Pith. sign in

REVIEW 3 major objections 5 minor 30 references

GRAND : Graph Reconstruction from potential partial Adjacency and Neighborhood Data

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read An adversary who sees only the common-neighbors matrix of an undirected graph can rebuild its adjacency matrix, up to co-squareness, and the GRAND pipeline does this exactly on real data.

desk verdict A promising high-level idea for reconstructing graphs from common-neighbor data, but the central topological algorithms as written are unsound, so the reported results do not follow from the presented methods. read the letter →

arxiv 2412.02329 v2 pith:7C22VQ6B submitted 2024-12-03 cs.CR cs.SI

classification cs.CRcs.SI
keywords graphreconstructioncommonneighborsmatrixadjacencyprivacyattacksecuremultipartycomputationco-squarenessspectraltopological
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

GRAND is a graph-reconstruction attack: given the matrix that counts common neighbors for every pair of vertices of an undirected graph, it rebuilds the graph's adjacency matrix. The paper proves that this is possible only up to a newly introduced equivalence, co-squareness, because distinct graphs can share exactly the same common-neighbors matrix. On real datasets the pipeline reconstructs the adjacency matrix exactly on Polblogs and achieves zero common-neighbors error on three of four networks even with no prior knowledge. The upshot is that a secure multiparty protocol computing this matrix leaks the underlying graph structure, since the output alone is enough to reproduce it.

What carries the argument

The central object is the common-neighbors matrix $G^2$, the square of the adjacency matrix, whose diagonal contains vertex degrees and whose off-diagonal entries count length-2 paths. The topological attacks rest on propositions linking row sums of $G^2$ to sums of neighbor degrees and linking matched degrees or common-neighbor counts to forced edge absences or presences; these run iteratively until no entry changes. The spectral attack then uses the fact that the eigenvalues of $G^2$ are the squares of the eigenvalues of $G$, so reconstruction reduces to fixing a sign for each square root, with the greedy choice guided by the Frobenius distance to a binary matrix and to known edges, applying the best low-rank approximation theorem. The introduced equivalence notion, co-squareness, identifies when two graphs share the same $G^2$ and thereby marks the unavoidable ambiguity of the problem.

What would settle it

Run the GRAND pipeline on the Polblogs dataset with the diagonal of $G^2$ treated as unknown and measure the reconstruction error: if the relative absolute error or the common-neighbors error rises, the diagonal of $G^2$ is load-bearing for the claimed attack.

Watch

Extended reading notes

Core claim

The paper claims that the common-neighbors matrix $G^2 = A^2$ of an undirected graph $G$ is, in practice, a near-complete blueprint of the adjacency matrix $A$. GRAND combines deterministic topological inferences — such as matching the degree implied by the diagonal of $G^2$ to the reconstructed neighborhood, or matching common-neighbor counts to infer absent edges — with a spectral heuristic that assigns the sign of each eigenvalue of $G^2$ (whose squares are the eigenvalues of $A$) so that the reconstructed matrix is binary and close to any partially known entries. The resulting pipeline recovers $A$ exactly on the Polblogs dataset and produces graphs whose square equals $G^2$ (zero common-neighbors error) on Netscience, Bio-diseasome and Polblogs without any prior knowledge. Because two co-square graphs can never be told apart from $G^2$ alone, the paper argues this is the theoretical ceiling for any reconstruction method.

Load-bearing premise

The attacker is assumed to have the exact, complete common-neighbors matrix $G^2$ including its diagonal, which reveals every vertex's degree; the paper explicitly leaves aside the scenario where the diagonal is unavailable.

Editorial extensions

If this is right

  • A protocol that privately outputs the common-neighbors matrix exposes the input graph: secure multiparty computation alone is not a sufficient privacy safeguard for graph data.
  • Without prior knowledge, no reconstruction algorithm can beat GRAND's zero common-neighbors error, because co-square graphs are indistinguishable by definition from $G^2$ alone.
  • Partial knowledge of existing and non-existing edges monotonically improves reconstruction, with relative error decreasing as the proportion of known entries grows on all four datasets.
  • Exact adjacency reconstruction is achievable on at least one real network (Polblogs) with no prior knowledge, so the threat is practical, not merely theoretical.

Reading between the lines

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

  • If the diagonal of $G^2$ were withheld or perturbed (e.g., by differential privacy), the degree-based topological attacks lose their main input; a plausible countermeasure is to protect the diagonal, and a version of GRAND that works from off-diagonal entries alone is a natural test.
  • The co-square equivalence class is a privacy ceiling: an ideal defense would randomize within the class so the attacker cannot know which co-square graph is real.
  • Because the spectral step is essentially a constrained sign-assignment problem, the success of the greedy heuristic suggests that the suspected NP-hardness may be concentrated in the sign choices, pointing toward approximation algorithms for larger graphs.
  • The same reconstruction logic likely transfers to directed graphs by counting common in-, out- or combined neighbors, which would extend the attack to a broader class of relational data.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies the problem of reconstructing an undirected graph G from its common-neighbors matrix G2, possibly combined with partial knowledge E0/E1. It introduces GRAND, a pipeline that first applies deterministic “topological attacks” derived from propositions about row sums and neighborhoods, then a spectral sign-assignment heuristic based on the SVD of G2, then a targeted error-forgetting step and a co-square instantiation. The paper also defines co-square equivalence and the Common Neighbors Error metric. It reports experiments on four datasets, claiming exact reconstruction on Polblogs and zero CNE on three datasets for an unknowledgeable adversary.

Significance. If the results held, GRAND would be a strong demonstration that secure multiparty protocols that release a common-neighbors matrix leak the graph structure, and the co-squareness notion would be a useful conceptual tool for graph equivalence. The topological propositions are simple and partially correct, and the decision to evaluate success against the square matrix (CNE) is methodologically appropriate. However, the two core deterministic attacks contain concrete errors (Algorithm 1 deletes true edges; Algorithm 5 does not implement Proposition 3), and the experimental claims are not backed by code, data, or variance information. The paper's main privacy conclusion is therefore not supported as written.

major comments (3)
  1. [Section 5.1, Algorithm 1 (lines 6 and 21–23)] The degree-combination rule is unsound. From Proposition 1, s = Σ_v G2(u,v) = Σ_{w∈Γ(u)} |Γ(w)|, hence s−d = Σ_{w∈Γ(u)}(|Γ(w)|−1). A neighbor of u can therefore have degree up to s−d+1, but the test G2(v,v) < s−d classifies every such neighbor as a non-candidate and line 22 sets the true edge to 0. In the graph consisting of a single edge (u,v), s=d=1, so s−d=0 and the true edge is deleted. More generally, every edge incident to a leaf is erased. This invalidates the claim in Section 5.1 that the topological attacks “recover G in an exact manner w.r.t. G2”.
  2. [Section 5.1, Algorithm 5] The pseudocode does not implement Proposition 3. The common-neighborhood part of Proposition 3 requires G2(u,v) − |Γ⋆(u)∩Γ⋆(v)| = |U|, where U is the set of unknown edges incident to u. Algorithm 5 instead tests G2(u,v) = |Γ⋆(u)|+|U| (and the symmetric test with v). These conditions are not equivalent; moreover, U as defined includes v itself when (u,v) is unknown, so the test is not the one stated in Proposition 3 and will fail to fire in many cases where the proposition applies. The NeighborCompletionAttack as written is therefore not the algorithm behind the reported completions.
  3. [Section 6 and Figures 11–13] The empirical claims (exact reconstruction of Polblogs, zero CNE on Netscience/Polblogs/Bio-diseasome) are averages over 10 runs but no standard deviations are reported, and no code or data are supplied. In view of the errors in Algorithms 1 and 5, the reported numbers cannot be attributed to the algorithm as described. A revision should provide the artifact and corrected algorithms, or substantially weaken the claims.
minor comments (5)
  1. [Section 3, paragraph 3] The sentence “This problem is more general than the one we solved in this paper, since we have more information” is unclear; additional information makes a problem easier, not more general.
  2. [Algorithm 5] The variable V is reused as both the set of all vertices and the local set of unknown neighbors of v; rename the local set to avoid confusion.
  3. [Conclusion, page 20] “Co-squarity” should be “co-squareness”.
  4. [References] References [14] and [15] are the same Motwani and Sudan paper; one duplicate should be removed.
  5. [Table 1] Γ2(x) is defined in the notation table but never used in the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CNE metric is the intended success criterion for the co-square reconstruction goal, and the exact Polblogs adjacency claim is evaluated against ground truth rather than against the method's own fitted values.

full rationale

The reconstruction task is explicitly defined as recovering G from the common-neighbors matrix G2 up to co-square equivalence; the CNE metric measures agreement between the reconstructed square G-hat^2 and the input G2, which is the correct objective for that task rather than a circular validation. The stronger experimental claim, exact adjacency reconstruction on Polblogs, is scored with FPR, FNR, and RAE against the original graph, so it is not forced by construction from G2. The spectral-attack weights alpha and beta and the binarization threshold t are manually chosen (beta from known-entry fraction, alpha as its complement, t from the prior work of Erdős et al.), not fitted to a held-out target, so no fitted parameter is renamed as a prediction. The only self-citation, Crypto'Graph [1], is used for motivating the protocol setting and is not load-bearing in the reconstruction derivation. The paper also explicitly flags the assumption that the diagonal of G2 is available and states that the missing-diagonal scenario is not studied; this is a data-availability limitation, not a circular step. Algorithmic soundness concerns raised against Algorithms 1 and 5 are correctness issues, not circularity, and are therefore outside this pass. Overall, the derivation chain is self-contained with respect to circularity.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard linear algebra (eigendecomposition of symmetric matrices, Eckart-Young) plus domain assumptions about the attacker's knowledge. The manually chosen spectral threshold and weights are the only free numerical parameters; they are not fitted to a held-out set. No new entities are introduced; co-squareness is a definition, not an entity.

free parameters (2)
  • binarization threshold t = 0.5
    Used in the spectral attack to convert real-valued matrix entries to 0/1. Chosen by hand, following the value reported in [8].
  • spectral weights alpha, beta = beta = 2*|E*|/|V|^2, alpha = 1 - beta
    Trade off binarization fidelity versus known-edge fidelity in the spectral sign choice. Set by a heuristic formula, not fitted to a held-out set.
assumptions (6)
  • domain assumption Graph G is undirected, simple, and has no self-loops.
    Section 2 defines the problem on such graphs; the common-neighbors matrix is the squared adjacency matrix.
  • domain assumption The attacker obtains the exact full common neighbors matrix G2, including its diagonal (degrees).
    The degree-based attacks rely on G2(u,u). The paper notes in Section 3 that cases without the diagonal are out of scope.
  • domain assumption Prior edge knowledge E0 and E1 is correct.
    Section 2 defines E0/E1 as true non-edges and edges in G; the experiments sample them from the true matrix.
  • standard math Since G is symmetric, G2 shares the same eigenvectors as G, with squared eigenvalues.
    Section 4.3.2 derives G = U*Sigma*U^T from the eigendecomposition of G2.
  • standard math Eckart-Young theorem gives the best rank-k Frobenius approximation.
    Used in Theorem 1 and the greedy sign-assignment logic of Section 5.2.
  • ad hoc to paper The greedy sign assignment with the chosen alpha/beta objective finds the true eigenvalues of G.
    Section 5.2 offers a heuristic with no guarantee; the early-iteration dominance is an empirical claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GRAND : Graph Reconstruction from potential partial Adjacency and Neighborhood Data." pith.science (2026). https://pith.science/paper/7C22VQ6B

@misc{pith2026241202329,
  author       = {Pith},
  title        = {Pith review of: GRAND : Graph Reconstruction from potential partial Adjacency and Neighborhood Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7C22VQ6B}},
  note         = {Machine review of arXiv:2412.02329}
}
read the original abstract

Cryptographic approaches, such as secure multiparty computation, can be used to compute in a secure manner the function of a distributed graph without centralizing the data of each participant. However, the output of the protocol itself can leak sensitive information about the structure of the original graph. In particular, in this work we propose an approach by which an adversary observing the result of a private protocol for the computation of the number of common neighbors between all pairs of vertices, can reconstruct the adjacency matrix of the graph. In fact, this can only be done up to co-squareness, a notion we introduce, as two different graphs can have the same matrix of common neighbors. We consider two models of adversary, one who observes the common neighbors matrix only, and a knowledgeable one, that has a partial knowledge of the original graph. Our results demonstrate that secure multiparty protocols are not enough for privacy protection, especially in the context of highly structured data such as graphs. The reconstruction that we propose is interesting in itself from the point of view of graph theory.

Figures

Figures reproduced from arXiv: 2412.02329 by the authors.

Figure 1
Figure 1. Reconstruction errors on Netscience, in terms of the adjacency matrix and the common [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Example of a graph G⋆ and its adjacency matrix. 4.1 Domains of study Consider the adjacency matrix of the graph G, denoted as MG. It can be expressed as: MG = UΛU T (1) where U is an orthogonal matrix and Λ is a diagonal matrix containing the eigenvalues of MG. This decomposition highlights that there are two equivalent representations of the graph information: the left-hand side of Equation (1), which we call the t… view at source ↗
Figure 3
Figure 3. Two co-square (and non co-spectral) graphs [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: DegreeCombinationAttack. The sum of the row of vertex 3 in G2 is equal to the sum of the degrees of 1, 2, 4 and 5. Thus, the edges (1, 3), (2, 3), (4, 3), (5, 3) do exist in G. DegreeCombinationAttack Proposition 1 allows the following inference : given a vertex u, if …
Figure 5
Figure 5. Figure 5: DegreeMatchingAttack. The degree of vertex 5 (1) is equal to the number of its currently known neighbors in G⋆ . Thus all unknown edges in the neighborhood of 5 cannot exist. edges in G and puts 0’s in G⋆ by comparing the degrees of vertices on G⋆ and G. An example of …
Figure 6
Figure 6. Figure 6: DegreeCompletionAttack. Vertex 1 has one known neighbor which is 3, and two unknown edges. Also, it should have 3 neighbors in total. Therefore all its unknown edges do exist in G. Algorithm 4 Identifies existing edges by comparing the neighborhoods of vertices on G an…
Figure 7
Figure 7. Figure 7: TriangleAttack. Edge (1, 2) is known. Also, vertices 1 and 2 have one common neighbor. Then, there exists one triangle containing 1 and 2. Moreover, 4 is the only vertex that has common neighbors with both 1 and 2. Therefore 4 is the vertex forming a triangle with 1 an…
Figure 8
Figure 8. Figure 8: BicliqueAttack. The neighbors of vertex 5 (which are 2 and 3) are all known, because its degree on G⋆ is equal to G2 (5, 5). Therefore, vertex 1, which has two common neighbors with 5, is also connected to 2 and 3. BicliqueAttack When the neighborhood of a certain vert…
Figure 9
Figure 9. Figure 9: Co-square sub-graphs in Netscience. The plain lines are reconstructed edges, the dotted [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: GRAND (54, 227) or by (53, 227) and (54, 226). Interestingly, the first graph of 5 vertices (hence detached from Netscience) and its two co-square graphs (which we just described) are the smallest co-square graphs. They are isomorphic, contrarily to the first co-squar…
Figure 11
Figure 11. Figure 11: False Positives Rate, False Negatives Rate, Relative Absolute Error and Common Neigh [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Relative Absolute Error (RAE) with respect to the amount of prior knowledge [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]
Figure 13
Figure 13. Figure 13: Common Neighbors Error with respect to the amount of prior knowledge [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: Number of modifications made by the topological attacks. [PITH_FULL_IMAGE:figures/full_fig_p020_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 17 canonical work pages

  1. [8]

    Reconstructing graphs from neighborhood data

    D´ ora Erd¨ os, Rainer Gemulla, and Evimaria Terzi. Reconstructing graphs from neighborhood data. In 2012 IEEE 12th International Conference on Data Mining , pages 231–240, 2012. doi:10.1109/ICDM.2012.154

  2. [1]

    Crypto’graph: Leveraging privacy-preserving distributed link prediction for robust graph learn- 20 ing

    Sofiane Azogagh, Zelma Aubin Birba, S´ ebastien Gambs, and Marc-Olivier Killijian. Crypto’graph: Leveraging privacy-preserving distributed link prediction for robust graph learn- 20 ing. 2024

  3. [2]

    Characterizing and recog- nizing exact-distance squares of graphs

    Yandong Bai, Pedro P Cort´ es, Reza Naserasr, and Daniel A Quiroz. Characterizing and recog- nizing exact-distance squares of graphs. Discrete Mathematics, 347(8):113493, 2024

  4. [3]

    Chv´ etal

    V. Chv´ etal. Recognizing intersection patterns. In M. Deza and I.G. Rosenberg, editors, Com- binatorics 79 Part I , volume 8 of Annals of Discrete Mathematics , pages 249–251. Elsevier,

  5. [4]

    Fast sequential creation of random realizations of degree sequences

    Brian Cloteaux. Fast sequential creation of random realizations of degree sequences. Internet Mathematics, 12(3):205–219, 2016

  6. [5]

    Applications of link prediction in social networks: A review

    Nur Nasuha Daud, Siti Hafizah Ab Hamid, Muntadher Saadoon, Firdaus Sahran, and Nor Badrul Anuar. Applications of link prediction in social networks: A review. Journal of Network and Computer Applications , 166:102716, 2020

  7. [6]

    Privacy-preserving link prediction

    Didem Demirag, Mina Namazi, Erman Ayday, and Jeremy Clark. Privacy-preserving link prediction. In Joaquin Garcia-Alfaro, Guillermo Navarro-Arribas, and Nicola Dragoni, editors, Data Privacy Management, Cryptocurrencies and Blockchain Technology , pages 35–50, Cham,

  8. [7]

    The approximation of one matrix by another of lower rank

    Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211–218, 1936

Show all 30 references
  1. [9]

    A survey of link prediction in social networks

    Mohammad Al Hasan and Mohammed J Zaki. A survey of link prediction in social networks. Social network data analytics , pages 243–275, 2011

  2. [10]

    Matrix analysis

    Roger A Horn and Charles R Johnson. Matrix analysis . Cambridge university press, 2012

  3. [11]

    Recognizing graphic degree sequences and generating all realizations

    Zolt´ an Kir´ aly. Recognizing graphic degree sequences and generating all realizations. E¨ otv¨ os Lor´ and University, Tech. Rep. Egres TR-2011-11, 2012

  4. [12]

    Deeprobust: A pytorch library for adversarial attacks and defenses

    Yaxin Li, Wei Jin, Han Xu, and Jiliang Tang. Deeprobust: A pytorch library for adversarial attacks and defenses. arXiv preprint arXiv:2005.06149 , 2020

  5. [13]

    The structural analysis of criminal networks

    Duncan McAndrew. The structural analysis of criminal networks. The Social Psychology of Crime, 2021

  6. [14]

    Computing roots of graphs is hard

    Rajeev Motwani and Madhu Sudan. Computing roots of graphs is hard. Discrete Applied Mathematics, 54(1):81–88, 1994

  7. [15]

    Computing roots of graphs is hard

    Rajeev Motwani and Madhu Sudan. Computing roots of graphs is hard. Discrete Applied Mathematics, 54(1):81–88, 1994. URL: https://www.sciencedirect.com/science/article/ pii/0166218X94000239, doi:10.1016/0166-218X(94)00023-9

  8. [16]

    Biological network analysis with deep learning

    Giulia Muzio, Leslie O’Bray, and Karsten Borgwardt. Biological network analysis with deep learning. Briefings in bioinformatics , 22(2):1515–1530, 2021

  9. [17]

    Using graph theory to analyze biological networks

    Georgios A Pavlopoulos, Maria Secrier, Charalampos N Moschopoulos, Theodoros G Soldatos, Sophia Kossida, Jan Aerts, Reinhard Schneider, and Pantelis G Bagos. Using graph theory to analyze biological networks. BioData mining , 4:1–27, 2011

  10. [18]

    Rossi and Nesreen K

    Ryan A. Rossi and Nesreen K. Ahmed. The network data repository with interactive graph analytics and visualization. In AAAI, 2015. URL: https://networkrepository.com

  11. [19]

    Silva, Ing-Ren Tsang, George D

    Nitai B. Silva, Ing-Ren Tsang, George D. C. Cavalcanti, and Ing-Jyh Tsang. A graph-based friend recommendation system using genetic algorithm. In IEEE Congress on Evolutionary Computation, pages 1–7, 2010

  12. [20]

    van Dam and Willem H

    Edwin R. van Dam and Willem H. Haemers. Which graphs are determined by their spec- trum? Linear Algebra and its Applications , 373:241–272, 2003. Combinatorial Matrix Theory Conference (Pohang, 2002). URL: https://www.sciencedirect.com/science/article/pii/ S002437950300483X, d...

  13. [21]

    Graph spectra for complex networks

    Piet Van Mieghem. Graph spectra for complex networks . Cambridge university press, 2023

  14. [22]

    Co-eigenvector graphs

    Piet Van Mieghem and Ivan Joki´ c. Co-eigenvector graphs. Linear Algebra and its Appli- cations, 689:34–59, 2024. URL: https://www.sciencedirect.com/science/article/pii/ S0024379524000521, doi:10.1016/j.laa.2024.02.008

  15. [23]

    Detection of drug-drug interactions by modeling interaction profile fingerprints

    Santiago Vilar, Eugenio Uriarte, Lourdes Santana, Nicholas P Tatonetti, and Carol Friedman. Detection of drug-drug interactions by modeling interaction profile fingerprints. PloS one , 8(3):e58321, 2013

  16. [24]

    Puttaswamy, and Ben Y

    Christo Wilson, Bryce Boe, Alessandra Sala, Krishna P.N. Puttaswamy, and Ben Y. Zhao. User interactions in social networks and their implications. In Proceedings of the 4th ACM European Conference on Computer Systems , EuroSys ’09, page 205–218, New York, NY, USA, 2009. Associ...

  17. [25]

    Ad- versarial examples for graph data: Deep insights into attack and defense

    Huijun Wu, Chen Wang, Yuriy Tyshetskiy, Andrew Docherty, Kai Lu, and Liming Zhu. Ad- versarial examples for graph data: Deep insights into attack and defense. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19 , pages 4816–4...

  18. [26]

    Gunter, and Bo Li

    Xiaojun Xu, Hanzhang Wang, Alok Lal, Carl A. Gunter, and Bo Li. Edog: Adversarial edge detection for graph neural networks. In 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML) , pages 291–305, 2023. doi:10.1109/SaTML54575.2023.00027

  19. [27]

    Predicting potential drug-drug interactions by integrating chemical, biological, phenotypic and network data

    Wen Zhang, Yanlin Chen, Feng Liu, Fei Luo, Gang Tian, and Xiaohong Li. Predicting potential drug-drug interactions by integrating chemical, biological, phenotypic and network data. BMC bioinformatics, 18:1–12, 2017

  20. [28]

    Adversarial attacks on neural networks for graph data

    Daniel Z¨ ugner, Amir Akbarnejad, and Stephan G¨ unnemann. Adversarial attacks on neural networks for graph data. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , KDD ’18, page 2847–2856, New York, NY, USA, 2018. Association...

  21. [1980]

    URL: https://www.sciencedirect.com/science/article/pii/S0167506008708835, doi:10.1016/S0167-5060(08)70883-5

  22. [2023]

    Springer International Publishing

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.