acceptodds
Under review as a conference paper at ICLR 2027

EchoGhost: A Frequency-Structural Backdoor Attack and Spectral-Counterfactual Defense for Graph Neural Networks

Abstract

Graph neural networks (GNNs) are increasingly deployed in security-sensitive domains, yet their message-passing mechanism can amplify hidden training-time perturbations and convert small poisoned patterns into systematic inference-time failures. Existing graph backdoor attacks typically rely on subgraph triggers, node injection, or local feature manipulation, which can leave detectable topological or feature-space artifacts. This article proposes EchoGhost, a frequency-structural backdoor framework for node classification in GNNs. EchoGhost embeds a universal echo trigger into selected node features through controlled frequency-band modulation and reinforces the trigger by injecting similarity-guided ghost edges among low-activity nodes. The resulting poisoned graph causes a trained GNN to map trigger-bearing nodes to an attacker-chosen target class while preserving clean-task utility. Unlike purely structural triggers, EchoGhost couples spectral feature evidence with graph propagation, creating a dual-domain attack surface that neither feature anomaly detection nor topology-only inspection fully captures. To study and mitigate this threat, we introduce a Spectral-Counterfactual Trigger Verification defense that identifies suspicious training nodes using robust spectral-energy statistics and verifies causal trigger dependence through counterfactual de-echoing. We sanitize confirmed nodes by removing them from the training set, then retrain from scratch on the original clean graph. We evaluate EchoGhost and the proposed defense on four benchmark graph datasets, such as CoauthorPhysics, CoauthorCS, Citeseer, and Cora, using three GNN architectures: GCN, GAT, and GraphSAGE, reporting all results as mean standard deviation over five random seeds. Across all 12 configurations, EchoGhost achieves an average attack success rate (ASR) of 98.04% with an average clean accuracy drop of 7.22%. The defense mechanism reduces the average ASR to 14.52% (an 85.21% relative reduction) while maintaining a false positive rate of 0.01% across all settings.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Graph neural networks, backdoor attacks, graph security, spectral analysis, frequency-domain triggers, node classification, counterfactual defense, data poisoning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_H8A1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents EchoGhost, a backdoor attack on graph neural networks that combines a frequency-domain feature trigger with structural modifications. The authors recognize that many existing graph backdoor attacks modify graph structure or raw features in ways that can be detected through structural statistics or feature analysis. EchoGhost injects a shared perturbation into selected frequency components of node features and adds similarity-based edges among nodes identified as having low activity. These ghost edges are intended to reinforce the trigger during message passing. The paper also proposes a defense that screens nodes using spectral statistics, checks whether their predictions depend on the trigger through counterfactual feature modification, and retrains the model after removing confirmed suspicious nodes from the training set. The evaluation covers four datasets and three GNN architectures across five random seeds, reporting high attack success rates and substantial reductions in attack success after defense.

Critical strengths
  • The paper is generally well organized, with a clear description of the attack and defense pipelines that makes the overall approach easy to follow.

  • Studying both the attack and a corresponding defense gives the work a broader scope and provides an opportunity to examine how the proposed trigger can be identified and mitigated.

  • The evaluation covers four datasets, three GNN architectures, and five random seeds, providing useful evidence of attack effectiveness across different settings.

Critical weaknesses
  • The benefit of ghost edges is not well supported quantitatively. Although the authors state that ghost edges improve clean-accuracy retention, Table 6 shows a larger clean-accuracy drop for the full method than for echo-only on all four datasets, with little or no improvement in attack success.

  • The defense evaluation does not isolate the contribution of the proposed detection method. Retraining uses the original clean features and graph structure, and Table 10 reports substantial defense success even when no poisoned nodes are identified.

  • The evidence for attack stealthiness is limited. Table 1 reports clean-accuracy drops exceeding 21 percentage points for GAT on two datasets, while the graph-level detectability metric only evaluates a specific threshold-based detector. These results do not sufficiently establish that the attack remains difficult to detect while preserving normal model performance.

Questions for authors
  • What benefit do ghost edges provide over echo-only? Table 6 appears inconsistent with the claim that they improve clean-accuracy retention. Please identify settings where they consistently improve attack success, preserve clean accuracy, or make the attack harder to defend against.

  • Could you evaluate the defense using only the poisoned data and the clean reference set available under the stated threat model? A separate clean-retraining baseline without detection would help establish how much of the reported improvement comes from restoring the original data rather than identifying poisoned nodes.

  • How does EchoGhost compare with prior attacks at similar clean-accuracy drops, and how well does it withstand established defenses such as DShield [1] and RigBD [2]? These comparisons should use comparable access to clean data and retraining budgets, with any necessary differences clearly explained.

[1] Yu, H., Ma, C., Wan, X., Wang, J., Xiang, T., Shen, M., & Liu, X. (2025). DShield: Defending against Backdoor Attacks on Graph Neural Networks via Discrepancy Learning. In NDSS.

[2] Zhang, Z., Lin, M., Xu, J., Wu, Z., Dai, E., & Wang, S. (2025, May). Robustness inspired graph backdoor defense. In International Conference on Learning Representations (Vol. 2025, pp. 1958-1984).

Reviewer_XS4qOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

EchoGhost proposes a dirty-label backdoor attack on GNN node classification. It injects a universal perturbation into the discrete Fourier transform of selected node features, relabels poisoned nodes to a target class, and adds similarity-guided edges between low-feature-norm nodes. The proposed defense identifies suspicious nodes using frequency-band energy and changes in target-class confidence after band removal, then excludes confirmed nodes from supervision and retrains on the original graph. Experiments on four datasets and three GNN architectures report 98.04% ASR and 14.52% ASR after defense.

Critical strengths
  1. The attack is simple, requires no new nodes or trigger generator, and achieves high ASR across three GNN architectures.
  2. Experiments cover four datasets, with ablations and sensitivity analyses. Key defense limitations are explicitly acknowledged.
Critical weaknesses
  1. Table 6 contradicts the claimed clean-accuracy benefit of ghost edges, as Echo-only achieves lower CAD on all four datasets. In Table 7, the five Cora/GCN CAD values are 4.30, 1.60, 9.30, 3.20, and 3.20. Their mean is 4.32 rather than the reported 3.56.
  2. The defense retrains on the original clean graph, but provides no restoration-only baseline to establish the detector’s contribution. Table 10 reports approximately 73.8% DSR even when no nodes are confirmed as poisoned. Appendix F acknowledges this issue, but a restoration-only baseline and comparisons using the same data and retraining procedure are needed to show the detector’s added value.
Questions for authors
  1. Could you please clarify what clean data, including labels, the defender has access to? Also, does detection lead to better mitigation than restoration-only retraining?
  2. Could you please explain the discrepancies in Tables 6, 7, and 10?
  3. Could you please explain why Echo-only achieves better CAD?
  4. Some icons in Figure 1 contain unclear details. Could you please clarify them?
Reviewer_Rw6BOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes EchoGhost, a backdoor attack on GNN node classification. It adds a fixed perturbation to three DFT frequency bins of selected training nodes’ features, flips their labels to a target class, and injects similarity-guided “ghost” edges among low-norm nodes. The authors also propose a paired Spectral–Counterfactual Trigger Verification defense: spectral-energy screening, then counterfactual “de-echoing” of candidates, then retraining. Experiments on Cora, Citeseer, CoauthorCS, and CoauthorPhysics with GCN, GAT, and GraphSAGE report 98.04% average ASR, and 14.52% ASR after defense with 0.01% FPR.

Critical strengths
  1. The authors run the official SBA, GTA, and UGBA code under their own protocol instead of relying only on numbers copied from other papers.
  2. The paper pairs the attack with a defense and openly acknowledges some limitations.
Critical weaknesses
  1. Results for the same configuration differ across tables.
  2. The evaluation is limited. Large-scale benchmarks such as ogbn-arxiv and ogbn-products [1] or Flickr [2], used by UGBA [3], are missing. Please run OGB datset.
  3. The presentation has many errors. Some are factual, some are internal contradictions, and some are typos (e.g., duplicate references):

[1] Hu et al., Open Graph Benchmark: Datasets for Machine Learning on Graphs, NeurIPS 2020 [2] Zeng et al., GraphSAINT: Graph Sampling Based Inductive Learning Method, ICLR 2020 [3] Dai et al., Unnoticeable Backdoor Attacks on Graph Neural Networks, WWW 2023 [4] Zhang et al., Robustness Inspired Graph Backdoor Defense, ICLR 2025 [5] Yu et al., DShield: Defending against Backdoor Attacks on GNNs via Discrepancy Learning, NDSS 2025

Questions for authors

See weaknesses. Please adress the scalability issues, and robustness to adaptive attacks.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.