acceptodds
Under review as a conference paper at ICLR 2027

Stealthy Yet Effective: In-Distribution Node Injection Attacks for Hyperedge Prediction

Abstract

Deep hypergraph models are widely adopted in real-world decision-making systems due to their powerful higher-order modeling capabilities for complex data, making their security and robustness increasingly critical. Recent studies have revealed adversarial vulnerabilities in deep hypergraph models on node-level tasks that predict individual node properties. However, their robustness in hyperedge prediction, a fundamental task for inferring missing or future group-level interactions among multiple entities, remains largely unexplored. To fill this gap, we systematically analyze mainstream injection-based attacks and identify two key challenges in applying them to hyperedge prediction: (1) Perturbation attenuation: we theoretically demonstrate that group-level aggregation in edge prediction weakens the adversarial influence of injected nodes, limiting attack effectiveness. (2) Distributional deviation: empirical results reveal pronounced shifts in membership- and hyperedge-level statistical distributions after injection, compromising attack stealthiness. To tackle these challenges, we propose DPHIA, the first effective and unnoticeable node injection attack tailored for hyperedge prediction. DPHIA employs an adaptive ranking-based objective to perturb the decision boundary between positive and negative hyperedges, thereby misleading the predictor. A dual-level adversarial regularizer constrains distributional shifts in higher-order relations through membership- and hyperedge-level adversarial training, improving attack stealthiness. Extensive experiments on five real-world datasets demonstrate that DPHIA achieves a superior trade-off between attack effectiveness and detectability compared with state-of-the-art baselines. Our code is available at https://anonymous.4open.science/r/DPHIA-063D.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Hyperedge Prediction, Node Injection Attacks, Deep Hypergraph Learning, Security

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_CuzbOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The authors propose a new framework of node injection attacks for hyperedge prediction, called Distribution-Preserving Hypergraph Injection Attack (DPHIA). The ultimate goal is to design effective and stealthy attacks on hypergraphs. To achieve this aim, the authors combine the choice of the uncertainty-guided injection with distribution-preserving feature optimization, by using an adaptive ranking-based attack objective and dual-level adversarial regularization. It is important to note that adversarial hypergraph learning is one of the rapidly developing directions over the last few years, but it is still substantially behind the conventional adversarial graph learning. This paper provides an interesting new contribution to this important, yet largely uncharted domain.

Critical strengths

The paper focuses on the important but largely under-investigated topic of adversarial learning of hypergraph-structured data. Most existing studies in this domain target node classification, while adversarial robustness of hyperedge prediction remains virtually unexplored. To the best of my knowledge, this paper is the first one to address adversarial hyperedge prediction and offers a new and important contribution to the existing body of knowledge.

The paper also presents a comprehensive study on node injection attacks for hyperedge prediction and their deficiencies, which is of interest on its own.

Critical weaknesses

My main criticism is that the considered datasets, i.e., Citeseer, Cora etc, are not inherently hypergraphs. Of course, other papers on adversarial hypergraph learning use these data, but from my perspective, it undermines the paper value. It will be much more interesting and valuable, especially in the context of hyper-edge prediction, to consider the data which are intrinsically described by a hypergraph. Probably, some knowledge graphs will be a good option to consider.

A related comment, it will be beneficial to have a deeper discussion comparing the scenarios of hyperedge vs. edge prediction, so that the peculiarities of hypergraphs are pronounced stronger.

Another major concern is that the numerical results are provided without assessment of variability, no standard deviations are reported. As such, it is hard to judge how much better (if better) DPHIA is, given uncertainties.

Questions for authors

See comments above

12 comments
AuthorsOpenReview ↗

Overall response to Reviewer Cuzb

Thank you for your thoughtful and encouraging feedback on our paper. We sincerely appreciate your recognition of the importance of adversarial hyperedge prediction and the novelty of our contribution to this underexplored area. We are also grateful for your positive assessment of our comprehensive analysis of existing node injection attacks and their limitations.

We address your concerns as follows.

[Preview]

  • Weakness #1: Evaluation on more intrinsic hypergraphs.
  • Weakness #2: Discussion about hyperedge prediction and edge prediction.
  • Weakness #3: Experimental variability and stability.
  • Conclusion.
AuthorsOpenReview ↗

Response to Weakness #1

W1. Evaluation on more intrinsic hypergraphs.

A1. Thank you for this thoughtful comment. Our original evaluation follows widely adopted benchmarks for hyperedge prediction, enabling comparison with prior work. We fully agree that incorporating additional datasets with more explicit real-world semantics of higher-order interactions would further strengthen our work.

Following your suggestion, we extend our evaluation to five additional intrinsic hypergraphs: (a) two knowledge hypergraphs and (b) three hypergraphs capturing real-world group interactions. All experiments follow the main-paper protocol, with HyGEN as the victim.

(a) Knowledge hypergraphs. We adopt WD50K (100) [1] from Wikidata and FB-AUTO [2] from Freebase, both dominated by higher-arity facts (arity > 2) and used in recent knowledge hypergraph learning [3]. Knowledge hypergraphs encode multi-entity factual relations. We represent each entity as a node and each fact as a hyperedge connecting its participating entities. Relation types and entity roles are discarded, duplicate entity sets are merged, and entity names and descriptions are encoded using Qwen3-Embedding-0.6B as node features.

Table A1(a) shows that DPHIA achieves the best overall attack performance, with the lowest AUC on both knowledge hypergraphs and the lowest AP on FB-AUTO. Specifically, DPHIA reduces AUC by 6.47 and 19.05 points on WD50K (100) and FB-AUTO, respectively. These results demonstrate DPHIA's effectiveness on intrinsic higher-order knowledge relations, beyond conventional hyperedge prediction benchmarks.

Table A1(a). Results on knowledge hypergraphs (AUC | AP, %).

DomainDatasetBenignRandomFGATH-AttackHyperNearDPHIA
Open-domainWD50K (100)65.11 | 62.1262.16 | 62.2064.12 | 60.4961.10 | 59.6061.15 | 58.4258.64 | 60.40
AutomotiveFB-AUTO74.18 | 73.3173.34 | 74.5062.61 | 63.2263.90 | 59.6268.70 | 65.2855.13 | 53.12

(b) Real-world group interactions. We further evaluate three hypergraphs from TAHB [4], whose hyperedges capture naturally occurring group relationships across diverse domains:

  • Amazon Pet Supplies (e-commerce): Nodes represent products, with each hyperedge connecting products purchased and reviewed by the same user.
  • TMDB (movies): Nodes represent movies, with each hyperedge grouping movies featuring the same actor.
  • House (politics): Nodes represent legislative bills, with each hyperedge connecting bills sponsored by the same legislator.

As shown in Table A1(b), DPHIA achieves the lowest AP on all three datasets and the lowest AUC on two of them. Notably, on House, DPHIA reduces AUC by 16.02 percentage points, compared with 11.78 percentage points for the strongest baseline. These results further support DPHIA's effectiveness across diverse real-world higher-order interactions.

Table A1(b). Results on real-world group interactions (AUC | AP, %).

DomainDatasetBenignRandomFGATH-AttackHyperNearDPHIA
E-commerceAmazon Pet Supplies67.65 | 69.8367.32 | 69.7666.68 | 68.8867.62 | 69.9161.62 | 63.6963.29 | 62.80
MoviesTMDB52.25 | 54.6552.57 | 54.8955.15 | 57.6452.24 | 54.4751.93 | 53.8350.27 | 51.64
PoliticsHouse64.25 | 64.1158.56 | 59.0158.47 | 59.4663.91 | 63.6652.47 | 54.4648.23 | 50.47

We have included the new results and analysis in Appendix G.10 (Tables 16–17) of the revised manuscript.

AuthorsOpenReview ↗

Response to Weakness #2

W2. Discussion about hyperedge prediction and edge prediction.

A2. Thank you for this valuable comment. Following your suggestion, we provide a deeper comparison between the two scenarios below.

(a) Comparison between edge and hyperedge prediction. Edge prediction aims to predict whether a pair of nodes forms an edge, whereas hyperedge prediction aims to predict whether a candidate set of multiple nodes forms a hyperedge. This distinction introduces two fundamental challenges:

  • Node aggregation. Edge prediction typically combines the representations of two endpoint nodes through pairwise scoring operations (e.g., dot product or concatenation). In contrast, hyperedge prediction requires aggregating variable numbers of member nodes with potentially different contributions to capture more complex group-level interactions.
  • Combinatorial candidate space. For nn nodes, edge prediction considers (n2)\binom{n}{2} possible undirected edges, whereas hyperedge prediction involves a potentially exponential candidate space of O(2n)O(2^n) node subsets, introducing substantially greater computational complexity and challenges in accurate candidate prediction.

(b) Differences in adversarial attack settings. Existing attacks on edge prediction primarily perturb graph edges to degrade prediction performance [5, 6] or inject malicious nodes to manipulate specific edge predictions [7]. To the best of our knowledge, untargeted node injection poisoning for edge prediction remains relatively underexplored. More importantly, directly adapting pairwise attacks to hyperedge prediction faces two distinct challenges:

  • Group-level perturbation attenuation. Pairwise attacks manipulate node representations or pairwise prediction scores without explicitly considering how group-level aggregation can attenuate node-level perturbations (Lemma 1 and Corollary 1). DPHIA addresses this through an adaptive ranking-based objective that directly optimizes post-aggregation scores to disrupt positive–negative hyperedge rankings.
  • Higher-order stealthiness constraints. Unlike attacks targeting pairwise relations, injecting a node into a hyperedge changes its interactions with an entire group, potentially introducing membership- and hyperedge-level distributional deviations. DPHIA addresses these deviations through dual-level adversarial regularization.

Together, these distinctions highlight the peculiarities of hypergraphs for adversarial attacks, and the need for task-specific attack designs beyond directly transferring existing edge prediction attacks.

We have incorporated this discussion into Appendix H of the revised manuscript to further clarify the distinctions between hyperedge and edge prediction.

Reviewer_p66POpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a node injection poisoning framework (i.e., DPHIA) for hyperedge prediction that combines uncertainty-guided target selection, adaptive ranking-based optimization, and dual-level adversarial regularization. Experiments across five datasets and five victim models demonstrate strong attack effectiveness and resilience to the evaluated filtering defenses.

Critical strengths
  1. The paper addresses an underexplored robustness problem and presents a coherent framework that considers both attack effectiveness and stealthiness.

  2. Evaluations across multiple victim models, surrogate architectures, and negative sampling strategies, together with component ablations, provide broad empirical support.

Critical weaknesses
  1. The results assume max–min pooling and a fixed prediction head, whereas victim models are retrained on poisoned data. Further evidence is needed to establish whether perturbation attenuation explains the observed behavior across architectures.

  2. The analysis primarily uses the proposed MD and HD measures and their corresponding filtering defenses. Independent detectors, quantitative detection metrics, and filtering-threshold sensitivity would better substantiate the in-distribution claim.

  3. Regularization weights are tuned separately for different defenses, but the available feedback is unspecified. Results using a fixed configuration selected without victim feedback would strengthen the transferability claims.

  4. The datasets use bag-of-words attributes, but constraints on generated features, including nonnegativity, sparsity, and magnitude, are unclear. Evaluations under realistic feature constraints would strengthen the practical relevance.

  5. Although experiments use five random splits, the tables provide only average results. Standard deviations or confidence intervals would help assess the reliability of improvements, particularly when differences between methods are small.

  6. Some relevant references are missing, such as [1–3]. Including these studies would strengthen the related-work discussion and better contextualize the proposed method.

[1] TDGIA: Effective Injection Attacks on Graph Neural Networks.

[2] GANI: Global Attacks on Graph Neural Networks via Imperceptible Node Injections.

[3] GZOO: Black-Box Node Injection Attack on Graph Neural Networks via Zeroth-Order Optimization.

Questions for authors

Please refer to the weakness section.

10 comments
AuthorsOpenReview ↗

Overall response to Reviewer p66P

Thank you for your detailed and constructive feedback on our paper. We sincerely appreciate your recognition of the significance of our attack setting and DPHIA's joint consideration of attack effectiveness and stealthiness. We are also grateful for your positive assessment of our comprehensive empirical evaluation.

We respond to your concerns as follows.

[Preview]

  • Weakness #1: Theoretical generality and further evidence.
  • Weakness #2: Independent detectors, quantitative detection metrics, and filtering-threshold sensitivity.
  • Weakness #3: Selection of regularization weights.
  • Weakness #4: Realistic constraints on generated features.
  • Weakness #5: Experimental variability and stability.
  • Weakness #6: Missing references on graph node injection attacks.
  • Conclusion.
AuthorsOpenReview ↗

Response to Weakness #1

W1. Theoretical generality and further evidence.

A1. Thank you for this insightful comment. We first clarify the assumptions in Section 3.1. (1) Max–min pooling is widely adopted in representative hyperedge prediction models (NHP, AHP, and HyGEN) to capture the coherence of member embeddings within a hyperedge. We therefore use it as a representative mechanism to analyze group-level perturbation attenuation. (2) The fixed prediction head serves to isolate the effect of node perturbations on candidate scores, rather than implying that victim models are not retrained.

To address your concern, we (a) generalize Lemma 1 and Corollary 1 to different aggregators, (b) extend the analysis to account for victim retraining, and (c) discuss the connection between perturbation attenuation and attack effectiveness.

(a) Generalized Lemma 1. For a candidate cc, let qc=agg({zv}v∈c)q_c=\mathrm{agg}(\lbrace z_v\rbrace_{v\in c}) and q^c=agg({zv+δv}v∈c)\hat q_c=\mathrm{agg}(\lbrace z_v+\delta_v\rbrace_{v\in c}). We decompose each node perturbation as δv=ac+rv\delta_v=a_c+r_v, where ac=1∣c∣∑v∈cδva_c=\frac{1}{|c|}\sum_{v\in c}\delta_v is the shared shift and rvr_v is the member-specific residual.

Proposition 1 (Aggregation sensitivity). Under the corresponding Lipschitz conditions, the aggregated representation shift satisfies ∥q^c−qc∥∞≤Bc\lVert\hat q_c-q_c\rVert_\infty\le B_c, with aggregation-specific bounds:

AggregatorBound BcB_c
Max–min2max⁡v∈c∥rv∥∞2\max_{v\in c}\lVert r_v\rVert_\infty (Lemma 1)
Mean∥ac∥∞\lVert a_c\rVert_\infty (exact)
Deep SetsLρLϕ∣c∣∑v∈c∥δv∥∞\frac{L_\rho L_\phi}{\lvert c\rvert}\sum_{v\in c}\lVert\delta_v\rVert_\infty
Attention∥W∥∞(1+2LsRc)max⁡v∈c∥δv∥∞\lVert W\rVert_\infty(1+2L_sR_c)\max_{v\in c}\lVert\delta_v\rVert_\infty

Here, LϕL_\phi and LρL_\rho are the Lipschitz constants of the Deep Sets mappings with normalized mean pooling; LsL_s bounds attention-score sensitivity, WW is the value projection, and Rc=max⁡v∈c∥zv∥∞R_c=\max_{v\in c}\lVert z_v\rVert_\infty.

Building on Proposition 1, we further generalize Corollary 1:

Corollary 2 (Aggregation-dependent ranking reversal). For an LaL_a-Lipschitz prediction head and a correctly ranked positive–negative pair (c+,c−)(c^+,c^-) with initial probability gap Δ>0\Delta>0, ranking reversal requires

Δ≤La(Bc++Bc−),\Delta\le L_a(B_{c^+}+B_{c^-}),

where BcB_c is defined in Proposition 1. For max–min pooling, Bc=2εcB_c=2\varepsilon_c, recovering Eq. (3).

(b) Extension to victim retraining. We further extend the ranking analysis to retrained victims. Let ϑ\vartheta and ϑ^\hat\vartheta denote the victim parameters trained on the clean hypergraph H\mathcal H and poisoned hypergraph H^\widehat{\mathcal H}, respectively. The prediction shift decomposes into injection and retraining effects:

\underbrace{p_c(\hat\vartheta,\widehat{\mathcal H})-p_c(\hat\vartheta,\mathcal H)}_ {\text{Injection effect}} + \underbrace{p_c(\hat\vartheta,\mathcal H)-p_c(\vartheta,\mathcal H)}_ {\text{Retraining effect } \xi_c},$$ where $p_c=p_c(\vartheta,\mathcal H)$ and $\hat p_c=p_c(\hat\vartheta,\widehat{\mathcal H})$. **Corollary 3 (Ranking reversal under victim retraining).** For a correctly ranked positive–negative pair $(c^+,c^-)$ with initial probability gap $\Delta=p_{c^+}-p_{c^-}>0$, ranking reversal after retraining requires $$\Delta\le \hat L_a(\hat B_{c^+}+\hat B_{c^-})+(\xi_{c^-}-\xi_{c^+}),$$ where $\hat L_a$ and $\hat B_c$ denote the prediction-head Lipschitz constant and aggregation-specific perturbation bound under fixed retrained parameters, respectively, and $\xi_c$ captures retraining-induced probability shifts. This result shows that **injection-induced score changes remain subject to aggregation-dependent bounds after retraining**, while retraining introduces additional shifts that may facilitate or hinder ranking reversal. **(c) Implications for attack effectiveness.** Proposition 1 and Corollary 2 show that **aggregation-dependent bounds can limit prediction-score changes**, since $|\hat p_c-p_c|\le L_aB_c$ under fixed parameters. Consequently, when $L_a(B_{c^+}+B_{c^-})<\Delta$, the perturbation is insufficient to reverse the candidate ranking. Corollary 3 further extends this analysis to retrained victims by separating injection and retraining effects. Together, these results **offer theoretical insights into how aggregation-dependent sensitivity can constrain attack effectiveness across different architectures**, further motivating DPHIA's adaptive ranking-based objective, which directly optimizes post-aggregation scores and prioritizes candidate pairs with smaller probability gaps. **We have incorporated these theoretical extensions into Section 3.1 and provided their formal statements and proofs in Appendix A.3-A.5 of the revised manuscript.**
AuthorsOpenReview ↗

Response to Weakness #2

W2. Independent detectors, quantitative detection metrics, and filtering-threshold sensitivity.

A2. Thank you for this thoughtful comment. Following your suggestion, we conducted additional experiments to further substantiate DPHIA's in-distribution stealthiness from two perspectives: (a) detectability under independent detectors, evaluated using quantitative detection metrics (AP and Recall@kk), and (b) sensitivity to different MD/HD filtering thresholds.

(a) Independent detectors with quantitative metrics. We evaluate five detectors unseen during attack training, covering feature-level (Isolation Forest [1], LOF [2]), membership-level (Hyper-SAGNN [3]), and hyperedge-level (HAD [4], HashNWalk [5]) anomaly metrics. Given the sparsity of injected anomalies, we report AP and Recall@kk, where kk equals the number of injected units at each level, to quantify detection effectiveness under class imbalance.

As shown in Table A2(a), while HyperNear and TH-Attack are highly detectable by certain independent detectors, DPHIA consistently maintains low detection AP (≤26.38%) and Recall@kk (≤37.50%) across all five detectors on both datasets. Notably, under Isolation Forest, HyperNear and TH-Attack achieve AP ≥97.99% and Recall@kk ≥94.44%, whereas DPHIA achieves substantially lower detection scores. These results support DPHIA's more consistent stealthiness across diverse detection mechanisms.

Table A2(a). Detection performance of independent detectors (AP | Recall@kk, %).

DatasetLevelDetectorHyperNearTH-AttackDPHIA
CiteseerFeatureIsolation Forest100.00 | 100.0097.99 | 94.4426.38 | 37.50
FeatureLOF3.22 | 0.0067.20 | 70.838.79 | 11.11
MembershipHyper-SAGNN2.38 | 0.0037.99 | 46.232.68 | 0.00
HyperedgeHAD20.05 | 24.0010.29 | 5.5619.26 | 23.61
HyperedgeHashNWalk4.45 | 0.009.32 | 1.396.82 | 0.00
CoraFeatureIsolation Forest100.00 | 100.00100.00 | 100.005.50 | 0.00
FeatureLOF2.98 | 0.0047.84 | 60.5618.73 | 25.35
MembershipHyper-SAGNN1.54 | 0.0010.88 | 11.383.05 | 2.82
HyperedgeHAD10.46 | 5.7110.71 | 14.087.08 | 1.41
HyperedgeHashNWalk7.23 | 0.006.79 | 0.005.86 | 1.41

(b) Filtering-threshold sensitivity. To examine whether DPHIA's stealthiness depends on a particular filtering threshold, we vary the filtering ratios of MD-filtering and HD-filtering in Section 5.3 over {1, 3, 5, 10, 15, 20}% on Citeseer, retraining HyGEN after filtering. We apply identical filtering ratios to clean hypergraphs for comparison.

As shown in Table A2(b), DPHIA remains effective with increasing filtering ratios, with AUC at least 13.27 points below the corresponding filtered clean results. Even under aggressive filtering, the maximum AUC recovery is only 5.27 points relative to the unfiltered attack. These results demonstrate DPHIA's resilience to structural filtering across different MD/HD thresholds.

Table A2(b). Sensitivity to MD/HD filtering ratios on Citeseer (AUC | AP, %).

RatioMD-filtering (Clean)MD-filtering (DPHIA)HD-filtering (Clean)HD-filtering (DPHIA)
0%83.31 | 83.3264.07 | 67.4583.31 | 83.3264.07 | 67.45
1%84.44 | 83.4263.80 | 67.5884.39 | 83.5163.75 | 66.39
3%84.00 | 83.4363.15 | 67.6283.48 | 82.6463.14 | 66.41
5%83.19 | 82.7664.34 | 67.2385.17 | 83.7263.28 | 66.28
10%82.74 | 81.2167.75 | 69.7383.94 | 82.6868.88 | 71.17
15%83.63 | 81.9266.55 | 68.8282.61 | 81.2569.34 | 71.19
20%83.78 | 81.3767.07 | 70.3282.66 | 80.3868.83 | 72.50

We have incorporated the independent detection experiments with quantitative metrics into Section 5.4 (Table 3), and the filtering-threshold sensitivity analysis into Appendix G.7 (Table 13) of the revised manuscript.

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

This paper proposes DPHIA, a node injection poisoning attack against hyperedge prediction, the task of inferring missing group links in hypergraphs. They motivate the method with two analyses. First, because the model scores a candidate group by pooling its members' embeddings, perturbing individual nodes has limited effect; the attacker must instead flip the ranking between true and false candidate groups. Second, injected nodes stand out statistically, both from their fellow hyperedge members and from clean hyperedges, which makes them detectable. DPHIA addresses both points: it injects nodes into hyperedges near the decision boundary, then trains a generator with a ranking-based attack objective, together with membership- and hyperedge-level discriminators that keep the injection hard to detect, via a min–max loop with an adaptive surrogate. Across five datasets, five victim models, and a black-box setting, DPHIA lowers AUC and AP more than the baselines.

Critical strengths
  1. The paper studies a new attack target.
  2. The design follows from the analysis. The attenuation result explains why the attack must flip candidate rankings. The deviation result explains why the attack must hide its injected nodes.
  3. The attack works where the baselines do not.
Critical weaknesses
  1. The stealth test is circular. The attack trains against the same deviation measures that the paper uses to detect it. The two defenses also use these measures. The paper does not test against a separate detector.
  2. The theory is weak. Lemma 1 is a simple bound, and it holds only for max–min pooling. Corollary 1 is a standard Lipschitz step.
  3. The baselines do not fit the task. HyperNear and TH-Attack are built for node classification. The paper applies them out of their setting.
Questions for authors
  1. Your attack trains against the membership and hyperedge deviation measures. The stealth test and the two defenses also use these measures. How well does the attack hide from a separate detector that you do not train against?
  2. Lemma 1 assumes max–min pooling. Does the attenuation result still hold for mean, attention, or deep-set aggregators? How sensitive are the results to the aggregator?
  3. You motivate the work with fraud. Does the attack still work on larger or fraud-domain hypergraphs, beyond the citation and authorship datasets?
8 comments
AuthorsOpenReview ↗

Overall response to Reviewer PFiN

Thank you for your careful and constructive review. We sincerely appreciate your recognition of the novelty of our attack target, the analysis-driven design of DPHIA, and its effectiveness in settings where existing baselines struggle. Your feedback has helped us further strengthen and clarify our work.

We respond to your concerns and questions as follows.

[Preview]

  • Weakness #1 and Question #1: Stealthiness against separate detectors and defenses.
  • Weakness #2 and Question #2: Theoretical generality and sensitivity to aggregation.
  • Weakness #3: Applicability of baselines to hyperedge prediction.
  • Question #3: Evaluation on larger and fraud-domain hypergraphs.
  • Conclusion.
AuthorsOpenReview ↗

Response to Weakness #1 and Question #1

W1/Q1. Stealthiness against separate detectors and defenses.

A1. Thank you for this valuable comment. We would like to first clarify that DPHIA does not directly optimize the MD or HD measures used for stealthiness evaluation. As discussed in Section 4.2, DPHIA employs learned discriminators to mitigate membership- and hyperedge-level distributional deviations, while the autoencoder used to compute HD is not involved in attack training. Thus, the evaluation measures are not directly used as training objectives.

We fully agree that evaluating stealthiness under independent detectors would strengthen our experiments. Following your suggestion, we have added experiments to assess (a) detectability under detectors not used during attack training. Furthermore, since the original MD/HD-filtering defenses also rely on the same deviation measures, we further investigate (b) attack effectiveness under independent defenses based on alternative mechanisms.

(a) Separate detectors. We evaluate membership-level (Hyper-SAGNN [1]) and hyperedge-level (HAD [2], HashNWalk [3]) detectors, and additionally include feature-level detectors (Isolation Forest [4], LOF [5]). We report AP and Recall@kk, where kk equals the number of injected units at each level.

As shown in Table A1(a), while HyperNear and TH-Attack are highly detectable by certain independent detectors, DPHIA consistently maintains low detection AP (≤26.38%) and Recall@kk (≤37.50%) across all five detectors on both datasets, demonstrating more consistent stealthiness across diverse detection mechanisms.

Table A1(a). Detection performance of separate detectors (AP | Recall@kk, %).

DatasetLevelDetectorHyperNearTH-AttackDPHIA
CiteseerFeatureIsolation Forest100.00 | 100.0097.99 | 94.4426.38 | 37.50
FeatureLOF3.22 | 0.0067.20 | 70.838.79 | 11.11
MembershipHyper-SAGNN2.38 | 0.0037.99 | 46.232.68 | 0.00
HyperedgeHAD20.05 | 24.0010.29 | 5.5619.26 | 23.61
HyperedgeHashNWalk4.45 | 0.009.32 | 1.396.82 | 0.00
CoraFeatureIsolation Forest100.00 | 100.00100.00 | 100.005.50 | 0.00
FeatureLOF2.98 | 0.0047.84 | 60.5618.73 | 25.35
MembershipHyper-SAGNN1.54 | 0.0010.88 | 11.383.05 | 2.82
HyperedgeHAD10.46 | 5.7110.71 | 14.087.08 | 1.41
HyperedgeHashNWalk7.23 | 0.006.79 | 0.005.86 | 1.41

(b) Separate defenses. We further evaluate two recent defenses: (1) HGNN-Shield [6], a purification defense for hypergraphs, and (2) Elasso-HyGEN, which adapts the robust structure refinement of ElassoGNN [7] to HyGEN following MeLA [8].

As shown in Table A1(b), DPHIA remains the strongest attack under both hypergraph defenses, achieving the lowest AUC and AP on both datasets. In particular, under HGNN-Shield, DPHIA achieves AUCs of 64.42% and 53.76% on Citeseer and Cora, compared with 82.91% and 75.53% for HyperNear, further demonstrating DPHIA's effectiveness against independent defenses not used during attack training.

Table A1(b). Attack performance against HyGEN under separate defenses (AUC | AP, %).

DatasetDefenseTH-AttackHyperNearDPHIA
CiteseerNone84.24 | 84.2667.39 | 68.4564.07 | 67.45
Elasso-HyGEN80.44 | 79.9574.73 | 76.1066.62 | 69.95
HGNN-Shield81.62 | 80.3582.91 | 81.1364.42 | 68.30
CoraNone73.98 | 70.1562.82 | 64.9452.11 | 47.92
Elasso-HyGEN75.65 | 73.0566.48 | 64.4552.94 | 56.71
HGNN-Shield76.62 | 73.8475.53 | 71.5653.76 | 52.18

We have incorporated these additional experiments into Section 5.4 (Tables 3–4) of the revised manuscript, further supporting DPHIA's stealthiness against separate detectors and attack effectiveness under independent defenses, beyond the original MD/HD-based evaluation.

AuthorsOpenReview ↗

Response to Weakness #2 and Question #2

W2/Q2. Theoretical generality and sensitivity to aggregation.

A2. Thank you for this valuable comment. We appreciate the reviewer's concern and would like to first clarify our choice of max–min pooling for theoretical analysis. Max–min pooling is widely adopted in representative hyperedge prediction models (NHP, AHP, and HyGEN), and NHP has demonstrated its ability to capture the maximum discrepancy among member embeddings, making it well-suited for modeling higher-order relationships.

Under this aggregation, Lemma 1 provides an intuitive explanation of group-level perturbation attenuation: even substantial node-level perturbations may have limited effects on candidate representations when nodes undergo similar changes. This highlights a key challenge in attacking hyperedge prediction models, motivating DPHIA's adaptive ranking-based optimization.

We fully agree that extending the theoretical analysis to different aggregation strategies would strengthen the generality of our findings. Following your suggestion, we first generalize Lemma 1 to different aggregators, including max–min pooling, mean pooling, Deep Sets, and attention. Then, we extend Corollary 1 to characterize aggregation-dependent ranking vulnerability. Finally, we empirically evaluate DPHIA's sensitivity to different aggregation strategies.

