regarding embedding and extraction, many previous methods have already addressed them. For example, GaussianMarker have already used a PointNet-based decoder to extract the watermarks. NGS-Marker explicitly addresses native verification from local Gaussian subsets under partial infringement.

Second, the figures show that this work directly embeds images into the 4DGS model. This demonstrates the proposed method's higher capacity. However, I do not think that it is necessary to achieve a high capacity in watermarking. Such a strategy makes this method look like an approach for information hiding rather than watermarking. As watermarking cares more about robustness, more focus should be on the robustness of the proposed approach. However, from the method, it is difficult to find enough contributions on this part.

Then, the comparisons in this approach are not very fair. Many results seem to be copied directly from the original paper. Some comparisions also do not follow the standard pipeline.

At last, Tables 16 and 17 describe Hide-in-Motion's recovery PSNR as “artificially” inflated or elevated because it uses a learned decoder. A learned decoder is part of that method, and its use does not by itself invalidate the reconstruction metric. Given the acknowledged differences in representation, payload specification, and evaluation protocol, the supported conclusion is that these numbers are not directly comparable; suggesting an artificial advantage requires additional evidence.

Questions for authors

See my weakness

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

Related papers

Discussion (0)

Sign in to comment.