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 r≥0.991) on all 150 tested models with a mean rendering penalty of ≤0.13~dB. Held-out evaluation yields perfect detection on 400 true trials and zero false positives across 1,600 wrong-subset null trials. We also strictly bound the method's robustness: SHield survives INT8 quantization and 30% opacity pruning for the image payload, and on aggressive SH-degree-1 truncation bit payloads are still decodable.
Figures & tables
Figure 1: File-centric ownership verification with SHield. The owner embeds a payload into the SH coefficients of key-selected Gaussians. Under adversarial file degradation, verification geometrically re-associates the key via canonical centres, enabling closed-form extraction directly from the parameters without learned decoders or rendering loops.
Figure 2: Overview of SHield. (a) Embedding: A zero-mean payload is tiled across the SH slots of key-selected Gaussians (red). Learnable offsets are applied to these SH coefficients while the scene geometry, unselected Gaussians, and deformation network remain strictly frozen (green). A joint loss anchors the offsets to the tiled target while preserving rendering fidelity (yellow). (b) Verification: Centre-based re-association locates the target Gaussians even if the file is reordered or pruned. Subtracting the stored reference SH cleanly isolates the perturbation, and the redundant copies are averaged to reconstruct the payload W^ . Ownership is verified if the correlation r≥τ .
Dataset (models)
PSNR clean
Δ PSNR [worst]
SSIM ( Δ )
LPIPS ( Δ )
WM-PSNR
min r
D-NeRF (35) ( Pumarola et al., 2021 )
35.31
− 0.048 [ − 0.113]
0.9842 ( − 0.0004)
0.0164 ( + 0.0002)
44.77
0.9990
HyperNeRF (85) ( Park et al., 2021 )
26.92
− 0.047 [ − 0.150]
0.7761 ( − 0.0012)
0.2949 ( + 0.0009)
36.27
0.9910
DyNeRF (30) ( Li et al., 2022 )
31.19
− 0.127 [ − 0.314]
0.9353 ( − 0.0011)
0.1099 ( + 0.0012)
36.59
0.9952
Table 1: Rendering cost and extraction ( α0=0.1 , five payloads per scene, 20 test views). Δ values are against the unwatermarked model, with the worst scene bracketed; min r is the worst payload recovery over models. The comparison against plain tiling at the same α0 and selection is in Appendix H .
Figure 3: Rendering residual ( α0=0.1 , wm1): each dataset’s median-residual scene at its median test view, to avoid cherry-picking; per-pixel RMS difference to the clean render in 8-bit levels (lv), one shared scale, zoom box at the largest local residual.
Opacity pruning
SH quantization (whole file)
WM-PSNR
r
r≥τ
WM-PSNR
r
r≥τ
0 %
38.32
0.998
150
FP32/16
38.32
0.998
150
10 %
23.38
0.965
150
INT12
38.25
0.998
150
30 %
18.21
0.892
150
INT10
37.35
0.998
150
50 %
15.38
0.790
150
INT8
32.08
0.994
150
80 %
12.70
0.544
52
Table 2: Post-processing of the model file ( α0=0.1 , all 150 image-payload models): mean WM-PSNR, mean r and passes at τ=0.6 .
Figure 4: Payloads recovered from the model file , for the scenes of Figure 3 : as distributed, after whole-file SH INT8, and after 30% opacity pruning, with Pearson r against τ=0.6 . On an unmarked model the recovery collapses to a flat block of the payload’s per-channel mean colour, whose residual correlation is payload-dependent.
r re-associated
r oracle indices
Fine-tuned
steps
Δ PSNR
mean
min
mean
min
matched (%)
passes
SH only
100
+ 0.01
0.9615
0.9282
0.9615
0.9282
100.0
5/5
SH only
500
− 0.04
0.8994
0.8458
0.8994
0.8458
100.0
5/5
SH only
2,000
− 0.18
0.7740
0.6634
0.7740
0.6634
100.0
5/5
SH only
5,000
− 0.39
0.5257
0.4312
0.5257
0.4312
100.0
1/2
SH only
10,000
− 0.63
0.3946
0.3109
0.3946
0.3109
100.0
0/2
Table 3: Adaptive fine-tuning attack. Δ PSNR: quality on held-out views against the distributed model, i.e. the adversary’s cost (training-view quality rises throughout). r after re-association and at the key’s stored indices ( oracle ); passes at τ=0.6 .
Table 4: Cross-threat-model reference with Table 1 of Mark4D ( Lee et al., 2026 ) . Baselines are quoted; SHield is measured on our base models with the same message lengths but a different D-NeRF protocol. Bold marks the best value in each column; the row groups answer different threat models, so this is a cross-threat-model reference rather than a ranking. Acc.: bit accuracy (%); “–”: unsupported. Baseline decoders: ( Ma et al., 2022 ; Zhu et al., 2018 ; Tancik et al., 2020 ; Fernandez et al., 2022 ; Zhang et al., 2019 ; Fernandez et al., 2024 ) .
