Pith. sign in

REVIEW 5 major objections 5 minor 70 references

KGMark: A Diffusion Watermark for Knowledge Graphs

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read KGMark embeds removable watermarks into knowledge-graph embeddings using a diffusion model, claiming detection AUC up to 0.99 and robustness near 0.95 under post-editing attacks.

desk verdict Legitimate first application of diffusion watermarking to KG embeddings with public code, but the isomorphism invariance claim rests on a non-canonical sort and Eq. (16) is the wrong CDF; worth reviewing, not ready to accept. read the letter →

arxiv 2505.23873 v2 pith:E6DWO7TF submitted 2025-05-29 cs.CR cs.AI

classification cs.CRcs.AI
keywords knowledgegraphwatermarkingdiffusionmodelwatermarkDDIMinversionembeddinglearnablemaskmatrixisomorphismrobustnessprovenance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes KGMark, a watermarking scheme for knowledge graphs (KGs) that hides a secret signature inside the latent embedding of a graph rather than in its visible structure. The authors aim to establish that a diffusion-based, embedding-level watermark can be detected with very high accuracy, cause only a small drop in downstream KG quality, and survive practical edits such as node relabeling, triple deletion, and relation alteration. If these claims hold, KGMark would be the first practical method for tracing ownership and provenance of AI-generated or dynamically evolving knowledge graphs.

What carries the argument

The central machinery is the DDIM-inversion-based latent diffusion watermark: a latent representation $Z_0$ of the KG is diffused forward to a noise vector $Z_T^{\text{INV}}$ via deterministic inversion, a Fourier-domain watermark $\Delta = F(Z_T^{\text{INV}}) \cdot (1-M) + F(S)\cdot M$ is injected under a learnable mask matrix $M$, and the watermarked vector is diffused back and decoded into a watermarked graph. Robustness is carried by two auxiliary mechanisms: a graph alignment procedure that reorders nodes by degree then clustering coefficient (Eq. 17, Algorithm 1) to make embeddings invariant to isomorphism, and a redundant embedding strategy that partitions the graph into communities and embeds the watermark in both community spectral profiles and vertex-level edge weights (Algorithm 2). Detection is a likelihood-ratio test using a noncentral chi-square statistic over the mask region.

What would settle it

Take any KG, apply a nontrivial automorphism (a relabeling that swaps nodes with equal degree and clustering coefficient), embed the watermark in the original, then attempt detection on the relabeled graph; if detection AUC falls below the claimed clean level, the isomorphism-invariance claim is falsified. A second test: generate graphs from one isomorphism class, embed with the same signature, and check whether extracted test statistics coincide.

Watch

Extended reading notes

Core claim

KGMark claims the first watermarking framework for knowledge graph embeddings (KGE) that combines a learnable mask matrix, a clustering-based graph alignment, and redundant community/vertex-level embedding to satisfy transparency, detectability, and robustness simultaneously. Concretely, it encodes a graph with a VAE graph encoder, inverts a latent diffusion (DDIM) trajectory to obtain a noise vector, embeds a Fourier-domain signature modulated by a learnable mask, and detects it via a likelihood-ratio test on the noncentral chi-square statistic. Experiments report watermark detection AUC up to 0.99, downstream performance loss between 0.02% and 9.7% across benchmarks, and AUC around 0.95 against relation alteration, triple deletion, isomorphism variation, Gaussian noise, smoothing, NEA, and L2 metric attacks. The paper also claims that its node reordering by degree and clustering coefficient makes the latent representation invariant to isomorphism, which is what supposedly underlies robustness to relabeling.

Load-bearing premise

The method assumes that sorting nodes by degree and then clustering coefficient produces the same node ordering for any two isomorphic graphs, so that isomorphic graphs map to the same aligned embedding and the watermark survives relabeling; in general this assumption is false because ties can be permuted by an isomorphism and no complete tie-breaker is specified.

Editorial extensions

If this is right

  • If KGMark works as claimed, a knowledge graph owner can prove provenance of a leaked or AI-generated graph by extracting a hidden signature from its embedding, even after structural edits.
  • The method implies that watermarking can be pushed from the text/image domain to structured relational data, where the watermark survives both spatial rearrangements (node relabeling) and temporal changes (incremental edge updates).
  • The reportedly low downstream-task degradation (0.02% to 9.7%) would mean that embedding-level watermarks can be transparent enough for production recommendation or retrieval systems built on KGs.
  • Robustness under isomorphism variation is the linchpin: if the degree/clustering-coefficient alignment truly canonicalizes isomorphic graphs, then KGMark is robust to node relabeling regardless of the label convention.
  • The design suggests an attack surface: an adversary aware of the alignment rule and the mask structure could potentially remove the watermark by targeted spectral or structural perturbations; the paper claims only generic post-editing robustness.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The isomorphism-invariance claim rests on a canonicalization assumption that is not generally valid: sorting by degree and clustering coefficient leaves ties that an isomorphism can permute, and the tie-breaking rule is unspecified, so two isomorphic graphs could produce different aligned embeddings and thus different extracted watermarks.
  • A direct test of the alignment claim would be to embed watermarks in multiple graphs from the same isomorphism class, extract after permuting node labels, and check whether the watermark detection statistic is stable; the paper does not report such a controlled experiment.
  • If the mask matrix is learned per graph and treated as a secret, then detection appears to require evaluating many candidate masks, so the paper's security model implicitly assumes the adversary does not know $M$; an adversary who does may be able to erase the signature by whitening the mask region in the frequency domain.
  • The claimed compatibility with dynamic KGs could be tested by embedding a watermark, performing sequential edge deletions/additions, and measuring detection AUC as a function of graph edit distance; the current experiments fix attack intensity percentages rather than model true temporal evolution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. KGMark proposes a watermarking framework for knowledge graphs (KGs) that embeds a signature into the latent space of a knowledge-graph embedding via a latent diffusion model. The method aligns the graph by sorting vertices by degree and clustering coefficient, redundantly embeds the watermark in community and vertex layers, and detects it by a likelihood-ratio test whose p-value is computed from a claimed noncentral chi-square CDF. Experiments on three datasets (LastFM, MIND, Alibaba-iFashion) report high detection AUC (up to 0.99), modest downstream performance loss, and robustness to relation alteration, triple deletion, Gaussian noise, smoothing, adversarial attacks, and isomorphism variation. The paper also introduces a learnable mask matrix (LAWMM) to balance detectability and transparency.

Significance. If the claims are correct, KGMark would be the first embedding-level watermarking scheme for knowledge graphs, and its combination of community- and vertex-level redundancy with latent diffusion inversion is a plausible design direction. The paper ships code, evaluates on multiple public datasets, and provides an ablation study of the proposed components, which is commendable. However, several load-bearing elements are not sound as written: the graph alignment does not in general produce an isomorphism-invariant representation, the detection statistic of Eq. (16) is not the noncentral chi-square CDF, the theoretical principles are asserted rather than derived, and the experimental protocol omits key details such as negative-sample definition and variance estimates. These issues undermine the credibility of the reported robustness and detectability numbers.

major comments (5)
  1. [§3.5, Eq. (17), Algorithm 1] The alignment procedure does not guarantee invariance under graph isomorphism. Sorting by degree and clustering coefficient leaves ties unresolved whenever two vertices have identical degree and clustering coefficient, which is common in real KGs (e.g., many degree-1 leaves with clustering 0). An isomorphism can permute such tied vertices, producing a different aligned adjacency matrix and hence a different latent representation, so the watermark extraction would not be invariant. There is also an internal contradiction: Eq. (17) sorts descending by degree and clustering (higher degree and higher clustering get smaller indices), while Algorithm 1 line 3 states 'ascending order.' No tie-breaking rule is specified. Consequently, the IsoVar results in Table 1 and the claim of being 'agnostic to node ordering' (Sec. 1) are not supported by the described method.
  2. [§3.4, Eq. (16)] Equation (16) is not the CDF of a noncentral chi-square distribution. The noncentral chi-square density contains a Bessel function factor and a (x/λ) term; the integrand shown here reduces to something like a gamma density with an extra exponential e^{-λ/2}, omitting the Bessel correction. As written, the p-value computation is statistically incorrect, and all detection decisions (including the AUC values in Tables 1 and 4) rely on this formula. This is a load-bearing error because the entire detectability evaluation depends on correct computation of p-values.
  3. [§4.1, Table 1] The paper reports AUC values without defining the negative-sample distribution or the number of runs, and no standard deviations or confidence intervals are given. AUC is only meaningful relative to a specific set of negative samples (e.g., graphs with a different/random signature); without this definition, the reported 'Clean' AUC and attack-robustness AUC are not reproducible or interpretable. This also applies to the ablation figures (Figures 3a–3c), where no error bars are provided despite the claims of stability.
  4. [§3.3, Principles 3.1 and 3.2] Principles 3.1 and 3.2 are presented as formal statements but are not derived or proved. Principle 3.1's inequality (Eq. 7) is simply asserted, and it is not clear under what conditions the decomposition of total deviation holds or how the learnable mask M is optimized to satisfy it. Principle 3.2 asserts an information-theoretic lower bound β and an effective capacity C(W), but no derivation connects the redundancy strategy (Eq. 21) to this bound. These principles are therefore not a theoretical foundation; they are assumptions. The paper should either provide rigorous arguments or clearly label these as design heuristics.
  5. [§4.1, Baselines] The four baselines are only vaguely adapted to the graph-embedding setting. Statements such as 'TreeRing and GaussianShading ... embedding watermarks by replacing 5% of nodes in the graph' do not explain how an image-diffusion watermarking technique is applied to a knowledge-graph embedding or how its detection statistic is modified. Similarly, DwtDct and DctQim are described as 'classical watermarking techniques that modify transformed coefficients,' but the transformation domain for a KGE is not specified. Without precise adaptation protocols, the comparative evaluation in Tables 1 and 4 is not reproducible, and the claimed superiority of KGMark may partly reflect implementation choices rather than method advantages.
minor comments (5)
  1. [Throughout] There are numerous typos and notation inconsistencies, e.g., 'thenormal' in Eq. (4), the misuse of '⇒' in Eq. (17) (it should express an order relation, not implication), and inconsistent use of Z w_T vs. Z^w_T. A careful proofread is needed.
  2. [§4.5 and Appendix B] The case study in Figure 4 is anecdotal and not quantified; the t-SNE visualization in Figure 5 is described as an 'Ablation study of the density' but appears to be a different experiment. Clarify what is being shown and add quantitative measures if this is meant to support transparency.
  3. [References] Some references are duplicated (e.g., TreeRing appears as both Wen et al. 2023 and Wen et al. 2024) and a few citations seem mismatched to the text (e.g., the isomorphism discussion cites Yan & Han 2002 for graph isomorphism, which is a substructure mining paper). Please clean up the bibliography.
  4. [Table 3] The heading 'KG Quality Metric @ 75 Steps' is confusing because columns GMR/HMR/AMR/Hits@10 are not obviously tied to a '75 steps' setting. Clarify whether these metrics are computed at the same DDIM steps used elsewhere and how the original KG baseline is defined.
  5. [§5.1, Eq. (23)] Equation (23) is an approximation of the representation shift under neighborhood perturbation, but the derivation is not given and it is not used in the experiments. Either provide a short derivation or omit it, as it currently reads as an unjustified side remark.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation chain: the watermark signature is random and the mask is optimized for reconstruction, not for the reported AUC; the only self-citations are background and non-load-bearing.

full rationale

KGMark's derivation does not reduce to its inputs by construction. The watermark signature S is sampled randomly as S ~ N(0, sigma^2 I) (Eq. 4), and the learnable mask M is trained to minimize the reconstruction loss between the original and watermarked latents (Eqs. 8-9), not to maximize the detectability or robustness numbers in Tables 1 and 4. Watermark extraction compares the Fourier-domain residual to the stored key S using a likelihood-ratio test (Eqs. 12-16), which is the standard, intended mechanism of a symmetric watermarking scheme, not a fitted prediction. The reported AUC values are empirical consequences of that mechanism on separate watermarked and unwatermarked samples. The graph alignment procedure in Sec. 3.5 and Algorithm 1 sorts vertices by degree and clustering coefficient (Eq. 17) and claims z0 invariance under node reordering. This is an unproven canonicalization and a correctness limitation, but it is not circular: isomorphism robustness is not assumed by defining the target AUC; rather, the reported IsoVar results depend on an unsupported property of the alignment. That belongs under correctness risk, not circularity. The paper contains two self-citations by coauthor Kun Wang (Wang et al., 2023 TPAMI and Wang et al., 2025 arXiv survey). Both appear only as general background in the introduction ('graphs (Han et al., 2025; Wang et al., 2023)' and 'harmful content ... (Jiang et al., 2024; Wang et al., 2025)'). No load-bearing theorem, ansatz, or uniqueness claim is imported from these works, and no central conclusion depends on them. The transparency metrics (cosine similarity, GMR, HMR, AMR, Hits@10) are partly aligned with the objective used to train M, but that is the intended design of a watermarking method, and downstream task metrics are evaluated with an external KGE model rather than directly optimized by the mask. No reported quantity is a renamed fit of the data used to produce it. Overall, I find no circular step that meets the evidentiary bar; the score reflects only the presence of minor, non-load-bearing self-citations.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central claim rests on two idealizations that are not established: exact DDIM inversion through a lossy VAE, and canonical graph alignment by degree and clustering sorting. Several constants (mask density, coefficient alpha, threshold delta) are chosen by hand or tuned on the benchmark, and the noncentral chi-square assumption is mis-stated. No new entities besides the learnable mask are introduced.

free parameters (5)
  • Learnable mask matrix M (LAWMM) = per-graph binary mask, e.g., density 0.015 in Appendix B
    Optimized per graph to minimize reconstruction loss (Eqs. 8-9); controls where the watermark is embedded and is selected for each evaluation.
  • Tunable correction coefficient alpha in Eq. (9) = not stated
    Scales the added S times M term in the sample-then-embed objective; no value or selection criterion is given.
  • Mask density rho = swept 0.015 to 0.1; 0.015 used in the case study
    Chosen as a trade-off knob; Fig. 3c explicitly shows density tuning changes AUC and cosine similarity.
  • Local/global balance alpha in divergence metric Eq. (18) = not stated
    Balances edge perturbations versus spectral shifts; the same symbol alpha is reused with a different meaning in Eq. (9).
  • Significance level alpha and threshold delta = alpha = 5e-5 in experiments
    Detection thresholds are chosen by the authors, and the reported AUC depends on them (Fig. 3a).
assumptions (5)
  • domain assumption DDIM inversion exactly recovers the original noise latent even after VAE encode and decode and adversarial edits.
    Sections 3.1 and 3.4 assume Z_INV_T from the watermarked graph matches the embedding-time latent; Table 3 shows reconstruction is lossy (cosine 0.94 to 0.97 even without watermark), so this is an idealization.
  • ad hoc to paper Sorting by degree and clustering coefficient is a canonical ordering that makes embeddings invariant under isomorphism.
    Algorithm 1 and Eq. 17; ties are not broken, so isomorphic permutations in tied classes can produce different orderings.
  • domain assumption The residual statistic (R divided by sigma) follows a noncentral chi-square distribution with the stated CDF.
    Eqs. 13-16; the CDF written is actually a central chi-square density times exp(-lambda/2), not a noncentral chi-square CDF.
  • ad hoc to paper Information-theoretic lower bound beta can be enforced by community and vertex encoding.
    Principle 3.2 and Eqs. 19-20 are asserted; no proof or construction shows the bound.
  • domain assumption Attacks are well modeled as additive structural perturbations with bounded norm.
    Eqs. 5 and 18; real post-editing attacks may not be additive or bounded on embeddings.
invented entities (1)
  • Learnable Adaptive Watermark Mask Matrix (LAWMM)
    purpose: Selects Fourier coordinates where the signature is embedded so that distortion is minimized.
    It is an introduced component evaluated only within this paper's ablations; no external falsifiable prediction outside the paper is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KGMark: A Diffusion Watermark for Knowledge Graphs." pith.science (2026). https://pith.science/paper/E6DWO7TF

@misc{pith2026250523873,
  author       = {Pith},
  title        = {Pith review of: KGMark: A Diffusion Watermark for Knowledge Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E6DWO7TF}},
  note         = {Machine review of arXiv:2505.23873}
}
read the original abstract

Knowledge graphs (KGs) are ubiquitous in numerous real-world applications, and watermarking facilitates protecting intellectual property and preventing potential harm from AI-generated content. Existing watermarking methods mainly focus on static plain text or image data, while they can hardly be applied to dynamic graphs due to spatial and temporal variations of structured data. This motivates us to propose KGMARK, the first graph watermarking framework that aims to generate robust, detectable, and transparent diffusion fingerprints for dynamic KG data. Specifically, we propose a novel clustering-based alignment method to adapt the watermark to spatial variations. Meanwhile, we present a redundant embedding strategy to harden the diffusion watermark against various attacks, facilitating the robustness of the watermark to the temporal variations. Additionally, we introduce a novel learnable mask matrix to improve the transparency of diffusion fingerprints. By doing so, our KGMARK properly tackles the variation challenges of structured data. Experiments on various public benchmarks show the effectiveness of our proposed KGMARK. Our code is available at https://github.com/phrara/kgmark.

Figures

Figures reproduced from arXiv: 2505.23873 by the authors.

Figure 1
Figure 1. Overview of our KGMark. KGMark implements a KGE watermarking scheme that preserves transparency, enables reliable detection, and remains robust against various post-editing attacks. swering (Yin et al., 2024; Zhou et al., 2024), and recom￾mendation systems (Fan et al., 2019; Wang et al., 2019). Deep learning-based models such as GraphRNN (You et al., 2018), GraphVAE (Simonovsky & Komodakis, 2018), Mol￾GAN (Cao & Kip… view at source ↗
Figure 2
Figure 2. Pipeline of the proposed KGMark. The target KGE undergoes community detection and alignment before watermark embedding, enabling robust watermark extraction under attacks (Relation Alteration, Subgraph Deletion, and Isomorphic Variation). DDIM (Song et al., 2022) for sampling. DDIM’s determin￾istic inversion process allows efficient recovery of the initial noise vector Z INV T , which is then used for watermark embe… view at source ↗
Figure 3
Figure 3. Ablation. (a) AUC under varying significance levels and DDIM inference steps. (b) Different DDIM inference steps during embedding and detection. (c) Effect of watermark mask density on all datasets, with DDIM steps set to 75 and significance level 5e−5. on par with the original KG and the non-watermarked ver￾sion. For example, in AliF, KGMark achieves a Hits@10 score of 0.8296, slightly lower than the non-watermarke… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Case Study. Workflow and representative results. The results, shown in [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Watermark Embedding in Frequency Domain. The theoretical basis of watermark embedding in the frequency domain is mainly based on the frequency domain transfor￾mation technology in signal processing. The essence of the Fourier transform we employ is to separate the time…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 58 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Waves: benchmarking the robustness of image watermarks

    An, B., Ding, M., Rabbani, T., Agrawal, A., Xu, Y., Deng, C., Zhu, S., Mohamed, A., Wen, Y., Goldstein, T., and Huang, F. Waves: benchmarking the robustness of image watermarks. In Proceedings of the 41st International Conference on Machine Learning. JMLR.org, 2025

  3. [3]

    Promark: Proactive diffusion watermarking for causal attribution

    Asnani, V., Collomosse, J., Bui, T., Liu, X., and Agarwal, S. Promark: Proactive diffusion watermarking for causal attribution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 10802--10811, 2024

  4. [4]

    R., Sharma, K., Aziz, A., Bajpai, S., Biswas, S., Sharma, V., Jain, V., Chadha, A., Sheth, A., and Das, A

    Barman, N. R., Sharma, K., Aziz, A., Bajpai, S., Biswas, S., Sharma, V., Jain, V., Chadha, A., Sheth, A., and Das, A. The brittleness of ai-generated image watermarking techniques: Examining their robustness against visual paraphrasing attacks. arXiv preprint arXiv:2408.10446, 2024

  5. [5]

    Comprehensive exploration of synthetic data generation: A survey, 2024

    Bauer, A., Trapp, S., Stenger, M., Leppich, R., Kounev, S., Leznik, M., Chard, K., and Foster, I. Comprehensive exploration of synthetic data generation: A survey, 2024

  6. [6]

    Adversarial Attacks on Knowledge Graph Embeddings via Instance Attribution Methods

    Bhardwaj, P., Kelleher, J., Costabello, L., and O'Sullivan, D. Adversarial attacks on knowledge graph embeddings via instance attribution methods. arXiv preprint arXiv:2111.03120, 2021

  7. [7]

    and G \"u nnemann, S

    Bojchevski, A. and G \"u nnemann, S. Adversarial attacks on node embeddings via graph poisoning. In International conference on machine learning, pp.\ 695--704. PMLR, 2019

  8. [8]

    Bouritsas, G., Frasca, F., Zafeiriou, S., and Bronstein, M. M. Improving graph neural network expressivity via subgraph isomorphism counting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (1): 0 657--668, 2023

Show all 70 references
  1. [9]

    Busbridge, D., Sherburn, D., Cavallo, P., and Hammerla, N. Y. Relational graph attention networks, 2019

  2. [10]

    Music mood dataset creation based on last

    C ano, E., Morisio, M., et al. Music mood dataset creation based on last. fm tags. In 2017 International Conference on Artificial Intelligence and Applications, Vienna, Austria, pp.\ 15--26, 2017

  3. [11]

    Cao, N. D. and Kipf, T. Molgan: An implicit generative model for small molecular graphs, 2022

  4. [12]

    A., and Shum, H

    Chang, Z., Koulieris, G. A., and Shum, H. P. H. On the design fundamentals of diffusion models: A survey, 2023

  5. [13]

    and Schockaert, S

    Charpenay, V. and Schockaert, S. Capturing knowledge graphs and rules with octagon embeddings. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, 2025

  6. [14]

    and Wornell, G

    Chen, B. and Wornell, G. A class of provably good methods for digital watermarking and information embedding. IEEE Trans. on Information Theory, 47 0 (4): 0 291--314, 2001

  7. [15]

    Differentiable neuro-symbolic reasoning on large-scale knowledge graphs

    Chen, S., Fang, H., Cai, Y., Huang, X., and Sun, M. Differentiable neuro-symbolic reasoning on large-scale knowledge graphs. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc

  8. [16]

    Pog: Personalized outfit generation for fashion recommendation at alibaba ifashion

    Chen, W., Huang, P., Xu, J., Guo, X., Guo, C., Sun, F., Li, C., Pfadler, A., Zhao, H., and Zhao, B. Pog: Personalized outfit generation for fashion recommendation at alibaba ifashion. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data ...

  9. [17]

    Digital watermarking and steganography

    Cox, I., Miller, M., Bloom, J., Fridrich, J., and Kalker, T. Digital watermarking and steganography. Morgan kaufmann, 2007

  10. [18]

    A decoder-only foundation model for time-series forecasting

    Das, A., Kong, W., Sen, R., and Zhou, Y. A decoder-only foundation model for time-series forecasting. In Proceedings of the 41st International Conference on Machine Learning. JMLR.org, 2024

  11. [19]

    A., Brown-Cohen, J., Bunel, R., Balle, B., Cemgil, T., Ahmed, Z., Stacpoole, K., Shumailov, I., Baetu, C., Gowal, S., Hassabis, D., and Kohli, P

    Dathathri, S., See, A., Ghaisas, S., Huang, P.-S., McAdam, R., Welbl, J., Bachani, V., Kaskasoli, A., Stanforth, R., Matejovicova, T., Hayes, J., Vyas, N., Merey, M. A., Brown-Cohen, J., Bunel, R., Balle, B., Cemgil, T., Ahmed, Z., Stacpoole, K., Shumailov, I., Baetu, C., Gowa...

  12. [20]

    On structural explanation of bias in graph neural networks

    Dong, Y., Wang, S., Wang, Y., Derr, T., and Li, J. On structural explanation of bias in graph neural networks. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 316–326, New York, NY, USA, 2022. Association for Computing Machinery

  13. [21]

    Deep social collaborative filtering

    Fan, W., Ma, Y., Yin, D., Wang, J., Tang, J., and Li, Q. Deep social collaborative filtering. In Proceedings of the 13th ACM Conference on Recommender Systems, pp.\ 305–313, New York, NY, USA, 2019. Association for Computing Machinery

  14. [22]

    Lossless data embedding—new paradigm in digital watermarking

    Fridrich, J., Goljan, M., and Du, R. Lossless data embedding—new paradigm in digital watermarking. EURASIP Journal on Advances in Signal Processing, 2002: 0 1--12, 2002

  15. [23]

    Monte carlo methods in financial engineering, 2004

    Glasserman, P. Monte carlo methods in financial engineering, 2004

  16. [24]

    A., Mukherjee, S., Tang, X., He, Q., Hua, Z., Long, B., Zhao, T., Shah, N., Javari, A., Xia, Y., and Tang, J

    Han, H., Wang, Y., Shomer, H., Guo, K., Ding, J., Lei, Y., Halappanavar, M., Rossi, R. A., Mukherjee, S., Tang, X., He, Q., Hua, Z., Long, B., Zhao, T., Shah, N., Javari, A., Xia, Y., and Tang, J. Retrieval-augmented generation with graphs (graphrag), 2025

  17. [25]

    A transfer attack to image watermarks

    Hu, Y., Jiang, Z., Guo, M., and Gong, N. A transfer attack to image watermarks. arXiv preprint arXiv:2403.15365, 2024

  18. [26]

    Ji, S., Pan, S., Cambria, E., Marttinen, P., and Yu, P. S. A survey on knowledge graphs: Representation, acquisition, and applications. IEEE Transactions on Neural Networks and Learning Systems, 33 0 (2): 0 494–514, 2022

  19. [27]

    KG - FIT : Knowledge graph fine-tuning upon open-world knowledge

    Jiang, P., Cao, L., Xiao, C., Bhatia, P., Sun, J., and Han, J. KG - FIT : Knowledge graph fine-tuning upon open-world knowledge. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

  20. [28]

    A watermark for large language models

    Kirchenbauer, J., Geiping, J., Wen, Y., Katz, J., Miers, I., and Goldstein, T. A watermark for large language models. In International Conference on Machine Learning, pp.\ 17061--17084. PMLR, 2023

  21. [29]

    Knowledge graphs can be learned with just intersection features

    Le, D., Zhong, S., Liu, Z., Xu, S., Chaudhary, V., Zhou, K., and Xu, Z. Knowledge graphs can be learned with just intersection features. In Salakhutdinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., and Berkenkamp, F. (eds.), Proceedings of the 41st Inter...

  22. [30]

    Clustering then propagation: Select better anchors for knowledge graph embedding

    LIANG, K., Liu, Y., Li, H., Meng, L., Liu, S., Wang, S., sihang zhou, and Liu, X. Clustering then propagation: Select better anchors for knowledge graph embedding. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

  23. [31]

    Liu, A., Pan, L., Lu, Y., Li, J., Hu, X., Zhang, X., Wen, L., King, I., Xiong, H., and Yu, P. S. A survey of text watermarking in the era of large language models, 2024

  24. [32]

    Fast and continual knowledge graph embedding via incremental lora

    Liu, J., Ke, W., Wang, P., Wang, J., Gao, J., Shang, Z., Li, G., Xu, Z., Ji, K., and Li, Y. Fast and continual knowledge graph embedding via incremental lora. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, 2025

  25. [33]

    Dynasemble: Dynamic ensembling of textual and structure-based models for knowledge graph completion

    Nandi, A., Kaur, N., Singla, P., and Mausam. Dynasemble: Dynamic ensembling of textual and structure-based models for knowledge graph completion. In Annual Meeting of the Association for Computational Linguistics, 2024

  26. [34]

    The non-central 2-and f-distribution and their applications

    Patnaik, P. The non-central 2-and f-distribution and their applications. Biometrika, 36 0 (1/2): 0 202--232, 1949

  27. [35]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., and Rombach, R. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

  28. [36]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning transferable visual models from natural language supervision. In Meila, M. and Zhang, T. (eds.), Proceedings of the 3...

