acceptodds
Under review as a conference paper at ICLR 2027

Is Decision Boundary a Reliable Signature for Graph Neural Network Ownership Verification?

Abstract

Graph neural networks (GNNs) have demonstrated superior performance in various applications, such as recommendation systems and financial risk management. However, deploying large-scale GNN models locally is particularly challenging for users, as it requires significant computational resources and extensive property data. Consequently, Machine Learning as a Service (MLaaS) has become increasingly popular, offering a convenient way to deploy and access various models, including GNNs. However, an emerging threat known as Model Extraction Attacks (MEAs) presents significant risks, as adversaries can readily obtain surrogate GNN models exhibiting similar functionality. Specifically, attackers repeatedly query the target model using subgraph inputs to collect corresponding responses. These input-output pairs are subsequently utilized to train their own surrogate models at minimal cost. Many techniques have been proposed to defend against MEAs, but most are limited to specific output levels (e.g., embedding or label) and suffer from inherent technical drawbacks. To address these limitations, we propose a novel ownership verification framework CITED which is a first-of-its-kind method to achieve ownership verification on both embedding and label levels. Moreover, CITED is a novel signature-based method that neither harms downstream performance nor introduces auxiliary models that reduce efficiency, while still outperforming all watermarking and fingerprinting approaches. Extensive experiments demonstrate the effectiveness and robustness of our CITED framework. Code is available at: https://anonymous.4open.science/r/CITED.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Machine Learning, Model Extraction Defense

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XhrVOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies whether decision boundaries can provide reliable signatures for GNN ownership verification. CITED selects signature nodes using prediction uncertainty, embedding distances, prediction distributions, and neighborhood label disagreement. Verification then compares either the embedding distributions through Wasserstein distance or the predicted labels through label agreement. The experiments cover 7 datasets and 5 GNN architectures, with further studies in the appendix. My main concerns are the selection formulas and the theoretical analysis. Some experimental choices also make the reported gains difficult to interpret.

Critical strengths

The paper addresses a relevant problem, and it is useful to study embedding-level and label-level verification within the same framework.

The verification procedure is straightforward and does not require a separate verification classifier. Figure 2 reports lower inference time than GrOVe in the evaluated embedding-level setting.

The experimental coverage is broad. In addition to the main comparisons, the appendix examines larger graphs, signature removal, adaptive filtering, and signature size. Reporting variation across random seeds is also helpful.

Critical weaknesses

1. Some terms in the selection formulas do not have the stated effect.

In Equation (1), let ZZ denote the logit matrix and vv a node. The paper defines pp and qq as the indices of the largest and second-largest logits for node vv, respectively. The two logit values therefore satisfy Zv,q≤Zv,pZ_{v,q}\leq Z_{v,p}, so

ReLU⁡(Zv,q−Zv,p)=0.\operatorname{ReLU}(Z_{v,q}-Z_{v,p})=0.

As written, the boundary score reduces to its negative-entropy term and does not include the claimed contribution from the logit gap.

In Equation (5), the heterogeneity term takes a maximum over a set. For a fixed maximization set, the maximum is the same for every candidate node. Subtracting a common value cannot change the candidate ranking or the resulting quantile-based selection. The ablation in Table 3 does not resolve the ambiguity: it lacks a margin-plus-thickness control that would isolate the added effect of heterogeneity.

Both issues could be notation errors, but they affect the definition of the signature set. Please provide the implemented expressions and clarify whether the experiments and ablations used the formulas printed in the paper.

2. The current proofs do not establish the claimed bounds.

The proof of Theorem 4.1 in Appendix A.2 relies on an incorrect event implication. Let SS denote the linear term in the embedding change caused by parameter perturbations, RR the higher-order remainder, ΔG\Delta_G the deterministic bound on the embedding change, and λ>0\lambda>0 the embedding-distance threshold. Using Euclidean norms, the proof treats

\qquad \|R\|_2\leq\Delta_G-\lambda$$ as sufficient for the total embedding change to satisfy $\|S+R\|_2<\lambda$. The implication fails even with zero remainder, $R=0$. For example, setting the threshold to $\lambda=\Delta_G/4$ and the linear-term norm to $\|S\|_2=\Delta_G/2$ satisfies both premises, but the total embedding change exceeds the threshold. The triangle inequality instead gives the sufficient condition $\|S\|_2+\|R\|_2<\lambda$. The probability bound needs a corrected proof. The proof of Theorem 4.2 in Appendix A.3 has a separate gap. Let $C$ denote the number of classes. The proof first obtains a union bound with coefficient $C$, then drops the negative-tail event for the originally predicted class to obtain $C-1$. A decrease in the originally predicted class's logit can change the prediction even when no competing logit crosses its positive-tail threshold, so the omitted event still matters. A coefficient of $C-1$, corresponding to the competing classes, may be recoverable by bounding pairwise logit differences. However, the argument in the paper does not establish that bound. The same proof assumes that the nonlinear remainder is negligible relative to the classification margin without giving a quantitative condition that relates the two. Small parameter perturbations alone do not justify ignoring the remainder when the margin can be arbitrarily small. The missing condition is especially relevant for a method that selects nodes near the decision boundary. There is also a question of scope. Small perturbations of corresponding parameters do not directly describe independently trained or cross-architecture surrogates. The experiments do compare extracted surrogates with independently trained models, but the theorems do not establish the corresponding separation or a false-positive guarantee. The distinction between parameter stability and ownership verification should be made explicit. **3. The extra fine-tuning needs a matched control.** Appendix E.4 describes 200 epochs of initial training, and Appendix E.5 adds 50 epochs of fine-tuning with the original task labels after signature generation. Appendix E.6 specifies 200 epochs for the baselines. Fine-tuning with the original labels does not by itself imply signature injection, but the extra optimization could explain part of the utility gains in Table 1. The reported training budgets need to be reconciled before attributing those gains to CITED. I would like to see CITED without fine-tuning, ordinary fine-tuning for the same additional 50 epochs, and the full CITED procedure, with utility and verification results under matched training budgets. Please also specify which labeled nodes and data split are used for fine-tuning, whether the fine-tuning objective or sampling depends on the signature set, and whether the signatures remain fixed or are recomputed afterward. **4. The results may depend on the boundary-focused extraction protocol.** Appendix E.2 constructs the query set using 20% boundary samples and 80% randomly selected samples. The protocol is reasonable for one extraction strategy, but it gives the surrogate access to the same type of boundary information that CITED uses for verification. It is therefore unclear how much the verification results depend on this query choice. A random-only query control and a sweep over the boundary-query fraction would help answer the question. Query budgets should be matched, and verification results should be compared at similar surrogate utility, or at least reported separately by utility level. Otherwise, lower verification performance could simply reflect a less successful extraction. Final signature sets selected randomly or by entropy alone, with the same number of nodes, would also help isolate the value of the proposed selection rule. The adaptive-filtering experiments are useful, but filtering an existing query set does not directly answer how verification changes under different initial query strategies. **5. The discussion of label-level fingerprinting is too broad.** Section 3.2 states that fingerprinting does not generalize to label-level verification. However, IPGuard (Cao et al., 2021), which the paper already cites, fingerprints a classifier using inputs near its decision boundary and checks agreement between predicted labels. Label-level boundary verification is therefore an existing idea. The contribution of CITED may lie in graph-specific node selection, neighborhood information, and support for both output levels. I would like the paper to explain those differences more clearly. A graph-adapted boundary-fingerprinting baseline would make the comparison more convincing. If such an adaptation is not appropriate, please explain the specific reason.
Questions for authors
  1. What are the implemented forms of Equations (1) and (5)? Which formulas were used in the experiments, and how does heterogeneity affect candidate selection?

  2. Can the authors correct the proofs of Theorems 4.1 and 4.2, including the missing control of the nonlinear remainder relative to the classification margin? How do the theoretical assumptions relate to the independently trained and cross-architecture surrogates used in the experiments?

  3. How do the results change when the additional fine-tuning is controlled for? Which labels are used, does fine-tuning depend on the signature set, and when are the final signatures computed?

  4. How does verification change under random-only queries and different boundary-query fractions at matched query budgets and comparable surrogate utility?

Reviewer_NPTAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CITED, a decision boundary-aware signature-based framework for verifying the ownership of Graph Neural Networks (GNNs) under Model Extraction Attacks (MEAs). Unlike conventional watermarking and fingerprinting methods that often focus on a particular output level, CITED aims to provide a unified ownership verification mechanism for both embedding-level and label-level outputs.

The proposed framework identifies signature nodes based on three characteristics: boundary margin, boundary thickness, and neighborhood heterogeneity. These nodes, along with their corresponding outputs from the protected model, are subsequently used to distinguish extracted surrogate models from independently trained models. The authors employ Wasserstein distance for embedding-level verification and prediction agreement for label-level verification, using ARUC and AUC to evaluate verification effectiveness.

The primary evaluation covers seven graph datasets and five GNN architectures under a transductive node classification setting. The authors compare CITED against several watermarking and fingerprinting baselines and provide theoretical analysis based on bounded parameter perturbations. The supplementary material additionally includes experiments on OGB-Arxiv, graph classification, regression, removal attacks, adaptive filtering, and signature-size sensitivity.

Critical strengths

--A key strength of the paper is its attempt to address ownership verification at both embedding and label levels within a single framework. This is a relevant problem for GNN-based MLaaS systems, where the outputs exposed to users can vary across applications. The proposed approach is also attractive because it attempts to avoid task-irrelevant watermark triggers and the auxiliary-model training overhead associated with some fingerprinting approaches.

This direction has practical relevance, and the idea of extracting model-specific information directly from the learned decision boundary is worth exploring.

--The authors evaluate CITED across 5 GNN architectures and datasets, comparing it against established watermarking and fingerprinting approaches. The reported ARUC results consistently favor CITED in the primary experiments.

The paper also includes useful supplementary experiments examining removal attacks, adaptive filtering, noisy labels, class imbalance, signature compression, and additional learning tasks. These demonstrate an effort to evaluate the framework beyond a single narrow configuration.

However, several aspects of these experiments require further clarification and stronger controls, as discussed below.

Critical weaknesses

--My primary concern is that the proposed signature-scoring formulation lacks sufficient theoretical motivation, and certain equations appear inconsistent with the intended functionality.

--In Equation (1), the boundary score includes the term

ReLU(Zv,q−Zv,p),{ReLU}(Z_{v,q}-Z_{v,p}),

where pp and qq correspond to the largest and second-largest logits, respectively. Since Zv,p≥Zv,qZ_{v,p}\geq Z_{v,q}, this term is always zero. Consequently, the boundary score, as written, reduces to a negative entropy score. It is unclear whether this is a typographical error or whether the implementation indeed performs entropy-based boundary selection.

Even setting aside these issues, the choice of combining boundary margin, thickness, and heterogeneity in this particular manner remains insufficiently justified. The individual components are intuitively motivated, but it is not clear why their weighted combination should produce signatures that are simultaneously distinctive across independently trained models and preserved in extracted surrogate models.

--My second major concern is that several aspects of the experimental evaluation make it difficult to determine how much of the reported improvement can be attributed to the proposed signature construction.

Table 1 lacks the utility of the original, undefended target model. This makes it difficult to establish the actual utility degradation introduced by each defense. More importantly, Appendix E.5 indicates that CITED fine-tunes the target model for an additional 50 epochs following signature generation. Without an additional-training-only control and matched optimization budgets across defenses, it is unclear whether some reported utility improvements result from the signature method or simply from additional training.

CITED achieves the highest downstream accuracy in most of configurations in Table 1, but there are several cases where other methods perform better, sometimes by multiple percentage points. A failure-case analysis would help explain these differences.

Second, although the main experiments demonstrate favorable verification results, the supplementary evaluations reveal important limitations. For example, Table 10 reports an ownership-verification AUC of approximately 0.574 on OGB-Arxiv, suggesting relatively weak discrimination between extracted and independent models.

Finally, the primary evaluation is restricted to transductive node classification. Although the appendix includes graph classification and regression experiments, the method's applicability to inductive graph learning and strongly heterophilous graphs remains insufficiently established.

Questions for authors

The following questions focus on clarifications and evidence that could materially influence my assessment.

--In Equation (1), why is the difference between the second-largest and largest logits passed through ReLU when this difference is necessarily non-positive? Was the intended subtraction order reversed?

Similarly, in Equation (5), is the heterogeneity term intended to depend on the candidate node rather than the maximum over a fixed set?

Could the authors clarify the exact scoring functions used in the implementation and whether the reported experimental results were generated using these expressions?

--What motivates the precise aggregation of margin, thickness, and heterogeneity in Equation (5)?

Although Appendix D.5 provides a coefficient sensitivity analysis, Table 11 reports three-component coefficient configurations, whereas Equation (5) explicitly introduces only two weighting coefficients.

Could the authors reconcile these formulations and provide the actual coefficient values used in the primary experiments, including the selection procedure and the normalization applied to individual scores? The selection of other score-related parameters, including λ\lambda and γ\gamma, should also be clarified.

--Could the authors report the utility of the original target models before applying any ownership protection, alongside the utility changes introduced by each defense?

Since CITED performs an additional 50 epochs of fine-tuning, it would be particularly useful to compare against an undefended model trained for an equivalent number of epochs and to clarify whether competing methods receive comparable optimization budgets.

Also, which node labels are used during CITED's additional fine-tuning stage, and are signature outputs recorded before or after this stage? These details are necessary to assess potential data leakage and the consistency of the stored signature.

--Table 10 reports an AUC of approximately 0.574 on OGB-Arxiv. What explains the substantial difference from the near-perfect AUC results reported on several smaller datasets? Which backbone and output level were used for this experiment?

Similarly, Table 12 indicates meaningful degradation under pruning, fine-tuning, and distribution shift, especially on larger graphs. Could the authors provide a more detailed analysis of these failure cases, together with the corresponding downstream utility of the modified surrogate models?

--Could the authors discuss whether CITED can be adapted to inductive node classification, where unseen nodes or entirely new graphs may become available after the protected model has been trained?

I would also be interested in understanding how the method behaves on heterophilous graphs. Since Equation (4) measures neighborhood prediction disagreement, high heterogeneity may correspond to natural graph structure rather than decision-boundary instability. Would this affect the selection and distinctiveness of signature nodes?

Reviewer_3gtXOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CITED for GNN ownership verification under model extraction. The motivation is that a model service may return either node embeddings or only labels, while most existing watermark or fingerprint methods handle just one of these levels. The proposed method first finds nodes near the decision boundary of the target model using the logit gap and prediction entropy. It then scores the remaining nodes by boundary margin, boundary thickness and neighbor label disagreement, and keeps the lowest-scoring ones as the signature set. Experiments cover seven node classification datasets and five GNN backbones with additional results on OGB-Arxiv, removal attacks and graph-level tasks.

Critical strengths

The paper addresses the relevant problem of unified GNN ownership verification across embedding and label outputs. Its decision boundary-aware signature construction combines representation, prediction, and structural information without relying on artificial watermark triggers. The theoretical analysis investigates output consistency under parameter perturbations, while experiments cover multiple datasets, architectures, and defense baselines

Critical weaknesses

W1. The comparison with IPGuard should be more accurate. The paper does discuss fingerprinting and argues CITED is different because fingerprinting relies on embedding distributions, struggles at the label level and needs reference models. But IPGuard, which the paper groups with embedding-distribution methods, actually fingerprints the classification boundary with points near it and checks label agreement, which is close to the label-level part of CITED.

W2. The utility and cost evaluations lack some reference points. Table 1 does not include an explicit row for the undefended model. Since GrOVe does not modify the target model, its row should be able to serve as this baseline? But the paper does not explicitly state this. Regarding cost, Fig. 2 only compares inference time with GrOVe, and Table 23 reports the signature generation time, but the 50-epoch fine-tuning of the target model described in E.5 is not included in any cost analysis.

W3. It's unclear how verification depends on the attacker's query budget. Assumption B.3 says surrogates need enough queries to be useful, but the experiments use one query setting. It's natural to ask how verification and surrogate accuracy change together as the budget grows.

W4 Some equations and notations need fixing. In Eq. 1, if p is the top-1 index and q is the top-2, the ReLU term is always 0. The max term in Eq. 5 doesn't depend on the node, so it's a constant. Eq. 5 has two weights while Table 11 lists three coefficients. The subscripts in Eq. 6 and Eq. 7 are broken.

Questions for authors

refer to weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.