Differentially Private Decoupled Graph Convolutions for Multigranular Topology Protection

GNNs can inadvertently expose sensitive user information and interactions through their model predictions. To address these privacy concerns, Differential Privacy (DP) protocols are employed to control the trade-off between provable privacy protection and model utility. Applying standard DP approaches to GNNs directly is not advisable due to two main reasons. First, the prediction of node labels, which relies on neighboring node attributes through graph convolutions, can lead to privacy leakage. Second, in practical applications, the privacy requirements for node attributes and graph topology may differ. In the latter setting, existing DP-GNN models fail to provide multigranular trade-offs between graph topology privacy, node attribute privacy, and GNN utility. To address both limitations, we propose a new framework termed Graph Differential Privacy (GDP), specifically tailored to graph learning. GDP ensures both provably private model parameters as well as private predictions. Additionally, we describe a novel unified notion of graph dataset adjacency to analyze the properties of GDP for different levels of graph topology privacy. Our findings reveal that DP-GNNs, which rely on graph convolutions, not only fail to meet the requirements for multigranular graph topology privacy but also necessitate the injection of DP noise that scales at least linearly with the maximum node degree. In contrast, our proposed Differentially Private Decoupled Graph Convolutions (DPDGCs) represent a more flexible and efficient alternative to graph convolutions that still provides the necessary guarantees of GDP. To validate our approach, we conducted extensive experiments on seven node classification benchmarking and illustrative synthetic datasets. The results demonstrate that DPDGCs significantly outperform existing DP-GNNs in terms of privacy-utility trade-offs.

Paper

Similar papers

Peer review

Reviewer fPNi5/10 · confidence 5/52023-07-01

Summary

Graph neural networks have privacy leakage in both their topology information and node attribute information. This paper proposes a differential privacy framework to protect both graph topology and node attributes. A model that decouples graph convolution and node attribute embedding is proposed.

Strengths

A graph differential privacy (GDP) framework is proposed for GNN models. Theoretical GDP guarantees are provided.

Weaknesses

1. A weakness of using the differential privacy (DP) metric is the significant deterioration of utility (in this paper, it is the node classification accuracy) for even a very generous privacy budget. As seen in Table 3 and Figure 3, \epsilon=16 has to be set to achieve reasonable accuracy (except for the simplest case of edge-level privacy). However, even with such a generous budget, the test accuracy drops significantly compared to the non-private case. One should question if DP is indeed the proper framework to use in GNN (despite its popularity in database privacy). For example, there are frameworks on *inference* privacy that specifically protect certain private attributes instead of the full "data" (graph topology + features), which are more applicable in practice. 2. By decoupling the graph adjacency information A from the node attributes X, the model can no longer benefit from graph aggregation and local node processing. This also explains why the proposed model does not perform well on homophily datasets.

Questions

1. It was not immediately clear by the end of Section 5 that the output of DP-MLP_W is also designed to be DP, hence the overall framework is DP due to the composition theorem. It caused some confusion for me and I would appreciate if this point is emphasized since DP-MLP_W is never discussed in detail throughout the paper. 2. How is the MLP in Table 3 trained to achieve GDP? If the MLP is GDP and protects graph information (i.e., using the individual outputs from the MLP applied to individual nodes, one cannot easily infer if there are edges between them), why does it perform significantly better than DPDGC or DP-SAGE? 3. On the heterophily datasets, the reduction in test accuracy is very significant compared to the non-private scenario. What is causing this? A detailed discussion should be added. 4. How tight are the bounds in the theoretical results in Section 5 and are these used directly in the DPDGC model or is the model tuned instead based on an empirical estimate of its GDP? 5. What is the computational complexity compared to baselines? Is a distributed implementation using message passing possible?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes

Reviewer ssFv4/10 · confidence 5/52023-07-04

Summary

This paper introduces a new framework called Differentially Private Decoupled Graph Convolutions (DPDGC) for graph learning settings that ensures both provably private model parameters and predictions. The framework is designed to protect sensitive user information and interactions in graph-structured data. The authors highlight the limitations of standard Differential Privacy techniques in graph learning settings and propose a novel notion of relaxed node-level data adjacency to establish guarantees for different degrees of graph topology privacy. The paper also includes an analysis of the framework and its performance compared to existing methods.

Strengths

1. This paper focuses on the important problem of graph differential privacy, which is critical in protecting sensitive user information and interactions in graph-structured data. 2. This paper conducts experimental evaluation on seven node classification benchmarking datasets.

Weaknesses

1. The paper's presentation is difficult to follow, which may make it challenging for readers, especially for those who do not have strong background knowledge on this topic hard to understand the proposed method and its contribution. 2. The proposed method has poor performance, which is reflected in the experimental results presented in the paper. 3. Some relevant literature has not been cited in the paper, which could suggest that the authors have not conducted a thorough review of the existing research in this area.

Questions

1. How can one determine the appropriate value of K for a given dataset, in order to achieve meaningful results considering both privacy protection and utility? 2. In the case of the Pubmed and Cora datasets, why do the results of graph-based machine learning methods remain unchanged across different values of K, and what implications does this have for the use of these datasets in research?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

1. The paper's presentation could be improved to make it easier for the reader to follow the content and understand the proposed method. 2. The utility of the proposed method is limited, as the privacy protection achieved with a privacy budget of eps=16 is too weak to be meaningful for many datasets. Furthermore, even with this level of privacy protection, the performance of the proposed method is significantly worse than that of non-private baseline models on most datasets, indicating poor utility. 3. The paper could benefit from a more comprehensive review of related works, as some relevant studies are not cited or discussed in the text, including [1]. [1] Zhang, Q., Ma, J., Lou, J., Yang, C., & Xiong, L. (2022). Towards Training Graph Neural Networks with Node-Level Differential Privacy. arXiv preprint arXiv:2210.04442.

Reviewer tLab4/10 · confidence 2/52023-07-07

Summary

The paper presents a well-written and easily understandable framework called Graph Differential Privacy (GDP) tailored for graph learning methods. The proposed framework aims to address the privacy challenges associated with GNNs by ensuring both model parameter and prediction privacy. The authors introduce the Differentially Private Decoupled Graph Convolution (DPDGC) model, which offers superior privacy-utility trade-offs compared to existing approaches. The paper includes theoretical analysis that provides a solid foundation for the proposed model. The authors evaluate the DPDGC and compare it with existing differentially DP-GNN methods, as well as non-private models. By achieving SOTA performance, the experimental results validate the effectiveness and utility of the proposed DPDGC model in graph learning tasks.

Strengths

1. The paper is its clarity and coherence. The authors have effectively conveyed complex concepts and ideas in a well-structured manner. 2. The theoretical analysis provided in the paper adds good value to the research, supporting the proposed model and enhancing its credibility. 3. The incorporation of the DPDGC model, which leverages decoupled graph convolution, achieves SOTA result.

Weaknesses

1. Lack of comparison with other methods: The paper focuses primarily on comparing the performance of the proposed GDP-based methods (including DPDGC) against other differentially private graph learning methods. However, it would be beneficial to include a comparison with SOTA non-private graph learning methods to better understand the tradeoffs between privacy and utility.   2. Limited evaluation on larger and more diverse datasets: The experimental evaluation of the proposed methods is conducted on a relatively small set of benchmark datasets. The generalizability and scalability of the methods to larger and more diverse datasets are not extensively explored. Including a broader range of datasets would provide a more comprehensive evaluation of the proposed methods' performance and generalizability.   3. Lack of detailed analysis on privacy guarantees: While the paper mentions the privacy guarantees provided by the GDP framework and the DPDGC model, the detailed analysis of these guarantees is not thoroughly discussed. Providing more in-depth analysis, proofs, and discussions of the privacy guarantees would strengthen the paper's claims about the privacy properties of the proposed methods.   4. Limited exploration of alternative privacy mechanisms: The paper primarily focuses on GDP as the privacy framework and DPDGC as the corresponding graph learning model. However, there are various other privacy mechanisms and techniques available in the field of differential privacy.

Questions

See Weaknesses

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The authors do not adequately acknowledge potential limitations in their work, which may indicate a lack of comprehensive understanding of the challenges and constraints associated with the proposed framework.

Reviewer XfoH7/10 · confidence 3/52023-07-27

Summary

The paper introduces a differentially private GNN model that allows for different privacy requirements for node attributes and graph structure. The model decouples graph convolutions from node attributes and graph topology and provides provable privacy guarantees. Experimental results are provided to show the proposed methodology's superiority.

Strengths

The problem and contributions discussed in the paper are clearly mentioned and the illustrations do a good job of conveying them. Graph differential privacy is an interesting topic and the idea of providing flexibility for node attributes and graph structure is promising.

Weaknesses

Some discussion regarding the questions mentioned in the next section would add to the paper greatly.

Questions

1. From line 310: "we also test (DP-)MLP and several DP-GNN baselines that can achieve GDP guarantees, including RandEdge+SAGE [29] and DP-SAGE [18] for edge and node GDP, respectively.". However, it is not clear what MLP refers to in Table 3. 2. There is a large jump in performance for the non-private setting for DPDGC but the other methods do not exhibit this behavior. Any insight into this phenomenon? 3. From line 341 - DPDGC starts to outperform GAP when privacy budget increases but lags behind when privacy budget is small. What is the typical scenario in real world situations? 4. Regarding the comment on line 330: does homophily alone decide for which datasets utility loss from privacy noise compensates graph structure information? Basically, what are the things that one has to consider before picking the right algorithm to achieve GDP? Minor: On line 109, what is T? Table 3: change "none" to "non"

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer WwFs4/10 · confidence 3/52023-07-27

Summary

The authors introduce a new model for Graph Differential Privacy (GDP) that ensures a parametric level of topological privacy through decoupling of the graph convolution mechanism (i.e., preventing direct neighborhood aggregation of features = standard $AXW$ aggregation). The key definition is that up to $k$ in- and out-neighbours of a randomly selected single node $r$ can be modified to create a new adjacency matrix. This makes this a hybrid between edge ($k=1) and node ($k=n) GDP, the parameter $k$ allows a tradeoff between topology privacy and accuracy. All GNN training is done using this modified graph $D’$ to create the model parameters for inference. The main contribution of the paper is three-fold 1) proving theoretical bounds on the Differential Privacy (DP) state of the art DPGNN called GAP [19] 2) Intuitions from analyzing the DP weakness of GAP to develop a novel Differentially Private Decoupled Graph Convolution (DPDGC) model, which benefits from decoupling graph convolution while providing GDP guarantees 3) theoretical bounds on the DP of their proposed DPDGC. The key intuition is that GAP has greater privacy leakage because they compute $A'H'$ where both adjacency matrix and features change. Motivated by this the authors propose DPDGC in which the $A'H'$ product is avoided thus providing more privacy than GAP.

Strengths

Sound theoretical analysis of the two GDP models Theoretical analysis of GAP i.e., the presence of the $A'H'$ product which is shown from Theorem 1 to be a contributing factor to the DP limitation of GAP. This leads to their derivation of DPGDC which applies a DP-MLP to adjacency matrix A using a non-linear operation on $AW^{(A)}$ to create the adjacency matrix embedding $Z$, where $W_A$ are fixed model weights. This is opposed to GAP which computes $AH$ for the $Z$ embeddings. By ensuring $W^{(A)}$ is DP, they only need to look at $A'W^{(A)}$ versus $A'H'$ in GAP.

Weaknesses

A big portion of the paper is spent defining the problem and setting up conventions for future Differential Privacy studies to be more suitable to the GNN field/setting. The novelty of the decoupling method is not very convincing. The idea of decoupling is previously found in this paper [1]"Large Scale Learning on Non-Homophilous Graphs: New Benchmarks and Strong Simple Methods", D. Lim, F. Hohme et. al, Neurips 2021, which defines a method called LINKX for non-homophilous graphs. LINKX separately embeds the adjacency $A$ to $h^A$ and the features $X$ into $h^X$ before mixing adjacency and feature information. The decoupling method proposed here seems to be a modification of the LINKX strategy. From Table 3, the proposed DPDGC model works well compared to the other differential privacy models when the graphs are heterophilic. This is not surprising because the decoupling idea is known to be beneficial for heterophilic graphs [1]. However, the other baseline methods work better in a homophily setting (columns in the right side of the table) because they are not specifically designed for that kind of setting. I would expect a more adaptive method that works in both settings to be more convinced about the utility of this method. Also from Table 3, simple MLP outperforms DPDGC in accuracy for higher $k$ on many datasets, especially homophilic ones. As pointed out by the authors themselves, protecting the graph information (higher privacy requirements) quite drastically reduces the utility of these decoupling methods. So the practical utility of this seems questionable.

Questions

What is the practical meaning of high-$k$ topology protection? in other words what are the additional security benefits obtained by going from $k$ to $k+1$. Its not clear to what is the marginal utility of increasing $k$ in terms of the increasing cost to an adversary who wants to break privacy. This is an important question as varying $k$ is the major difference between this work and GAP. How well will the decoupling method DPDGC work on large scale homophilous graphs with moderately high privacy budgets. I would like to see more extensive experimental results to justify the cost of decoupling versus simple MLP methods. Fig 1 is hard to understand (contextualize) with a lot of undefined terms (e.g., k-neighbor level adjacency) especially as it comes so early in the paper. Consider moving to later or defining better in-text. Def 4.3 seems ambiguous: $k$ entries of $A_{rj}$ and $A_{lj}$ are modified but what is the size of $ j + l $( Is $j+l=k$ or $2k$?). Text only says “some” $j$ and $l$. There are multiple variations of row normalization, are you doing Euclidean norm normalization of each row, it is not clear from the text and obviously it makes a big difference in the proof as it is known that Euclidean row normalization dampens the effect of outliers [1] “Sign and rank covariance matrices” J. of Statistical Planning and Inference, Dec. 2000.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The authors have addressed the limitations of this work in the appendix. "we do not believe that the current DPDGC model is the ultimate solution for GDP-aware graph learning methods. To support this claim, we note that the nonprivate state of-the-art performance for learning on large-scale homophilic graphs is achieved by standard graph convolution models [47, 48]. The authors have stated that the proposed topic doesn't have any negative societal impacts, instead GDP can potentially protect user data and is beneficial. However this is a generic statement and needs to be proved (to what extent will breaking GDP of a graph impact individual users?)

Reviewer fPNi2023-08-12

Thank you for your clarifications. I appreciate the authors' efforts in responding to the reviews. I have no further questions.

Authorsrebuttal2023-08-14

Thank you for the acknowledgment!

Reviewer WwFs2023-08-13

Thank you for the clarifications. No further questions as of now.

Authorsrebuttal2023-08-14

Thank you for the notification and response!

Reviewer XfoH2023-08-13

Thank you for the thorough response

Changed the score to 7

Authorsrebuttal2023-08-14

Thank you for the positive feedback and raising the score! We really appreciate the fruitful discussion with reviewer XfoH.

Reviewer ssFv2023-08-14

Thanks for the clarifications. It's worth noting that there's relevant literature missing, specifically reference [1]: [1] Epasto, Alessandro, et al. "Differentially Private Graph Learning via Sensitivity-Bounded Personalized PageRank." Advances in Neural Information Processing Systems 35 (2022): 22617-22627.

Authorsrebuttal2023-08-14

About the additional reference

We thank reviewer ssFv for providing the additional reference [ref 4]. However, we would like to emphasize that [ref 4] is not about DP-GNN but specifically the DP PageRank algorithm. We believe that there can be much more literature about DP graph algorithms. Due to the space limitation, we choose to focus our discussion on DP-GNNs in the related work section, which is the most relevant to our manuscript. Note that neither our DPDGC nor discussed DP-GNN baselines leverage PageRank algorithms. The only exception is the reference Zhang et al. [ref 1] provided by the reviewer ssFv, where their work is a privatized version of APPNP and thus relevant to [ref 4]. We will try to include [ref 4] along with [ref 1] if there is still space in our revision. ### Reference [ref 1] Towards Training Graph Neural Networks with Node-Level Differential Privacy. Zhang et al. arXiv preprint arXiv:2210.04442. [ref 4] Epasto, Alessandro, et al. "Differentially Private Graph Learning via Sensitivity-Bounded Personalized PageRank." Advances in Neural Information Processing Systems 35 (2022): 22617-22627.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC