acceptodds
Withdrawn from ICLR 2027

Adversarially Robust Graph Neural Networks via Adaptive Multi-Frequency Filtering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph Neural Networks (GNNs) have achieved remarkable success in various graph learning tasks. However, their vulnerability to adversarial attacks limits their trustworthy deployment in real-world applications. In recent years, spectral-based defense methods have provided a new perspective for enhancing the adversarial robustness of GNNs. However, existing methods typically adopt a single, fixed frequency response, which makes them difficult to adapt to spectral variations induced by adversarial attacks. Moreover, adversarial perturbations are often localized, whereas existing spectral filtering methods mainly rely on global spectral responses and lack the capability to capture localized graph variations, which compromises their adversarial robustness. To address these issues, we propose AdaMF, a dual-perspective robustness framework that captures localized graph variations and adaptively adjusts the spectral responses. Specifically, we construct learnable low-, high-, and mid-frequency filter bases with adaptive weighting coefficients to characterize diverse spectral shifts caused by adversarial perturbations, enabling flexible global spectral adaptation. Furthermore, we introduce framelet-based multi-resolution filtering to learn representations across different graph scales, enabling the model to preserve local structural details and global semantic information. Additionally, we employ prototype contrastive learning guided by pseudo-labels to further enhance the robustness of node representations. Extensive experiments on real-world datasets demonstrate that the proposed method achieves superior robustness against various adversarial attacks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Graph Neural Networks, Adversarial Robustness, Adaptive Multi-Frequency Filtering, Spectral Filtering

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_e9cyOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes AdaMF, a robust GNN for node classification under evasion attacks. It has three components: (1) a hybrid spectral filter that combines low-, high- and mid-pass filters with learnable weights on the probability simplex, (2) a framelet-based multi-scale filtering branch to capture localized variations, and (3) a pseudo-label-guided prototype contrastive loss between two views augmented with 1% feature and structure perturbations. The paper provides bounds on the sensitivity of the hybrid filter to eigenvalue and Laplacian perturbations. Experiments on Cora, Citeseer, Pubmed and ogbn-arxiv against ten baselines and seven evasion attacks (five graph modification and two node injection attacks) show that AdaMF achieves the best accuracy in almost all settings, with further results under varying perturbation strength, a 10% target-node setting, ablations and a sensitivity study.

Critical strengths

The method is simple and well motivated by the observation that attacks shift the whole spectrum rather than one band, and the evaluation covers several datasets (including ogbn-arxiv), ten baselines and seven attacks, with consistent improvements and an ablation of each component.

Critical weaknesses

1. The evaluation does not convincingly show adversarial robustness.

  • Mismatch between the attacks and the theory. For the graph modification attacks, the paper perturbs 2% of the feature dimensions of 1% target nodes, i.e., only features are changed and the graph structure is untouched. However, the motivation and Theorem 2 concern perturbations of the graph Laplacian, which do not change under feature-only attacks, so the theory does not apply to most of the reported results. Classic structure attacks such as Metattack, Nettack or PGD, which are discussed in the related work, are not evaluated.
  • No adaptive attacks. The paper does not state whether the attacks are generated against AdaMF itself or transferred from other models. Prior work (e.g., Mujkanovic et al., 2022) shows that many GNN defenses break under attacks designed against them, so evaluating only against generic attacks likely overestimates robustness.
  • The attacks appear weak. Many models barely degrade under attack; for example, AutoGCN on Cora drops only from 85.40 to 85.03 under GC-RWCS. AdaMF sometimes performs better under attack than on clean data (Cora RWCS: 88.35 vs. 88.17 clean; Pubmed AFGSM: 90.00 vs. 88.56 clean), which suggests the attacks are not strong enough to test robustness.
  • Unfair comparison. AdaMF is trained with 1% feature and structure perturbations as augmentation, which acts as a mild form of adversarial training, while the baselines are not. It is unclear how much of the gain comes from this augmentation rather than from the proposed filters.
  • Incomplete statistics. The paper reports paired t-tests over 10 runs, but Table 1 gives no standard deviations. AFGSM uses only 50 fixed target nodes, so accuracy per run moves in steps of 2%, and differences between methods may not be meaningful.

2. Limited novelty and some inconsistencies.

  • Limited novelty. Adaptive combinations of multiple frequency filters (e.g., AutoGCN, and HyperAIM cited by the authors) and framelet-based GNNs (Zheng et al., 2021) already exist. The paper mainly combines these components and applies them to adversarial robustness.
  • Inconsistencies. ξ = 1 is used in all experiments, but the sensitivity analysis only covers ξ ∈ {0.1, 0.3, 0.5, 0.7, 0.9}, which excludes the default value. On ogbn-arxiv, some baselines have unusually low clean accuracy (GAD 33.66, Elastic 25.86), which suggests they may not be properly tuned.
Questions for authors
  1. How does AdaMF perform under structure attacks such as Metattack, Nettack and PGD, which are the setting the motivation and Theorem 2 address?
  2. Are the attacks generated against AdaMF itself? Can you report results under adaptive white-box attacks that target the full AdaMF model?
  3. Can you explain why AdaMF sometimes performs better under attack than on clean data, and report results with stronger attack budgets on all methods?
  4. How do the baselines perform when trained with the same 1% perturbation augmentation, and how does AdaMF perform without it?
  5. Can you report standard deviations for Table 1 and the results of the t-tests per setting?
  6. Why does the sensitivity study exclude ξ = 1, and how were the baselines tuned on ogbn-arxiv?
Reviewer_sQL6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Adversarial attacks can degrade the node classification performance of graph neural networks by perturbing graph structures and node features, posing a challenge to their adversarial robustness. To mitigate this problem, the paper proposes AdaMF, which combines learnable multi-frequency filtering, framelet-based multiscale analysis, and prototype contrastive learning to enhance the robustness of GNNs against evasion attacks. The paper also theoretically analyzes the stability of the filtering mechanism under perturbations and empirically evaluates the proposed framework on multiple benchmark datasets under various adversarial attack scenarios. The reported results suggest that AdaMF can improve adversarial robustness, and ablation studies are conducted to assess the contributions of its individual components.

Critical strengths

1.The combination of frequency and multiscale information is reasonably motivated. 2.The empirical results are promising, and the ablation studies support the contributions of individual components.

Critical weaknesses

1.The incremental contribution over existing methods such as AutoGCN and ASF-GCN remains unclear. The ablations do not isolate the benefit of learned frequency weights from other components. 2.The attack protocol is insufficiently specified. It is unclear whether attacks are generated directly against AdaMF, limiting the conclusions about robustness against defense-adaptive attacks.

Questions for authors

1.What are the main technical differences from existing adaptive spectral methods? How much do learned frequency weights improve over fixed weights? 2.Are the frequency weights fixed after training? How does the model adapt to different perturbations at inference time? 3.Are attacks generated directly against AdaMF? Please clarify attacker knowledge, perturbation budgets, and attack procedures.

Reviewer_PvYHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AdaMF, a framework for improving the adversarial robustness of graph neural networks in node classification. It combines learnable low-, high-, and mid-frequency filters with a framelet branch for multiscale representation learning. Features from both branches are fused for classification, while confidence-weighted class prototypes support contrastive learning across feature- and topology-perturbed views. The paper characterizes the spectral response family and derives perturbation bounds for the frequency-filtering branch. Experiments compare AdaMF with ten baselines across four citation datasets, using seven attacks overall, and report improved accuracy under the evaluated evasion attacks.

Critical strengths
  1. The integration of spectral filtering, multiscale framelets, and prototype contrastive learning provides a coherent architecture for investigating complementary sources of robustness in graph learning.
  2. The reported results are promising, particularly on PubMed under GZOO.
  3. The evaluation includes component ablations, perturbation-strength sweeps, and an extreme-budget setting with ten-run statistics, providing useful evidence beyond a single attack configuration.
Critical weaknesses
  1. At ε=0.1, Equation (6) gives |g_M(0)|=0.8 and |g_M(1)|=0.01, so the filter favors endpoint frequencies in magnitude. This seems inconsistent with the mid-pass interpretation, and the ε values used in experiments are not reported.
  2. Table 1 reports 73.01% accuracy for GCN on Cora under VoteAttack, but Figure 4 shows approximately 57% at perturbation strength 1. I could not find an explanation for this difference, which needs clarification.
  3. Removing entire branches shows that these components help, but does not establish why. Comparisons with fixed frequency mixtures and a single framelet scale would help determine whether adaptivity and multiple scales actually explain the gains.
  4. The stability bounds describe the filter family, but it remains unclear whether training learns less sensitive filters. The condition α=β used for the special cancellation result is neither enforced nor checked experimentally.
Questions for authors

See the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.