Adversarial Training for Graph Neural Networks: Pitfalls, Solutions, and New Directions

Despite its success in the image domain, adversarial training did not (yet) stand out as an effective defense for Graph Neural Networks (GNNs) against graph structure perturbations. In the pursuit of fixing adversarial training (1) we show and overcome fundamental theoretical as well as practical limitations of the adopted graph learning setting in prior work; (2) we reveal that more flexible GNNs based on learnable graph diffusion are able to adjust to adversarial perturbations, while the learned message passing scheme is naturally interpretable; (3) we introduce the first attack for structure perturbations that, while targeting multiple nodes at once, is capable of handling global (graph-level) as well as local (node-level) constraints. Including these contributions, we demonstrate that adversarial training is a state-of-the-art defense against adversarial structure perturbations.

Paper

References (54)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer bw8r5/10 · confidence 4/52023-07-05

Summary

This paper presents a comprehensive analysis of the transductive setting, highlighting the limitations of achieving perfect robustness solely by memorizing the clean graph. The authors argue that the observed improvements in robustness from previous adversarial training methods stem from unintentional leakage of information about the clean test nodes. To address this issue, the authors explore the inductive setting, which mitigates these drawbacks. They propose a defense mechanism by combining GPRGNN and ChebNetⅡ with adversarial training, effectively countering both globally and locally constrained attacks. For the implementation of the locally constrained attack, the authors develop an approximate solution to a relaxed multi-dimensional knapsack problem. The experimental results presented in the paper showcase the efficacy of the proposed attack and defenses.

Strengths

1. It is meaningful to discuss the adversarial training on graph models, as this is still an open problem. 2. I highly value the investigation conducted in the section dedicated to locally constrained attacks, and I find the implementation of LR-BCD to be inherently rational. 3. Experiments are sufficient enough to demonstrate the efficacy of the proposed methods.

Weaknesses

1. The primary concern with this work lies in the disparity between the initial analysis and the subsequent defense method design. Although the analysis extensively examines the limitations of transductive settings, there is a noticeable absence of a correlation with the utilization of an adaptive spectral GNN to enhance robustness. The motivation of the proposed defense is not very clear, and it is a combination of existing methods. While adaptively assigning weights to local neighborhoods can be an effective strategy for improving robustness, it appears to be unrelated to the previous discussions on the flaws of transductive settings and adversarial training. 2. The authors claim that GPRGNN offers interpretability, but Section 3 primarily focuses on presenting the learned message passing weights and filter kernels of GPRGNN. While this aligns with the intuition of enhancing GNN robustness by incorporating high-frequency information and higher-order neighborhoods, it remains challenging to argue that this alone qualifies as interpretability. 3. The discussion of graph adversarial attacks in the transductive setting remains valuable. On one hand, many methods in the transductive setting do not involve leakage of clean local structures of test nodes, and conventional adversarial training does not include a self-training module. On the other hand, in real-world scenarios, memorizing the graph is not certainly feasible because the structure and labels of graph nodes can be dynamically changing. In fact, even without the addition of new nodes to the graph, the defense cannot classify the current nodes based on the remembered graph structure. For example, in graph-based spammer or fraud detection, benign users and fraudulent users can switch roles as time goes on. 4. The theoretical part of this work, Section 2.1.1 and Appendix C.1, seems to be needless and trivial. It is evident that in evasion attacks, accuracy will not be affected if clean graphs are used for testing.

Questions

1. On homophilous graphs, a high-pass filter is likely to decrease accuracy. Figure 8 indicates that adversarial training transforms the GNN into a high-pass filter. However, why does this not result in a trade-off between accuracy and robustness? Can you provide some possible analyses? 2. Can the authors provide more examples of information leakage in the transductive setting? As far as I know, this is not common. Overall, this paper is intriguing and attempts to address some of the key challenges in graph adversarial attacks. If the authors address my concerns well, I am willing to consider raising the score.

Rating

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

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

yes

Reviewer fJD87/10 · confidence 4/52023-07-05

Summary

This paper shows fundamental limitations of transductive adversarial training of GNNs for node classification and advocates that adversarial training of GNNs should be in an inductive setting. Therefore, this paper proposes a more flexible message-passing scheme using learnable diffusion to gain robust representations under adversarial training, and uses additional local constraints for structural perturbation and an efficient attack (LR-BCD) that contains both local and global perturbations.

Strengths

This paper studies an important problem on the GNN adversarial robustness and provides some interesting observations and insights on this topic. In particular, the analysis and discussions of theoretical and empirical limitations of adversarial training in GNN transductive learning are insightful and useful for the community. The proposed robust diffusion framework for adversarial training appears efficient in a fully inductive GNN setting with flexible GNN architectures and refined local/global topology perturbations.

