acceptodds
Under review as a conference paper at ICLR 2027

Membership Inference Attacks in Simplicial Complexes Neural Networks

Abstract

Simplicial Complex Neural Networks (SCNs) learn from relational data with group-level structure, including triangles, cliques, loops, and higher-order interactions. Boundary operators and Hodge Laplacians support this representation learning and encode membership-relevant cues when constructed from training simplices. These construction-dependent cues create a distinctive privacy risk alongside model memorization. We present the first security study of membership inference attacks on SCNs. Our exposure-aware analysis connects higher-order utility to membership signals carried by operator construction and model responses. We introduce a Structure-Aware Black-Box attack (SA-BB) that compares a queried simplex's prediction probability with those of neighbouring candidates in a public topology. A membership classifier trained on shadow models uses these local response patterns to infer target membership through a prediction API. To mitigate this risk, we propose Dynamic Topological Pruning and Injection (DTPI), which perturbs selected high-impact topological relations during training while preserving useful higher-order structure. Our evaluation examines membership leakage and the privacy–utility trade-off across datasets, backbones, and defenses.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
membership inference, black-box attack, privacy-preserving machine learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TcuyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies membership inference attacks (MIAs) against Simplicial Complex Neural Networks (SCNs), focusing on privacy leakage arising from higher-order topological structures. The authors propose a structure-aware black-box attack (SA-BB) that exploits local prediction patterns to infer training membership. They further propose Dynamic Topological Pruning and Injection (DTPI), a defense that modifies topological relations during training to mitigate privacy leakage. This paper evaluates the effectiveness of SA-BB and the privacy–utility trade-offs achieved by experiments across different datasets, model architectures, and attack and defense baselines.

Critical strengths
  1. This paper proposes the SA-BB attack, which exploits local prediction patterns to infer training membership without requiring access to the target model's internal parameters. Experimental results demonstrate its effectiveness compared with some existing membership inference baselines.

  2. This paper proposes DTPI defense, which dynamically modifies topological structures to mitigate membership leakage while preserving model utility. Experiments demonstrate favorable privacy–utility trade-offs across different attacks, and ablation studies validate the contributions of its key components.

  3. This paper is well-organized and easy to follow.

Critical weaknesses
  1. This paper evaluates several membership inference attacks and defenses; however, the comparisons mainly rely on conventional baselines, leaving the advantages of the proposed methods insufficiently validated. For attack evaluation, stronger approaches such as LiRA [1], RMIA [2], and recent topology-aware attacks [3] are not considered. Similarly, the defense evaluation lacks comparisons with established methods such as SELENA [4] and more directly related topology-based defenses [5]. Including them would provide a more comprehensive and convincing evaluation of SA-BB and DTPI.

  2. The proposed SA-BB attack shows limited advantages over existing methods at low FPR. On SEMANTIC, SA-BB achieves a TPR of only 2.14% at 1% FPR, showing almost no improvement over the conventional NN attack (2.01%). Similarly, the proposed DTPI defense still suffers from substantial membership leakage. On CITESEER, SIA achieves a TPR of 41.54% at 1% FPR even after applying DTPI, indicating that considerable membership information remains exposed. These results raise concerns about the effectiveness of both proposed methods.

  3. The core design of SA-BB relies on relatively straightforward techniques, including prediction confidence, neighborhood-based statistics, and a logistic regression classifier trained on shadow models. Previous studies, such as SIA and SMIA, have already explored exploiting structural information for membership inference in graph neural networks. While SA-BB extends this direction to simplicial complexes under a different access setting, its methodological novelty beyond this adaptation remains unclear. Moreover, Table 10 shows that degree-matched neighbors outperform topology-based neighbors on CITESEER, raising further questions about whether the improvement primarily comes from simplex-specific topological information or more general structural statistics.

  4. The core design of SA-BB relies on relatively conventional techniques, including prediction confidence, neighborhood statistics, and a logistic regression classifier trained on shadow models. Previous studies have already explored structural inference attacks on GNNs [6], while SMIA [7] specifically exploits model predictions and shadow training to infer higher-order subgraph structures. Although SA-BB extends this research direction to simplicial complexes, its technical novelty beyond this adaptation remains unclear. Furthermore, Table 10 shows that degree-matched neighbors without shared boundary faces outperform topology-based neighbors on CITESEER, raising questions about the unique contribution of the proposed face-sharing neighborhood design.

[1] "Membership Inference Attacks From First Principles", Carlini et al., IEEE S&P 2022.

[2] "Low-Cost High-Power Membership Inference Attacks", Zarifzadeh et al., ICML 2024.

[3] "Attention-Based Membership Inference Attacks on Graph Neural Network Through Topological Features", Guan et al., IEEE TDSC 2025.

[4] "Mitigating Membership Inference Attacks by Self-Distillation Through a Novel Ensemble Architecture", Tang et al., USENIX Security 2022.

[5] "Topology Modification against Membership Inference Attack in Graph Neural Networks", Guan et al., Knowledge-Based Systems, 2024.

[6] "Inference Attacks Against Graph Neural Networks", Zhang et al., USENIX Security 2022.

[7] "Subgraph Structure Membership Inference Attacks against Graph Neural Networks", Wang et al., PoPETs 2024.

Questions for authors
  1. How do the proposed SA-BB attack and DTPI defense compare with more recent and stronger membership inference attack and defense methods? Please refer to Weakness 1 for details.

  2. What are the key technical differences between SA-BB and existing structure-aware membership inference attacks? Please further clarify the methodological novelty of SA-BB and provide additional evidence supporting the benefits of its simplex-specific design.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Tcuy (Part 1)

We thank the reviewer for the positive comments on SA-BB, DTPI, and the paper's presentation, as well as for suggesting recent attack and defense baselines [1–5] and a closer comparison with [6, 7].

W1 & Q1. Comparison with recent attacks [1–3] and defenses [4–5]

We extend both benchmarks on SCCNN under the 3-target-seed ×\times 3-shadow protocol (balanced 1:1 membership split, mean ±\pm std):

1.Attacks (Table 1). Alongside NN and Loss from submitted Table 9, we evaluate offline/online LiRA [1], RMIA [2], and Att-Topo (adapting [3] from multi-class node posteriors to a scalar simplex API via a shadow attention aggregator over face-sharing neighbor predictions).

2.Defenses (Table 2). Selecting configs at validation Model AUC and reporting held-out Test Model AUC, we compare DTPI with No Defense, SHAP, GEP, SELENA [4] (Split-AI sub-models with strictly isolated simplicial operators plus self-distillation), and Topo-Mod [5] (degree-constrained edge edits rebuilt into a valid complex), retraining all black-box attacks on three identically defended shadows.

Table 1. Undefended black-box attack comparison on SCCNN (NN and Loss from submitted Table 9; remaining rows over 3 target seeds ×\times 3 shadows).

AttackShadowsCiteSeer AUCCiteSeer TPR@5%Semantic AUCSemantic TPR@5%
NN30.8320.1440.7090.093
Loss30.8320.1630.6840.091
LiRA (offline) [1]30.744 ± 0.0830.2570.661 ± 0.0300.133
LiRA (online) [1]30.842 ± 0.0550.3320.694 ± 0.0170.151
RMIA [2]30.833 ± 0.0210.3260.702 ± 0.0260.137
Att-Topo (adapted [3])30.869 ± 0.0070.3750.718 ± 0.0150.123
SA-BB (B=20B=20)30.885 ± 0.0100.3770.729 ± 0.0290.133
RMIA + SA-BB30.891 ± 0.0040.4150.736 ± 0.0250.130

SA-BB achieves the strongest standalone black-box performance on both datasets, leading online LiRA, RMIA, and Att-Topo in AUC on CiteSeer (0.885) and Semantic (0.729) as well as CiteSeer TPR@5% (0.377). Explicit rank/gap calibration outperforms Att-Topo's scalar attention under a single-probability API, and combining SA-BB with RMIA further boosts CiteSeer to 0.891 AUC and 0.415 TPR@5% (0.736 AUC on Semantic).

Table 2. Matched-utility defense comparison on CiteSeer (SCCNN; 3 target seeds ×\times 3 defended shadows; configs selected at validation Model AUC 0.8340±0.010.8340 \pm 0.01).

DefenseConfigTest Model AUCSA-BB AUC / TPR@5%LiRA-on AUC / TPR@5%RMIA AUC / TPR@5%SIA-WB AUC / TPR@5%
No Defense–0.857 ± 0.0200.885 ± 0.010 / 0.3770.842 ± 0.055 / 0.3320.833 ± 0.021 / 0.3260.927 ± 0.013 / 0.621
SHAPb=0.1b=0.10.856 ± 0.0220.864 ± 0.019 / 0.4610.792 ± 0.075 / 0.2970.813 ± 0.042 / 0.3130.925 ± 0.007 / 0.618
GEPb=0.1b=0.10.858 ± 0.0200.862 ± 0.020 / 0.4430.794 ± 0.022 / 0.2510.803 ± 0.038 / 0.2770.922 ± 0.012 / 0.616
SELENA [4]K=25,L=10K=25, L=100.829 ± 0.0530.856 ± 0.055 / 0.5110.810 ± 0.011 / 0.3010.813 ± 0.034 / 0.2980.924 ± 0.010 / 0.610
Topo-Mod [5]ρ=0.3\rho=0.30.842 ± 0.0090.850 ± 0.020 / 0.2480.825 ± 0.044 / 0.2460.768 ± 0.010 / 0.2500.882 ± 0.004 / 0.577
DTPI20% active pool0.855 ± 0.0090.844 ± 0.017 / 0.2060.770 ± 0.011 / 0.2370.794 ± 0.024 / 0.2680.864 ± 0.011 / 0.463

DTPI delivers the strongest overall defense at matched utility, retaining 0.855 Test Model AUC while achieving the lowest AUC and TPR@5% against SA-BB, online LiRA, and white-box SIA. Topo-Mod lowers RMIA further (0.768 / 0.250 vs. 0.794 / 0.268) at the cost of lower utility (0.842) and weaker protection on the remaining three attacks, whereas SHAP, GEP, and SELENA (24$\times$ slower) fail to curb SIA (>0.92 AUC) or SA-BB TPR@5%.

[1] "Membership Inference Attacks From First Principles", Carlini et al., IEEE S&P 2022.

[2] "Low-Cost High-Power Membership Inference Attacks", Zarifzadeh et al., ICML 2024.

[3] "Attention-Based Membership Inference Attacks on Graph Neural Network Through Topological Features", Guan et al., IEEE TDSC 2025.

[4] "Mitigating Membership Inference Attacks by Self-Distillation Through a Novel Ensemble Architecture", Tang et al., USENIX Security 2022.

[5] "Topology Modification against Membership Inference Attack in Graph Neural Networks", Guan et al., Knowledge-Based Systems, 2024.

[6] "Inference Attacks Against Graph Neural Networks", Zhang et al., USENIX Security 2022.

[7] "Subgraph Structure Membership Inference Attacks against Graph Neural Networks", Wang et al., PoPETs 2024.

AuthorsOpenReview ↗

Response to Reviewer Tcuy (Part 3)

W3, W4 & Q2. Differences from SIA [6] and SMIA [7], classifier choice, and Table 10

(Since W3 and W4 raise the same points on [6, 7] and Table 10, we address them together.)

Technical differences from SIA [6] and SMIA [7]. Table 3 summarizes the distinction. SIA [6] and SMIA [7] target node-classification GNNs whose API returns multi-class posterior vectors per node; in particular, SMIA infers subgraph structure from pairwise cosine-similarity rankings across multi-dimensional node posteriors (or embeddings). In kk-simplex prediction, the black-box API returns only one scalar probability pν∈[0,1]p_\nu \in [0,1] per candidate simplex, so pairwise node-posterior similarity matrices cannot be constructed from black-box outputs. Evaluating SIA and SMIA on simplex prediction in submitted Table 7 therefore required granting them white-box access to internal edge or node embeddings.

Table 3. Comparison with structure-aware GNN attacks.

MethodTarget Model & TaskModel Access & Output GranularityStructural Signal
SIA [6]GNN, node classificationNode multi-class posteriors (white-box edge embeddings in Table 7)Pairwise node-output / embedding similarity for 1-skeleton links
SMIA [7]GNN, node classificationNode multi-class posteriors on subgraph (white-box node embeddings in Table 7)Pairwise cosine-similarity rankings across node posteriors
Guan et al. [3]GNN, node classificationNode multi-class posteriors of node + 1/2-hop neighborsAttention aggregation over node-posterior neighborhoods
SA-BB (ours)SCN, kk-simplex predictionBlack-box simplex API: one scalar pν∈[0,1]p_\nu \in [0,1] per candidate simplexScalar calibration against public (k−1)(k-1)-face / coface candidates

Classifier choice. Holding standardized logistic regression and balanced shadow splits fixed across NN, Loss, and SA-BB (with a fixed B=20B=20 neighbor budget cap for SA-BB; Appendix D.2 and E.5) is an intentional control so that gains can be attributed strictly to the topological calibration features rather than classifier capacity. Within the overall paper, SA-BB serves as the black-box counterpart to our Hodge-Laplacian exposure analysis (Proposition 4.2) and cascade-guided defense (DTPI).

Neighbor selection in Table 10. In Appendix E.5, Matched aligns the mean and standard deviation of candidate-triangle coface counts across σ\sigma's three boundary edges (not 1-skeleton vertex degree). Also, fallback in Section 4.3 triggers only when ∣N(σ)∣=0|\mathcal{N}(\sigma)| = 0 ("B+1B+1 prediction calls" in Appendix E.5 is the budget cap), so sparse CiteSeer queries with 1≤∣N(σ)∣<201 \le |\mathcal{N}(\sigma)| < 20 use fewer than 20 neighbors under unpadded Topology. Table 4 adds a 1-skeleton Node-degree matched control and Topology + padding (full B=1,20B=1, 20 table in Reviewer 7PWn Table 1).

Table 4. SA-BB neighbor-selection breakdown (B=20B=20, SCCNN; AUC / TPR@5% on Semantic and CiteSeer, AUC on Cora).

Selection RuleSemantic (B=20B=20)CiteSeer (B=20B=20)Cora (B=20B=20)
Random0.700 ± 0.028 / 0.1040.849 ± 0.005 / 0.2060.930 ± 0.012
Node-degree matched (1-skeleton vertex degree)0.706 ± 0.026 / 0.0810.875 ± 0.005 / 0.3410.929 ± 0.013
Matched (2-simplex coface count)0.702 ± 0.022 / 0.1010.897 ± 0.005 / 0.4630.937 ± 0.011
Topology (face-sharing)0.729 ± 0.029 / 0.1330.885 ± 0.010 / 0.3770.950 ± 0.012
Topology + padding0.728 ± 0.026 / 0.1310.899 ± 0.005 / 0.4860.946 ± 0.011

2-simplex calibration yields SA-BB's strongest results on every dataset: Topology leads on Semantic (0.729 / 0.133) and Cora (0.950), and Topology + padding leads on sparse CiteSeer (0.899 / 0.486, comparable to Matched at 0.897 / 0.463 once partial neighborhoods are padded to B=20B=20). In contrast, 1-skeleton Node-degree matched trails 2-simplex rules on CiteSeer and Cora (matching Random on Cora) and stays near Random alongside Matched on dense Semantic.

[6] "Inference Attacks Against Graph Neural Networks", Zhang et al., USENIX Security 2022.

[7] "Subgraph Structure Membership Inference Attacks against Graph Neural Networks", Wang et al., PoPETs 2024.

AuthorsOpenReview ↗

Response to Reviewer Tcuy (Part 2)

W2. Low-FPR behavior on Semantic (2.14%) and SIA on CiteSeer (41.54%)

Low-FPR behavior of SA-BB. Compared with CiteSeer (average degree 2.736, Appendix D.1), Semantic is a dense co-authorship complex (average degree 8.375). Heavy message passing across shared edges smooths local representations, causing high-confidence members and well-connected non-members to overlap in the extreme 1% FPR tail (2.14% TPR for SA-BB vs. 2.01% for NN and 1.43% for Loss in submitted Table 9). At 5% FPR on Semantic (submitted Table 9 and Table 1), local calibration separates these regimes clearly: SA-BB reaches 13.3% TPR, a 43% relative improvement over NN (9.3%) and 47% over Loss (9.1%), outperforming Att-Topo (12.3%) and matching RMIA (13.7%) while achieving the highest overall AUC (0.729). On sparser CiteSeer, where face-sharing propagation creates sharper local contrast around training triangles, SA-BB improves substantially at both 1% FPR (18.4% vs. 6.4% for NN and 5.0% for Loss in submitted Table 9) and 5% FPR (37.7% vs. 14.4% for NN, 33.2% for online LiRA, 32.6% for RMIA, and 41.5% for RMIA + SA-BB in Table 1).

Residual white-box SIA (41.54%) on CiteSeer. Two factors account for this figure. First, as noted in Appendix D.2, SIA is a white-box audit with access to internal edge embeddings and ground-truth task labels, which are unavailable through a prediction API (where DTPI reduces black-box NN TPR@1% to 5.95% in submitted Table 8, and reduces SA-BB and online LiRA TPR@5% to 20.6% and 23.7% in Table 2). Consistent with the utility–exposure coupling in Proposition 4.2, preserving predictive utility on sparse CiteSeer relies on message passing over core training triangles, and empirically internal edge embeddings retain structural correlations unless utility collapses; in submitted Table 8, even DP-SGD at collapsed utility (0.6593 Model AUC) still leaves SIA TPR@1% at 48.98% (vs. 50.67% undefended in Table 7). Second, relative to all evaluated defenses, DTPI provides the strongest mitigation against SIA: in submitted Table 8, DTPI achieves the lowest SIA TPR@1% (41.54%, while all six baselines remain between 46.89% and 51.35%) at the highest Model AUC (0.7807), and in Table 2 at matched utility (0.855 Test Model AUC) it achieves the largest reduction in both SIA AUC (0.927 to 0.864) and TPR@5% (0.621 to 0.463).

Thank you for the detailed response and additional experiments. I appreciate the new comparisons with recent attack and defense methods, as well as the clarifications regarding the differences from existing approaches. These additions address some of my concerns.

However, I still have reservations about the technical novelty of SA-BB, as its core methodology remains largely based on conventional membership inference techniques. In addition, the attack performance at low FPR on SEMANTIC remains less convincing. Specifically, at 1% FPR, SA-BB achieves a TPR of only 2.14%, showing almost no improvement over the conventional NN attack (2.01%). While the results at 5% FPR are better, I consider 5% FPR to be well beyond the low-FPR regime, which, in my view, should focus on FPRs no greater than 1%, or even as low as 0.1%.

Therefore, I will maintain my original rating.

Reviewer_7PWnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates membership inference in simplicial complex neural networks, distinguishing direct membership exposure through released topological operators from leakage through prediction outputs. Its contributions include an information decomposition, a structure-aware black-box attack using neighboring candidates’ prediction statistics, and a defense based on dynamic topological pruning and injection. Validation includes simplex-prediction experiments on Semantic, Cora, and CiteSeer.

Critical strengths

1.The distinction between direct operator exposure and prediction-based leakage provides a useful framework for examining privacy risks in simplicial networks.

2.The empirical evaluation includes independent shadow-model transfer, low-FPR results, and attack and defense ablations, providing evidence beyond aggregate attack accuracy.

Critical weaknesses
  1. The paper identifies membership cues in simplicial operators, but does not clearly establish whether higher-order topology introduces a distinct privacy mechanism beyond structural leakage already studied in GNNs. PSA detects membership directly encoded in released operators, while SA-BB’s gains may partly reflect generic confidence and degree-related signals. Importantly, the degree-matched, non-face-sharing control outperforms face-sharing neighbors on CiteSeer (Table 10). Consequently, the contribution appears better supported as an extension and empirical characterization of structural membership inference than as evidence of a fundamentally new vulnerability.

  2. DTPI includes nontrivial edge-selection mechanisms and supporting ablations, but its privacy rationale remains largely heuristic. Theorem 5.1 bounds prediction-loss changes rather than membership distinguishability, so it does not establish why the selected edits should specifically reduce privacy leakage.

Questions for authors

Relative to the cited work on simplicial-network privacy and structural membership attacks, what precise technical contribution supports the novelty claim beyond changing the prediction unit to a simplex? Beyond the existing screening ablations, can you connect the cascade score to reduced membership distinguishability and demonstrate a consistent DTPI advantage at matched utility against attackers retrained on defended shadow models?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 7PWn (Part 1)

We thank the reviewer for the thoughtful feedback and for noting the value of separating direct operator exposure from output-based leakage. Below we clarify our simplicial-specific technical contributions, the definition of Matched in Table 10, the connection between Theorem 5.1 and membership advantage in Appendix B.3, and the defended-shadow evaluation at matched utility.

W1 & Q (Part 1). Simplicial-specific contributions and Table 10 breakdown

Prior work on simplicial-network privacy studies 1-skeleton graph reconstruction from learned embeddings (Zhan et al., 2023) or application-driven representation protection on text-attributed networks (Zhan & Sheng, 2023), rather than simplex membership inference, operator-versus-output leakage decomposition, or cascade-aware topological defense. Relative to both those studies and graph-level structural MIAs (SIA, SMIA), three mechanisms are specific to SCNs beyond changing the prediction unit to kk-simplices:

1.Hodge-Laplacian support equivalence. Even when BkB_k is withheld and only the (k−1)(k-1)-Hodge Laplacian is released, any two distinct (k−1)(k-1)-faces share at most one kk-coface. The off-diagonal support of Lk−1,u=BkBk⊤L_{k-1,u} = B_k B_k^\top therefore identifies every materialized kk-simplex exactly (matching direct BkB_k inspection in Appendix E.1, Figure 11), whereas Lk−1,ℓ=Bk−1⊤Bk−1L_{k-1,\ell} = B_{k-1}^\top B_{k-1} provides only weak indirect cues (Proposition 4.2).

2.Calibration under a scalar simplex API. Structural GNN attacks (SIA, SMIA) target node classification and rely on multi-class node posterior vectors or internal embeddings. A simplex-prediction API returns only one scalar probability pν∈[0,1]p_\nu \in [0,1] per candidate simplex, so SA-BB calibrates scalar scores across public (k−1)(k-1)-face and coface neighborhoods.

3.Closure-coupled defense (DTPI). Because a simplicial complex is closed under taking faces, editing an edge mask m^e\hat{m}e propagates to all incident higher-order simplices via m^σ=∏α∈∂σm^α\hat{m}\sigma = \prod_{\alpha \in \partial\sigma} \hat{m}\alpha. Toggling one edge simultaneously perturbs T(e)T(e) cofaces and both Lk−1,ℓL{k-1,\ell} and Lk−1,uL_{k-1,u}, requiring joint screening of higher-order cascade size and lower-order degree sensitivity.

What Matched measures in Table 10. In Appendix E.5, "matches local degree statistics" means matching the mean and standard deviation, across σ\sigma's three boundary edges, of incident candidate-triangle counts (d△(e)d_\triangle(e), a 2-simplex coface count rather than 1-skeleton vertex degree). Also, fallback sampling in Section 4.3 triggers only when ∣N(σ)∣=0|\mathcal{N}(\sigma)| = 0 (while "B+1B+1 prediction calls" in Appendix E.5 is the budget cap), so many sparse CiteSeer queries average over fewer than 20 neighbors under unpadded Topology. Under the Table 9/10 protocol (3 target seeds ×\times 3 shadows), Table 1 adds a 1-skeleton Node-degree matched control (zero triangle information) and Topology + padding (filling partial face-sharing neighborhoods up to BB with coface-matched candidates).

Table 1. SA-BB attack AUC across neighbor-selection rules on SCCNN (3 target seeds ×\times 3 shadows; mean ±\pm std).

Selection RuleUses 2-Simplex Info?Semantic (B=1B=1)Semantic (B=20B=20)CiteSeer (B=1B=1)CiteSeer (B=20B=20)Cora (B=20B=20)
Random (no shared face)No0.6885 ± 0.0160.6996 ± 0.0280.8366 ± 0.0200.8490 ± 0.0050.9303 ± 0.012
Node-degree matched (1-skeleton vertex degree only)No0.6949 ± 0.0280.7064 ± 0.0260.8480 ± 0.0150.8745 ± 0.0050.9294 ± 0.013
Matched (candidate-triangle coface count)Yes0.6876 ± 0.0230.7023 ± 0.0220.8590 ± 0.0120.8972 ± 0.0050.9372 ± 0.011
Topology (shares a (k−1)(k-1)-boundary face)Yes0.6954 ± 0.0210.7286 ± 0.0290.8537 ± 0.0220.8853 ± 0.0100.9503 ± 0.012
Topology + padding (face-sharing first, then matched)Yes0.6954 ± 0.0210.7275 ± 0.0260.8588 ± 0.0180.8989 ± 0.0050.9456 ± 0.011

Across all three datasets, SA-BB is strongest when calibrated with 2-simplex structure: face-sharing Topology leads on Semantic (0.7286) and Cora (0.9503), and Topology + padding leads on sparse CiteSeer (0.8989, comparable to Matched at 0.8972 after removing unpadded Topology's reference-size gap), with 2-simplex rules also gaining the most from B=1B=1 to 2020. By contrast, 1-skeleton Node-degree matched trails both 2-simplex rules on CiteSeer and Cora (matching Random on Cora), and on dense Semantic both non-adjacent controls stay near Random.

AuthorsOpenReview ↗

Response to Reviewer 7PWn (Part 2)

W2 & Q (Part 2). Link from Theorem 5.1 to membership advantage, and defended-shadow evaluation

Connection to threshold membership distinguishability. Theorem 5.1 bounds the high-probability magnitude of query-level loss variation under toggling an edge ee given (G,p)(G, p) in terms of T(e)T(e). Because standard MIAs threshold a scalar attack score S(σ;S~)S(\sigma; \widetilde{S}) rather than observing loss changes alone, submitted Appendix B.3 (Corollary B.10 and Section B.3.1) translates this bound into the local sensitivity of balanced-prior threshold membership advantage Advt(S~):=12∣Pr⁡(At=1∣M=1,G,p,Y)−Pr⁡(At=1∣M=0,G,p,Y)∣\mathrm{Adv}t(\widetilde{S}) := \frac{1}{2}|\Pr(A_t=1 \mid M=1,G,p,Y) - \Pr(A_t=1 \mid M=0,G,p,Y)|. Under score Lipschitzness (Assumption B.9) and linear margin mass qm,t(η)≤κtηq{m,t}(\eta) \le \kappa_t \eta, the per-edge advantage sensitivity Δe(t):=∣Advt(S~(1,e))−Advt(S~(0,e))∣\Delta_e^{(t)} := |\mathrm{Adv}_t(\widetilde{S}^{(1,e)}) - \mathrm{Adv}_t(\widetilde{S}^{(0,e)})| satisfies

Δe(t)≤at,δ,T(e)+bt,δT(e)+ct,δ.\Delta_e^{(t)} \le a_{t,\delta},T(e) + b_{t,\delta}\sqrt{T(e)} + c_{t,\delta}.

Averaging over observed edges with retention probabilities pep_e and applying Jensen's inequality (Section B.3.1) bounds the pep_e-weighted mean local threshold-advantage variation by

at,δ,∑e∈EobspeT(e)Pobs+bt,δ∑e∈EobspeT(e)Pobs+ct,δ,a_{t,\delta},\frac{\sum_{e\in\mathcal{E}{\mathrm{obs}}} p_e T(e)}{P{\mathrm{obs}}} + b_{t,\delta}\sqrt{\frac{\sum_{e\in\mathcal{E}{\mathrm{obs}}} p_e T(e)}{P{\mathrm{obs}}}} + c_{t,\delta},

which increases monotonically with 1Pobs∑e∈EobspeT(e)\frac{1}{P_{\mathrm{obs}}}\sum_{e\in\mathcal{E}{\mathrm{obs}}} p_e T(e) (where Pobs=∑e∈EobspeP{\mathrm{obs}} = \sum_{e\in\mathcal{E}{\mathrm{obs}}} p_e). While Corollary B.10 bounds two-sided local sensitivity rather than providing a closed-form directional guarantee on end-to-end retrained attack AUC, it gives a principled sensitivity surrogate for edge selection and regularization: because Eq. (10) penalizes observed edges proportionally to Tˉ(e)\bar{T}(e), optimization suppresses mem_e more strongly on high-T(e)T(e) edges than on low-T(e)T(e) edges, shifting the normalized retention weights toward smaller-T(e)T(e) edges and lowering 1Pobs∑e∈EobspeT(e)\frac{1}{P{\mathrm{obs}}}\sum_{e\in\mathcal{E}_{\mathrm{obs}}} p_e T(e), while training under randomized masks (plus λ2\lambda_2 background decoys) weakens reliance on individual training simplices. We will add an explicit forward reference from Section 5.1 to Corollary B.10 and Section B.3.1 in the main text.

Matched-utility evaluation with defended shadow models. To test DTPI at matched utility against attackers adapted to the defense, we select each defense's hyperparameter on the validation split, train three independent shadow SCCNNs under that identical defense setting, fit SA-BB, online LiRA, and RMIA on those defended shadows, and report held-out test utility (Test Model AUC) and attack metrics across three target seeds alongside refitted white-box SIA in Table 2.

Table 2. Matched-utility defense comparison on CiteSeer (SCCNN; 3 target seeds ×\times 3 defended shadows; configs selected at validation Model AUC 0.8340±0.010.8340 \pm 0.01).

DefenseConfigTest Model AUCSA-BB AUC / TPR@5%LiRA-on AUC / TPR@5%RMIA AUC / TPR@5%SIA-WB AUC / TPR@5%
No Defense–0.857 ± 0.0200.885 ± 0.010 / 0.3770.842 ± 0.055 / 0.3320.833 ± 0.021 / 0.3260.927 ± 0.013 / 0.621
SHAPb=0.1b=0.10.856 ± 0.0220.864 ± 0.019 / 0.4610.792 ± 0.075 / 0.2970.813 ± 0.042 / 0.3130.925 ± 0.007 / 0.618
GEPb=0.1b=0.10.858 ± 0.0200.862 ± 0.020 / 0.4430.794 ± 0.022 / 0.2510.803 ± 0.038 / 0.2770.922 ± 0.012 / 0.616
SELENA [1]K=25,L=10K=25, L=100.829 ± 0.0530.856 ± 0.055 / 0.5110.810 ± 0.011 / 0.3010.813 ± 0.034 / 0.2980.924 ± 0.010 / 0.610
Topo-Mod [2]ρ=0.3\rho=0.30.842 ± 0.0090.850 ± 0.020 / 0.2480.825 ± 0.044 / 0.2460.768 ± 0.010 / 0.2500.882 ± 0.004 / 0.577
DTPI20% active pool0.855 ± 0.0090.844 ± 0.017 / 0.2060.770 ± 0.011 / 0.2370.794 ± 0.024 / 0.2680.864 ± 0.011 / 0.463

DTPI achieves the best overall privacy–utility trade-off: it preserves 0.855 Test Model AUC (matching undefended 0.857) while attaining the lowest AUC and TPR@5% on three of the four attacks (SA-BB, online LiRA, and white-box SIA). Topo-Mod yields lower RMIA (0.768 / 0.250 vs. 0.794 / 0.268), but drops Test Model AUC to 0.842 and is weaker on the other three attacks; SHAP, GEP, and SELENA (24$\times$ slower from 25 sub-models) leave Hodge-Laplacian propagation intact, keeping white-box SIA above 0.92 AUC.

[1] "Mitigating Membership Inference Attacks by Self-Distillation Through a Novel Ensemble Architecture", Tang et al., USENIX Security 2022.

[2] "Topology Modification against Membership Inference Attack in Graph Neural Networks", Guan et al., Knowledge-Based Systems, 2024.

Thank you for the detailed clarification and additional experiments. The simplicial-specific contributions are clearer. My remaining concern is that the sensitivity bound limits how much membership advantage can change, but does not establish why prioritizing high-cascade edges should reduce it.

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

This paper studies membership inference for candidate simplices processed by simplicial complex neural networks. It distinguishes direct exposure through released, query-aligned operators from leakage through model responses. SA-BB combines a query's prediction with local response statistics from public candidate simplices and trains a membership classifier on shadow models. The proposed DTPI defense prunes and injects topology using cascade size- and degree-sensitive screening.

Critical strengths
  • The motivation is clear, and the research perspective is novel. The paper examines how simplicial representations and training-only operator construction can expose membership information in SCNs.
  • The paper combines theoretical and empirical analysis of SCN vulnerabilities. The theorem bounds query-loss changes under edge toggling, motivating the cascade-based screening in DTPI.
  • Extensive experiments demonstrate the vulnerability of SCNs to membership inference attacks and the effectiveness of the proposed defense under the evaluated settings.
Critical weaknesses
  • The paper lacks a clear explanation of how face-sharing neighbors provide additional membership information beyond the target simplex’s own prediction in SA-BB.
  • The protocol for the defense results in Figure 8 is not sufficiently clear. Please clarify whether the attacks were trained against models using the same DTPI configuration as the target.
Questions for authors
  • Please see the Weaknesses.
  • How are predictions for held-out simplices computed when those simplices have no explicit column in the training-only operators? Please specify the SCN readout and confirm that it does not branch on membership.
2 comments
AuthorsOpenReview ↗

Response to Reviewer wYvG (Part 1)

We thank the reviewer for the positive assessment of our motivation, analysis, and empirical study, and address each question below.

W1. Why face-sharing neighbors add membership information beyond pσp_\sigma

In SCCNN, a candidate triangle σ=e1,e2,e3\sigma = {e_1, e_2, e_3} is scored as pσ=sigmoid(MLP([he1∣he2∣he3]))p_\sigma = \mathrm{sigmoid}(\mathrm{MLP}([h_{e_1} | h_{e_2} | h_{e_3}])). A single probability pσp_\sigma conflates training memorization with the baseline predictability of the local graph region (e.g., feature homophily or dense local connectivity). Querying face-sharing public candidates τ∈N(σ)\tau \in \mathcal{N}(\sigma) adds membership signal in two ways:

1.Local baseline calibration. Any candidate τ\tau sharing boundary edge e1e_1 reuses the learned edge embedding he1h_{e_1} and sits in the same local neighborhood, so pττ∈N(σ){p_\tau}{\tau \in \mathcal{N}(\sigma)} serves as a local reference for non-memorized confidence around σ\sigma. The relative rank rσr\sigma and probability-gap statistics (μσ,sσ,dσmax⁡)(\mu_\sigma, s_\sigma, d_\sigma^{\max}) (along with binary entropy H(pσ)H(p_\sigma) in Eq. (7)) measure how much σ\sigma stands out against this local baseline. Consistent with this view, pairing pσp_\sigma with rσr_\sigma in submitted Figure 7 gives the largest two-feature improvement (from 0.8413 to 0.8676 AUC, and 0.8853 with all six features).

2.Asymmetric coface coupling in L1,uL_{1,u}. When σ\sigma is a positive training member, it appears in B2,trainB_{2,\mathrm{train}}, so L1,u=B2,trainB2,train⊤L_{1,u} = B_{2,\mathrm{train}} B_{2,\mathrm{train}}^\top couples all three of its boundary edges during message passing while task supervision updates [he1∣he2∣he3][h_{e_1} | h_{e_2} | h_{e_3}] jointly. A held-out neighbor sharing only e1e_1 receives partial spillover through he1h_{e_1} without upper-Laplacian closure across its remaining two edges, widening the gap between pσp_\sigma and pττ∈N(σ){p_\tau}_{\tau \in \mathcal{N}(\sigma)} for members relative to non-members.

W2. Attack training protocol in Figure 8

Yes. At every hyperparameter point of DTPI and of every baseline defense in Figure 8, the evaluated attack is trained on features extracted from a model trained under that exact defense configuration.

Concretely, each subplot of Figure 8 fixes one attack method (NN, Loss, SA-BB, Gradient, SIA, or SMIA) and sweeps seven defenses (DTPI, Adv-Reg, GEP, SHAP, DP-SGD, KD, and Early-Stop) on the same SCCNN backbone on CiteSeer (Adam, learning rate 10−310^{-3}, hidden dimension 32, up to 1000 epochs with validation-AUC checkpoint selection). Following Appendix D.3, we train an independent defended SCCNN at each hyperparameter value along each defense's curve: DTPI sweeps λ1∈10−5,10−4,10−3,5⋅10−3,10−2,5⋅10−2,10−1,0.5,1.0\lambda_1 \in {10^{-5}, 10^{-4}, 10^{-3}, 5\cdot 10^{-3}, 10^{-2}, 5\cdot 10^{-2}, 10^{-1}, 0.5, 1.0} with λ2=0.1λ1\lambda_2 = 0.1\lambda_1 and fixed lexicographic screening; Adv-Reg sweeps λ∈[0,500]\lambda \in [0, 500]; GEP and SHAP sweep noise scale ∈[0,1]\in [0, 1]; DP-SGD sweeps σ∈[10−4,200]\sigma \in [10^{-4}, 200]; KD sweeps α∈[0.1,1.0]\alpha \in [0.1, 1.0]; and Early-Stop sweeps the epoch budget ∈[10,1000]\in [10, 1000].

For each trained (Defense, hyperparameter) model, attack features are extracted from that defended model, and the attack classifier is fitted from scratch on a stratified 70/30 split of its member and non-member instances across repeated random seeds. We never apply an attacker trained on an undefended model (or on another hyperparameter setting) to a defended target. In addition, in our new matched-utility experiment (see Reviewer 7PWn Table 2 / Reviewer Tcuy Table 2), we retrain black-box attackers on independent shadow SCCNNs using the exact same defense configuration as the target, and observe the same privacy–utility advantage for DTPI. We will state this protocol explicitly in Section 6.3 and Appendix D.3.

AuthorsOpenReview ↗

Response to Reviewer wYvG (Part 2)

Q2. Readout for held-out simplices and confirmation of no membership branching

Held-out simplices do not need an explicit column in Bk,trainB_{k,\mathrm{train}}, and the readout has no membership-dependent branch.

In our k=2k=2 triangle-prediction tasks, every candidate triangle (both training members and held-out non-members) consists of three boundary edges (ei,ej,ek)(e_i, e_j, e_k) from the shared public 1-skeleton S1S_1 (incidence matrix B1B_1). During the forward pass, the SCN computes edge representations over S1S_1 using L1,ℓL_{1,\ell} (from B1B_1) and L1,uL_{1,u} (from materialized training triangles B2,trainB_{2,\mathrm{train}}, together with cross-order projections in models such as SCCNN, Eq. (1)), producing a single edge-embedding matrix H∈R∣S1∣×dH \in \mathbb{R}^{|S_1|\times d}. Any queried candidate σ=ei,ej,ek\sigma = {e_i, e_j, e_k} is scored by indexing those three edge rows in HH and passing their concatenation through the same two-layer MLP head:

pσ=sigmoid(W2,sigmoid(W1,[hei∣hej∣hek]+b1)+b2).p_\sigma = \mathrm{sigmoid}\big(W_2,\mathrm{sigmoid}(W_1,[h_{e_i} | h_{e_j} | h_{e_k}] + b_1) + b_2\big).

Because the readout only indexes rows of HH by edge IDs and never consults columns of B2,trainB_{2,\mathrm{train}}, score differences between members and held-out non-members come solely from (i) parameter updates on training queries and (ii) training simplices' contribution to L1,uL_{1,u} (the internal operator representation RR) during message passing, which shapes the API responses OO in Proposition 4.2 (see also Figure 1(b) and Appendix E.1).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.