acceptodds
Under review as a conference paper at ICLR 2027

FlyMark: Training-Free Invisible Watermarking of 3D Gaussian Splatting via a Fruit Fly Connectome

Abstract

A trained 3D Gaussian Splatting (3DGS) scene ships as a portable parameter array that can be copied, pruned, requantized, or repackaged outside its training pipeline, so ownership evidence is most useful when it lives in the released parameters and remains checkable long after the embedding tooling is gone. Existing 3DGS watermarks typically tie embedding or extraction to scene optimization, a learned decoder, or rendered views, so the evidence survives only as long as a second trained artifact does. FlyMark instead writes a keyed message into the parameters a 3DGS file already stores. Its carrier directions are derived from the photoreceptors of a published connectome, a citable versioned artifact that fixes the geometry exhaustively and leaves nothing to tune per scene. A virtual observer reads cone-wise apparent luminance along a scene-normalized orbit from stored centers, colors, and opacities; a keyed dithered quantization-index-modulation code replicates each message bit across these observations; and one sparse bounded least-squares solve realizes the targets through achromatic shifts of existing degree-zero colors under a hard per-channel linear-RGB bound. All geometry and higher-order appearance parameters are preserved bit-identically, and extraction needs only cone queries, rounding, and majority voting. Under a model-domain threat model on synthetic and real scenes, FlyMark attains high clean bit accuracy and visual fidelity while cleanly separating matched from wrong keys.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
3D Gaussian Splatting, Digital Watermarking, Fruit Fly Connectome

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_xjxGOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes FLYMARK, a training-free watermarking framework for 3D Gaussian Splatting (3DGS) scenes that embeds ownership evidence directly into the stored scene parameters rather than relying on a separately trained encoder, decoder, or rendered views. The method uses the photoreceptor layout of a published fly connectome as a fixed, public, and reproducible carrier geometry. A virtual observer orbits the scene and computes cone-wise apparent luminance from the stored Gaussian centers, colors, and opacities. A keyed, dithered quantization-index modulation scheme redundantly encodes message bits across these observations, while majority voting enables deterministic extraction without a learned decoder.

To realize the watermark, FLYMARK formulates a sparse bounded least-squares problem that modifies only the existing degree-zero color parameters under hard per-channel linear-RGB constraints. The method preserves Gaussian positions, point counts, covariances, opacities, and higher-order spherical harmonics bit-identically. The authors evaluate FLYMARK on synthetic and real 3DGS scenes using a configuration fixed before evaluation, measuring clean bit recovery, visual fidelity, robustness to model-domain transformations, and separation between matched and wrong keys. The reported results demonstrate high bit accuracy and visual fidelity, robustness to the evaluated transformations, and effective wrong-key separation. An additional natural-image experiment is conducted to show that the retinal carrier and keyed QIM design can also support bounded watermarking of ordinary 2D images.

Critical strengths
  1. Training-independent and potentially long-term verifiable watermarking. A key strength is that the watermark is embedded directly into the released 3DGS parameters and can be extracted without a separately trained encoder or decoder. This makes the ownership evidence less dependent on specific training artifacts or proprietary embedding models, potentially allowing the watermark to remain verifiable over the long term and to be integrated with existing 3DGS training and asset-generation pipelines.

  2. Effective parameter-space design with strong preservation properties. The proposed bounded least-squares formulation modifies only the degree-zero color parameters while preserving geometry, opacities, covariances, and higher-order appearance parameters bit-identically. This provides a practical balance between watermark embedding, visual fidelity, and preservation of the original 3DGS representation.

Critical weaknesses
  1. The experimental evaluation appears incomplete. Lines 280–283 state that the method is evaluated on eight NeRF-Synthetic scenes and nine Mip-NeRF 360 scenes, but the main paper provides only three qualitative examples in Fig. 3 (Lego, Hotdog, and Bonsai). No additional qualitative results are provided, and I could not find supplementary materials. More representative visual results are needed to support the claims and demonstrate the method's generalization.

  2. The ablation study needs to be more comprehensive. Please provide ablation results for each major design component, including cone activation, keyed QIM, and parity voting. For example, why is cone activation preferable to alternative activation strategies? Why is keyed QIM used instead of other coding mechanisms, and what is the performance improvement? Similarly, what is the contribution of parity voting? These ablations should include both quantitative and qualitative results, together with an analysis of model complexity and runtime.

  3. The claim that the watermark preserves 3DGS geometry requires stronger evidence. Although the abstract claims that the method preserves the geometry of the Gaussian Splatting representation, the current evaluation appears to focus mainly on appearance-based metrics. Please provide geometric evaluation, such as reconstructed geometry compared with the ground truth or rendered depth comparisons and quantitative depth/geometric metrics, to substantiate this claim.

  4. Several important technical details remain unclear. For example, the paper should explain the motivation and implementation of repeated readout, the purpose and mechanism of parity voting, and how cone activation and keyed QIM are defined and implemented. More details about the relevant feature representations, dimensions, and tensor shapes would also improve the reproducibility and clarity of the method.

  5. Robustness and practical deployment need further clarification. How robust is the proposed watermark to noise or perturbations applied to the Gaussian Splatting parameters? If it is not robust to certain perturbations, the corresponding limitations and failure cases should be clearly discussed. In addition, please clarify whether watermark embedding and extraction must be performed jointly with the Gaussian Splatting training of a specific scene, or whether the method can be directly applied to an already-trained 3DGS scene without retraining.

  6. Evaluation on more complex 3DGS scenes would strengthen the generalization claim. It would be valuable to evaluate the method on larger and more complex scenes, such as Tanks and Temples or larger Mip-NeRF 360 scenes, including challenging indoor and outdoor environments. This would provide stronger evidence that the proposed watermarking method generalizes beyond relatively small or simple scenes.

Questions for authors
  1. Can the authors clarify whether the reported evaluation is complete? The paper states that eight NeRF-Synthetic and nine Mip-NeRF 360 scenes are evaluated, but only three qualitative examples are shown in the main paper. Are additional qualitative results or supplementary materials available that demonstrate the method across the full set of evaluated scenes?

  2. What is the contribution of each major design component to the final performance? In particular, how much do cone activation, keyed QIM, repeated readout, and parity voting contribute individually? Are there ablations comparing these choices with reasonable alternatives?

  3. Can the authors provide evidence supporting the claim that the proposed method preserves 3DGS geometry? Since the current evaluation appears to focus mainly on appearance-based metrics, how does the geometry or depth of the watermarked scene compare with the original scene quantitatively and qualitatively?

  4. Can the authors clarify the practical deployment setting of FLYMARK? Can the method be applied directly to an already-trained 3DGS scene by embedding and subsequently extracting the watermark, or must watermark embedding be integrated into the scene-specific Gaussian Splatting training process?

  5. How robust is the watermark to parameter-space perturbations or common modifications of a 3DGS model? In particular, what types and magnitudes of noise, pruning, quantization, or other model transformations can the watermark tolerate before reliable extraction fails?

10 comments
AuthorsOpenReview ↗

Response to Reviewer xjxG (1/4)

We thank the reviewer for the detailed feedback and suggestions for clarifying the evaluation and method.

We have uploaded the revised manuscript, with changes marked in red. Below, we respond to each question.

W1 / Q1. Evaluation coverage and qualitative results

“Can the authors clarify whether the reported evaluation is complete and whether additional qualitative results or supplementary materials are available that demonstrate the method across the full set of evaluated scenes?”

Original Table 2 and Section B.1, Table 6 report held-out Synthetic and all nine Mip-NeRF 360 scenes, with their evaluation roles. The original submission also includes three registered views each of Lego, Hotdog and Bonsai in original Section H, Figures 7–9, beyond the examples in Figure 3. The Supplementary Materials contain the implementation and reproduction README.

To make full visual coverage easy to inspect, revised Section H, Figures 12–17 now adds all 23 evaluated scenes: eight NeRF Synthetic, nine Mip-NeRF 360 and six Tanks and Temples scenes, each with a reference, marked output and amplified difference. Cameras and crops match within each comparison, with bit accuracy and displayed-view PSNR annotated. These figures are integrated into the revised PDF.

AuthorsOpenReview ↗

Response to Reviewer xjxG (2/4)

W2 / Q2. Component contributions, alternatives and cost

1) “why is cone activation preferable to alternative activation strategies?”

Cone activation provides a local angular measurement that pools multiple Gaussians rather than selecting a single element. Dividing by the cone membership count makes it an average rather than a sum that grows directly with that count. Most importantly, with memberships and opacities fixed, the mean opacity-weighted luminance is linear in the Gaussian color shifts. This gives the sparse linear measurement operator used by the bounded color solve in original Sections 3.3–3.5 and C.3. The same functional can be recomputed directly from stored model parameters, without rendering or a learned decoder.

Original Table 1 compares a parameter-space alternative, SpatialHash-QIM, which replaces cone grouping with keyed spatial assignment under the same message length and color budget.

2) “What is the contribution of each major design component to the final performance? In particular, how much do cone activation, keyed QIM, repeated readout, and parity voting contribute individually? Are there ablations comparing these choices with reasonable alternatives?”

Original Table 4 reports the following component comparisons on Lego, Chair, Ficus and Materials with key 17. All variants below use the same 0.02 per-channel linear-RGB edit bound; the measured fidelity is shown alongside bit accuracy.

VariantBit accuracyDirect PSNR (dB)
Full method100.00%46.02
One orbit stop89.06%50.13
No keyed dither91.41%46.29

Cone activation. Across ten scenes and three keys in original Table 1, cone activation gives 99.27% bit accuracy and 44.21 dB direct PSNR, versus 83.12% and 44.41 dB for keyed spatial assignment.

Keyed QIM. The no-keyed-dither control in original Table 4 isolates the dither component of keyed QIM: bit accuracy decreases from 100.00% to 91.41%, a difference of 8.59 percentage points. For an alternative coding mechanism, revised Section F.6, Table 19 matches the carrier, orbit, eligibility, bounded solve, color budget and RMS target displacement for QIM and blind additive spread spectrum. The control follows the standard SS formulation reviewed in Section II of [r3] and decodes by centered correlation of received readings with keyed signs. Neither decoder requires unmarked luminance readings. Clean bit accuracy is 100.00% for FlyMark and 70.31% for the blind additive control. On these four scenes, mean bit-accuracy Macro-AUC over the ten registered attack families is 0.8978 for FlyMark and 0.6895 for blind additive SS. Full-test direct PSNR is 46.02 dB and 45.93 dB, respectively.

Repeated readout. Extraction repeats the cone-activation computation on the received model. The orbit-stop ablation in original Table 4 tests the contribution of collecting these observations at multiple orbit locations: reducing the orbit from 24 stops to one lowers bit accuracy from 100.00% to 89.06%, a difference of 10.94 percentage points.

Parity voting. Each observation's integer parity gives a binary decision, and majority aggregation combines the repeated decisions. The same-marked-asset control in revised Section F.6, Table 19 compares one uniformly selected eligible observation per bit with all-vote majority aggregation. Expected bit accuracy is 64.84% with a single observation and 100.00% with majority aggregation, a gain of 35.16 percentage points. This comparison changes only decoding, not the marked model.

3) “These ablations should include both quantitative and qualitative results”

Revised Section F.5, Figure 7 shows matched outputs and equal-size amplified differences for the full, one-stop, no-dither and spatial-assignment variants on all four component-study scenes. Revised Section F.6, Figure 8 provides the corresponding QIM/additive comparison.

4) “an analysis of model complexity and runtime”

Revised Section F.5, Table 18 reports the actual design sizes and recorded CPU costs on those same four scenes. Median embedding time is 62.41 s for the full method, 2.48 s for one stop, 59.72 s without dither and 61.62 s for additive coding. All variants configure eight CPU threads. One stop reduces mean matrix nonzeros from approximately 850,000 to 35,000; removing dither or changing to additive coding preserves the design size.

All these variants have zero learned decoder parameters. Revised Table 5 and Section B report implementation costs and count all required decoder weights, including frozen feature encoders. Extraction constructs one spatial index per orbit stop and then performs cone queries, arithmetic and votes, without rendering or neural inference.

AuthorsOpenReview ↗

Response to Reviewer xjxG (3/4)

W3 / Q3. Geometry and depth preservation

1) “Can the authors provide evidence supporting the claim that the proposed method preserves 3DGS geometry?”

The preservation claim concerns embedding-induced change relative to the original reconstruction. Original Section 3.1, Equation (1), Section 3.5, Section 4.2.1 and Section C.3 specify the structural contract, degree-zero-only writes and field-level checks. We have independently repeated the serialized-field comparison on all 51 scene/key assets across eight Synthetic and nine real scenes. Count, positions, scales, rotations, opacity and every non-DC SH coefficient are exactly unchanged, as reported in revised Section I.3.

2) “how does the geometry or depth of the watermarked scene compare with the original scene quantitatively and qualitatively?”

We render the color-independent depth defined in revised Section I.3 at three fixed views each of Lego, Hotdog and Bonsai. Across nine pairs and 5,946,228 valid pixel comparisons, MAE, RMSE and maximum absolute depth error are zero; complete depth and accumulated-opacity arrays are bit-identical. Revised Section I.3, Figure 18 displays the measured depths and differences and defines the valid-opacity mask.

W4. Repeated readout, parity and tensor shapes

1) “the motivation and implementation of repeated readout”

Repeat readout evaluates the same cone-luminance functional on the received model to obtain the post-embedding readings. It supplies the measurements to the decoder using the public carrier and orbit, without rendering. Original Sections 3.3 and 3.6 and revised Section C.2 specify this computation.

2) “the purpose and mechanism of parity voting”

Rounding a keyed normalized reading gives an integer; its even or odd parity encodes one binary vote. Majority aggregates the eligible votes assigned to each message bit with a deterministic tie rule. Original Section 3.6 and revised Section C.2 distinguish this observation-level decision from aggregation across repeated observations.

3) “how cone activation and keyed QIM are defined and implemented.”

Cone activation averages the opacity-weighted luminance of its member Gaussians: atj=ntj−1∑i∈Ctjℓia_{tj}=n_{tj}^{-1}\sum_{i\in\mathcal C_{tj}}\ell_i, where ntj=∣Ctj∣n_{tj}=|\mathcal C_{tj}| and ℓi=αiw⊤ri\ell_i=\alpha_i\mathbf w^\top\mathbf r_i. Keyed QIM generates utju_{tj} from the key and chooses the nearest point utj+Δzu_{tj}+\Delta z whose integer parity z mod 2z\bmod 2 equals the assigned message bit. These points form the targets for the bounded color solve, as defined in original Sections 3.4–3.5.