Weaknesses

1. The technical contributions of the work seem a bit limited. It appears the proposed method is a combination of existing works, although it is a simple yet effective solution. The claimed novelty lies in the handling of the new local constraints with LR-BCD, but it does not seem to be technically different from PR-BCD. In addition, there is a lack of theoretical justification why the new projection imposes sensible global and local constraints. 2. Given the limited technical contributions of the methodologies, I would like to see more extensive experimental evaluations of the proposed framework. More evaluations on other homophilic and heterophilic datasets with more powerful adversarial attacks (e.g., adaptive attacks [1]) under inductive setting should be conducted to demonstrate the universality of the proposed methods. In addition, the comparison with other robust GNN models (e.g., ProGNN [2], GRAND [3], GARNET [4]). [1] Felix Mujkanovic, Simon Geisler, Stephan Günnemann, and Aleksandar Bojchevski. Are defenses for graph neural networks robust? NeurIPS’22. [2] Wei Jin, Yao Ma, Xiaorui Liu, Xianfeng Tang, Suhang Wang, and Jiliang Tang. Graph structure learning for robust graph neural networks. SIGKDD’22. [3] Wenzheng Feng, Jie Zhang, Yuxiao Dong, Yu Han, Huanbo Luan, Qian Xu, Qiang Yang, Evgeny Kharlamov, and Jie Tang. Graph random neural network for semi-supervised learning on graphs. ICML’21. [4] Deng et al. GARNET: Reduced-Rank Topology Learning for Robust and Scalable Graph Neural Networks. LoG'22. 3. Some minor points: * It does not seem the equation (2) correctly reflects the principles of self-training, because theta’ appears predefined in the expression, while it is actually trained from labelled data. * In Equation-(5), the sum operation in the constraint is not well defined. * The authors may want to make the title more specific to better reflect the focus of this work, avoiding potential misleading and vagueness.

Questions

1. Please clarify how L/PR-BCD corresponds to adaptive attacks. 2. Please justify how the additional local constraints could preserve graph semantics.

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

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

N/A

Reviewer zNuu7/10 · confidence 3/52023-07-08

Summary

This paper focuses on adversarial training techniques for Graph Neural Networks (GNNs) in the context of inductive semi-supervised learning. It addresses the shortcomings of previous approaches in graph learning: perfect robustness is possible for transductive learning setting under evasion attack. The authors propose a more adaptable GNN model equipped with a learnable robust diffusion mechanism, capable of accommodating adversarial perturbations. Furthermore, the paper introduces a novel method for attacking GNNs by perturbing their structures, which consider local constraints in global attack settings. The study demonstrates the efficacy of adversarial training as a defensive strategy for GNNs and offers valuable insights into the field of adversarial learning for GNNs.

Strengths

- The paper provides a comprehensive summary of existing research and knowledge on adversarial training techniques for Graph Neural Networks (GNNs). - The authors elaborate the limitations of previous approaches in graph learning settings, particularly emphasizing the challenges associated with transductive learning settings for evasion attacks. It also introduces a more realistic fully inductive setting. - The authors propose a new adversarial training method combining a learnable graph diffusion, with discussion on the interpretability of the learned robust representation. Based on the revealed importance of local constrained, they further proposed an attack method LR-BCD, extending previous PR-BCD attack with a projection incorporating local constraints. - The paper presents extensive empirical evidence demonstrating the robustness of the proposed adversarial training techniques and the proposed attack method. - The paper offers interesting insights for the community regarding the role of local constraints in adversarial training for GNNs.

Weaknesses

- Limited to evasion attack, should clarify this in a early stage of presentation - Better presentation if the authors shorten section 2.1 and section 2.1.1, which introduce commonly-used settings and quite obvious propositions, and add more model framework details and training details in the main text. Minor issue: - the abbreviation GPRGNN is quite confusing while just appears with no explanation in the intro section

Questions

- Curious how to apply adversarial training under poisoning attack settings. it is okay if this is out of the scope of this paper. - How would the adversarial trained model generalize to other types of attacks? Are the proposed robust diffusion model able to defend other types of attacks? E.g., gray-box/black-box attacks, poisoning attacks.

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 ZTec7/10 · confidence 4/52023-07-11

Summary

The paper aims to fix adversarial training in graph neural networks (with respect to perturbations) by revealing limitations in prior work (from theoretical insufficiencies to experiment flaws), proposing a novel attack scheme and a superior architecture (robust diffusion) that makes adversarial training a leading defense method against such attacks.

Strengths

- Paper is well written and easy to understand. - Strong theoretical analysis supported by elegant intuition and solid experimental results. - High-quality claim is supported by code artifacts.

Weaknesses

- Readers unfamiliar with adversarial perturbations may have a hard time understanding the paper, which may warrant a short paragraph explaining the threat model for both numerical data and graph-structured data.

Questions

- What is the intuition behind the "diffusion" nomenclature?

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

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The authors have not addressed the potential negative societal impact of robust diffusion, and a short analysis in discussion may help.

Reviewer bw8r2023-08-17

Thanks for the rebuttal

I thank the authors for the reply. Most of my concerns are well addressed, and I raise my score. I sincerely suggest that the author include the answer to the first weakness in the paper. However, I still think the theoretical part (Section 2.1.1 and Appendix C1) is redundant. It is valuable to point out that there can be information leakage when using clean structure of test nodes during adversarial training. But reasoning on a clean, memorized graph is obviously not affected by perturbations and does not require a mathematical form to illustrate this point.

Authorsrebuttal2023-08-19

Thanks for the reply

We want to thank the reviewer for the positive reply and for raising the score! ## I: Including the answer to W1 in the paper >I sincerely suggest that the author include the answer to the first weakness in the paper. We plan to include our answer to W1 in the revised paper and hence, better incorporate the motivation for our model choice and its link to the discussion in §2. In particular, we plan the following changes highlighted in **bold** to the paper to make these points more clear: 1) In the *GNN architecture* paragraph in the Introduction, we plan to expand on the statement "approximate any spectral graph filter" starting at line 34/35 and adapt the text as follows: >*(ii) GNN Architectures:* [...] Instead, we propose to use flexible message-passing schemes based on learnable diffusion. **The main motivation behind this choice is the ability** to approximate any spectral graph filter. **Thus, adversarial training may choose the most robust filter that achieves a competitive training loss.** **Thereby, we** significantly improve robustness compared to previous work, while yielding an interpretable message-passing scheme, **and making the evaluation bias in the transductive setting (i) apparent.** 2) In Section 2.1.2, we plan to adapt the sentence on lines 162/163 as follows: > [...] In other words, it [GPRGNN] finds a trivial solution to this learning setting, **making its limitations evident** similar to the theoretic solution discussed in Section 2.1.1. [...] 3) In Section 2.1.2, we plan to exchange the last two sentences on lines 169-172 with the following text (bold writing omitted): > We find that flexible architectures like GPRGNN combined with adversarial training (see robust diffusion in subsequent Section 3) not only solve the transductive setting to perfection showcasing their adaptability, but also learn to robustly generalize to unseen nodes far better than previously studied models - while considering an unbiased inductive setting (see Section 5). 4) In the beginning of Section 3 *Robust Diffusion*, we plan to adapt the first two sentences and swap their order as follows: > We propose to use learnable graph diffusion models **able to approximate any graph filter** in conjunction with adversarial training to obtain a *robust diffusion*. **The key motivation is to** use more flexible GNN architectures for adversarial training than previous studies. [...] We are looking forward to further suggestions, if the proposed changes do not fully address the reviewer's point. ## II: On Section 2.1.1 >However, I still think the theoretical part (Section 2.1.1 and Appendix C1) is redundant. It is valuable to point out that there can be information leakage when using clean structure of test nodes during adversarial training. But reasoning on a clean, memorized graph is obviously not affected by perturbations and does not require a mathematical form to illustrate this point. We want to thank the reviewer for recognizing the value of pointing out the information leakage in Section 2.1.1. We took a lot of effort to present the conceptual differences between the inductive and transductive learning setting in Section 2.1 and 2.1.1 in a way, to make the pitfalls for robustness evaluations as intuitively clear as possible. While it is evident that a memorized graph is not affected by perturbations, the mathematical form serves other purposes: * Concerning adversarial training, it showcases how memorization is a valid and optimal solution to the so far studied saddle-point problem arising in the related adversarial training works [A,O]. Thus, it provides important intuition for the *perfect robustness* result presented in Section 2.1.2. * It further provides intuition, as to why information leakage is **not** an issue in the inductive setting. Thus, we think §2.1.1 is important to aid the conceptual shift from a transductive towards an inductive setting. This shift is crucial for further progress in the field of adversarial training for graph learning (and relevant for the graph robustness field in general). For a revised version of §2.1.1, we plan to better work out (and focus) these intuitions and lessen emphasis on Proposition 1.

Reviewer fJD82023-08-19

I appreciate the authors' responses to my comments, and they appear convincing. Therefore I would like to raise my rating to "accept". BTW, the new title looks much more appealing to me.

Authorsrebuttal2023-08-21

We thank the reviewer for the positive reply and for raising the score! As mentioned, we will change the title to *Adversarial Training for Graph Neural Networks: Pitfalls, Solutions and New Directions* in the revised paper version.

Reviewer ZTec2023-08-19

Thank you for your clarifications. My score remains the same since my questions did not constitute a significant factor in a score change.

Authorsrebuttal2023-08-21

We thank the reviewer for acknowledging our rebuttal!

Reviewer zNuu2023-08-21

Thank you for your response

Thank you for your detailed response. My concerns are well addressed. Overall I think the paper is of good quality and I will keep recommending the acceptance of the paper.

Authorsrebuttal2023-08-21

We want to thank the reviewer for acknowledging our rebuttal and the positive reply!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC