acceptodds
Withdrawn from ICLR 2027

MeshGuard: Hidden 3D Watermarking in Spectral Space through Laplacian Beltrami Operator and Geometry Fusion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent advances in 3D generation and reconstruction enable rapid creation of high-fidelity assets such as point clouds and meshes, increasing the need for reliable intellectual property protection. However, unlike well-studied 2D watermarking, 3D methods remain limited and often embed signals directly into surface geometry, making them vulnerable to distortion and adversarial attacks. We propose a robust mesh watermarking framework based on spectral geometry. Our method leverages the Laplace–Beltrami operator to extract intrinsic spectral representations and embeds watermark signals into mid-band components via eigen-decomposition. The watermark is applied in the spatial domain through predicted vertex-level offsets regressed from updated eigenvectors, achieving imperceptible yet stable embedding. To enhance robustness, we incorporate adversarial training with diverse mesh perturbations, enabling reliable decoding under severe distortions. We further introduce a flexible encoding scheme that converts text or image inputs into binary watermark sequences for fusion and evaluation. Extensive experiments demonstrate improved robustness against a wide range of geometric attacks while preserving high geometric fidelity, outperforming existing approaches by a large margin.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Watermarking, Mesh Processing, Spectral Geometry, Laplace–Beltrami Operator, Adversarial Robustness, Geometric Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Bi6COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MeshGuard, targeting the intellectual property protection of 3D mesh assets. Text or logos are first encoded into a 128-bit message by BLIP-2 and an MLP, then element-wise modulated on the mid-frequency eigenbasis of the Laplace–Beltrami operator, and vertex offsets computed by a network are applied to the vertices to complete the embedding. On the decoding side, a 3D CNN recovers the bits from potentially attacked meshes. During training, a learnable attack layer and a discriminator are introduced for adversarial training. The authors conduct experiments on multiple datasets and report results superior to existing methods.

Critical strengths

Learned watermarking for meshes has previously been relatively scarce. Using the mid-frequency eigenbasis of the Laplace–Beltrami operator to embed watermarks into meshes is also a reasonable design approach. Combined with various subsequent perturbations, this establishes an overall complete watermark embedding method.

Critical weaknesses

The experimental evidence in the paper is insufficient. The paper does not clearly specify the training and test sets for the messages. On this basis, one can only assume that both are the collection of these 200 logos, which means that for a new watermark, the network can only be retrained, clearly limiting the practicality of the method. Furthermore, can the decoded message be reconstructed back into the original input image, and if so, what is the quality? If not, why introduce the BLIP message encoder instead of directly using random messages?

Questions for authors

Answer the questions raised by the critical weakness.

Reviewer_RGkPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

MeshGuard proposes a learned watermarking pipeline for triangle meshes. It maps a logo or text input to a binary sequence using frozen BLIP-2 features and a projection head, modulates selected Laplace-Beltrami eigenvectors, and predicts small vertex displacements. An attack module applies geometric corruptions, and a neural decoder predicts the embedded bits. The paper reports experiments on ShapeNet categories and an Objaverse++ subset, comparisons with learned and non-learning baselines, and ablations of spectral bands, message encoders, losses, and decoders. Reported combined-attack bit accuracy is 0.892, compared with 0.686 for Deep3DMark in main Table 1. The supplement adds topology/cropping, attack-severity mismatch, false-positive, and adaptive-attack evaluations.

Critical strengths

Mid-band spectral conditioning offers a plausible way to distribute small displacements over a mesh. Main Table 4 reports 0.892 combined-attack bit accuracy for mid-band fusion, versus 0.831 for low-band fusion, 0.803 for high-band fusion, and 0.735 without LBO eigendecomposition. Within the reported protocol, these results support investigating the spectral design.

Main Table 2 and supplementary Tables 2, 5, 6, and 8 include topology changes, stronger test attacks, negative controls, adaptive removal, and crop holes. This breadth is relevant to practical watermarking and provides useful evidence, although its operational interpretation needs clarification below.

Critical weaknesses

In Secs. 3.2 and 4 (main pp. 4 and 6, lines 167-175 and 270-286), the projection producing the target bits is trainable, while the implementation description also says that the binary encoding is kept fixed. The paper needs to distinguish these possibilities. If the target mapping and decoder are jointly optimized without independent targets or a diversity constraint, a constant target sequence and a decoder that ignores geometry can drive the message loss toward zero; zero displacement also minimizes reconstruction loss. This establishes an ambiguity in what the objectives guarantee, rather than evidence that the reported run actually collapsed. Moreover, the described 200-logo pool and mesh train/test split do not specify a held-out-logo or held-out-message test. A deterministic codebook of 200 messages carries at most log2(200), approximately 7.64, bits of identity information, regardless of its 128-bit codeword length. That can be a valid identification task, but it needs explicit characterization and matched comparisons. Main Table 5's random/hash encoder variants do not report these controls; supplementary Table 6's collision test uses 32-bit messages rather than the main 128-bit configuration. Required modification: state when target codes are fixed, report bit balance, inter-message distances and unique codes for the actual 128-bit setup, and provide a geometry-ignoring decoder control plus held-out-message evaluation. Compare methods on the same message set, payload length, and distortion budget, or explicitly restrict the claim to identification within the fixed logo pool.

Sec. 3.6 (main p. 5, Eqs. 13-14) assigns discriminator label 1 to watermarked meshes and 0 to clean meshes. Minimizing the embedding loss -log D(G') therefore pushes the output toward the watermarked class, whereas the text says it makes the mesh indistinguishable from clean geometry. The claimed learnable attack module also lacks an explicit objective that maximizes decoding error: minimizing the common message loss and a positive attack regularizer would not define an adversarial attack update. Separately, supplementary Table 6 does provide adaptive attacks, but reports mean bit accuracies of 0.71 for white-box optimization and 0.58 for autoencoder laundering, while Table 5 uses matching thresholds of 0.90 and 0.95. Mean bit accuracy alone does not determine detection success at either threshold, so those numbers do not establish the claimed reliable recoverability. The attack budget remains the literal placeholder "CD = X", and the key used by the learned method is not defined sufficiently to interpret the mismatched-key test. Required modification: supply the actual alternating update objectives, consistent discriminator labels, and attack constraints; then report true-positive/detection or identification success alongside false positives at one explicitly fixed verification rule, with concrete distortion budgets, message length, key construction, attack steps, and test counts. Revise the robustness claim if the adaptive attacks remove operational detectability.

Questions for authors
  1. Are the 128-bit targets precomputed and frozen, or can the projection head change them during training? How many distinct codes are produced by the 200 logos, what are their per-bit frequencies and pairwise distances, and are test logos/messages disjoint from training? A geometry-ignoring predictor and held-out-message results would directly resolve the first weakness.

  2. Which exact losses update the embedder, discriminator, decoder, message head, and attack parameters? Are the labels in Eq. 13 reversed in the implementation, or is Eq. 14 incorrect? What objective makes a learned attack increase decoding difficulty rather than reduce it? This determines whether the adversarial-training ablations evaluate the proposed mechanism.

  3. At the same threshold used for the false-positive tests, what fraction of attacked meshes still verifies correctly for every attack in supplementary Table 6? Please specify the key, payload length, actual value of X, attack optimization settings, and evaluation counts. These results could materially change the robustness assessment.

  4. Please give a dimensionally consistent spectral-to-vertex computation and a method-by-method decoder/preprocessing specification. Eq. 1 is area-normalized and generally nonsymmetric in the ordinary Euclidean inner product, whereas Eq. 4 writes a symmetric spectral reconstruction; Sec. 3.3 specifies k-by-3 offsets despite n* = 4096 vertices and k = 256 modes. Main lines 289-303 and supplementary Sec. 2 also describe different decoding protocols. This matters because main Table 4 changes combined-attack accuracy from 0.685 with the Deep3DMark bin-voting decoder to 0.892 with the proposed decoder. Clarification and a comparison with decoder choice controlled would help establish which component accounts for the gain.

Reviewer_oJqCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents MeshGuard, a learning-based watermarking framework for 3D meshes. It uses a BLIP-2-based module to convert text or logo inputs into binary messages, modulates selected eigenvectors of the Laplace–Beltrami operator, and predicts vertex offsets from the resulting spectral features. A neural decoder recovers the watermark from potentially distorted meshes. The framework also incorporates attack simulation during training. Experiments evaluate geometric fidelity and watermark recovery under geometric and topology-changing attacks, together with ablations of the message encoder, spectral fusion, attack training, and decoder.

Critical strengths
  1. The paper is easy to follow.
  2. Robust watermarking of 3D mesh assets is an important problem. Using spectral features to guide learned geometric modifications is a promising design direction for balancing watermark recovery and geometric fidelity.
  3. The experiments cover geometric perturbations and topology-changing operations, including rotation, smoothing, simplification, subdivision, and remeshing.
Critical weaknesses
  1. The first claimed contribution appears to be a relatively straightforward adaptation of an existing multimodal encoder. According to Section 3.2 and the implementation details, the module consists of a frozen BLIP-2 model followed by learned projection layers and binary quantization. Supporting text and logo inputs offers functional flexibility, but the paper does not clearly identify a new encoding principle beyond this adaptation. In particular, it remains unclear how the semantic representation is specifically designed for reliable transmission through the mesh watermarking process. The empirical benefit of this component is worth acknowledging, but its status as an independent methodological contribution is insufficiently justified.

  2. Spectral mesh watermarking has a substantial history, including Ohbuchi et al., Watermarking 3D Polygonal Meshes in the Mesh Spectral Domain (GI 2001), and K. Wang et al., Blind and Robust Mesh Watermarking Using Manifold Harmonics (ICIP 2009). The latter explicitly uses Laplace–Beltrami/manifold-harmonics analysis for blind mesh watermarking. Thus, applying LBO eigen-decomposition and embedding information through spectral representations is not itself a new contribution. MeshGuard’s potential advance lies in modulating spectral features and mapping them to vertex offsets through a learned regressor. However, the paper does not sufficiently explain which limitation of existing spectral embedding methods this design resolves. A more explicit methodological comparison and evaluation against directly related LBO-based watermarking methods are needed to establish the contribution.

  3. Table 5 compares BLIP-2 with CLIP, hash-based encoding, and random encoding, and reports better performance for BLIP-2. However, these comparisons do not sufficiently explain whether the improvement arises from semantic information or from differences in the resulting codewords and their training. Relevant factors include bit balance, correlations between bits, and distances between watermark codewords. Moreover, the experiments use 200 logos without clearly specifying whether watermark identities are disjoint between training and testing. The reported results therefore leave both the source of the improvement and generalization to unseen watermark inputs unresolved. Further controls are needed to support the interpretation that semantic encoding improves robustness to mesh distortions.

  4. Figure 5 defines the low-, mid-, and high-frequency bands using 25%, 50%, and 25% of the spectrum, respectively. The manuscript does not explain whether the comparisons in Table 4 match the number of active eigenvectors, message length, and regressor capacity. Without these controls, the contribution of frequency location cannot be clearly separated from that of representation size. Furthermore, the final vertex perturbations are produced by a learned regressor, so selecting mid-band input features does not automatically establish that the resulting perturbations retain the claimed spectral properties. A controlled band comparison and an analysis of the output perturbations would strengthen the explanation of why the mid-band design improves robustness.

Questions for authors
  1. What is the methodological contribution of the message encoder beyond adapting a frozen multimodal model (e.g. BLIP) through projection and quantization?

  2. Which specific limitation of existing LBO/manifold-harmonics watermarking does the proposed feature modulation and learned vertex regression resolve?

  3. What objective is used to update the learnable attack parameters, and how are these updates coordinated with the embedding and decoding networks? What is the main contribution of this mechanism from existing attack-layer training?

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

Related papers

Discussion (0)

Sign in to comment.