- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper introduces Bayesian Membership Privacy (BMP) for GNNs, incorporating node-dependent membership priors and asymmetric privacy bounds. It develops theoretical properties and a Bayesian auditing procedure based on a LiRA-style attack. Experiments evaluate the effects of sampling strategies and model configurations on the estimated privacy risks.
Critical strengths
- The motivation is meaningful. I agree with the authors’ argument that existing GNN privacy analyses largely inherit assumptions from non-graph settings, motivating a closer examination of graph structure.
- The paper combines theoretical analysis with empirical analysis to study sampling-dependent privacy risks. Incorporating membership priors provides a useful perspective on interpreting attack evidence.
Critical weaknesses
-
The limitations attributed to existing privacy analyses need clearer justification. The authors argue that prior analyses overlook structural dependence and stochastic graph sampling. Which specific assumptions fail, and what misleading conclusions do they produce?
-
Although the BMP formulation allows dependent membership states, the practical estimator in Section 4.2 still uses independence assumptions. Please clarify how the audit handles structural dependence.
-
The additional benefit of the Bayesian formulation is insufficiently explained. The paper acknowledges that BMP also applies to non-graph settings with nonuniform inclusion probabilities. What additional insight does the framework provide beyond incorporating these priors into an existing likelihood-ratio test?
-
The comparison with existing attacks is limited. Experiments compare prior-aware and prior-unaware variants of the same LiRA-style attack. However, the practical benefit over existing attack methods remains unclear.
Questions for authors
Please see the Weaknesses.
2 comments
Pointwise Response to Reviewer DrjR (W4)
W4 : The comparison with existing attacks is limited.
Response: As we stated in the 3rd bullet point of the second half of our common response, our objective is not to develop the strongest possible attack or certify individual GNNs, but to establish a theoretical foundation for BMP auditing and demonstrate its practical feasibility. Our auditing framework in Section 4 can use any membership inference attack (MIA) of choice; so a specific attack such as LIRA is not indispensable for the proposed methodology. Specifically, in Algorithm 1 in Appendix A.1, lines 7-9 can be replaced with any more powerful MIA. Stronger attacks are certainly preferable and may change the estimates, but not the BMP definition or its theoretical characterization.
Action taken: In hindsight, reading Section 4.1, we admit the reader may have the impression of LiRA being described as the only choice for MIA. However, in reality, the auditing pipeline can accommodate any MIA, an important point not emphasised sufficiently. We would like to prevent that misunderstanding; therefore, in our revised version, we have added the following sentence at the top of Section 4.1, before explaining LiRA: We present a specific MIA, which we use in the experiments. However, note that the presented auditing pipeline is not restricted to a particular MIA and can accommodate any other MIA.
Pointwise Response to Reviewer DrjR (W1–W3)
We thank the Reviewer for their comments that helped improve the paper. The reviewer’s comments are largely addressed in a broader context in our common response; here we provide our point-to-point responses. We hope these clarifications address the remaining concerns and respectfully ask the Reviewer to reconsider and improve their score accordingly.
W1: Prior work: Which specific assumptions fail, and what misleading conclusions do they produce?
Response: In graph based learning, structural dependencies challenge statistical exchangeability of train-test records (as shown in Khosla (2026, section 3.3)), while graph-dependent sampling induces heterogeneous membership probabilities for train nodes. The loss of exchangeability can invalidate the assumptions underlying existing MIA-based DP auditing methods, limiting their ability to establish formal privacy guarantees for graph-based learning . Furthermore, assuming uniform membership priors overlooks the fact that graph sampling can make some nodes substantially more likely to be included than others. Consequently, identical MIA error rates can correspond to different posterior membership risks, while aggregate attack metrics can conceal substantial node-level privacy heterogeneity. Our BMP formulation directly addresses these limitations by incorporating node-specific membership priors and establishing a prior-dependent characterization of privacy guarantees through MIA error probabilities (Theorem 4).
W2: BMP formulation allows dependent membership… but estimator uses independence… Clarify how the audit handles structural dependence
Response: We would like to separate two dependencies, which may be conflated in the discussion.
W3 The additional benefit of the Bayesian formulation is insufficiently explained. What additional insight … beyond incorporating these priors into an existing likelihood-ratio test?
Response: We agree that incorporating prior odds into a likelihood-ratio test follows directly from Bayesian decision theory; however, we do not claim the prior-aware test itself as the principal contribution.
A test makes a membership decision for a particular node and threshold. BMP instead bounds the adversary’s posterior membership odds for every node. We show that this definition is equivalent to node-prior-dependent constraints on the type-I and type-II errors of any decision rule—not only the likelihood-ratio attack used in our experiments. This provides a way to analyze heterogeneous, asymmetric privacy risk and its implications for attack errors and decision costs.
There is yet another, distinct Bayesian component in the audit. From finite, noisy attack-error counts, our hierarchical model is used to estimate a posterior distribution over the BMP parameters, rather than reporting a single plug-in value. Graph sampling motivates the heterogeneous priors, but the definition and audit apply equally to non-graph settings with nonuniform inclusion probabilities.
We hope our response distinguishes the three different instances of Bayesian formulation in our work: Prior-aware attack, BMP definition, and Bayesian parameter estimation. We claim our contribution mostly in the latter two (corresponding to the first three contribution points at the end of Section 1), and not the first one.