acceptodds
Under review as a conference paper at ICLR 2027

Resolving Surrogate Target Ambiguity in Black-Box 3D Gaussian Splatting Watermark Attacks

Abstract

Watermarking 3D Gaussian Splatting (3DGS) assets requires modifying selected Gaussian primitives. When watermark carriers are concentrated in a few spatial regions, those regions must carry more of the embedding burden, potentially requiring stronger local perturbations that degrade rendering fidelity. However, existing methods often lack explicit mechanisms for spatially balanced carrier allocation. In this paper, we introduce MPP-GS, a high-fidelity 3DGS watermarking framework with two stages. In carrier construction, Multi-Projection Packing (MPP) organizes Gaussians on a UV grid and assigns each Gaussian multiple candidate destinations, reducing congestion and overflow while producing a more balanced carrier. In watermark embedding, Complementary Pair Modulation (CPM) exploits this carrier to distribute parent–auxiliary pairs and coordinate their updates, reducing rendering interference while preserving decodability. Across three benchmarks (Blender, LLFF, and Mip-NeRF 360), MPP reduces average carrier overflow from 17.75% to 0.08%, retaining substantially more Gaussian primitives. Together, balanced carrier construction and low-interference watermark embedding enable MPP-GS to achieve higher rendering fidelity with comparable or better decoding accuracy across multiple payloads. At 64 bits, MPP-GS achieves 35.15 dB PSNR and 95.95% bit accuracy, compared with 32.79 dB and 91.17% for W2M. Code will be made publicly available on GitHub.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
3D Gaussian Splatting, Digital Watermarking, Black-Box Attack, Privacy

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_NV7rOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to attack 3DGS watermarks using a public surrogate decoder without access to the owner's decoder, message, or verification queries. The authors argue that surrogate gradients do not directly determine an effective attack target and propose AACT. It flips exactly half of the surrogate message bits and selects the subset using a balanced D-optimal criterion based on the strength of gradients with respect to Gaussian appearance parameters and the complementarity of their carriers. The attack updates only bounded SH appearance parameters while preserving geometry and the Gaussian set. SNSS terminates optimization based on the convergence trajectory of the surrogate objective.

Critical strengths
  1. The method achieves a practically useful trade-off among attack effectiveness, rendering quality, and computational cost under a strict setting with no owner-side queries.

  2. The attack uses publicly available gradients without modifying geometry, making it a useful tool for watermark security evaluation.

  3. The supplementary experiments in the appendices substantively strengthen the conclusions rather than merely repeating the main results.

Critical weaknesses
  1. The “alignment” metric and controlled experiment do not fully support the proposed mechanistic explanation of the relationship between models. Section 4.1 measures alignment by bitwise agreement between decoder outputs, and Section 4.2 reduces it from 100% to 50% by applying a fixed sign mapping to the surrogate logits. However, this is a relabeling of the output code and does not necessarily reduce the functional or gradient relationship between the decoders in image space. In particular, for BI, if z′ = qz with q ∈ {−1, +1}, the initial pseudo-message flips accordingly, while the signed gradient toward the opposite decision remains unchanged. Its approximate invariance in the controlled experiment therefore has a direct algebraic explanation. The experiment demonstrates stability to code polarity changes, but cannot by itself establish that target ambiguity is resolved for genuinely independent decoders.

  2. The additional contribution of the full target-selection procedure over simpler targets requires clearer statistical and computational evidence. In Table 3, the full method does improve target-boundary crossing success and iteration counts, but the paper does not report per-scene paired differences for the main metrics, uncertainty across repeated runs, or the separate overhead of computing per-bit gradients and constructing the carrier kernel. Across different message lengths, AACT does not consistently outperform BI in ΔBAR and IDS, although it performs better in verification pass rates and rendering quality. A better-supported conclusion is therefore that it offers a stable and practical compromise, rather than a universally stronger target construction. The value of the full D-optimal selection should be assessed under equal total runtime or equal image-distortion budgets.

Questions for authors
  1. Could the authors distinguish output-code agreement from attack-gradient transferability? Please explain why changing only output polarity supports the current interpretation of surrogate–owner alignment, and clarify whether the owner and public surrogate decoders for GM/3D-GSW use the same checkpoint or share training provenance.
  2. Please provide per-scene paired results and uncertainty for Full versus Random and Strength, and clarify whether the reported total runtime includes all carrier-gradient computations and D-optimal selection. Does the advantage persist under equal total runtime or equal distortion?
  3. The inversion-aware verification in Appendix C is valuable. Could GMEA also be evaluated under the same protocol, with the number of random test messages reported? This would help determine whether the IDS advantage in the main text translates into consistently stronger disruption of owner verification.
4 comments
AuthorsOpenReview ↗

Response to Reviewer NV7r W1/Q1: Output Agreement and Gradient Transfer

Response to Reviewer NV7r W1/Q1: Output Agreement and Gradient Transfer

Thank you for this careful observation. GaussianMarker and 3D-GSW share the same publicly released HiDDeN weights with our original surrogate, so their 100% agreement reflects a matched-decoder setting. We agree that output agreement and gradient transfer are distinct, and that the polarity experiment alone does not establish transfer between independently trained decoders. We added independent-decoder evaluations, post-hoc gradient diagnostics, and checkpoint provenance in Appendix F.1–F.3 (Tables 13–15).

Why BI is unchanged under polarity relabeling. We agree with your algebraic explanation. For BI, flipping a logit sign also flips its target label, and BCE(−z,1−y)=BCE(z,y)\mathrm{BCE}(-z,1-y)=\mathrm{BCE}(z,y), so the loss and its gradient are unchanged. Accordingly, Fig. 4 should not be read as evidence of general cross-decoder transfer. We have revised Sec. 4.2 to describe it as a controlled code-polarity study, while the independent-decoder experiments in Appendix F provide separate evidence under actual decoder differences.

Tests with independent decoders. We evaluate five independently pretrained public surrogates: two Stable Signature (Fernandez et al., ICCV, 2023) HiDDeN48 extractors, TrustMark100 (Bui et al., ICCV, 2025), CIN30 (Ma et al., ACM MM, 2022), and MBRS30 (Jia et al., ACM MM, 2021). Each watermarked asset and owner decoder stays fixed, and CIN30 and MBRS30 use up to 240 optimization steps with SNSS monitoring (Appendix F.1). Over 15 scenes and three seeds, AACT reaches low inversion-aware pass rates on GS48 with all five surrogates (P2=0.42P_2=0.42–2.47%2.47\%). On GM48, the independent HiDDeN variants transfer much better (P2=3.19%P_2=3.19\% and 16.18%16.18\%) than TrustMark, CIN, and MBRS (P2=93.16P_2=93.16–94.34%94.34\%).

Checking whether gradients transfer. Output agreement measures whether the surrogate and owner return the same bits, not whether optimizing the surrogate objective increases the owner's recovery error. We therefore measure (i) the cosine between the negative surrogate target-loss gradient and the positive owner recovery-loss gradient in the same Gaussian SH appearance-parameter space, and (ii) the owner-loss change after a small step along the surrogate attack direction. Table 15 uses the initial assets, two fixed views per scene, seed 0, and 15 equally weighted scenes.

Public surrogateGM cosine ↑GM loss transfer ↑GS cosine ↑GS loss transfer ↑
Original HiDDeN48+0.68024+6.959e-4+0.00625+4.053e-4
Independent HiDDeN48+0.30616+3.092e-4+0.05371+2.607e-3
Augmented HiDDeN48+0.17593+1.699e-4+0.04839+2.501e-3
CIN30+0.00075+3.378e-6+0.05430+2.907e-3
TrustMark100-0.00108+2.488e-6+0.04925+2.539e-3
MBRS30-0.00102+2.488e-6+0.00724+4.161e-4

On GM48, the independent HiDDeN variants show much larger cosine and owner-loss transfer than CIN, TrustMark, and MBRS, consistent with their stronger disruption (IDS 80.61/75.12 versus 14.79–21.53). On GS48, independent HiDDeN, CIN, and TrustMark show owner-loss transfer of about 2.52.5–2.9×10−32.9\times10^{-3} and strong attacks (IDS 83.06–85.45). Because loss scales differ across owners, we compare these values within each owner. These results indicate that transferable surrogate guidance, rather than output agreement alone, helps explain the owner-dependent outcomes. Owner messages and gradients are used only for this analysis, never during attack optimization.

Beyond the matched setting. With the independent and augmented HiDDeN surrogates, AACT still outperforms GMEA on GaussianMarker in all four metrics: ΔBAR 3.15/12.24 vs. 18.71, IDS 80.61/75.12 vs. 62.60, P2P_2 3.19/16.18% vs. 20.42%, and PSNR 24.91/24.32 vs. 23.33 dB (Tables 7 and 14).

Summary. These analyses separate output agreement from transferable attack guidance. AACT addresses the remaining surrogate target-selection ambiguity using public information, without the owner's bit mapping or gradients. Sec. 4.1 now states the shared decoder weights, and Appendix F.3 details the distinction between output agreement and gradient transfer.

AuthorsOpenReview ↗

Response to Reviewer NV7r W2/Q2 (1/2): What Full Bit Selection Adds

Response to Reviewer NV7r W2/Q2 (1/2): What Full Bit Selection Adds

Thank you for this suggestion. Full selection lowers the inversion-aware pass rate P2P_2 on GS48 to a three-seed mean of 5.52%, compared with 8.68% for Random and 8.16% for Strength. We added per-scene paired comparisons (Appendix H.1, Table 18) and three-seed mean ± population SD (Table 19). All three strategies share the same half-Hamming geometry, appearance optimization, and SNSS, and differ only in which bits they select.

Overall results. The table summarizes GS48 over 15 Blender/LLFF scenes with seed 0, where the strategies differ more than on the near-saturated GM48 setting.

SelectionΔBAR ↓IDS ↑P₂ (%) ↓PSNR (dB) ↑
Random12.3575.228.7529.01
Strength10.4876.838.0228.54
Full9.3179.905.7328.60

Full lowers P2P_2 by 3.02 percentage points relative to Random and 2.29 points relative to Strength, with PSNR slightly above Strength and below Random. Combining carrier strength and complementarity therefore improves actual owner-verification disruption, beyond reaching the surrogate target.

Per-scene comparison. The table counts the scenes where Full performs better, ties, or performs worse than each alternative.

ComparisonMetricFull betterTieFull worse
Full vs. RandomIDS ↑1203
Full vs. RandomP₂ ↓4101
Full vs. StrengthIDS ↑1302
Full vs. StrengthP₂ ↓591

These paired results indicate broadly consistent IDS improvements across scenes. For inversion-aware verification, many scenes already achieve zero passes, and among the non-tied scenes, Full generally performs better. These findings support the contribution of carrier-aware selection while acknowledging metric saturation.

Repeated runs. Table 19 reports three-seed mean ± population SD. On GS48, P2P_2 is 8.681±1.0228.681\pm1.022 for Random, 8.160±0.7728.160\pm0.772 for Strength, and 5.521±0.6135.521\pm0.613 for Full.

Revisions. The per-scene results, three-seed variability, and inversion-aware P2P_2 of the selection strategies are already reported in Appendix H.1 (Tables 18–19), and Sec. 4.3 now points to Appendix H. We will also report the mean ±\pm population standard deviation of Δ\DeltaBAR, IDS, and PSNR in Table 1, and replace the target-crossing statistics in Table 3(b) with P2P_2. This provides a more direct assessment of whether carrier-aware selection improves actual owner-side disruption, beyond surrogate optimization behavior.

AuthorsOpenReview ↗

Response to Reviewer NV7r W2/Q2 (2/2): Computational Cost and Equal-Time Results

Response to Reviewer NV7r W2/Q2 (2/2): Computational Cost and Equal-Time Results

Thank you for raising the cost question. The reported totals include all per-bit gradient, kernel, and selection costs, and Full retains an advantage under equal runtime once sufficient optimization time is available. We added a stage-wise runtime breakdown (Appendix H.2, Tables 20–21) and equal-runtime comparisons (Appendix H.3, Table 22).

Where the cost comes from. The table reports average wall-clock seconds over 15 Blender/LLFF scenes with seed 0 and a fixed 80-step schedule. The total includes preparation, optimization, and output saving.

OwnerSelectionPer-bit Grad.KernelSelectionTotal
GM48Random000.03048.40
GM48Strength4.25100.03152.77
GM48Full4.7502.9290.04356.88
GS48Random000.03135.55
GS48Strength3.51900.03739.07
GS48Full3.8232.3700.04642.11

Full adds 8.48/6.56 s over Random and 4.11/3.04 s over Strength on GM48/GS48. Most of this comes from per-bit gradients and the carrier kernel, while greedy D-optimal selection takes less than 0.05 s.

Results under the same time budget. We compare all three strategies at total budgets of 11, 22, 32, and 43 s (Appendix H.3, Table 22). We highlight GS48 at 32 s, where optimization has not yet reached its final effectiveness.

SelectionIDS ↑P₂ (%) ↓PSNR (dB) ↑
Random74.79710.52129.298
Strength75.6979.89628.807
Full77.7028.02128.960

At 32 s, Full lowers P2P_2 by 2.500 and 1.875 percentage points relative to Random and Strength and raises IDS by 2.905 and 2.005 points, with comparable PSNR. It also exceeds Strength in PSNR by 0.153 dB. These gains already include the full target-construction overhead. Random benefits from its lower preparation cost under very short budgets (11–22 s), and at 43 s the results coincide with the unconstrained comparison.

Comparison at similar image quality. Full at 32 s and Random at 43 s reach nearly identical PSNR (28.960 vs. 29.013 dB), while Full yields a lower P2P_2 (8.021% vs. 8.750%). At 43 s, Full also improves both PSNR and P2P_2 over Strength (28.597 vs. 28.537 dB, and 5.729% vs. 8.021%).

Different owner message lengths. We agree that BI achieves slightly better Δ\DeltaBAR and IDS in several mismatched-length settings in Table 12 of the revised Appendix (Table 10 of the original submission). Across all three settings, AACT achieves equal or lower P2P_2 than both CC and BI, and higher PSNR than BI. This supports AACT's favorable balance between disruption and image quality rather than superiority on every individual metric.

Summary. Full carrier-aware selection provides additional owner-verification disruption at a measurable computational cost, and the equal-runtime results show that this benefit persists after accounting for gradient probing and kernel construction.

AuthorsOpenReview ↗

Response to Reviewer NV7r Q3: GMEA Under Inversion-Aware Verification

Response to Reviewer NV7r Q3: GMEA Under Inversion-Aware Verification

Thank you for this suggestion. Under the same protocol, AACT achieves lower P2P_2 than GMEA on GM48, while GMEA achieves lower P2P_2 on GS48 at a substantially higher rendering cost. We added GMEA to Appendix Tables 7 and 8 for single-view and multi-view verification, and added per-scene comparisons between GMEA and AACT in Appendix C.3, Tables 9–10.

Single-view verification. The table uses 64 held-out novel views per scene.

OwnerAttackP₁ (%) ↓P₂ (%) ↓
GM48GMEA20.4220.42
GM48AACT0.000.00
GS48GMEA0.100.10
GS48AACT5.525.52

Message protocol. GM48 uses one fixed 48-bit owner message across all 15 scenes, while GS48 uses 15 messages randomly sampled from GuardSplat's message set, one per scene. The three seeds share the same messages, and all attacks use the same owner messages and evaluation views. GMEA follows the same three-seed protocol (seeds 0, 1, and 2) as AACT. Table 7 reports the three-seed average, while Tables 9–10 report seed 0.

Interpretation. Although GMEA achieves stronger owner-verification disruption on GS48 (P2P_2 = 0.10% versus 5.52%), it incurs substantially greater rendering distortion and computational cost. Accordingly, AACT's main advantage is its more favorable balance among attack effectiveness, rendering fidelity, and efficiency, rather than uniformly superior watermark removal.

Reviewer_nHaDOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies black-box attacks against 3D Gaussian Splatting watermarking when the owner decoder and embedded message are inaccessible, but an independently available differentiable surrogate decoder can be used. It identifies surrogate target ambiguity: surrogate gradients are available, but the optimal target depends on the unknown bitwise alignment between the surrogate and owner decoders.

The authors propose Alignment-Agnostic Carrier Targeting (AACT). AACT uses a half-Hamming target displacement, carrier-aware bit selection with a weighted D-optimal criterion, bounded appearance-only optimization, and Self-Normalized Saturation Stopping (SNSS). Experiments on GaussianMarker, GuardSplat, and 3D-GSW across 15 Blender and LLFF scenes show strong watermark disruption while preserving Gaussian primitives, improving rendering fidelity, and reducing attack time compared with GMEA.

Critical strengths
  1. The paper identifies a meaningful problem. Surrogate target ambiguity is a relevant issue that is distinct from merely obtaining surrogate gradients. The paper correctly points out that the best target can change substantially with the unknown surrogate-owner alignment.

  2. AACT combines several components coherently. The half-Hamming target geometry, carrier-aware bit selection, bounded appearance optimization, and SNSS stopping rule form a consistent attack framework. The ablations indicate that carrier strength and complementarity both contribute to the final performance.

  3. The evaluation covers different alignment regimes. GaussianMarker and 3D-GSW represent high-alignment settings, while GuardSplat provides a lower-alignment setting. The controlled alignment experiment further supports the central motivation.

  4. The practical improvements are substantial. Compared with GMEA, AACT preserves all Gaussian primitives, improves PSNR/SSIM, and reduces attack time from thousands of seconds to tens of seconds. This is especially valuable for evaluating the security of explicit 3D representations.

Critical weaknesses
  1. The evaluation relies on a narrow surrogate and baseline set. All surrogate-guided attacks use a fixed pretrained HiDDeN decoder, and the main comparison is against GMEA. It is unclear whether the conclusions remain valid for other public decoders, surrogate architectures, or stronger recent black-box attacks.

  2. The alignment study models only a narrow type of decoder mismatch. To simulate different surrogate-owner alignments, the authors keep the same surrogate logits and simply flip the sign of selected bit coordinates. This corresponds to changing whether an individual bit is interpreted as 0 or 1, but it does not model other realistic differences between two decoders, such as a permutation of bit positions, correlated errors across bits, or different output semantics. Therefore, the controlled experiment supports robustness to bit-polarity changes, but does not establish robustness to general surrogate-owner decoder mismatch.

  3. The choice of a half-Hamming target is supported only by the evaluated experiments. The authors test several fractions of flipped bits and find that flipping 50% performs best on their evaluated watermarking systems. However, they do not provide a theoretical reason why 50% should be preferable in general, nor do they test whether this choice remains effective when the watermark length, decoder architecture, or watermarking method changes. Thus, the paper demonstrates that half-Hamming displacement works in the reported settings, but does not establish it as a generally optimal target geometry.

Questions for authors
  1. The evaluation relies on a narrow surrogate and baseline set. All surrogate-guided attacks use a fixed pretrained HiDDeN decoder, and the main comparison is against GMEA. It is unclear whether the conclusions remain valid for other public decoders, surrogate architectures, or stronger recent black-box attacks.

  2. The alignment study models only a narrow type of decoder mismatch. To simulate different surrogate-owner alignments, the authors keep the same surrogate logits and simply flip the sign of selected bit coordinates. This corresponds to changing whether an individual bit is interpreted as 0 or 1, but it does not model other realistic differences between two decoders, such as a permutation of bit positions, correlated errors across bits, or different output semantics. Therefore, the controlled experiment supports robustness to bit-polarity changes, but does not establish robustness to general surrogate-owner decoder mismatch.

  3. The choice of a half-Hamming target is supported only by the evaluated experiments. The authors test several fractions of flipped bits and find that flipping 50% performs best on their evaluated watermarking systems. However, they do not provide a theoretical reason why 50% should be preferable in general, nor do they test whether this choice remains effective when the watermark length, decoder architecture, or watermarking method changes. Thus, the paper demonstrates that half-Hamming displacement works in the reported settings, but does not establish it as a generally optimal target geometry.

5 comments
AuthorsOpenReview ↗

Response to Reviewer nHaD W1/Q1: Tests with Other Decoders and Attacks

Response to Reviewer nHaD W1/Q1: Tests with Other Decoders and Attacks

Thank you for this suggestion. AACT remains effective with independently pretrained decoders of different architectures and message lengths, and GMEA remains the most directly comparable external baseline under our protocol. We added cross-decoder experiments with five independently pretrained public decoders (Appendix F.2, Table 14), together with decoder provenance and adaptation details (Appendix F.1, Table 13).

Which decoders we tested. We added two HiDDeN-family extractors from Stable Signature (Fernandez et al., ICCV, 2023), trained with standard and augmented settings, plus TrustMark100 (Bui et al., ICCV, 2025), CIN30 (Ma et al., ACM MM, 2022), and MBRS30 (Jia et al., ACM MM, 2021). They cover different training settings, architectures, and message lengths. The original surrogate shares its weights with GaussianMarker and 3D-GSW, while the new surrogates use independent weights. The watermarked assets and owner decoders stay fixed, and only the public surrogate changes, with the input and loss adaptations listed in Appendix F.1. CIN30 and MBRS30 use up to 240 optimization steps with SNSS monitoring.

Results. The table reports means over 15 Blender/LLFF scenes and three seeds. GM48 and GS48 denote GaussianMarker and GuardSplat with 48-bit owner messages. P2P_2 is the owner's verification pass rate when full bit inversion is also allowed, so lower values mean stronger disruption.

OwnerPublic surrogateΔBAR ↓IDS ↑P₂ (%) ↓PSNR (dB) ↑
GM48Independent HiDDeN483.1580.613.1924.91
GM48Augmented HiDDeN4812.2475.1216.1824.32
GM48TrustMark10039.2121.5394.3425.10
GM48CIN3042.5714.7993.4723.48
GM48MBRS3039.5920.7493.1624.92
GS48Independent HiDDeN484.8085.450.4227.11
GS48Augmented HiDDeN484.6986.080.4926.17
GS48TrustMark1006.0383.061.3527.62
GS48CIN305.0083.280.6625.76
GS48MBRS306.9681.712.4727.36

On GS48, all five surrogates reach IDS above 81% and P2P_2 below 2.5%, including 30-bit and 100-bit decoders attacking a 48-bit owner. On GM48, the two independent HiDDeN surrogates reach IDS 80.61 and 75.12, with P2P_2 of 3.19% and 16.18%. These results demonstrate that AACT can achieve effective watermark disruption across independently pretrained decoders with different architectures and message lengths, even without predefined bitwise correspondence, provided that surrogate gradients offer transferable guidance.

Owner-dependent transfer. TrustMark, CIN, and MBRS transfer weakly to GM48, where CC and BI also achieve limited disruption. The gradient diagnostics in Appendix F.3 (Table 15) relate this difference to how well surrogate-guided perturbations transfer to each owner. Target construction therefore does not require a known bitwise alignment, while owner-side disruption depends on transferable surrogate guidance.

Scope of attack baselines. We use GMEA as the external baseline because it is a dedicated black-box 3DGS watermark attack previously evaluated on GaussianMarker and can be assessed under our bitstream-recovery protocol. GSPure (Huang et al., AAAI, 2026) mainly targets scene-hiding 3DGS watermarks through watermark-related Gaussian pruning. Its released implementation and protocol do not directly cover GaussianMarker, GuardSplat, or 3D-GSW, so a faithful comparison would require substantial adaptation. Image-domain removal attacks modify rendered images and would first require a validated multi-view refitting procedure to produce an attacked 3DGS asset, so we treat them as complementary 2D attacks rather than direct baselines (see also our response to Reviewer GdS6). Within our framework, CC and BI share the same optimization and SNSS protocol as AACT, while Random and Strength isolate the contribution of carrier-aware selection.

Summary. The new experiments extend the evaluation beyond a single decoder family and characterize where AACT transfers across independently pretrained surrogates.

AuthorsOpenReview ↗

Response to Reviewer nHaD W2/Q2: Beyond Output-Bit Polarity Changes

Thank you for this observation. We agree that Fig. 4 isolates bit-polarity remapping while keeping the same decoder architecture and weights, so it tests a controlled change in output-code correspondence rather than general decoder mismatch. We therefore added experiments with independently pretrained decoders (Appendix F.2, Table 14) and revised Sec. 4.2 accordingly.

What the new experiments change. Unlike polarity remapping, these experiments change the model weights, architectures, training objectives, and output spaces. The two independent HiDDeN variants use different pretrained weights, while TrustMark100, CIN30, and MBRS30 come from different decoder families. The 30-bit and 100-bit surrogates also differ in output length from the 48-bit owners, so no one-to-one bit correspondence is available, and AACT uses no predefined bit mapping.

Results. As detailed in our response to W1/Q1, AACT reaches IDS 83.06, 83.28, and 81.71 on GS48 with TrustMark, CIN, and MBRS, with inversion-aware pass rates of 1.35%, 0.66%, and 2.47%. AACT therefore remains effective under genuine cross-architecture differences, beyond polarity changes. The weaker transfer on GM48 matches the gradient diagnostics in Appendix F.3 (Table 15), which show differences in transferable surrogate guidance across owners.

The controlled study and the cross-decoder experiments provide complementary evidence. AACT constructs its target without a known surrogate–owner bitwise correspondence, and its effectiveness extends to independently trained decoders when transferable surrogate guidance is available.

AuthorsOpenReview ↗

Response to Reviewer nHaD W3/Q3: Why We Flip Half the Bits

Response to Reviewer nHaD W3/Q3: Why We Flip Half the Bits

Thank you for this question. We choose 50% because it is equidistant from the initial surrogate code and its full inversion, which gives a balanced target without owner feedback. We added a theoretical analysis (Appendix G.1) and flip-ratio experiments across watermarking systems (Table 16) and independent surrogate decoders (Table 17). This is a principled fixed choice rather than a universally optimal flip ratio.

Geometric motivation. Let m^∈{0,1}L\hat m\in\{0,1\}^{L} be the initial surrogate pseudo-code and mtm^t an attack target differing in kk bits. Its distances from the initial code and its full inversion satisfy

dH(mt,m^)=k,dH(mt,¬m^)=L−k,max⁡0≤k≤Lmin⁡{k,L−k}=L2.d_H(m^t,\hat m)=k,\quad d_H(m^t,\neg\hat m)=L-k,\quad \max_{0\le k\le L}\min\{k,L-k\}=\frac{L}{2}.

For even LL, k=L/2k=L/2 maximizes the minimum distance to these two endpoints, so it does not favor either one.

Link to owner verification. Consider an idealized fixed bit-permutation and polarity correspondence between surrogate and owner outputs. If the owner initially recovers its message correctly and the kk surrogate flips are exactly realized, the owner recovers K=L−kK=L-k correct bits, giving

max⁡(K,L−K)=max⁡(L−k,k)≥L2,\max(K,L-K)=\max(L-k,k)\geq \frac{L}{2},

with equality at k=L/2k=L/2. In our 48-bit inversion-aware test (Appendix C, t1=t2=36t_1=t_2=36), half-Hamming leaves 24 matches to both the original and inverted messages, below the acceptance threshold. By contrast, 25% and 75% leave 36 matches to one endpoint, and 100% leaves 48 matches to the inverted message.

This derivation motivates half-Hamming as a symmetric, owner-independent target geometry under an idealized distance-preserving correspondence. It does not imply optimality under arbitrary decoder relationships, which we evaluate empirically below.

Across watermarking systems. With the original HiDDeN48 surrogate, we vary the flip ratio on GM48 (matched decoder weights) and GS48 (unmatched decoder). Table 16 reports means over 15 Blender/LLFF scenes and three seeds.

Flip ratioGM48 P₂ (%) ↓GM48 PSNR (dB) ↑GS48 P₂ (%) ↓GS48 PSNR (dB) ↑
25%100.0026.018.5129.65
50%0.0025.485.5228.60
75%89.8325.475.6228.68
100%95.8025.394.6227.89

On GM48, 50% is the only ratio with zero inversion-aware passes, consistent with the analysis above. On GS48, P2P_2 varies less across ratios, and full inversion gives the lowest P2P_2 at the cost of the lowest PSNR. These per-owner results complement the owner-averaged ablation in Table 3(a).

Across decoder families and lengths. We next test TrustMark100 and MBRS30 against the same GS48 owner, again over 15 scenes and three seeds (Table 17). For the 30-bit MBRS30 surrogate, nominal 25% and 75% flip 8 and 22 bits.

SurrogateFlip ratioP₂ (%) ↓PSNR (dB) ↑
TrustMark10025%1.7027.73
TrustMark10050%1.3527.62
TrustMark10075%1.2827.53
TrustMark100100%1.8127.70
MBRS3025%6.8427.45
MBRS3050%2.4727.36
MBRS3075%2.4327.09
MBRS30100%2.5026.78

For both surrogates, 50% lowers P2P_2 relative to 25% and stays within 0.04–0.07 percentage points of 75%, while preserving higher PSNR than 75%. Full inversion does not consistently improve verification disruption. These results support half-Hamming as a competitive balance between disruption and image quality across architectures and code lengths.

Across owner message lengths. Table 12 in the revised Appendix (Table 10 of the original submission) attacks 32- and 64-bit owners with the same 48-bit surrogate. The half-Hamming target stays within 0.13 percentage points of the best ΔBAR and achieves equal or lower P2P_2 than CC and BI in all three settings (GS32, GM32, and GM64).

Summary. Half-Hamming has a clear geometric and verification-based motivation, since it balances the two complementary code endpoints without owner feedback. The experiments show that it is an effective fixed choice across the evaluated settings, while carrier-aware selection determines which bits to flip within that geometry (Appendix G.1–G.3, Tables 16–17).

Revisions. Per-owner P2P_2 for the flip-ratio ablation is already reported in Appendix G.2 (Table 16). We will also add P2P_2 to Table 3(a).

Reviewer_GdS6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies black-box watermark attacks against 3D Gaussian Splatting (3DGS), focusing on an insufficiently formalized problem: how to construct an effective attack target when an attacker can access a public surrogate decoder and a differentiable renderer but cannot access the owner's decoder, embedded message, or verification process.

The authors formulate this challenge as Surrogate Target Ambiguity and propose Alignment-Agnostic Carrier Targeting (AACT). AACT first constructs a target at a fixed half-Hamming distance from the initial surrogate pseudo-code by flipping half of its bits. It then selects the specific bits to flip using a balanced, greedy D-optimal criterion that considers both bit-flip strength and the complementarity of their Gaussian carrier signatures. Finally, bounded appearance-space optimization disrupts the watermark while preserving Gaussian geometry and the primitive set. Self-Normalized Saturation Stopping (SNSS) determines when optimization should terminate.

The experiments cover three watermarking systems—GaussianMarker, GuardSplat, and 3D-GSW—and 15 scenes from Blender and LLFF. A fixed pretrained 48-bit HiDDeN decoder serves as the public surrogate, and experiments use three random seeds. Additional evaluations examine Mip-NeRF 360, different owner-message lengths, held-out novel views, and multi-view verification. The results show strong watermark disruption on GaussianMarker and 3D-GSW, alongside substantial improvements in rendering fidelity, primitive retention, and runtime compared with GMEA. On GuardSplat, AACT improves rendering fidelity but is slightly less effective at watermark disruption than some alternative target strategies.

The main contribution is a clearly formulated attack-target construction problem and an integrated attack framework combining surrogate gradients, carrier-aware bit selection, and adaptive stopping. However, the experiments have not yet established the generality of AACT across a broad range of surrogate-model differences and arbitrary unknown decoder alignments. The empirical contribution is meaningful, but its generality and methodological novelty require further substantiation.

Critical strengths
  1. The problem formulation is clear and practically relevant. The paper correctly distinguishes the availability of surrogate gradients from the ability to determine an effective attack target. This is a meaningful problem for black-box watermark security evaluation and provides a clear basis for investigating the role and limitations of surrogate models.
  2. The experimental design is reasonably comprehensive. The authors compare AACT with GMEA and use CC and BI under a shared optimization protocol to isolate the effects of target construction. Ablations of bit selection, target Hamming distance, and SNSS are supplemented by novel-view evaluation, multi-view verification, and experiments with different message lengths. These experiments provide empirical support for several design choices.
Critical weaknesses

Weakness 1: The evidence does not yet support the full generality implied by “alignment-agnostic”. This is the most important limitation of the paper. The experiments use a single fixed 48-bit HiDDeN decoder as the public surrogate. GaussianMarker and 3D-GSW exhibit 100% bitwise agreement with this surrogate, whereas GuardSplat exhibits 54.02% agreement. These results cover high- and low-alignment cases, but the range of actual decoder differences remains limited. Moreover, the controlled alignment experiments vary alignment by applying bitwise polarity flips to the surrogate outputs. This primarily changes the sign correspondence between bits and does not adequately represent differences in architecture, training data, feature extraction, or nonlinear decoding behavior.

The authors do not need to prove that AACT works for every conceivable decoder relationship, but they should precisely define the scope of “alignment-agnostic.” The current evidence supports the claim that AACT can achieve reasonably stable attack performance without explicitly estimating alignment under the evaluated systems and constructed polarity mappings. It does not establish comparable effectiveness for arbitrary unknown surrogate–owner relationships.

The GuardSplat results also illustrate this qualification. In Appendix Table 7, BI achieves a (\Delta BAR) of 8.65 and an IDS of 82.53, compared with 9.52 and 80.32 for AACT. BI disrupts the watermark slightly more effectively, while AACT achieves higher PSNR: 28.72 dB versus 27.89 dB. AACT is therefore more accurately characterized as a target-construction strategy that emphasizes cross-setting robustness and the disruption–fidelity trade-off, rather than an attack that is strongest under every alignment condition. Required improvement: Evaluate multiple public surrogate decoders or architectures and include surrogate–owner differences beyond simple bit-polarity flips. If such experiments are not feasible, narrow the claims of generality.

Weakness 2: The efficiency comparison does not fully isolate the contribution of AACT. The paper compares AACT with GMEA and reports a reduction in attack time from thousands of seconds to tens of seconds. This is an attractive result, but the two methods differ fundamentally in attack information and optimization mechanisms: GMEA uses derivative-free evolutionary search, whereas AACT exploits gradients from a differentiable renderer and a public surrogate decoder. Consequently, the runtime difference reflects a change in the overall attack paradigm and cannot, by itself, establish a large efficiency gain attributable to AACT's target-construction strategy. The CC and BI comparisons under a shared optimization protocol provide a more appropriate way to isolate target construction, but the resulting gains are comparatively modest.

The paper also discusses related 3DGS watermark-removal work, including GSPure, without directly comparing it in the main results. Different attacks may use different access assumptions and threat models, so direct ranking is not necessarily appropriate. Nevertheless, the authors should explain why these methods are outside the comparison setting and limit cross-method efficiency claims accordingly. Required improvement: Separate the overall paradigm advantage over GMEA from the target-construction gains over CC and BI. Specify the hardware, timing protocol, and treatment of surrogate preparation and rendering overhead. Explain why other recent attacks are excluded from direct comparison.

Questions for authors

Q1. Generality of surrogate-owner alignment The controlled alignment experiments primarily use bitwise polarity flips. Can the authors evaluate AACT with public surrogate decoders from different architectures or training settings, and analyze whether its effectiveness depends on an approximately bitwise correspondence between surrogate and owner decoders? Q2. Attribution of efficiency gains How much of the runtime difference between AACT and GMEA comes from the differentiable surrogate-optimization paradigm, and how much comes from AACT's target construction? Please clarify the timing protocol and emphasize the gains over CC, BI, and random bit selection under a shared optimization protocol. Q3. Fidelity under difficult scenes The paper acknowledges local color artifacts in the Room scene. Can the authors provide per-scene fidelity results, worst-case statistics, and perceptual metrics such as LPIPS to establish that improvements in average PSNR and SSIM do not conceal substantial local distortions?

6 comments
AuthorsOpenReview ↗

Response to Reviewer GdS6 W1/Q1: Tests with Independent Decoders

Response to Reviewer GdS6 W1/Q1: Tests with Independent Decoders

Thank you for this suggestion. AACT remains effective with independently pretrained public decoders when the surrogate provides transferable guidance to the owner, and we agree that the original polarity-flip experiment alone cannot support such a claim. We therefore added cross-decoder evaluations (Appendix F.2, Table 14), gradient-transfer diagnostics (Appendix F.3, Table 15), and decoder provenance details (Appendix F.1, Table 13).

Which decoders we tested. We added two HiDDeN-family extractors from Stable Signature (Fernandez et al., ICCV, 2023), trained with standard and augmented settings, plus TrustMark100 (Bui et al., ICCV, 2025), CIN30 (Ma et al., ACM MM, 2022), and MBRS30 (Jia et al., ACM MM, 2021). They differ in weights, training settings, architectures, and message lengths. The original surrogate shares its weights with GaussianMarker and 3D-GSW, while the new surrogates use independent weights. The watermarked assets and owner decoders stay fixed. Input and loss adaptations are listed in Appendix F.1, and CIN30 and MBRS30 use up to 240 optimization steps with SNSS monitoring.

Results. The table reports means over 15 Blender/LLFF scenes and three seeds. GM48 and GS48 denote GaussianMarker and GuardSplat with 48-bit owner messages. P2P_2 is the owner's verification pass rate when full bit inversion is also allowed, so lower values mean stronger disruption.

OwnerPublic surrogateΔBAR ↓IDS ↑P₂ (%) ↓PSNR (dB) ↑
GM48Independent HiDDeN483.1580.613.1924.91
GM48Augmented HiDDeN4812.2475.1216.1824.32
GM48TrustMark10039.2121.5394.3425.10
GM48CIN3042.5714.7993.4723.48
GM48MBRS3039.5920.7493.1624.92
GS48Independent HiDDeN484.8085.450.4227.11
GS48Augmented HiDDeN484.6986.080.4926.17
GS48TrustMark1006.0383.061.3527.62
GS48CIN305.0083.280.6625.76
GS48MBRS306.9681.712.4727.36

All five surrogates reach P2P_2 below 2.5% on GS48, including 30-bit and 100-bit decoders attacking a 48-bit owner without any predefined bit mapping. On GM48, the two independent HiDDeN surrogates transfer strongly, while TrustMark, CIN, and MBRS transfer weakly. CC and BI also transfer weakly in these three settings, so the difference comes from the surrogate guidance rather than from AACT's target.

Why transfer differs across owners. We measure transfer post hoc with the initial gradient cosine and the owner-loss change after a small surrogate-guided step (Table 15). On GM48, the independent and augmented HiDDeN surrogates give cosines of 0.306 and 0.176 and owner-loss increases of 3.09e-4 and 1.70e-4, while TrustMark, CIN, and MBRS stay near zero, with increases of only 2–3e-6. On GS48, independent HiDDeN, TrustMark, and CIN give positive cosines near 0.05 and owner-loss increases of 2.5–2.9e-3. Compared within each owner, these measurements match the attack outcomes. Owner information is used only for this analysis, never for constructing or optimizing the attack.

Why target selection still matters. Even when surrogate guidance transfers, the attacker does not know whether CC or BI is the better target. With the original surrogate on GM48, CC and BI reach IDS 96.28 and 8.30 (Table 7). With the independent and augmented HiDDeN surrogates, the order reverses: 64.14 vs. 79.39 and 46.87 vs. 74.47. AACT avoids this choice by constructing a balanced half-Hamming target with carrier-aware bit selection from public surrogate gradients alone, without estimating the owner's bit mapping or querying owner verification, and reaches IDS 80.61 and 75.12. On GS48, it has the lowest mean P2P_2 among CC, BI, and AACT for each independent surrogate. In AACT, "alignment-agnostic" means that the attack target is constructed without knowing or estimating how the surrogate's output bits correspond to the owner's bits. Whether the attack then succeeds depends on how well surrogate gradients transfer to the owner, which the diagnostics above examine.

Revisions. Sec. 4.1 now states the shared decoder weights, and Sec. 4.2 describes Fig. 4 as a controlled polarity-remapping study.

AuthorsOpenReview ↗

Response to Reviewer GdS6 W2/Q2: Where the Speedup Comes From

Response to Reviewer GdS6 W2/Q2: Where the Speedup Comes From

Thank you for this suggestion. The speedup over GMEA mainly comes from differentiable surrogate optimization, while carrier-aware selection adds a measurable cost and improves attack effectiveness. We added a unified runtime comparison (Appendix H.2, Table 20), a stage-wise timing breakdown (Table 21), and equal-runtime evaluations (Appendix H.3, Table 22).

Timing protocol. GMEA and AACT run on the same NVIDIA RTX 4090 hardware. Timing starts when the attack process launches and ends after the attacked PLY is saved. It includes model and asset loading, rendering preparation, initial code estimation, target construction, optimization, stopping checks, and saving. Post-attack evaluation is excluded for both methods.

Speedup over GMEA. In Table 1, attack time drops from 5728.9 to 45.9 s on GaussianMarker, from 2181.2 to 35.1 s on GuardSplat, and from 2531.8 to 41.7 s on 3D-GSW, about 125×, 62×, and 61× faster. This gain mainly comes from the optimization paradigm. GMEA uses derivative-free evolutionary search with repeated candidate evaluation, whereas AACT optimizes Gaussian appearance parameters directly with surrogate gradients. We do not attribute this speedup to D-optimal selection.

Cost of each target strategy. To isolate target construction, we run CC, BI, Random, Strength, and Full (AACT) with the same surrogate and a fixed 80-step schedule. The times below average 15 scenes with seed 0 and include all preparation and saving.

OwnerCCBIRandomStrengthFull
GM4852.06 s54.25 s48.40 s52.77 s56.88 s
GS4837.82 s37.20 s35.55 s39.07 s42.11 s

Full costs more than the simpler strategies but stays within the same order of runtime. Table 1 reports AACT with SNSS, which typically stops earlier, so Full takes 56.88 s on GM48 here compared with 45.9 s in Table 1.

Where the added cost comes from. Table 21 separates the target-construction stages under the same setting, in seconds.

OwnerSelectionRef./CodePer-bit Grad.KernelSelection/Target
GM48Random4.541000.030
GM48Strength4.5054.25100.031
GM48Full4.5064.7502.9290.043
GS48Random3.752000.031
GS48Strength3.7403.51900.037
GS48Full3.7613.8232.3700.046

Most of the added cost comes from per-bit gradients and the carrier kernel. Greedy D-optimal selection itself takes less than 0.05 s. All of these costs are included in the totals above.

Results under the same time budget. We compare Random, Strength, and Full at total budgets of 11, 22, 32, and 43 s, including all preparation (Table 22). At 32 s on GS48, where all three strategies remain below their unconstrained results, Full reaches P2P_2 = 8.021%, compared with 10.521% for Random and 9.896% for Strength. Its IDS is 77.702, compared with 74.797 and 75.697, with comparable PSNR. Thus, Full achieves stronger owner-verification disruption even after its additional preparation cost is charged to the same budget. At 43 s, the results coincide with the unconstrained comparison in Table 18, where Full reaches P2P_2 = 5.729% versus 8.750% and 8.021%. Random can be advantageous under very short budgets because it avoids carrier analysis, while Full becomes more effective once sufficient optimization time is available.

Summary. Differentiable surrogate optimization provides the large speedup over GMEA, and carrier-aware selection improves attack effectiveness at a measurable added cost.

Response to Reviewer GdS6 W1: Attack Strength and Image Quality

We agree with this observation. On GS48 in Appendix Table 7, BI gives slightly stronger disruption than AACT (IDS 82.53 vs. 80.32), while AACT preserves higher image quality (PSNR 28.72 vs. 27.89 dB). This reflects a trade-off between disruption and image quality rather than a uniform advantage in attack strength.

AACT is designed to choose an effective target without knowing which endpoint suits the hidden owner. In Table 2, CC works better under high alignment (GaussianMarker and 3D-GSW), and BI works better under low alignment (GuardSplat). Across the three systems, AACT achieves the best average ΔBAR/IDS of 4.203/90.978, compared with 6.686/82.139 for CC and 27.314/43.775 for BI. AACT therefore delivers competitive disruption and image quality without choosing an endpoint based on alignment.

Response to Reviewer GdS6 W2: Why Other Attacks Are Not Directly Compared

Thank you for this comment. We use GMEA as the external baseline because it is a black-box 3DGS watermark attack, was previously evaluated on GaussianMarker, and can be assessed under our shared watermark-recovery protocol.

GSPure. GSPure (Huang et al., AAAI, 2026) mainly targets scene-hiding 3DGS watermarks and removes watermark-related Gaussians by pruning, whereas our evaluation focuses on rendered bitstream watermark verification. Its released implementation and evaluation protocol do not directly cover GaussianMarker, GuardSplat, or 3D-GSW, so a faithful comparison would require substantial adaptation.

Image-domain attacks. Regeneration (Zhao et al., NeurIPS, 2024), averaging (Yang et al., NeurIPS, 2024), and UnMarker (Kassis & Hengartner, IEEE S&P, 2025) modify individual rendered images and do not define how to map these changes back to a 3DGS asset. An asset-level comparison would first require designing and validating a multi-view refitting procedure for each method. We therefore treat them as complementary 2D attacks rather than direct 3DGS baselines.

Response to Reviewer GdS6 Q3: Image Quality in Difficult Scenes

Thank you for this suggestion. AACT preserves higher image quality than GMEA in every scene, and the largest remaining drop occurs on GS48 Room. We added per-scene PSNR, SSIM, and LPIPS for all 15 Blender/LLFF scenes (Appendix C.3, Tables 9–10) and updated Fig. 8 with measurements for the displayed Room views.

Per-scene results. AACT improves PSNR, SSIM, and LPIPS over GMEA on all 15 scenes for both GM48 and GS48, so the average advantage is not driven by a few favorable scenes. The table reports scene averages with seed 0, measured against dataset ground-truth images.

OwnerMethodPSNR (dB) ↑SSIM ↑LPIPS ↓
GM48No attack27.070.89930.0874
GM48GMEA23.330.74770.2356
GM48AACT25.570.84310.1656
GS48No attack32.400.96600.0277
GS48GMEA22.310.73380.2276
GS48AACT28.600.90200.0995

AACT improves average PSNR over GMEA by 2.24 dB on GM48 and 6.29 dB on GS48.

Worst cases and local artifacts. The largest scene-level PSNR drops from the unattacked asset are 2.47 dB on GM48 Lego and 7.60 dB on GS48 Room. On GS48 Room, AACT gives PSNR/LPIPS of 28.51/0.1763, compared with 20.90/0.4365 for GMEA. AACT keeps the Gaussian geometry and primitive set fixed but modifies SH appearance coefficients, which can introduce local color deviations. The updated Fig. 8 shows matched Room views and enlarged regions for GM48 and GS48, annotated with PSNR, LPIPS, ΔBAR, and IDS. Together with Tables 9–10, it shows both the remaining local artifacts and the clear quality advantage over GMEA.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response to Reviewers

We thank Reviewers GdS6, nHaD, and NV7r for their careful reviews. Their comments helped us sharpen the scope of AACT and strengthen the supporting evidence. The main changes are summarized below, and all new material is highlighted in blue in the revised paper.

  1. Independent public decoders (GdS6, nHaD, NV7r). We added five independently pretrained surrogates with different architectures, training settings, and message lengths. All five achieve low owner-verification pass rates on GuardSplat, and the two independent HiDDeN variants also transfer strongly to GaussianMarker. Gradient diagnostics relate these outcomes to how well surrogate guidance transfers to each owner. In AACT, "alignment-agnostic" means that the attack target is constructed without knowing or estimating how the surrogate's output bits correspond to the owner's bits. The original surrogate shares publicly released weights with GaussianMarker and 3D-GSW, and with independent HiDDeN surrogates AACT still outperforms GMEA on GaussianMarker in all four metrics (Appendix F, Tables 13–15).

  2. Why we flip half the bits (nHaD). We added a geometric motivation and flip-ratio tests across watermarking systems, decoder families, and owner message lengths. A 50% flip is equidistant from the initial surrogate code and its full inversion, which gives a balanced fixed target without owner feedback (Appendix G, Tables 16–17).

  3. Value and cost of carrier-aware selection (GdS6, NV7r). We added per-scene comparisons, three-seed variability, a stage-wise timing breakdown, and equal-runtime comparisons. The speedup over GMEA mainly comes from differentiable surrogate optimization. Full selection adds a measurable cost and lowers the verification pass rate on GuardSplat once sufficient optimization time is available (Appendix H, Tables 18–22).

  4. Owner verification and rendering fidelity (GdS6, NV7r). We evaluated GMEA under the same inversion-aware verification protocol and added per-scene PSNR, SSIM, and LPIPS. AACT achieves stronger verification disruption on GaussianMarker, GMEA does so on GuardSplat, and AACT preserves higher fidelity in every tested scene. The updated Room example shows the remaining local artifacts (Appendix C, Tables 7–10, Fig. 8).

  5. Scope of baseline comparisons (GdS6, nHaD). GMEA remains the external baseline because it attacks bitstream 3DGS watermarks under our verification protocol. We explain why GSPure (Huang et al., AAAI, 2026) and image-domain removal attacks require additional adaptation for an asset-level comparison.

Main-text revisions. Without changing any experimental results, Sec. 4.1 now states the shared decoder weights, Sec. 4.2 characterizes Fig. 4 as a controlled polarity-remapping study, and Sec. 4.3 points to Appendices G and H. We will also bring the appendix results into Table 1 (mean ± population SD) and Table 3 (P₂).

We appreciate the reviewers' time and welcome further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.