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Condition
exact
mis-assigned
removed entries falsely matched
Δ WM-PSNR / r
bits
A Unmodified file
192/192
0
–
0 / 0
100%
B Rows permuted
192/192
0
–
0 / 0
100%
C 30% lowest opacity deleted, permuted
117/192
0
123 (0.008%)
0.006 dB / 1.5×10−4
100%
D 20% random deleted, permuted
28/192
0
1,189 (0.116%)
0.017 dB / 2.1×10−4
100%
Appendix
Table 5: Geometric re-association performance across all 192 models (150 image, 42 bit payloads). Our centre-based matching algorithm demonstrates mathematically flawless structural recovery: across all adversarial conditions, surviving marked Gaussians were never assigned to an incorrect row (0 mis-assignments). Metrics: Δ denotes the maximum observed change in extraction fidelity (WM-PSNR and Pearson r ) relative to theoretical, perfectly aligned in-place removal. Baseline Alignment: Under Condition C (30% opacity pruning), the geometrically recovered mean r=0.892 and WM-PSNR of 18.21 dB perfectly reproduce the in-place baseline established in Table 2 , proving our re-association module incurs virtually zero mathematical penalty.
Figure 5: Stress-testing geometric re-association under spatial noise. We evaluate the breakdown point of the matcher across all 192 models by injecting isotropic Gaussian noise into the canonical centres. The fixed matching tolerance ( ϵ=0.01snn ) is indicated by the dotted line. (a) Matching Dynamics: The proportion of marked Gaussians that are correctly matched, safely left unmatched, or falsely matched to an incorrect Gaussian. Crucially, the false-match rate remains strictly near zero across all noise levels. (b) Extraction Fidelity: The resulting image-payload Pearson r (mean and minimum over 150 models) and bit accuracy (mean and minimum over 42 runs) following the noisy re-association. Mathematical Bound: The mean image r becomes mathematically undefined beyond σ=0.1snn . At this extreme noise level, every single entry is safely left unmatched, causing the greyscale payload’s recovery to collapse into a flat mean-colour block with exactly zero variance.
Condition
n
min
median
max
decisions at τ
H1 , retrospective
correct key, distributed model
80
0.9910
0.9991
1.0000
80/80 detected
after INT8
80
0.9766
0.9963
0.9995
80/80 detected
after 30% pruning
80
0.8147
0.9103
0.9737
80/80 detected
H0 , retrospective
other payload, same scene
320
− 0.0951
0.0060
0.3170
0 false positives
wrong subset
320
0.0382
0.1417
0.4858
0 false positives
unmarked model (64 defined)
80
0.0285
0.1400
0.3681
0 false positives
Appendix
Table 6: Ownership verification on the held-out evaluation set (16 scenes, 80 models, threshold τ=0.6 ). SHield achieves a perfect 100% detection rate across all true positive ( H1 ) conditions while successfully registering zero false positives across all null ( H0 ) conditions, including 1,600 fresh adversarial trials. Decision Rule: An undefined Pearson r strictly defaults to a non-detection. Statistical Bounds & Transparency: Because multiple trials within a single model inherently share payloads and selections, they are statistically correlated. Consequently, naive per-trial bounds (e.g., 0.19% for the 1,600 fresh nulls) mathematically underestimate the true false-positive rate. By strictly treating models or scenes as independent units, we derive nominal one-sided 95% Clopper–Pearson upper bounds of 3.68% per model ( n=80 ) and 17.07% per scene ( n=16 ). We explicitly report these boundaries, acknowledging their imperfect independence assumptions, to ensure absolute statistical rigor (§ C ).
Figure 6: Cost and capacity of Gaussian parameter groups as watermark carriers. This ablation systematically isolates the rendering penalty of embedding into different parameter spaces, decisively justifying our selection of higher-order SH bands. Ablation Setup: The identical tiled, zero-mean payload is injected into each respective parameter group for the exact same key-selected subset ( K=10% ). To ensure a mathematically fair comparison across disparate parameter ranges, the perturbation is normalized such that its per-slot RMS strictly equals ε times that specific group’s natural standard deviation. (a) Empirical Rendering Cost: PSNR measured against the clean render (where higher indicates a less visually destructive carrier), averaged over five scenes (four D-NeRF, one DyNeRF) with min–max variance shaded. At ε=0.01 , perturbing spatial positions triggers a catastrophic rendering penalty ( 38.7 dB). In stark contrast, the higher-order SH bands achieve a near-lossless 89.2 dB, massively outperforming the DC band ( 76.1 dB) and geometric attributes like scale, rotation, and opacity ( 68 – 73 dB). (b) Embedding Capacity: The total available scalar slots per Gaussian, highlighting the immense capacity advantage ( 45 slots) of the higher-order SH bands.
