acceptodds
Under review as a conference paper at ICLR 2027

Latent Angular Watermarking with Provable Marginal Gaussianity and Characterized Correlation

Abstract

Latent domain watermarking for diffusion models embeds watermarks directly into the latent prior, enjoying non-intrusiveness to model parameters and seamless integration with the generation process. However, due to the violation of latent marginal Gaussianity or sensitivity to normal and malicious perturbations during latent inversion, existing methods are prone to watermark detection or removal attacks. A further overlooked problem is the violation of the i.i.d. latent condition after watermarking, which leads to latent correlation degradation and generation fidelity loss. Although this has been externally measured by FID, the internal correlation structure has yet to be rigorously characterized. To address the above issues, and motivated by the rotation-invariant property of isotropic Gaussian, we propose Latent Angular Watermarking (LAW), which encodes watermark bits as antipodal angles ( relative to reference pairs) between disjoint pairs of latent elements while preserving marginal Gaussianity. The antipodal (-separation) encoding maximizes geometric separation between bit values, providing maximal tolerance against normal and malicious latent distortions. We further propose a magnitude-driven variant, LAW-M, which anchors watermark bits in the most geometrically stable latent pairs, yielding additional robustness gains. Theoretically, we prove that the decoding angular error is inversely proportional to the norm of the latent pair, i.e., , and provide a rigorous characterization of the induced correlation degradation by deriving the closed-form covariance matrix of the watermarked latent, demonstrating that correlations are confined to a sparse, structured set of off-diagonal elements with fixed values. Extensive experiments demonstrate state-of-the-art robustness against a broad range of post-processing and regeneration attacks, together with the superior image fidelity among existing methods.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Model Watermarking, IP Protection, AIGC Forensics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GdNVOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces LAW, a training-free watermarking method for latent diffusion models. The initial latent vector is partitioned into encoding pairs, corresponding reference pairs, and an unmodified remainder. Each bit is embedded by rotating an encoding pair so that its signed angle relative to the corresponding reference pair is either +π/2 or −π/2. Following DDIM inversion, the bit is recovered from the sign of this relative angle. A variant, LAW-M, embeds bits in the pairs with the largest magnitudes and retains the per-image sorted indices as side information.

The paper proves that each coordinate of the watermarked latent vector has a standard normal marginal distribution, N(0,1), and derives a closed-form expression for its covariance matrix, with cross-covariance entries of ±π/4 between the relevant paired coordinates. It also establishes a first-order angular-error bound that scales inversely with the pair magnitude ρ.

Experiments use Stable Diffusion v2.1 with prompts from COCO and DiffusionDB and compare LAW against 12 baseline methods. The evaluation covers image quality, measured by FID, CLIP score, and IS; robustness to 11 post-processing and regeneration attacks; classifier-based watermark detectability; and runtime. Additional ablation studies examine the effects of model version, scheduler, inversion step count, and guidance scale.

Critical strengths

S1. Simple and efficient encoding with competitive empirical robustness. The method embeds watermarks through norm-preserving rotations of latent pairs and requires no training. The runtime results in Fig. 5 indicate that it is orders of magnitude faster than PRC under the evaluated settings. The closed-form covariance expression in Property 2 is correct. The results across 11 attacks and the ablation studies in Tables 3–6 suggest competitive robustness that is reasonably consistent across the tested configurations. In particular, LAW-M performs within approximately 1.5 percentage points of the best-performing baseline on 10 of the 11 attacks. These results constitute the paper’s principal strength, although they do not sufficiently address the concerns raised in W1 and W2.

Critical weaknesses

W1. The theory does not support the central stealthiness claim and reveals substantial distributional deviations. Property 1 establishes Gaussian one-dimensional marginals, not an i.i.d. Gaussian joint distribution. Eqs. (2)–(3) impose exact orthogonality between encoding and reference pairs; Property C.2 identifies a two-dimensional marginal with a zero-probability quadrant; and Property 2 gives correlations of magnitude π/4≈0.785\pi/4 \approx 0.785. With R=7R=7 and L=512L=512, 87.5% of latent coordinates participate in these dependencies. Covariance sparsity does not imply a small distributional deviation. The main formulation is keyless (App. I), and the experimental partition is fixed and identifiable by the attacker (App. G). An algorithm-aware adversary can therefore run the public extraction procedure. Moreover, the mean of cos⁡2Δϕi\cos^2\Delta\phi_i is exactly zero for original watermarked latents but has expectation 1/21/2 under i.i.d. Gaussian latents. Its effectiveness after inversion requires testing, but Fig. 3’s ResNet-18 and MLP results do not exclude such structure-aware detection. Public embedding also permits attempted chosen-message forgeries through inversion and regeneration, an authenticity limitation outside the stated threat model. App. I’s secret permutation hides the pairings without removing their dependencies. Reusing a key and message preserves signed correlations, potentially exposing pairings through covariance estimation. Balanced random messages cancel these correlations in expectation, but fourth-order dependencies remain: E[(x′i,2)2ri,12]=3/2\mathbb{E}[(x'{i,2})^2r{i,1}^2]=3/2, versus 11 for independent Gaussian coordinates. Neither attack is evaluated. Furthermore, if Spherical Watermark matches all joint moments through order three, it already matches covariance, making Sec. 1’s criticism misleading. These issues undermine a principal motivation and Table 1’s unqualified stealthiness claim. The current encoding does not preserve the joint Gaussian distribution; practical stealthiness requires evaluation against structure-aware detectors.

W2. The robustness claims exceed the evidence, and LAW-M requires a different verification setting. Contrary to the abstract and Sec. 4.2, LAW trails the best latent-domain baseline in all 11 attack columns and ROBIN in nine. LAW-M is strictly best only on Diff (96.74 versus 96.52). The concern is the claimed superiority, not the absence of uniformly state-of-the-art performance. LAW-M stores 2L2L ordered indices from 8,192 pairs. At L=512L=512, straightforward storage requires 13,312 bits per image—about 38 times a 256-bit key plus 96-bit nonce, which the paper characterizes as substantial overhead for Gaussian Shading. Fig. 8 demonstrates dependence on this metadata, yet the paper does not explain how verifiers retrieve it for unknown images or control false positives when searching candidate records. LAW-M’s robustness and TPR at 1% FPR are therefore conditional on side information and are not directly comparable to blind verification without a common retrieval protocol. Finally, Sec. 4.1’s stated 512-bit payload conflicts with Table 1’s lower-capacity and zero-bit methods, while Table 2 does not clarify whether entries report bit accuracy or detection TPR. These ambiguities prevent a consistent comparison. Small FID differences (43.39, 43.55, and 44.36), without repeated-run variability, likewise do not establish a reliable fidelity advantage.

Questions for authors

Q1. Detection by an informed adversary. Please report image-level AUC and TPR@1%FPR using the extractor or mean cos⁡2Δϕi\cos^2\Delta\phi_i to distinguish watermarked from clean images for (a) keyless LAW and (b) keyed LAW (App. I), after estimating the hidden pairing from NN images sharing a key and message, for several values of NN. Evaluate detection on held-out images. Near-perfect detection in (a) would undermine the stealthiness claims in Table 1 and Sec. 4.2 under Sec. 3.1’s algorithm-aware threat model.

Q2. LAW-M verification protocol. How does the verifier obtain KK for a query image? If verification searches stored index sets, please report the search cost, aggregate false-positive rate, and TPR at an aggregate FPR of 1%.

Q3. Interpretation of “0%” accuracy. Are App. G’s “0%” extraction accuracy and Sec. 4.3’s “nearly zero” LAW-M-D accuracy bit accuracies or detection TPRs? For binary messages, 0% bit accuracy implies complete inversion and permits exact recovery by flipping every bit. Please specify the adaptive attack’s coordinate modifications and regeneration procedure, including the prompt.

Q4. Source of LAW-M’s FID improvement. Sec. 4.2 attributes the improvement to R=1R=1, but LAW and LAW-M also differ in magnitude-based sorting. Please report LAW’s FID at R=1R=1 to separate these effects.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3

We thank the reviewer for the helpful suggestions.

For W1:

Regarding watermark removal attacks, in Section 4.2 (Main Results: Classifier Indistinguishability), we empirically demonstrate that watermarked and non-watermarked samples cannot be distinguished by the classifiers, making classifier-based surrogate attacks practically infeasible. Furthermore, in Appendix G (Adaptive Attack), we show that an attacker with full knowledge of the LAW watermarking algorithm cannot successfully remove the watermark without significantly degrading visual fidelity.

Regarding stealthiness, we acknowledge that our method does not preserve the joint Gaussian distribution. However, distributional deviations do not necessarily imply a lack of practical stealthiness. Detecting such higher-order statistical dependencies or distributional deviations may require collecting a substantial number of generated images, which can be challenging in realistic scenarios. We have therefore refined our threat model and narrowed our stealthiness claims to scenarios where attackers cannot collect a large number of watermarked images. Moreover, even with sufficient watermarked images to detect the watermark, an attacker cannot remove it without significantly degrading image quality, as shown in Appendix G. Notably, Spherical Watermark only establishes the preservation of joint moments up to the third order, meaning that higher-order statistical deviations may still be exploited for detection.

Perfectly preserving the joint Gaussian distribution is inherently challenging, as watermarked latents would be statistically indistinguishable from randomly sampled Gaussian latents. Although Gaussian Shading preserves Gaussianity, it essentially stores watermark evidence in image-specific keys, making practical deployment difficult.

Regarding forgery attacks, we acknowledge this limitation of LAW, which is also a broader challenge for latent-domain watermarking. We will discuss this limitation and potential defenses inspired by SEAL in the revised manuscript.

Regarding the covariance properties of Spherical Watermark, we acknowledge the error and will correct the corresponding discussion in the revised manuscript.

For W2:

Regarding robustness, we acknowledge that LAW performs slightly worse than RingID, HSQR, and ROBIN under some of the 12 evaluated attacks, although the performance gaps are small. We will revise our claims of state-of-the-art robustness accordingly. Notably, LAW achieves state-of-the-art robustness among Gaussianity-preserving watermarkings. Moreover, its primary limitation lies in rotation robustness. Excluding rotation attacks, LAW achieves state-of-the-art performance among the evaluated latent-domain watermarkings. Rotation robustness could be further improved by embedding watermarks in symmetric patterns, such as rings.

Regarding storage overhead, we acknowledge that LAW-M requires more per-image storage than Gaussian Shading. However, as stated in the original manuscript, LAW-M is designed "for high-stakes scenarios to protect important single images, such as those of public figures." Such scenarios typically involve relatively few images, whereas Gaussian Shading requires storing an image-specific key for every generated image. Therefore, considering their different application scenarios, we believe our characterization of storage overhead is reasonable.

Regarding metadata retrieval, during embedding, the selected indices (K) can be stored in a server-side database and associated with a unique image identifier. During extraction, these indices can be retrieved using the identifier to reconstruct the original encoding and reference pairs. Thus, assuming the image identifier is available, verification does not require searching through candidate records, avoiding additional false positives caused by such searches. We will clarify this verification setting in the revised manuscript.

Regarding watermark capacity, the 11-bit, 72-bit, and 256-bit capacities in Table 1 correspond to the original configurations of RingID, HSQR, and Gaussian Shading, respectively. For fair comparison, we standardized the watermark length to 512 bits in our main experiments. We will clarify this distinction in the revised manuscript. Table 2 reports detection TPR, which will also be clearly stated in the revised manuscript.

Regarding FID, we agree that repeated-run variability should be reported to establish the reliability of fidelity comparisons. We have therefore additionally evaluated the FID of LAW and LAW-M across three different random seeds in the next table.

Table1: FID results across three random seeds.

MethodSeed 1Seed 2Seed 3
LAW44.3944.2544.18
LAW-M43.3943.3543.41
AuthorsOpenReview ↗

Response to Reviewer 3

For Q1: Regarding (a), we have already evaluated classifier indistinguishability in Section 4.2 (Main Results: Classifier Indistinguishability). Regarding (b), we have addressed the stealthiness concerns in our response to W1.

For Q2: We have addressed this concern in the Metadata Retrieval discussion of our response to W2. Notably, Gaussian Shading also relies on image-specific keys but does not specify how these keys are retrieved during verification. We consider this a database retrieval issue, where the computational cost, retrieval time, and retrieval accuracy depend on the database implementation rather than the watermarking method itself.

For Q3: The reported 0% accuracy refers to detection TPR, rather than bit accuracy. For the adaptive attack, we randomly select half of the encoding pairs and negate both elements in each selected pair, which means that their angles are rotated by 180°, thereby disrupting watermark extraction. To evaluate the worst-case scenario, we assume the attacker has access to the original prompt and uses the same prompt and Stable Diffusion configuration, including the guidance scale, for image regeneration. We will clarify these details in the revised manuscript.

For Q4: We acknowledge that the contributions of magnitude-based sorting and the repetition factor have not been sufficiently disentangled. We will include additional ablation experiments in the revised manuscript to independently evaluate their effects on FID.

We welcome any further questions from the reviewer and would be pleased to provide additional clarifications in further discussion.

Thank you for the clarifications on Q3, the Table 2 metric, the message lengths, and Spherical Watermark. My main concerns remain:

  1. Detection (W1, Q1). The main LAW has no key and a fixed partition, so under the algorithm-aware threat model of Sec. 3.1 the paper's own extractor can be applied to a single image; the many-images argument applies only to the keyed variant in App. I. Resistance to removal (App. G) is a separate property from indistinguishability, and the requested AUC and TPR@1%FPR are not provided in the response.

  2. Robustness (W2). In Table 2, LAW is below the best latent-domain baseline in all 11 reported columns, none of which is rotation (34.97 points on Drop, 11.27 on Diff), so "state-of-the-art excluding rotation" does not match the table. LAW is also below Gaussian Shading in 9 of the 11 columns.

  3. LAW-M verification (W2, Q2). Verification assumes the image identifier is available. For an image of unknown origin, such as a synthetic image of a public figure found online, the identifier must travel as strippable metadata or be embedded as another watermark, so LAW-M's TPR relies on side information that blind methods do not use.

The new FID runs cover only LAW and LAW-M, so the baselines are still single-run; Q4 is deferred to the revision.

I keep my score.

We thank the reviewer for their response. We would like to further clarify the following points.

1. Detection (W1, Q1).

We reiterate that we have already discussed the secret key for LAW and LAW-M in Appendix I, Discussion on Watermark Secret Key, of our original submission. Since introducing a secret key does not affect the core conclusions of this paper, we consider both LAW and LAW-M are equipped with a secret key by default.

Regarding Q1, (a) in practical deployment, a secret key shared across images can be introduced into LAW and LAW-M to randomly permute the encoding and reference regions. Consequently, an attacker cannot directly identify the encoding pairs and determine the presence of a watermark from a single image. (b) The proposed evaluation requires collecting multiple watermarked images sharing the same user, which contradicts our threat model.

In your response, you stated that "the many-images argument applies only to the keyed variant in App. I," which indicates that you have already acknowledged the stealthiness of both LAW and LAW-M when equipped with secret keys. Therefore, we believe that reporting AUC and TPR@1%FPR is unnecessary in this case.

2. Robustness (W2).

In the following two tables, we report the average TPR of each watermarking method across all 11 attacks, as well as the average TPR across the 10 attacks excluding random drop. We emphasize that, despite LAW-M's substantially weaker performance against the Drop attack compared with existing robust baselines, LAW-M still achieves highly competitive overall robustness among all latent-domain watermarking methods. Furthermore, LAW demonstrates superior robustness compared with all existing Gaussianity-preserving methods. Across the 10 attacks excluding random drop, LAW-M achieves state-of-the-art robustness among all latent-domain watermarking methods. We reiterate that, although LAW underperforms Gaussian Shading in 9 out of the 11 attacks, LAW still achieves higher average robustness than Gaussian Shading across all 11 attacks. We believe that the reviewer's expectation that our proposed method should outperform the strongest existing baseline under every individual attack is excessively demanding and overly critical.

Table2: Average Robustness Performance across 11 Attacks

MethodAverage (11 Attacks)
Tree-Ring63.45
RingID97.26
HSTR77.53
HSQR97.44
ROBIN98.47
Gaussian Shading92.72
PRC61.46
Spherical Watermark83.17
LAW93.43
LAW-M96.11

Table3: Average Robustness Performance across 10 Attacks

MethodAverage (10 Attacks)
Tree-Ring67.40
RingID98.11
HSTR82.31
HSQR98.52
ROBIN98.53
Gaussian Shading92.79
PRC65.95
Spherical Watermark85.23
LAW96.48
LAW-M98.86

3. LAW-M Verification (W2, Q2).

In LAW-M, the image identifier can be stored as image metadata for subsequent verification. We also emphasize that the retrieval of per-image information is primarily a database management and metadata robustness issue rather than an issue specific to the watermarking method itself. Similarly, Gaussian Shading also relies on per-image information for watermark decoding, yet its paper does not provide any algorithm for retrieving the corresponding secret keys. Nevertheless, Gaussian Shading was accepted at CVPR 2024.

4. FID Experiments and Q4.

We are currently conducting additional experiments and will include the results in the appendix before the final acceptance decision.

Overall, we would like to reiterate that our claims regarding stealthiness and robustness remain valid, as does the use of per-image information in LAW-M.

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

This paper proposes two latent-domain watermarking methods for diffusion models, LAW and LAW-M. LAW encodes watermark bits through antipodal angular relationships between latent vector pairs, preserving marginal Gaussianity while improving robustness against perturbations. The authors further characterize the induced covariance structure and propose LAW-M, which selects high-magnitude latent pairs to enhance robustness. Extensive experiments demonstrate competitive performance in watermark robustness and image fidelity.

Critical strengths
  1. Simple and effective methodology.
    LAW introduces a straightforward angular encoding mechanism without requiring additional training or modifications to the generative model. The method is computationally efficient and achieves competitive experimental performance.

  2. Thorough theoretical analysis and comprehensive validation.
    The paper provides detailed theoretical derivations of marginal Gaussianity, covariance structure, and angular perturbation errors. These analyses are supported by extensive comparisons and ablation studies, making the work informative and well-substantiated.

Critical weaknesses
  1. Unclear and AI-alike abstract.
    The abstract has too much technical details that are not important for abstract and beyond common readers' knowledge. E.g., "antipodal angles" and parameters and equations in brackets. It should be concise with the main motivation, problem, and contribution, stressing why this work is important.
  2. Differences with existing methods are not clear.
    The proposed method follows existing methods like Spherical Watermark, but its difference and improvements over closely related methods are not clear or prominent.
  3. Limited performance improvements over existing methods.
    Although LAW-M achieves competitive performance, its improvements over existing methods, particularly HSQR, are relatively modest. For example, its FID only improves from 43.55 to 43.39, while its robustness remains inferior to HSQR under several attacks.
  4. Insufficient evidence supporting the watermark stealthiness claims.
    Although LAW preserves marginal Gaussianity, Equations (2)–(3) enforce an orthogonal relationship between each encoding pair and its reference pair. Such exact orthogonality occurs with probability zero for independent Gaussian vectors. Under the threat model in Section 3.1, where the adversary knows the watermarking algorithm, this relationship could enable a statistical detector when the latent pairing is known. The existing classifier experiments do not sufficiently rule out such structured detection, particularly after DDIM inversion.
  5. Unclear relationship between watermark capacity and experimental configurations.
    Table 1 reports watermark capacities of 11, 72, and 256 bits for RingID, HSQR, and Gaussian Shading, respectively, whereas Section 4.1 states that all these methods use 512-bit watermarks.
Questions for authors
  1. Regarding the image-specific indices K in LAW-M.
    As indicated in Table 1, LAW-M requires additional image-specific information K for watermark extraction. How is K stored and retrieved in practical deployment, especially when the original image information is unavailable? Could the authors consider embedding K into the watermark itself to enable self-contained extraction?
  2. Regarding watermark capacity.
    Could the authors clarify the difference between the watermark information capacity reported in Table 1 and the 512-bit watermark length used for the baselines in Section 4.1? Additionally, could the authors clarify whether the LAW-M results reported in the main experiments correspond to a watermark capacity of 4096 bits?
2 comments
AuthorsOpenReview ↗

Response to All Reviewers and the Area Chair

We hope that the reviewer and the AC can read our response to this comment below. We would particularly like to clarify the concerns raised by Reviewer vmSE regarding the use of AI-generated text.

  1. On “AI-alike”. This is an unsupported accusation. The Abstract is manually written and proofread by all authors of this article. The reviewer gives no evidence for the claim, and speculation about authorship is not a valid criterion for judging a scientific submission. We ask that this remark be disregarded in the evaluation.

  2. On “too much technical detail”. We are surprised that the amount of technical content in an abstract is offered as a weakness. Every technical element in our abstract is there for a reason. The antipodal encoding is the core design principle. Δφ ∝ 1/ρ is our main theoretical result. The ±π/4 covariance structure is the first rigorous characterization of latent correlation degradation, a problem the abstract shows existing work has missed. Stating main results precisely is standard practice at ICLR. Even if the reviewer prefers a different style, that is a presentation preference. It says nothing about the paper’s technical soundness, novelty, or significance, and should not weigh on the assessment. We note that the review raises no concern about the correctness of our theory or the validity of our experiments. If the reviewer has such concerns, we would be glad to address them.

  3. On “antipodal angles” being beyond common knowledge. “Antipodal” is standard terminology in signal processing, coding theory, and geometry. It has been used for decades, for example in antipodal signaling and BPSK. We do not think it is beyond the knowledge of the ICLR audience. It is also the core design principle of LAW: antipodal encoding maximizes the geometric separation between bit values. Removing it from the abstract would hide the main idea of the paper.

  4. On “unclear.” The abstract states the motivation, the problem, and the contribution explicitly: Motivation: Latent-domain watermarking is non-intrusive to model parameters and integrates seamlessly with generation. Problem: Existing methods violate marginal Gaussianity or are fragile under latent inversion. They also break the i.i.d. latent condition, causing correlation degradation that has never been rigorously characterized. Contribution: We propose LAW and LAW-M, which encode bits as antipodal angles and preserve marginal Gaussianity. We prove the decoding error bound and derive the closed-form covariance of the watermarked latent. Significance: LAW achieves state-of-the-art robustness and fidelity. We would appreciate it if the reviewer could point to a specific sentence that is unclear rather than offering a general impression.

  5. On parameters and equations in brackets. The relation Δφ ∝ 1/ρ and the fixed ±π/4 correlation structure are the paper’s main theoretical results. Stating main results concisely in the abstract is common practice at ICLR and similar venues. That said, we are willing to move the parenthetical notation, (±π/2 …) and (π-separation), to the introduction, which shortens the abstract without losing content.

We respectfully ask the reviewer to base the assessment on the technical merits of the work. We remain happy to address any concrete technical concerns.

AuthorsOpenReview ↗

Response to Reviewer 2

For W2: We have explicitly clarified the differences between LAW, LAW-M, and existing methods in Lines 57–64. Although our method and Spherical Watermark both exploit the rotational invariance of Gaussian distributions, their core watermark embedding mechanisms are fundamentally different. Furthermore, across 11 attacks, LAW and LAW-M achieve average TPRs of 93.43% and 96.11%, respectively, significantly outperforming Spherical Watermark (83.17%). Both LAW and LAW-M also achieve better FID scores than Spherical Watermark, demonstrating improvements in both watermark robustness and visual fidelity.

For W3: Overall, LAW-M and HSQR achieve comparable robustness. The main limitation of our method is its relatively lower robustness against rotation attacks. However, this could potentially be improved by embedding watermarks into latent dimensions corresponding to circular regions or by employing additional repetition. More importantly, we emphasize that, unlike HSQR, our method provides a theoretical guarantee of marginal Gaussianity and explicitly characterizes the induced correlations. Therefore, our contribution extends beyond empirical improvements in robustness and visual fidelity to providing a theoretically grounded watermarking.

For W4: We have already discussed this scenario in Appendix G (Adaptive Attack), where we consider an adversary who is aware of the watermarking algorithm and exploits the known latent partitioning structure, further supporting our claim of watermark stealthiness.

For W5: The watermark capacities of 11, 72, and 256 bits reported in Table 1 correspond to the original configurations of RingID, HSQR, and Gaussian Shading, respectively. To ensure a fair comparison, we standardized the watermark length to 512 bits in our main experiments. We will add further clarification in the revised manuscript to avoid confusion.

For Q1: Regarding the storage and index retrieval procedure of LAW-M, for a latent dimension of (D=16,384) and a watermark length of (L=512), a total of (2L=1,024) indices must be stored. Since there are (D/2=8,192) latent pairs, each index requires 13 bits. Therefore, the total storage overhead is 13,312 bits per image. During watermark embedding, the selected indices (K) can be stored in a server-side database and associated with a unique image identifier. During extraction, the corresponding indices can be retrieved using this identifier and used to restore the original encoding and reference pair arrangement before watermark decoding. For comparison, Gaussian Shading also requires image-specific information when using independent keys and nonces for different images. Under the standard IETF ChaCha20 configuration, a 256-bit encryption key and a 96-bit nonce require a total of 352 bits per image. However, the Gaussian Shading paper does not explicitly describe how the corresponding key and nonce are retrieved during watermark decoding. We acknowledge that LAW-M requires more per-image information than Gaussian Shading. However, as emphasized in our manuscript, LAW-M is specifically designed for high-stakes scenarios involving the protection of important individual images, such as those of public figures. In such scenarios, the number of protected images is expected to be relatively limited, making the additional storage overhead manageable. In contrast, Gaussian Shading is designed for broader deployment, where image-specific key management may need to scale to a much larger number of generated images, potentially introducing greater database management complexity. Therefore, we believe that, for the targeted high-stakes scenarios, the additional storage of approximately 1.625 KiB per image represents a reasonable trade-off for achieving improved watermark robustness and image quality. Regarding the suggestion of embedding (K) into the watermark itself, we have explored this approach experimentally. However, we observed a substantial decrease in TPR. The main reason is that errors in recovering the image-specific information (K) can lead to incorrect latent pair reconstruction, resulting in watermark decoding failure. Therefore, we retain (K) as separately stored information to ensure reliable watermark extraction.

For Q2: The LAW-M results reported in the main experiments correspond to a 512-bit watermark, rather than 4096 bits. We have already addressed the remaining questions in our response to W5.

We welcome any further questions from the reviewer and would be pleased to provide additional clarifications in further discussion.

Reviewer_79G3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LAW, a training-free watermarking method that embeds bits into the relative angles between pairs of elements in the initial latent noise. It preserves the marginal Gaussian distribution of each element and analyzes the covariance structure induced by the embedding. A variant, LAW-M, leverages element pairs with larger magnitudes to enhance robustness at the cost of storing image-specific sorting indices. DDIM inversion is employed for extraction.

Critical strengths
  1. The problem formulation and design are clear. Focusing on the shift in latent distribution caused by watermarking and utilizing the rotational invariance of Gaussians to preserve marginal distributions is a positive and sound approach.

  2. The scope of distribution preservation is analyzed with clear distinctions. Beyond the marginal distribution preservation of LAW, the paper also presents the covariance structure at R=1R=1 and the non-Gaussianity of certain multi-dimensional marginal distributions. This is valuable in that it distinguishes marginal distribution preservation from full joint distribution preservation.

Critical weaknesses
  1. The evaluation supporting the claim of stealthiness is limited. While the paper demonstrates marginal distribution preservation and the failure of specific classifiers to distinguish watermarked latents, it also acknowledges inter-element dependencies and a non-Gaussian joint distribution. It remains unverified whether stealthiness holds against an algorithm-aware attacker utilizing structural statistics such as relative angles. Furthermore, it should be clarified whether the latent classification experiments use the noise immediately after embedding or the actual inversion results. Drawing a conclusion of more general stealthiness from the current results alone is difficult.

  2. The experimental scope and comparisons do not sufficiently substantiate performance and applicability. While the comparisons across SD v1.4, v2.0, and v2.1 in the appendix are useful, applicability to DiT-based diffusion or flow matching models such as SD3 and FLUX has not been verified. Given that related methods utilizing initial Gaussian noise evaluate such extensions, and LAW relies on inversion, it is necessary to examine performance across variations in architecture and generative processes. Additionally, in Table 2, LAW exhibits lower robustness than several baselines, and the FID improvement in LAW-M is not sufficiently disentangled, as both sorting and repetition factors vary simultaneously. The specific storage overhead and index retrieval procedure for LAW-M must also be clarified.

Questions for authors
  1. Are the latent inputs in Figure 3 the noise immediately after embedding or the results of DDIM inversion? Can you provide detection results using relative angle statistics on the actual inverted latents? Please also specify whether a secret key is applied and its configuration.

  2. What is the per-image index storage overhead for LAW-M, and what is the retrieval method for matching the index to a query image? To disentangle the cause of FID improvement, please provide the FID/CLIP/IS scores for LAW (R=1R=1).

  3. Do the extraction performances in Table 2 and Appendix G represent ACC or TPR? If 0% in Appendix G corresponds to bit-wise extraction accuracy (ACC), how do you distinguish between full bit inversion and complete watermark removal? Please also clarify the perturbation and regeneration procedures for the attacks.

  4. What is the rationale for excluding RingID, HSTR, HSQR, ROBIN, and LAW-M from Figure 3? Please add results for the key methods necessary for a comprehensive stealthiness comparison.

  5. Can you provide the embedding/extraction performance, generation quality, and attack robustness of LAW on DiT-based diffusion and flow matching models such as SD3 and FLUX? Please also specify the inversion procedure used for each model and any necessary modifications. These results are critical for evaluating the broader applicability of the method beyond comparisons within the SD family.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 1

We thank the reviewer for the helpful suggestions.

For W1: In Appendix G, we consider an algorithm-aware attacker who attempts to disrupt the watermark by directly modifying the dimensions used for watermark encoding to randomly flip the embedded bits. Our experimental results show that such modifications lead to severe degradation in visual fidelity, further supporting the stealthiness of our method. Regarding the latent classification experiments, we clarify that they are conducted on the noise immediately after watermark embedding, rather than the inversion results. We have added this clarification to the revised manuscript.

It is worth noting that a strict i.i.d. Gaussian distribution after watermarking is infeasible (nothing is changed so no watermark can be embedded). Our proposed method is to date the embedding that is most close to it, i.e., achieving marginal distribution preservation, whose benefits is improved visual fidelity. As such, our method has better stealthiness over prior works.

For W2:

Regarding the generalizability to different model architectures, we acknowledge that the applicability of our method to DiT-based diffusion models or flow matching models, has not yet been verified. Following the experimental settings of Gaussian Shading, we evaluated the generalizability of our method across three versions of Stable Diffusion (SD v1.4, v2.0, and v2.1). It is also worth noting that another closely related method, PRC, was evaluated only on SD v2.1. We have acknowledged this limitation in the revised manuscript and included evaluations on DiT-based diffusion and flow matching models as future work.

Regarding robustness, we would like to emphasize that LAW achieves state-of-the-art robustness among Gaussian-preserving methods. Across 11 attacks, Gaussian Shading, PRC, and Spherical Watermark achieve average robustness scores of 92.72%, 61.46%, and 83.17%, respectively, whereas LAW achieves 93.43%, and LAW-M further improves this result to 96.11%.

Regarding the FID improvement of LAW-M, we acknowledge that the contributions of magnitude-based sorting and the repetition factor have not been sufficiently disentangled. We will include additional ablation experiments in the revised manuscript to independently evaluate their effects on FID.

Regarding the storage overhead and index retrieval procedure of LAW-M, for a latent dimension of (D=16,384) and a watermark length of (L=512), a total of (2L=1,024) indices must be stored. Since there are (D/2=8,192) latent pairs, each index requires 13 bits, the total storage overhead is 13,312 bits per image. Note that this is a contant value, not dependent on the number of images protected. During watermark embedding, the selected indices (K) can be stored in a server-side database and associated with a unique image identifier. During extraction, the corresponding indices can be retrieved using this identifier and used to restore the original encoding and reference pair arrangement before watermark decoding.

For comparison, Gaussian Shading requires image-specific information when using independent keys and nonces for different images. Under the standard IETF ChaCha20 configuration, a 256-bit encryption key and a 96-bit nonce require a total of 352 bits per image, which grows linearly with the number of images. If Gaussian Shading protects more than 37 images, then its storage overhead will exceed LAW-M’s.

It is also important to note that LAW-M is created to trade storage overhead for additional robustness over LAW, so if offers a more flexible solution to deal with different scenarios. With LAW, there is zero storage overhead.

For Q1: The latents used in Figure 3 are the initial noise immediately after watermark embedding. It is worth noting that Figure 3 aims to demonstrate the classifier-based indistinguishability between watermarked and standard Gaussian noise. Therefore, we believe that using the noise immediately after watermark embedding is a more appropriate experimental setting for this purpose. In our main experiments (Table 2), watermark detection is indeed performed on the actual DDIM-inverted latent using relative angle statistics. Regarding the secret key and its configuration, we have provided a discussion in Appendix I.

For Q2: Please see answer to W2.

For Q3: The watermark extraction performance reported in Table 2 and Appendix G is measured using the true positive rate (TPR), rather than bit-wise extraction accuracy (ACC). We have clarified this distinction in the revised manuscript. We have also added further experimental details in Appendix G. Specifically, we randomly perturb half of the encoded pairs to flip the corresponding watermark bits, thereby reducing the bit-wise extraction accuracy to 50% and the TPR to 0%.

AuthorsOpenReview ↗

Response to Reviewer 1

For Q4: We followed the experimental settings of PRC and Spherical Watermark for evaluating computational indistinguishability. RingID, HSTR, HSQR, and ROBIN explicitly modify the structure of the initial noise and therefore do not satisfy the Gaussian-preserving property targeted by this evaluation, so they are out of scope for this test. LAW-M follows the same Gaussian-preserving principle as LAW.

For Q5: Please refer to answer to W2.

We welcome any further questions from the reviewer and would be pleased to provide additional clarifications in further discussion.

Comments on the paper

2 comments

Dear Area Chair and Reviewers,

We have uploaded a revised version of our manuscript. Compared with the original submission, we have made the following revisions to address the reviewers' comments and improve the clarity and completeness of the paper.

  1. Abstract and Section 1 (Introduction): We have refined our claim of state-of-the-art robustness to "state-of-the-art robustness among Gaussianity-preserving methods" to more accurately reflect the scope of our contributions.

  2. Section 1 (Introduction, Line 63): We have corrected the inaccurate claim that Spherical Watermark does not characterize the covariance properties of watermarked latents.

  3. Table 1 and Section 4 (Experiments): We have clarified that the watermark capacities reported in Table 1 follow the original configurations of the respective methods, whereas the main experiments use a unified watermark length of 512 bits.

  4. Section 3.1 (Threat Model): We have refined the threat model by explicitly assuming that attackers cannot collect a large number of watermarked images from the same user, thereby providing stronger support for our stealthiness claims.

  5. Section 4.2 (Main Results: Classifier Indistinguishability): We have clarified that the latent-level classification experiments use the initial noise immediately after watermark embedding, rather than DDIM-inverted latents.

  6. Table 2: We have explicitly clarified that the reported watermark detection performance is measured using the true positive rate (TPR), rather than bit-wise extraction accuracy (ACC).

  7. Appendix G (Adaptive Attack): We have added further experimental details to improve reproducibility and reported additional evaluation metrics. These results further demonstrate that the evaluated adaptive attack cannot successfully remove the watermark without significantly degrading visual fidelity.

  8. Appendix J (Discussion on Storage Overhead and Index Retrieval of LAW-M): We have added a detailed discussion of the storage overhead and index retrieval procedure of LAW-M. We further emphasize that LAW-M is intended for high-stakes scenarios involving a limited number of important images, where its overall storage overhead can be significantly lower than that of Gaussian Shading under the corresponding deployment assumptions.

  9. Appendix K (Limitations): We have expanded the limitations discussion to acknowledge the vulnerability of our method to forgery attacks, as well as the lack of evaluation on DiT-based diffusion models and flow matching models.

We thank the reviewers for their constructive feedback, which has helped us improve the clarity and completeness of our manuscript. We welcome any further questions or comments and would be pleased to provide additional clarifications during the discussion.

Best regards,

Paper Authors

Area_Chair_5nWMOpenReview ↗

[Urgent] Please check the authors' rebuttal soon

Dear Reviewers,

    The authors have posted their responses to the issues you raised. As the discussion period will end soon, please take a look and have a discussion with the authors at your earliest convenience. Thank you very much!

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.