4) “More details about the relevant feature representations, dimensions, and tensor shapes would also improve the reproducibility and clarity of the method.”

The numerical objects are explicit parameter arrays. Original Sections 3.2–3.6 and Section C.1–C.3 define luminance, membership, sparse realization, parity and majority. Revised Section C.4 adds the shapes: for NN Gaussians, centers and linear RGB are N×3N\times3, opacity is length NN, and the C=825C=825 directions form C×3C\times3. The T=24T=24 stops produce T×CT\times C readings, with the same indexing for eligibility, dither and target arrays. Flattened observation index modulo message length BB assigns each observation to a bit. With MM eligible observations and VV writable color shifts, the realization matrix is sparse M×VM\times V, and the unknown shift vector has length VV.

Reviewer_9cCTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FlyMark proposes a parameter-domain watermarking scheme for static 3DGS that does not require decoder training. It converts photoreceptor coordinates from a public fruit fly connectome into 825 fixed viewing directions and aggregates the apparent luminance of Gaussians within cones at 24 positions along a scene-normalized orbit. Repeated message bits are encoded using keyed dithered QIM, and existing DC colors are modified through a single bounded sparse least-squares solve. Changes to each writable linear-RGB color channel are bounded by 0.02, while geometry, opacity, and higher-order SH coefficients remain unchanged. Extraction uses parameter queries, quantization, and majority voting, requiring the key and a verification record containing the embedding-time eligibility mask.

Critical strengths
  1. The method combines parameter-domain readout, redundant QIM, and constrained color modification into a complete, auditable system. Geometry and higher-order appearance remain bit-identical, DC modifications satisfy a hard bound, and extraction requires neither a renderer nor a neural network. This deployment model has value independently of the biological origin of the carrier.

  2. The method achieves favorable results in several experiments. Its parameters are frozen before evaluation on held-out and external real scenes, with 96.64% clean BitAcc reported across nine real scenes.

Critical weaknesses
  1. In Table 1, GuardSplat achieves only 62.11% clean BitAcc and 3D-GSW achieves 86.67%. It is necessary to confirm that these baselines have converged properly under the current data, message, and extraction protocols; otherwise, the magnitude of the reported advantage cannot be interpreted reliably. Scene coverage and payload lengths are also not fully matched across methods.

  2. RandomCone-QIM is a useful control, but it is insufficient to attribute the gains to the fruit fly connectome itself. The effects of connectome provenance, nonuniform sampling density, the number of Gaussians supporting each cone, and the conditioning of the linear system need to be disentangled. A reproducible but otherwise arbitrary fixed geometry should not be conflated with a biological mechanism.

  3. There remains a substantial gap between high bit recovery and practical ownership verification, and the method's synchronization requirements need further validation.

  4. Section 3.3 defines the orbit center as the midpoint of the bounding box of all Gaussians. Adding a small number of distant, nearly transparent Gaussians could therefore change the bounding box and all readout directions. Although the paper tests random injection, it does not isolate an attack targeting this coordinate frame. The current results thus support clean detection and robustness to some transformations, but are insufficient to establish stable ownership evidence under general model-file tampering.

Questions for authors
  1. With the number of directions, angular density, support per cone, and solver budget matched, how do regular sampling, multiple blue-noise seeds, or sampling that mimics the connectome's density perform? Are the per-scene differences between FlyMark and RandomCone consistent?
  2. At the preselected threshold of 30 out of 32 correct bits, please report detection rates on external real scenes and after each attack, rather than only average BitAcc. The five million wrong-key trials change both the dither and the expected message. Are the false-positive results consistent when the registered message is held fixed and only the key is changed?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 9cCT (1/4)

We thank the reviewer for the careful assessment and specific experimental suggestions.

We have uploaded the revised manuscript, with changes marked in red. Below, we respond to each question.

W1. Baseline implementation, convergence, camera protocol and message length

1) “It is necessary to confirm that these baselines have converged properly under the current data, message, and extraction protocols”

The original 3D-GSW experiments followed the default optimization configuration of the official training script. In this revision, we retain its fixed decoder, FGD, objective and optimizer updates, use native black-background inputs, and rerun with a predetermined recovery-stability rule. Starting at epoch 8, we stop at the first epoch whose last three training accuracies and last three monitoring accuracies each span at most 0.5 percentage points, with an 80-epoch upper budget. Final test images do not select checkpoints. The updated 30-cell bit accuracy is 97.38%; the stopping epochs range from 8 to 16. Revised Table 1 and Section J contain only the updated 3D-GSW result and its protocol.

GuardSplat completes its official 100-epoch watermark schedule; mean final training-view bit accuracy is 97.37% over 30 runs. Its Synthetic training uses the prescribed ring distribution. Revised Table 1 evaluates the same trained assets using common held-out images (GuardSplat) and the original camera/readout protocol (GuardSplat (original)), giving 61.67% and 92.29% bit accuracy. The original-readout result reaches 99.09% on the Synthetic subset. Fidelity uses the same common held-out views for both rows.

2) “Scene coverage and payload lengths are also not fully matched across methods.”

We added the missing Bonsai and Garden GuardSplat runs. Revised Table 1 uses the same ten scenes and three keys for every method.

For message length, the additional FlyMark 48-bit row embeds GaussianMarker's identical complete message in all 30 paired cells, scoring all 48 positions rather than a prefix. FlyMark uses a 0.03 linear-RGB color budget, with its carrier, orbit, QIM spacing and solver unchanged. Bit accuracy is 99.72% for FlyMark and 98.75% for GaussianMarker; direct PSNR is 40.37 and 30.11 dB on the same full test views. Revised Table 1 and Section J.1 give the settings and paired results.

AuthorsOpenReview ↗

Response to Reviewer 9cCT (2/4)

W2 / Q1. Carrier density, support, conditioning and per-scene consistency

1) “A reproducible but otherwise arbitrary fixed geometry should not be conflated with a biological mechanism.”

Our biological inspiration is the retinal spatial sampling layout, which supplies a fixed anatomical prior for the carrier directions. We distinguish this prior from the watermarking mechanism: the main technical contributions are cone-based keyed QIM and bounded color realization. The connectome instantiates the measurement geometry within that framework.

2) “With the number of directions, angular density, support per cone, and solver budget matched, how do regular sampling, multiple blue-noise seeds, or sampling that mimics the connectome's density perform?”

The matched-support clean comparison in revised Section F.3, Table 16 fixes 825 directions, 24 orbit stops, the connectome-derived cone radius, eight Gaussian members per eligible cone and 32 observations per message bit. All carriers share key 17, QIM step 0.02 and the same solver tolerances and iteration caps. We compare Fibonacci sampling, three independent blue-noise realizations (seeds 1103, 2207 and 3301), and Density matched 1 (seed 1103). The density control independently perturbs each anatomical direction within half its original nearest-neighbor chord distance, restoring points that leave the forward cap. At each common linear-RGB color bound, mean clean bit accuracy on the test data is:

Color boundConnectomeFibonacciBlue noise 1Blue noise 2Blue noise 3Density matched 1
0.01100.00%100.00%98.44%96.88%100.00%98.44%
0.00592.19%82.81%84.38%87.50%85.94%89.06%

3) “The effects of connectome provenance, nonuniform sampling density, the number of Gaussians supporting each cone, and the conditioning of the linear system need to be disentangled.”

Revised Section F.3 fixes support at eight Gaussian members per cone and 32 observations per message bit, giving 1,024 observations. A separate original-color-bound diagnostic gives mean condition numbers of 1.21 to 1.26 across the controlled carriers, computed on the 32-row, row-mean message-aggregated design. These values are close to 1, indicating good and comparable numerical conditioning of those message-aggregated designs.

4) “Are the per-scene differences between FlyMark and RandomCone consistent?”

Yes. With the same three keys averaged within each scene, FlyMark matches or exceeds RandomCone-QIM in clean bit accuracy across all ten evaluated scenes, with a mean paired gain of 3.33 percentage points. Revised Section F.2, Table 15 reports:

SceneFlyMarkRandomCone-QIMGain (pp)
Chair100.00%100.00%0.00
Drums100.00%100.00%0.00
Ficus97.92%87.50%+10.42
Hotdog100.00%100.00%0.00
Lego100.00%100.00%0.00
Materials100.00%94.79%+5.21
Mic95.83%87.50%+8.33
Ship100.00%100.00%0.00
Bonsai100.00%100.00%0.00
Garden98.96%89.58%+9.38
Mean99.27%95.94%+3.33
AuthorsOpenReview ↗

Response to Reviewer 9cCT (3/4)

W3 / Q2. Practical decisions, synchronization and fixed-message wrong keys

1) “At the preselected threshold of 30 out of 32 correct bits, please report detection rates on external real scenes and after each attack, rather than only average BitAcc.”

We have added detection rates at the preselected 30/32 threshold. Revised Section I.1, Table 21 reports 85.19% on Mip-NeRF 360 and 100.00% on Tanks and Temples, with per-scene results. Revised Section I.4, Table 23 reports detection rates for each standard attack family alongside bit accuracy. For example, detection remains 100.00% under row reordering, color quantization and SH truncation. Under DC noise, increasing the linear-RGB standard deviation from 0.005 to 0.02 reduces detection from 100.00% to 96.67%.

2) “Are the false-positive results consistent when the registered message is held fixed and only the key is changed?”

In our verification, the extracted message is compared with the expected message generated by the candidate key. Original Section 4.2.3 and Figure 5 calibrate this rule over five million trials, with no acceptance at 30/32 matching bits. For the fixed-message diagnostic, embedding key seed 17 and the registered message remain fixed while 500,000 wrong-key seeds are tested per input. The mean wrong-key match rate over the test samples is 0.0120% at the same cutoff (revised Section I.6).

3) “the method's synchronization requirements need further validation.”

Verification state and coordinate registration are now distinguished explicitly in revised Section 3.6 and Section I. In registered verification, the owner retains the key and embedding-time mask; the evaluated expected message is key-derived, and carrier geometry and remaining settings are public. A fresh strict key-only path reconstructs those public quantities and recomputes eligibility. Revised Section I.2, Table 22 gives 98.96% recovery on the comparison corpus and 96.88% across all nine real scenes.

For unknown rotations, we now establish a deterministic canonical frame from Gaussian centers at both embedding and extraction, then reconstruct the orbit in that frame. No original model or attack rotation matrix is supplied. On Hotdog, Bonsai and Garden, three keys and six rotations give 99.31% bit accuracy across 54 rotated scene/key cases. Revised Section 3.3 and Section I.5 specify the geometry-only alignment and give its derivation and pose-recovery measurements.

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

This paper proposes FlyMark, a training-free, model-domain watermarking method for static 3D Gaussian Splatting assets. Rather than training an additional decoder or extracting a watermark from rendered images, FlyMark embeds a keyed binary message directly into the degree-zero color parameters already stored in a 3DGS model.

The method constructs a deterministic set of carrier directions from the photoreceptor layout of a published fruit-fly connectome. A virtual observer follows a scene-normalized orbit and measures cone-wise aggregate luminance from Gaussian centers, colors, and opacities. Message bits are redundantly assigned to these observations and encoded using keyed dithered QIM. The desired observation changes are then realized by a single sparse bounded least-squares solve that modifies only existing degree-zero colors under a hard linear-RGB perturbation bound. Extraction is deterministic and does not require a learned decoder or rendering.

The evaluation covers NeRF Synthetic and Mip-NeRF 360 scenes, with separate development, validation, held-out, and external-real-scene roles. FlyMark is compared with published 3DGS watermarking methods as well as matched parameter-space controls. The paper evaluates clean recovery, visual fidelity, robustness to model transformations and targeted attacks, wrong-key separation, overwrite/collusion behavior, efficiency, and several design components.

Critical strengths
  1. The parameter-domain, training-free formulation is a meaningful contribution. The strongest aspect of the paper, in my view, is not simply that it achieves high bit accuracy, but that it constructs the watermark directly from quantities already stored in the 3DGS asset and requires no learned decoder. The hard per-channel color bound and preservation of geometry, opacity, Gaussian count, and higher-order appearance parameters also make the modification contract unusually explicit. This is a useful alternative to watermarking systems whose verification depends on retained neural-network weights or a particular rendering pipeline.

  2. The evaluation is unusually comprehensive and thoughtfully controlled. The paper separates development, validation, held-out synthetic scenes, and external real scenes, rather than tuning and evaluating on the same small collection. It also includes matched parameter-space controls, several published baselines, multiple model-domain attacks, damage-matched evaluation, wrong-key calibration over a very large number of trials, overwrite/collusion experiments, and efficiency measurements. These experiments substantially increase my confidence that the basic QIM/cone-aggregation construction is real rather than an artifact of a particular scene or parameter setting.

Critical weaknesses
  1. The paper does not yet convincingly isolate the technical value of the fruit-fly connectome itself. This is the main issue affecting my assessment of novelty. The most relevant matched control, RandomCone-QIM, uses the same orbit, eligibility rule, QIM construction, and bounded solve while replacing only the carrier directions. It already performs quite strongly. FlyMark improves clean bit accuracy and the aggregate attack AUC, but the margin is not large in all settings, and the random carrier is slightly better in the reported damage-matched comparison.

More importantly, Appendix F characterizes other deterministic equal-count lattices such as Fibonacci and regular-hex arrangements geometrically, but does not evaluate them as watermark carriers using the complete pipeline. Since being deterministic, public, reproducible, and non-scene-specific is not unique to a biological connectome, I think the paper needs stronger evidence that the particular nonuniform connectome geometry contributes to the watermarking performance, rather than mainly providing an appealing externally specified lattice. Without such evidence, I would view the central technical contribution more as the cone-based keyed-QIM construction and bounded realization solver, with the connectome being one possible carrier instantiation.

  1. The "released model + secret key alone" verification claim is stronger than the protocol currently demonstrated. The introduction motivates FlyMark partly by arguing that verification should not depend on an auxiliary artifact that must be retained alongside the asset. However, the primary extraction protocol uses a private verification record containing the expected message and carrier information, and, importantly, the embedding-time eligibility mask. The paper does report strong results when eligibility is recomputed from the received model, which is encouraging, but the relationship between that experiment and the stronger "secret key alone" claim is not completely clear. I do not consider this fatal: retaining a small record is qualitatively different from retaining a trained neural decoder. Still, this distinction is central to the motivation, and the paper should state precisely what information an auditor must preserve and which parts can be deterministically reconstructed from only the suspect model, the public specification, and the secret key.
Questions for authors
  1. How much of the improvement is specifically due to the connectome-derived carrier geometry? Could the authors run the full watermarking pipeline using the equal-count Fibonacci and regular-hex lattices already analyzed geometrically in Appendix F, ideally together with several independent random/blue-noise lattices? I would also be interested in a paired statistical comparison against RandomCone-QIM. This would materially affect my view of whether the biological carrier is a substantive algorithmic contribution or primarily one convenient deterministic instantiation of the broader cone-QIM framework.

  2. What is the minimum information that must be retained for verification? The introduction suggests verification from the released asset and a secret key alone, whereas Section 3.6 describes a private verification record containing the expected message, carrier information, cone radius, and embedding-time eligibility mask. Can the authors define a strict "model + public algorithm + key only" protocol and report its performance separately? In particular, can the eligibility mask be recomputed and the expected message be deterministically derived from the key, or must either of these be stored externally?

  3. How does FlyMark behave under global changes of the scene coordinate frame? The scene center and scale are normalized, but the observer orbit and carrier directions appear to retain an orientation relative to the model coordinate axes. A global rigid rotation or axis permutation is a plausible asset-processing operation and can preserve the represented scene when the associated coordinate convention is transformed consistently. Does such a transformation preserve the watermark, and if not, is there a canonical orientation that can be recovered deterministically from the model? I think this is relevant to the paper's portability/repackaging motivation.

5 comments
AuthorsOpenReview ↗

Response to Reviewer akrY (1/4)

We thank the reviewer for the thoughtful review and constructive suggestions.

We have uploaded the revised manuscript, with changes marked in red. Below, we respond to each question.

W1 / Q1. Contribution of the connectome carrier and complete lattice comparisons

1) “Without such evidence, I would view the central technical contribution more as the cone-based keyed-QIM construction and bounded realization solver, with the connectome being one possible carrier instantiation.”

Our core technical contributions are the parameter-domain cone-based keyed QIM construction and the bounded realization of its targets using colors of existing Gaussians. The connectome provides a biologically grounded carrier instance whose anatomical sampling layout determines the spatial measurement operator for embedding and extraction.

The biological inspiration is a spatially organized retinal sampling layout used as a multichannel measuring instrument. The published visual-system connectome [r1] documents retinotopic organization; the MaleCNS resource [r2] supplies our anatomical source. In original Sections 3.3 and F.1, Figure 6, our specified projection converts the photoreceptor coordinates into 825 distinct directions. This gives an externally specified, nonuniform layout fixed independently of the protected scene. Replaying it along the observer orbit produces localized cone measurements. The biological element is the source sampling geometry; the orbit, cone functional and watermark code are the computational construction built around it.

In original Sections 3.3–3.5 and C.3, let ntj=∣Ctj∣n_{tj}=|\mathcal C_{tj}|, and let W\mathcal W be the writable Gaussians. An achromatic color shift δi\delta_i changes the observation by

=\sum_{i\in\mathcal C_{tj}\cap\mathcal W}\frac{\alpha_i}{n_{tj}}\delta_i =(A\boldsymbol\delta)_{tj}.$$ Here $A$ is the sparse realization matrix, with rows indexed by eligible $(t,j)$. Directions determine cone membership, support, overlap and hence those rows. They therefore affect which QIM targets can be realized within the color bounds and how many observations support each bit. The observations are indexed by scene geometry rather than serialized Gaussian row numbers. Hold the scene frame, cone memberships and opacities fixed, and suppose per-Gaussian luminance perturbations $e_i$ are independent, zero-mean and have variance $\sigma^2$. Since $0\leq\alpha_i\leq1$, $$\operatorname{Var}\left(\frac{1}{n_{tj}}\sum_{i\in\mathcal C_{tj}}\alpha_i e_i\right) =\frac{\sigma^2}{n_{tj}^{2}}\sum_{i\in\mathcal C_{tj}}\alpha_i^2 \leq\frac{\sigma^2}{n_{tj}}.$$ A supported cone mean therefore attenuates independent color noise under these assumptions. Nonuniform sampling changes support and overlap; overlapping observations are coupled by the bounded solve. **2)** *“How much of the improvement is specifically due to the connectome-derived carrier geometry? Could the authors run the full watermarking pipeline using the equal-count Fibonacci and regular-hex lattices already analyzed geometrically in Appendix F, ideally together with several independent random/blue-noise lattices?”* **Original Table 1** already compares RandomCone-QIM and SpatialHash-QIM with the proposed readout. We have now added the requested complete Fibonacci and regular-hex pipelines to **revised Section F.2, Table 14**, on ten scenes and three keys per carrier: | Carrier | Clean bit accuracy | | --- | ---: | | Connectome | 99.27% | | Fibonacci | 95.73% | | Regular hex | 96.04% | | RandomCone-QIM | 95.94% | **Revised Section F.3, Table 16** adds three independent blue-noise realizations (carrier seeds 1103, 2207 and 3301) and an IID-random carrier. For the matched-support clean comparison, every carrier uses 825 directions, 24 orbit stops, eight Gaussian members per cone and 32 observations per message bit. We fix key 17, the common cone radius, QIM step 0.02 and solver budget, and evaluate linear-RGB color bounds of 0.01 and 0.005. Mean clean bit accuracy on the test dataset is: | Color bound | Connectome | Fibonacci | Blue noise 1 | Blue noise 2 | Blue noise 3 | Density matched 1 | IID random | | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | | 0.01 | 100.00% | 100.00% | 98.44% | 96.88% | 100.00% | 98.44% | 98.44% | | 0.005 | 92.19% | 82.81% | 84.38% | 87.50% | 85.94% | 89.06% | 87.50% |
AuthorsOpenReview ↗

Response to Reviewer akrY (2/4)

W1 / Q1. Contribution of the connectome carrier and complete lattice comparisons

3) “I would also be interested in a paired statistical comparison against RandomCone-QIM.”

In revised Section F.2, Table 15, we pair the methods on the same ten scenes, averaging the same three keys within each scene. FlyMark matches or exceeds RandomCone-QIM in clean bit accuracy on all ten evaluated scenes, with a mean paired gain of 3.33 percentage points. The scene-wise results are:

SceneFlyMarkRandomCone-QIMGain (pp)
Chair100.00%100.00%0.00
Drums100.00%100.00%0.00
Ficus97.92%87.50%+10.42
Hotdog100.00%100.00%0.00
Lego100.00%100.00%0.00
Materials100.00%94.79%+5.21
Mic95.83%87.50%+8.33
Ship100.00%100.00%0.00
Bonsai100.00%100.00%0.00
Garden98.96%89.58%+9.38
Mean99.27%95.94%+3.33
AuthorsOpenReview ↗

Response to Reviewer akrY (3/4)

W2 / Q2. Minimum retained information and strict key-only verification

1) “What is the minimum information that must be retained for verification?”

Retain the key and embedding-time eligibility mask for registered-mask verification. In the evaluated profile, the expected message is derived from the key. Carrier coordinates, enumeration and projection, the cone-radius rule, orbit and remaining settings are public; copies in a convenience record are not additional owner-specific secrets. Original Section 3.6 describes the record, and original Sections 4.2.1, B.2 (Table 7) and C.2 distinguish fixed-mask and recomputed eligibility.

For an arbitrary owner-selected message, the owner additionally retains the expected message for comparison. It is not an input to the extracted parity votes. Revised Sections 3.6 and C.4 now separate public, key-derived, recomputed and retained information.

2) “Can the authors define a strict "model + public algorithm + key only" protocol and report its performance separately?”

Retain only the key. Reconstruct the public carrier and radius, derive the expected message and dithers, and recompute eligibility from the received model without loading the embedding-time mask or private record.

The strict key-only implementation in revised Section I.2, Table 22 achieves 98.96% bit accuracy on the original 30-cell cross-method corpus, versus 99.27% with the registered mask. Across all nine real scenes and three keys, it achieves 96.88% bit accuracy, versus 96.64% with registration.

Comments on the paper

2 comments

Dear Reviewers and Area Chair,

Thank you again for your time and constructive feedback. We have posted our point-by-point responses and revised manuscript with changes marked in red.

The main revisions include:

  • Full-pipeline carrier comparisons and density and support controls (Sections F.2–F.3).
  • Strict key-only verification and canonical-frame rotation experiments (Sections I.2 and I.5).
  • Further baseline validation, component ablations, geometry and depth checks, and expanded real-scene evaluation, as detailed in our individual responses.

We would appreciate your feedback on whether our responses and additional results address your main concerns. We would be happy to discuss any further questions you may have.

Thank you for your time and consideration.

The Authors of #38877

AuthorsOpenReview ↗

Still Awaiting Reviewer xjxG’s Response to Our Rebuttal

Dear all,

We have been waiting for Reviewer xjxG’s response to our rebuttal and follow-up questions, but have not received any reply so far. Our concern is not the numerical ratings themselves, but comments in the review that appear inconsistent with its own summary and stated strengths.

Our questions concern the lowest ratings for Presentation, Soundness and Informativeness in relation to the review’s detailed summary and its recognition of our effective parameter-space design and training-independent formulation. We have also asked the reviewer to explain why joint training is questioned after training independence is identified as a strength.

In our rebuttal, we provided additional ablations, geometry preservation checks, and broader scene evaluations. These additions respond directly to the reviewer’s stated requests.

We believe our responses and additions substantively address the stated concerns. We would appreciate the Area Chair’s consideration of the paper’s acknowledged strengths and the additional evidence provided in our rebuttal.

Thank you for your time and consideration.

The Authors of #38877

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.