Figure 7: Defeating targeted SH band erasure via coprime stride interleaving. This figure mathematically demonstrates how our coprime stride prevents the catastrophic erasure of bit payloads when adversaries systematically truncate higher-order SH bands. In our flattened 1D formulation, slot s inherently maps to SH coefficient smodCtot (where Ctot=16 ). (a) Vulnerability vs. Defense: We visualize a 32-bit message embedded into the bouncingballs scene (dark slots denote at least one embedded copy). Naive Tiling: Because Ctot=16 perfectly divides all baseline message lengths ( L∈{32,48,64} ), plain tiling collapses all ≈1,360 redundant copies of a specific bit into a single SH coefficient. Consequently, truncating the model to SH degree 1 (red) deterministically erases every single copy of 24 of the 32 bits. Coprime Stride: By shifting each tile t by tσ (where σ is strictly coprime to both L and Ctot ), SHield uniformly distributes the redundant copies across all 16 coefficients. Under the identical degree 1 truncation attack, the stride merely sacrifices 75% of the copies globally, safely preserving ≈340 redundant copies per bit to cleanly resolve the final sign decision. (b) Empirical Resilience: Validating this mathematical guarantee on bouncingballs : SHield maintains a flawless 100% measured bit accuracy across all truncation levels and message lengths. In stark contrast, a simulated noise-free naive tiling degrades proportionally with band erasure, plunging to near-chance recovery ( 52 – 55% ) when truncated down to the DC band (degree 0).
Image payload wm1 (30 scenes)
All image payloads
Bit payloads (42 runs)
SH degree
Δ PSNR [worst]
r aware
passes aware
passes naive
passes aware (naive)
bit accuracy
margin
PSNR vs. clean ( L=32 )
3
0.00
0.9997
30/30
30/30
150 (150)
100%
0.998
44.40
2
− 0.76 [ − 3.41]
0.749
30/30
19/30
150 (51)
100%
0.561
38.56
1
− 1.36 [ − 4.54]
0.500
0/30
0/30
36 † (0)
100%
0.249
36.69
0
− 2.27 [ − 9.66]
0.251
0/30
0/30
0 (0)
100%
0.062
33.75
Appendix
Table 7: SH-degree truncation: attack cost and survival, by payload type ( α0=0.1 ). Image payload wm1 ( 30 scenes): Δ PSNR against the untruncated render (mean [worst scene]); mean r and threshold passes with truncation-aware extraction (same bands zeroed in the reference) and naive extraction. All five payloads : passes of 150 models, aware (naive); † the 36 passes at degree 1 are 30 wm2 and 6 wm3, both accounted for by the tiling index map and, for wm2, its high null floor (Appendix E ); none establishes image recovery. Bit payloads : 42 runs, accuracy, decoding margin (1 = unattacked) and PSNR of the 14 truncated L=32 models against the untruncated clean render. Per dataset: Appendix F .
Dataset
degree
Δ PSNR (dB)
r aware / naive
passes aware / naive
bit models: PSNR (dB)
D-NeRF
2
− 0.523
0.749 / 0.628
7/7 / 5/7
43.24
1
− 1.142
0.500 / 0.302
0/7 / 0/7
40.89
0
− 2.410
0.251 / 0.076
0/7 / 0/7
37.40
HyperNeRF
2
− 0.243
0.749 / 0.619
17/17 / 10/17
–
1
− 0.785
0.500 / 0.300
0/17 / 0/17
–
0
− 1.420
0.251 / 0.079
0/17 / 0/17
–
Appendix
Table 8: Dataset-level breakdown of robustness against SH-degree truncation ( α0=0.1 ). This detailed ablation isolates the extraction fidelity and rendering penalties across D-NeRF, HyperNeRF, and DyNeRF under progressive SH band erasure. Image Payloads (wm1): We report the rendering penalty ( Δ PSNR) against ground truth relative to the untruncated render, alongside the Pearson r and threshold passes for both truncation-aware and naive extraction. Artifact Isolation: Crucially, when an unmarked clean model is truncated in the exact same manner, its rendering score falls within a negligible 0.07 dB of the watermarked model. This mathematically proves that the observed visual degradation is fundamentally caused by the loss of higher-order SH detail, rather than the adversarial exposure of watermarking artifacts. Bit Payloads: We report the PSNR of the truncated L=32 models evaluated against the clean untruncated render (using Mark4D test views). Empowered by the coprime stride, bit decoding accuracy remains a flawless 100% on every single run across all datasets and all truncation degrees.
Scene / Split
PSNR clean
PSNR wm
Δ PSNR
SSIM wm
LPIPS wm
WM-PSNR
Pearson r
D-NeRF (7 synthetic monocular scenes)
Bouncingballs
40.378
40.265
− 0.113
0.9940
0.0112
41.39
0.9994
Hellwarrior
28.731
28.709
− 0.022
0.9720
0.0296
42.54
0.9995
Hook
32.774
32.742
− 0.032
0.9760
0.0201
45.29
0.9998
Jumpingjacks
35.497
35.462
− 0.035
0.9855
0.0159
45.20
0.9998
Mutant
37.523
37.458
− 0.065
0.9875
0.0106
46.28
0.9998
Appendix
Table 9: Comprehensive per-scene evaluation of SHield ( α0=0.1 ). This granular breakdown demonstrates SHield’s exceptional consistency, achieving near-perfect payload recovery (Pearson r≥0.997 ) across all individual scenes and datasets regardless of topological complexity. Aggregation Setup: Each row reports the mean performance across all five distinct payloads; HyperNeRF rows further aggregate the scenes within each respective capture split. Metric Isolation: Absolute metrics (PSNR, SSIM, and LPIPS) are measured against ground-truth images, primarily reflecting the natural reconstruction difficulty of each specific scene. Crucially, Δ PSNR isolates the pure mathematical rendering penalty incurred by the watermark, cleanly separating our defense’s visual footprint from pre-existing baseline artifacts.
Figure 8: Visual payload recovery under worst-case rendering conditions. This figure qualitatively demonstrates that SHield successfully reconstructs highly recognizable images across a massive range of payload capacities, even on the dataset’s most challenging scene. Worst-Case Evaluation: We deliberately visualize extractions from bouncingballs , which represents the mathematical worst-case D-NeRF scene for Watermark PSNR ( 41.4 dB, compared to the 44.8 dB dataset average). Each payload is recovered directly from the distributed .ply utilizing the registered key. Error Metric: The Error column visualizes the absolute per-pixel RMS difference across RGB channels between the ground-truth payload and the reconstructed extraction, normalized to a single shared intensity scale for fair comparison. Capacity vs. Redundancy: The spatial resolution of the payload strictly dictates its spread-spectrum redundancy. The two compact 50k-element payloads securely embed two redundant copies per element, whereas the three massive 217k-element payloads are constrained to a single copy, directly accounting for the observed spread in WM-PSNR. Furthermore, the high-contrast, flat text payload naturally recovers 3 – 4 dB higher than the dense, high-frequency images of identical size.
D-NeRF (7)
HyperNeRF (17)
DyNeRF (6)
Payload ( H×W )
Content
WM-PSNR
r
WM-PSNR
r
WM-PSNR
r
η2
wm1 ( 128×394 )
ICLR logo, flat colour
47.81
0.9999
42.02
0.9996
42.48
0.9997
0.029
wm2 ( 194×259 )
Road/bridge at dusk
48.61
0.9999
42.28
0.9995
42.56
0.9996
0.368
wm3 ( 350×621 )
Bold “SMILE” on white
44.02
0.9998
32.41
0.9963
32.97
0.9972
–
wm4 ( 386×563 )
Detailed colour poster
41.70
0.9995
32.59
0.9956
32.85
0.9961
0.214
wm5 ( 350×621 )
Anime group scene
41.71
0.9995
32.06
0.9955
32.10
0.9959
0.067
Appendix
Table 10: Extraction fidelity across diverse payload capacities. This ablation demonstrates that SHield maintains near-perfect payload recovery ( r≥0.995 ) regardless of the semantic content, spatial resolution, or complexity of the embedded 2D image. Evaluation Scope: Mean Watermark PSNR (WM-PSNR) and Pearson correlation r evaluated at the standard operating point ( α0=0.1 ) across all 150 models. The Capacity-Redundancy Trade-off: The mathematically lower WM-PSNR observed for the three massive 217k-element payloads (wm3, wm4, wm5) is a direct consequence of tile repetition limits. Because our adaptive selection rule strictly caps the marked subset K at ρmaxN to preserve rendering fidelity, these larger payloads permit only nt=1 redundant copy per element on sparser datasets like D-NeRF, compared to nt=2 copies for the compact 50k-element payloads (wm1, wm2). The Unmarked Null ( η2 ): We explicitly report the payload-dependent null floor, the theoretical correlation of a flat, per-channel mean-colour recovery against the target payload, as formally established in our baseline analysis (§ C ).
Scene
wm1
wm2
wm3
wm4
wm5
ratio
coffee_martini
− 0.067 / − 0.061
− 0.066 / − 0.076
− 0.268 / − 0.184
− 0.152 / − 0.146
− 0.252 / − 0.146
1.31 ×
cook_spinach
− 0.111 / − 0.106
− 0.098 / − 0.076
− 1.084 / − 0.395
− 0.539 / − 0.316
− 0.557 / − 0.337
1.94 ×
cut_roasted_beef
− 0.220 / − 0.009
− 0.041 / + 0.037
− 0.593 / − 0.092
− 0.728 / − 0.037
− 0.549 / − 0.045
9.64 ×
flame_salmon_1
− 0.087 / − 0.017
− 0.062 / − 0.020
− 0.286 / − 0.037
− 0.176 / − 0.006
− 0.491 / − 0.055
8.14 ×
flame_steak
− 0.130 / − 0.026
− 0.083 / − 0.011
− 0.862 / − 0.031
− 0.457 / + 0.010
− 0.705 / − 0.062
15.96 ×
sear_steak
− 0.262 / − 0.181
− 0.138 / − 0.133
− 1.257 / − 0.462
− 0.756 / − 0.388
− 0.667 / − 0.403
1.96 ×
Appendix
Table 11: Empirical justification of anchored optimization on DyNeRF ( α0=0.1 ). This pairwise ablation demonstrates that SHield’s optimization framework strictly dominates naive plain tiling, drastically reducing the visual rendering penalty across diverse dynamic scenes and payload capacities. Metric Format: Evaluated under identical subset selection and testing protocols, we report the rendering penalty ( Δ PSNR in dB) for each scene–payload combination formatted as: naive tiling / SHield optimization . Performance Multiplier: The final column aggregates the per-scene improvement ratio of the mean absolute rendering penalty ( ∣Δ PSNR ∣ ). Crucially, the anchored optimization yields massive visual fidelity improvements, suppressing the watermark’s visual footprint by up to 15.96× on highly complex scenes like flame_steak and improving 29 of the 30 scene–payload pairs; coffee_martini /wm2 is the single exception ( −0.076 against −0.066 dB).
α0
Δ PSNR
WM-PSNR
Pearson r
0.01
− 0.011
26.82
0.9770
0.05
− 0.022
36.16
0.9969
0.10
− 0.063
38.32
0.9980
0.20
− 0.250
40.17
0.9986
0.50
− 1.377
43.07
0.9994
Appendix
Table 12: Ablation of the perturbation scale ( α0 ) on image payloads. This ablation maps the fundamental trade-off between visual imperceptibility and extraction fidelity across all 150 evaluated models. Optimization Target: The hyperparameter α0 establishes both the initial injection magnitude and the strict anchoring target for the watermarking loss ( Lwm ). Trade-off Dynamics: As the perturbation scale increases, extraction fidelity (WM-PSNR and Pearson r ) naturally strengthens, but at the direct expense of the rendering penalty ( Δ PSNR). Selected Operating Point: The shaded row denotes our standard operating point ( α0=0.10 ), deliberately chosen as the optimal equilibrium. It guarantees near-perfect extraction ( r=0.9980 ) while bounding the average visual penalty to a highly stable and mathematically minimal −0.063 dB.
Metric
hook (D-NeRF)
tamping (HyperNeRF)
flame_salmon_1 (DyNeRF)
Image
Δ PSNR (dB)
−0.0226±0.0034
−0.0145±0.0027
−0.0038±0.0135 [ −0.0171 , +0.0152 ]
WM-PSNR (dB)
48.69±0.21
44.15±0.52
43.84±0.23
Pearson r
0.99994±0.00000
0.99981±0.00003
0.99979±0.00001
r after 30% pruning
0.9331±0.0024
0.9604±0.0014
0.9512±0.0011
r after 80% pruning
0.5951±0.0107 [ 0.5875 , 0.6133 ]
0.7550±0.0080
0.7071±0.0049
r at SH degree 2 (aware)
0.7490±0.0000
0.7489±0.0000
0.7489±0.0000
Appendix
Table 13: Algorithmic stability across random seed initializations. This ablation confirms that SHield’s performance is fundamentally grounded in its rigorous mathematical design (anchored optimization and coprime stride) rather than a statistically fortunate initialization. Metric Format: We evaluate five independent seeds ( 42,1,2,3,4 ), reporting the mean ± standard deviation, alongside specific [min, max] bounds where relevant. Image Payloads (wm1): Across three representative scenes, we observe exceptionally tight variance in both baseline rendering penalties ( Δ PSNR) and extraction fidelity under the severe post-processing attacks established in Table 2 . This strictly proves that our optimizer stably converges to an optimal equilibrium regardless of the specific Gaussian subset initially selected by the key. Bit Payloads ( L=32 ): Evaluated under the Mark4D comparison framework. Strikingly, across the entire suite of extreme adversarial attacks detailed in Appendix J (including INT8 quantization, up to 80% opacity pruning, catastrophic truncation down to SH degree 0, 5α0 parameter noise, and geometric distortion), SHield maintains a flawless 100% decoding accuracy on every single seed .
α0
bit acc.
PSNR (dB)
SSIM
LPIPS
per-slot RMS
file RMS
energy
σ=0.3
σ=1.0
0.01
100%
62.48 [56.99, 67.40]
0.99994
0.00006
0.00998
0.00316
1.2%
72.4%
57.8%
0.05
100%
50.32 [47.81, 53.63]
0.99875
0.00095
0.04988
0.01577
6.0%
99.2%
79.9%
0.10
100%
44.29 [41.57, 47.62]
0.99564
0.00373
0.09981
0.03156
12.0%
100%
95.1%
Appendix
Table 14: Ablation of the perturbation scale ( α0 ) for bit payloads. This ablation rigorously quantifies the fundamental trade-off between absolute visual imperceptibility and resilience against catastrophic structural distortion. Evaluation Scope: Aggregated across 42 independent runs (spanning 8 D-NeRF and 6 DyNeRF scenes, with message lengths L∈{32,48,64} at seed 42). Visual and Physical Footprint: We report standard rendering metrics (PSNR, SSIM, LPIPS) against the original clean renders following the Mark4D protocol. To explicitly bound the physical magnitude of the modification, we additionally track the absolute per-slot RMS across all SH entries and the total relative perturbation energy ( ∥Δh∥/∥h∥ ). Extreme Stress Testing: The rightmost columns evaluate extraction under our aggressive Mark4D-style distortion, which simultaneously injects massive absolute SH noise ( σ ) while catastrophically pruning 95% of the Gaussians. (Note: Under milder σ≤0.1 noise, all configurations successfully decode at 100% with the sole exception of α0=0.01 , which drops to 95.1% ). Operating Point Justification: While the weakest perturbation ( α0=0.01 ) achieves an astonishingly invisible 62.48 dB PSNR, its robustness fractures under heavy noise (plunging to 57.8% accuracy at σ=1.0 ). By contrast, our selected operating point ( α0=0.10 , shaded) deliberately accepts a mild visual penalty to establish an incredibly durable payload—maintaining flawless 100% decoding up to σ=0.3 , and a highly robust 95.1% even when completely submerged in extreme σ=1.0 noise.
Hide-in-Motion (ICRA 2025)
SHield
Representation
E-D3DGS with an extra composite per-Gaussian attribute that also drives the deformation
unchanged 4DGS model
When embedded
jointly during reconstruction training
post hoc on a trained model
Carrier
learned composite feature
existing canonical SH of key-selected Gaussians
Extraction
feature map rendered at a check viewpoint, private U-Net decoder
closed-form arithmetic against stored reference SH; no rendering or learned decoder
Payloads
2D images; text, QR codes, audio, video via modality-specific decoders
RGB images, bit messages
Attacks evaluated
none reported
quantization, reordering, removal, SH truncation, SH noise (bits), centre perturbation
Appendix
Table 15: Architectural and security paradigm comparison with Hide-in-Motion. This conceptual breakdown contrasts SHield against the recent Hide-in-Motion (ICRA 2025) baseline, highlighting fundamental differences in embedding mechanics, extraction requirements, and adversarial rigor. Architectural Simplicity: Hide-in-Motion requires altering the underlying 3DGS data structure (adding new composite attributes) and must embed the watermark jointly during the initial, expensive scene reconstruction. In stark contrast, SHield operates strictly post hoc on a fully trained, completely unmodified 4DGS asset. Extraction Mechanics: Hide-in-Motion relies on rendering customized feature maps and passing them through a private, learned U-Net decoder. SHield completely eliminates the need for rendering or proprietary network architectures, instead executing mathematically exact closed-form arithmetic directly against the static parameters. Adversarial Rigor: While Hide-in-Motion supports exotic multimodal payloads (audio, video), it reports no evaluation against adversarial post-processing. SHield deliberately trades modality diversity for robust security, providing a rigorously stress-tested defense against severe white-box file manipulations.
Dynamic rendering
Recovery
Method
Base
PSNR
SSIM
LPIPS
Δ PSNR
PSNR
SSIM
E-D3DGS ( Bae et al., 2024 )
–
25.641
0.714
0.283
–
–
–
Hide-in-Motion
E-D3DGS
25.573
0.709
0.297
− 0.068
50.341
0.998
4DGS ( Wu et al., 2024 )
–
25.000
0.676
0.394
–
–
–
SHield (ours)
4DGS
24.950
0.674
0.395
− 0.049
36.345
0.921
Appendix
Table 16: Quantitative reference comparison against Hide-in-Motion on HyperNeRF. This comparison isolates the specific visual penalty of watermarking across differing 3D representation paradigms. Evaluation Protocol: Hide-in-Motion and its E-D3DGS baseline metrics are quoted directly from Liu et al. (2025) . SHield and its unmodified 4DGS baseline metrics are our empirical measurements (averaged across the four shared scenes and five payloads at α0=0.1 , evaluating the first 20 test views). Relative Visual Penalty ( Δ PSNR): Absolute rendering metrics (PSNR, SSIM, LPIPS) against ground truth inherently reflect the varying baseline reconstruction capacities of E-D3DGS versus 4DGS. Crucially, Δ PSNR cleanly isolates the specific rendering cost imposed by the watermark itself. Here, SHield imposes a strictly lighter visual footprint ( −0.049 dB) than Hide-in-Motion ( −0.068 dB). Extraction Incomparability: As established in Table 15 , the recovery metrics are fundamentally incomparable; Hide-in-Motion leverages a heavy, learned U-Net decoder to reconstruct payloads from rendered feature maps (artificially inflating raw PSNR), whereas SHield strictly reports the fidelity of closed-form, parameter-space arithmetic without neural post-processing.