(a) Generalized Lemma 1. For a candidate cc, let qc=agg({zv}v∈c)q_c=\mathrm{agg}(\lbrace z_v\rbrace_{v\in c}) and q^c=agg({zv+δv}v∈c)\hat q_c=\mathrm{agg}(\lbrace z_v+\delta_v\rbrace_{v\in c}). We decompose each node perturbation as δv=ac+rv\delta_v=a_c+r_v, where ac=1∣c∣∑v∈cδva_c=\frac{1}{|c|}\sum_{v\in c}\delta_v is the shared shift and rvr_v is the member-specific residual.

Proposition 1 (Aggregation sensitivity). Under the corresponding Lipschitz conditions, the aggregated representation shift satisfies ∥q^c−qc∥∞≤Bc\lVert\hat q_c-q_c\rVert_\infty\le B_c, with aggregation-specific bounds:

AggregatorBound BcB_c
Max–min2max⁡v∈c∥rv∥∞2\max_{v\in c}\lVert r_v\rVert_\infty (Lemma 1)
Mean∥ac∥∞\lVert a_c\rVert_\infty (exact)
Deep SetsLρLϕ∣c∣∑v∈c∥δv∥∞\frac{L_\rho L_\phi}{\lvert c\rvert}\sum_{v\in c}\lVert\delta_v\rVert_\infty
Attention∥W∥∞(1+2LsRc)max⁡v∈c∥δv∥∞\lVert W\rVert_\infty(1+2L_sR_c)\max_{v\in c}\lVert\delta_v\rVert_\infty

Here, LϕL_\phi and LρL_\rho are the Lipschitz constants of the Deep Sets mappings with normalized mean pooling; LsL_s bounds attention-score sensitivity, WW is the value projection, and Rc=max⁡v∈c∥zv∥∞R_c=\max_{v\in c}\lVert z_v\rVert_\infty.

(b) Aggregation-dependent ranking reversal. Based on Proposition 1, we extend Corollary 1 as follows.

Corollary 2 (Aggregation-dependent ranking reversal). For an LaL_a-Lipschitz prediction head and a correctly ranked positive–negative pair (c+,c−)(c^+,c^-) with initial probability gap Δ>0\Delta>0, reversing their ranking requires Δ≤La(Bc++Bc−),\Delta\le L_a(B_{c^+}+B_{c^-}), where BcB_c is the aggregation-specific bound in Proposition 1. For max–min pooling, Bc=2εcB_c=2\varepsilon_c, which recovers Eq. (3).

These results extend our analysis in Section 3.1 by establishing aggregation-dependent bounds on candidate representation and prediction-score changes. Specifically, max–min eliminates shared shifts, mean preserves average perturbations, Deep Sets aggregates transformed member representations, and attention exhibits weight-dependent sensitivity. Together, these bounds characterize how group-level aggregation can limit the influence of node-level perturbations on candidate rankings, further motivating DPHIA's adaptive ranking-based objective, which prioritizes positive–negative pairs with smaller probability gaps.

(c) Sensitivity to aggregation strategies. We further evaluate DPHIA on Citeseer by fixing the HyGEN victim backbone and DPHIA's max–min surrogate, while varying the victim's candidate aggregator among max–min, mean, Deep Sets, and attention.

As shown in Table A2, DPHIA consistently reduces AUC and AP under all four aggregators, with AUC drops ranging from 6.59 to 19.24 percentage points, although the benign performance varies substantially across aggregators (AUC: 66.36–83.31%). These results demonstrate that the effectiveness of DPHIA does not depend on a particular aggregation strategy or on matching the aggregator of the victim.

Table A2. Sensitivity to victim candidate aggregators (AUC | AP, %).

AggregatorBenignDPHIAAUC / AP drop
Max–min83.31 | 83.3264.07 | 67.4519.24 | 15.87
Mean75.26 | 76.1957.26 | 52.9018.00 | 23.29
Deep Sets70.64 | 72.6160.44 | 58.5510.20 | 14.06
Attention66.36 | 62.0459.77 | 57.316.59 | 4.73

We have added a brief discussion of the theoretical generalizations to Section 3.1, the formal statements and proofs to Appendix A.3-A.5, and the aggregator sensitivity experiments to Appendix G.4 (Table 11) of the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.