  29. [37]

    High-resolution image synthesis with latent diffusion models, 2021

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models, 2021

  30. [38]

    N., Bloem, P., van den Berg, R., Titov, I., and Welling, M

    Schlichtkrull, M., Kipf, T. N., Bloem, P., van den Berg, R., Titov, I., and Welling, M. Modeling relational data with graph convolutional networks, 2017

  31. [39]

    Toward degree bias in embedding-based knowledge graph completion

    Shomer, H., Jin, W., Wang, W., and Tang, J. Toward degree bias in embedding-based knowledge graph completion. In Proceedings of the ACM Web Conference 2023, pp.\ 705–715, New York, NY, USA, 2023. Association for Computing Machinery

  32. [40]

    and Komodakis, N

    Simonovsky, M. and Komodakis, N. Graphvae: Towards generation of small graphs using variational autoencoders. In K u rkov \'a , V., Manolopoulos, Y., Hammer, B., Iliadis, L., and Maglogiannis, I. (eds.), Artificial Neural Networks and Machine Learning -- ICANN 2018, pp.\ 412--...

  33. [41]

    and Borghuis, T

    Smits, J. and Borghuis, T. Generative AI and Intellectual Property Rights, pp.\ 323--344. T.M.C. Asser Press, The Hague, 2022

  34. [42]

    Denoising diffusion implicit models, 2022

    Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models, 2022

  35. [43]

    Rotate: Knowledge graph embedding by relational rotation in complex space

    Sun, Z., Deng, Z.-H., Nie, J.-Y., and Tang, J. Rotate: Knowledge graph embedding by relational rotation in complex space. arXiv preprint arXiv:1902.10197, 2019

  36. [44]

    Graph attention networks, 2018

    Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., and Bengio, Y. Graph attention networks, 2018

  37. [45]

    Cuts: customizable tabular synthetic data generation

    Vero, M., Balunovi\' c , M., and Vechev, M. Cuts: customizable tabular synthetic data generation. In Proceedings of the 41st International Conference on Machine Learning. JMLR.org, 2024

  38. [46]

    Digress: Discrete denoising diffusion for graph generation, 2023

    Vignac, C., Krawczuk, I., Siraudin, A., Wang, B., Cevher, V., and Frossard, P. Digress: Discrete denoising diffusion for graph generation, 2023

  39. [47]

    Dkn: Deep knowledge-aware network for news recommendation

    Wang, H., Zhang, F., Xie, X., and Guo, M. Dkn: Deep knowledge-aware network for news recommendation. In Proceedings of the 2018 world wide web conference, pp.\ 1835--1844, 2018

  40. [48]

    Brave the wind and the waves: Discovering robust and generalizable graph lottery tickets

    Wang, K., Liang, Y., Li, X., Li, G., Ghanem, B., Zimmermann, R., Zhou, Z., Yi, H., Zhang, Y., and Wang, Y. Brave the wind and the waves: Discovering robust and generalizable graph lottery tickets. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46 0 (5): 0 3388...

  41. [49]

    A comprehensive survey in llm (-agent) full stack safety: Data, training and deployment

    Wang, K., Zhang, G., Zhou, Z., Wu, J., Yu, M., Zhao, S., Yin, C., Fu, J., Yan, Y., Luo, H., et al. A comprehensive survey in llm (-agent) full stack safety: Data, training and deployment. arXiv preprint arXiv:2504.15585, 2025

  42. [50]

    Neural graph collaborative filtering

    Wang, X., He, X., Wang, M., Feng, F., and Chua, T.-S. Neural graph collaborative filtering. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, pp.\ 165–174, New York, NY, USA, 2019. Association for Computing Machinery

  43. [51]

    Tree-rings watermarks: Invisible fingerprints for diffusion images

    Wen, Y., Kirchenbauer, J., Geiping, J., and Goldstein, T. Tree-rings watermarks: Invisible fingerprints for diffusion images. In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S. (eds.), Advances in Neural Information Processing Systems, pp.\ 58047--580...

  44. [52]

    Tree-rings watermarks: Invisible fingerprints for diffusion images

    Wen, Y., Kirchenbauer, J., Geiping, J., and Goldstein, T. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024

  45. [53]

    MIND : A large-scale dataset for news recommendation

    Wu, F., Qiao, Y., Chen, J.-H., Wu, C., Qi, T., Lian, J., Liu, D., Xie, X., Gao, J., Wu, W., and Zhou, M. MIND : A large-scale dataset for news recommendation. In Jurafsky, D., Chai, J., Schluter, N., and Tetreault, J. (eds.), Proceedings of the 58th Annual Meeting of the Assoc...

  46. [54]

    and Han, J

    Yan, X. and Han, J. gspan: Graph-based substructure pattern mining. In 2002 IEEE International Conference on Data Mining, 2002. Proceedings., pp.\ 721--724. IEEE, 2002

  47. [55]

    Guise: Graph gaussian shading watermark, 2024

    Yang, R. Guise: Graph gaussian shading watermark, 2024

  48. [56]

    A review on the reliability of knowledge graph: from a knowledge representation learning perspective

    Yang, Y., Chen, J., and Xiang, Y. A review on the reliability of knowledge graph: from a knowledge representation learning perspective. World Wide Web, 28 0 (1): 0 4, 2024 a

  49. [57]

    Gaussian shading: Provable performance-lossless image watermarking for diffusion models

    Yang, Z., Zeng, K., Chen, K., Fang, H., Zhang, W., and Yu, N. Gaussian shading: Provable performance-lossless image watermarking for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 12162--12171, 2024 b

  50. [58]

    Rethinking complex queries on knowledge graphs with neural link predictors

    Yin, H., Wang, Z., and Song, Y. Rethinking complex queries on knowledge graphs with neural link predictors. In The Twelfth International Conference on Learning Representations, 2024

  51. [59]

    G raph RNN : Generating realistic graphs with deep auto-regressive models

    You, J., Ying, R., Ren, X., Hamilton, W., and Leskovec, J. G raph RNN : Generating realistic graphs with deep auto-regressive models. In Dy, J. and Krause, A. (eds.), Proceedings of the 35th International Conference on Machine Learning, pp.\ 5708--5717. PMLR, 2018

  52. [60]

    Artificial fingerprinting for generative models: Rooting deepfake attribution in training data

    Yu, N., Skripniuk, V., Abdelnabi, S., and Fritz, M. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International conference on computer vision, pp.\ 14448--14457, 2021

  53. [61]

    An improved gnn using dynamic graph embedding mechanism: A novel end-to-end framework for rolling bearing fault diagnosis under variable working conditions

    Yu, Z., Zhang, C., and Deng, C. An improved gnn using dynamic graph embedding mechanism: A novel end-to-end framework for rolling bearing fault diagnosis under variable working conditions. Mechanical Systems and Signal Processing, 200: 0 110534, 2023

  54. [62]

    Data poisoning attack against knowledge graph embedding

    Zhang, H., Zheng, T., Gao, J., Miao, C., Su, L., Li, Y., and Ren, K. Data poisoning attack against knowledge graph embedding. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19 , pp.\ 4853--4859. International Joint Conferen...

  55. [63]

    V., Bearfield, C

    Zhang, L., Liu, X., i Martin, A. V., Bearfield, C. X., Brun, Y., and Guan, H. Attack-resilient image watermarking using stable diffusion. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024 a

  56. [64]

    Editguard: Versatile image watermarking for tamper localization and copyright protection

    Zhang, X., Li, R., Yu, J., Xu, Y., Li, W., and Zhang, J. Editguard: Versatile image watermarking for tamper localization and copyright protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11964--11974, 2024 b

  57. [65]

    Provable robust watermarking for ai-generated text, 2023 a

    Zhao, X., Ananth, P., Li, L., and Wang, Y.-X. Provable robust watermarking for ai-generated text, 2023 a

  58. [66]

    Invisible image watermarks are provably removable using generative ai

    Zhao, X., Zhang, K., Su, Z., Vasan, S., Grishchenko, I., Kruegel, C., Vigna, G., Wang, Y.-X., and Li, L. Invisible image watermarks are provably removable using generative ai. arXiv preprint arXiv:2306.01953, 2023 b

  59. [67]

    A recipe for watermarking diffusion models

    Zhao, Y., Pang, T., Du, C., Yang, X., Cheung, N.-M., and Lin, M. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023 c

  60. [68]

    Less is more: One-shot subgraph reasoning on large-scale knowledge graphs

    Zhou, Z., Zhang, Y., Yao, J., quanming yao, and Han, B. Less is more: One-shot subgraph reasoning on large-scale knowledge graphs. In The Twelfth International Conference on Learning Representations, 2024

  61. [69]

    Genimage: A million-scale benchmark for detecting ai-generated image

    Zhu, M., Chen, H., Yan, Q., Huang, X., Lin, G., Li, W., Tu, Z., Hu, H., Hu, J., and Wang, Y. Genimage: A million-scale benchmark for detecting ai-generated image. Advances in Neural Information Processing Systems, 36, 2024 a

  62. [70]

    Watermark-embedded adversarial examples for copyright protection against diffusion models

    Zhu, P., Takahashi, T., and Kataoka, H. Watermark-embedded adversarial examples for copyright protection against diffusion models. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 24420--24430, 2024 b

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.