Dynamic rendering
Recovery
Method
Base
PSNR
SSIM
LPIPS
Δ PSNR
PSNR
SSIM
E-D3DGS ( Bae et al., 2024 )
–
31.500
0.938
0.149
–
–
–
Hide-in-Motion
E-D3DGS
31.471
0.935
0.153
− 0.029
50.075
0.998
4DGS ( Wu et al., 2024 )
–
31.192
0.936
0.109
–
–
–
SHield (ours)
4DGS
31.065
0.935
0.110
− 0.127
36.591
0.927
Appendix
Table 17: Quantitative reference comparison against Hide-in-Motion on DyNeRF. This evaluation extends our comparative analysis to the highly dynamic DyNeRF (Neural 3D Video) dataset, contrasting SHield’s parameter-space approach against Hide-in-Motion’s rendering-based paradigm. Evaluation Protocol: As established in Table 16 , baseline and Hide-in-Motion metrics are quoted directly from Liu et al. (2025) . SHield metrics represent our exact empirical measurements averaged across all six shared scenes and five distinct payloads. Relative Visual Penalty ( Δ PSNR): While Hide-in-Motion reports a marginally smaller rendering penalty on this specific dataset ( −0.029 dB vs. SHield’s −0.127 dB), SHield achieves this highly competitive, near-lossless visual fidelity entirely post hoc . We secure the model without structurally altering the underlying 3D representation or forcing an expensive joint reconstruction training phase. Extraction Incomparability: We strictly reiterate that recovery metrics are fundamentally incomparable; Hide-in-Motion’s recovery PSNR is artificially elevated by passing rendered feature maps through a proprietary, learned neural decoder, whereas SHield guarantees recovery via rigorous, closed-form arithmetic directly in the parameter space.
Figure 9: Temporal stability and time-independent extraction. This figure demonstrates that SHield successfully avoids the temporal flickering artifacts common to render-domain watermarks, while completely decoupling the extraction process from the time dimension. Qualitative Consistency (Top): Rendered frames spanning the watermarked hook sequence, demonstrating smooth, artifact-free temporal dynamics. Quantitative Temporal Stability (Bottom Left): We track the PSNR between the watermarked and clean renders across 60 sequential timestamps from a fixed camera viewpoint (evaluated on the median scene of each dataset). The exceptionally tight standard deviation ( 0.015 – 0.78 dB around means of 48 – 55 dB) mathematically proves the absence of high-frequency temporal flickering. Crucially, this stability holds even for DyNeRF, where the underlying deformation network actively injects time-dependent SH offsets during rendering. Time-Invariant Extraction (Bottom Right): The recovered image payload. Because SHield securely embeds the signal directly into the canonical parameter space, extraction is executed exactly once from the static .ply file and is fundamentally independent of the continuous time variable t .
D-NeRF (7)
HyperNeRF (17)
DyNeRF (6)
Method
View type
WM-PSNR
r
WM-PSNR
r
WM-PSNR
r
2D per-frame
Same-view (training)
14.34
0.917
15.28
0.943
15.30
0.943
2D per-frame
Novel-view (test)
6.49
0.010
7.78
0.000
8.30
0.011
SHield (ours)
model file
44.77
0.9997
36.27
0.9973
36.59
0.9977
Appendix
Table 18: The catastrophic novel-view failure of post-hoc 2D frame watermarking. This baseline comparison mathematically demonstrates why applying standard 2D image watermarks to training frames fails to secure underlying 3D/4D representations. Extraction Paradigms: The 2D baseline injects watermarks directly into the training images; “same-view” extraction is evaluated on these explicitly watermarked training cameras, while “novel-view” extraction targets unseen test cameras. View-Synthesis Vulnerability: While the 2D baseline achieves marginal recovery on explicitly memorized training views ( r≈0.92 – 0.94 ), it completely collapses during novel-view synthesis (Pearson r plunging to near-zero noise floors of 0.000 – 0.011 ). Because 4DGS inherently entangles geometry and view-dependent colour, naive 2D signals do not structurally propagate to novel camera trajectories. The Parameter-Space Advantage: In stark contrast, SHield extracts directly from the static model file. By embedding directly into the canonical SH parameters, SHield renders the extraction process fundamentally view-agnostic, guaranteeing near-perfect recovery ( r≥0.997 ) completely independent of the rendered camera path. Evaluation Scope: Averages are reported across the full suite of 30 scenes and 5 payloads.
Attribute
D-NeRF (7)
HyperNeRF (17)
DyNeRF (6)
SH DC ( fdc )
0.0084
0.0083
0.0085
SH rest ( frest )
0.0191
0.0123
0.0088
SH combined
0.0086
0.0083
0.0093
Appendix
Table 19: Statistical imperceptibility of the parameter distribution. This analysis proves that SHield’s parameter-space injection does not introduce detectable statistical anomalies into the canonical 4DGS asset, effectively thwarting distribution-based steganalysis. Distributional Stealth: We quantify the global distributional shift by computing the 3D Sinkhorn distance (optimal transport cost) between the clean and watermarked spherical harmonic (SH) attributes. Across all 150 models evaluated at the α0=0.1 operating point, the distributional deviation is microscopic (strictly <0.02 ). This confirms that the watermarked SH coefficients remain statistically indistinguishable from their natural distribution. Geometric Isolation: SHield writes only into the SH coefficients, the DC band included: that band holds 1/Ctot of the tiled payload and carries its own learned offset Δdc , which is why it appears as a separate row above. The structural attributes (positions, opacities, scales and rotations) are untouched by construction, incur exactly zero distributional penalty, and are omitted from the table.
Scene
params
steps
Δ PSNR
r re-assoc.
r oracle
matched (%)
verifies
bouncingballs
sh
100
+ 0.04
0.9808
0.9808
100.0
yes
bouncingballs
sh
500
− 0.03
0.9257
0.9257
100.0
yes
bouncingballs
sh
2,000
− 0.26
0.8035
0.8035
100.0
yes
bouncingballs
sh
5,000
− 0.59
0.6202
0.6202
100.0
yes
bouncingballs
sh
10,000
− 0.98
0.4782
0.4782
100.0
no
bouncingballs
sh
20,000
− 1.51
0.3701
0.3701
100.0
no
Appendix
Table 20: Adaptive fine-tuning attack, per run (wm1, α0=0.1 ).
Attack
Matcher
Matched (%)
r
Passes
Duplicate 10%
shipped
89.96
0.9795
30/30
lowest-numbered
100.00
0.9997
30/30
Duplicate 50%
shipped
49.91
0.8544
30/30
lowest-numbered
100.00
0.9997
30/30
Duplicate 100%
shipped
0.00
0.0285
0/30
lowest-numbered
100.00
0.9997
30/30
Appendix
Table 21: Attacks on key re-association (wm1, α0=0.1 , 30 models). Matched is pooled over key entries; r is the mean over models; passes at τ=0.6 . Shipped : the matcher of the main results; lowest-numbered : ambiguous entries take the lowest-numbered bit-identical candidate; aligned : the suspect file is first aligned to the original model.
While 4D Gaussian Splatting (4DGS) has revolutionized high-fidelity dynamic reconstruction, safeguarding the intellectual property of these assets remains an open challenge. Conventional steganographic techniques often neglect the underlying kinematic manifolds, triggering non-physical artifacts such as severe temporal flickering and "FVD collapse". To address this, we propose 4D-GSW, a kinematic-aware watermarking framework designed to embed robust copyright information while preserving high spatio-temporal consistency. Unlike prior 4D steganography that primarily focuses on opacity-guided invisibility, our approach explicitly addresses the physical coherence of motion trajectories. We introduce a Spatio-Temporal Curvature (STC) metric to identify "Dynamic Instants," adaptively gating watermark gradient injection to shield critical motion manifolds from non-physical perturbations. To ensure global coherence across complex deformations, we formulate a joint HMM-MRF energy minimization model that synchronizes watermark phases within both temporal trajectories and spatial neighborhoods. Furthermore, an anisotropic gradient routing mechanism ensures that watermark embedding remains strictly decoupled from photometric reconstruction fidelity. Extensive experiments have demonstrated the superior performance of our method in robustly hiding watermarks while resisting various attacks and maintaining high rendering quality and spatiotemporal consistency.
Sifan Zhou, Hang Zhang, Yuhang Wang +1
Southeast University · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) · University of Chinese Academy of Sciences
As 3D Gaussian Splatting becomes the de facto representation for interactive 3D assets, robust yet imperceptible watermarking is critical. We present a representation-native framework that separates where to write from how to preserve quality. A Trio-Experts module operates directly on Gaussian primitives to derive priors for carrier selection, while a Safety and Budget Aware Gate (SBAG) allocates Gaussians to watermark carriers, optimized for bit resilience under perturbation and bitrate budgets, and to visual compensators that are insulated from watermark loss. To maintain fidelity, we introduce a channel-wise group mask that controls gradient propagation for carriers and compensators, thereby limiting Gaussian parameter updates, repairing local artifacts, and preserving high-frequency details without increasing runtime. Our design yields view-consistent watermark persistence and strong robustness against common image distortions such as compression and noise, while achieving a favorable robustness-quality trade-off compared with prior methods. In addition, decoupled finetuning provides per-Gaussian attributions that reveal where the message is carried and why those carriers are selected, enabling auditable explainability. Compared with state-of-the-art methods, our approach achieves a PSNR improvement of +0.83 dB and a bit-accuracy gain of +1.24%.
Mingshu Cai, Jiajun Li, Osamu Yoshie +2
Waseda University · Southeast University · Nanyang Technological University
With the rapid development and adoption of 3D Gaussian Splatting (3DGS), the need for effective copyright protection has become increasingly critical. Existing watermarking techniques for 3DGS mainly focus on protecting rendered images via pre-trained decoders, leaving the underlying 3D Gaussian primitives vulnerable to misuse. In particular, they are ineffective against Partial Infringement, where an adversary extracts and reuses only a subset of Gaussians. In this paper, we propose NGS-Marker, a novel native watermarking framework for 3DGS. It integrates a jointly trained watermark injector and message decoder, and employs a gradientbased progressive injection strategy to ensure full-scene coverage. This enables robust ownership decoding from any local region. We further extend NGS-Marker with hybrid protection (combining native and indirect watermarks) and support for multimodal watermarking. Extensive experiments demonstrate that NGS-Marker effectively defends against partial infringement while offering practical flexibility for real-world deployment.
Hao Qin, Yukai Sun, Luyuan Chen +5
Zhejiang University · Zhejiang Key Laboratory of Geographic Information Science