acceptodds
Under review as a conference paper at ICLR 2027

Stealth Still Leaves Traces: Exposing and Reducing Trigger Footprints in Graph Backdoor Attacks

Abstract

Graph neural networks (GNNs) have been widely applied to various graph learning tasks, yet they remain vulnerable to backdoor attacks. By implanting specific triggers, an adversary can cause a model to misclassify nodes with attached triggers into a designated target class while preserving normal predictive performance on clean nodes. Although existing generative subgraph-based backdoor attacks have continuously improved the stealthiness of individual triggers, we find that the generated triggers still leave detectable multilevel trigger footprints. These footprints manifest as anomalous trigger node features, recurring local structures, and abnormally high feature similarity between corresponding nodes across different triggers. Based on these findings, we propose footprint reducing graph backdoor attack, which reduces these footprints by tailoring trigger generation to different types of node features, adopting a minimal single-node trigger to limit structural perturbation, and constraining excessive feature similarity across different triggers. Extensive experiments across multiple real-world graph datasets and representative GNN architectures demonstrate that the proposed attack achieves high attack success rates, exhibits robustness against multiple graph backdoor defenses, and maintains stable classification performance on clean nodes. The source code is available at https://anonymous.4open.science/r/FRGBA-702C.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor Attack, Graph Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_ySoNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors compare triggers produced by existing graph backdoor attacks and find that nodes in corresponding positions often have unusually similar features. They use this observation, together with local attachment patterns, to build MTFD. Their attack, FRGBA, constructs each trigger from the victim node's features and limits similarity across triggers while preserving similarity to the victim. On four citation networks, tests with GCN, GraphSAGE, and GAT show that MTFD suppresses the baseline attacks, whereas FRGBA retains average attack success rates above 90% with clean accuracy close to the corresponding clean models.

Critical strengths
  1. Figure 2 identifies unusually high similarity between corresponding nodes across triggers, and Tables 2 and 7 show that MTFD can use this pattern to detect or suppress the evaluated attacks.
  2. FRGBA retains 90.06%-96.70% average ASR under MTFD across four datasets and three GNNs, while the ablation in Figure 5 supports the contribution of its cross-trigger similarity constraint.
Critical weaknesses
  1. FRGBA's single-node triggers still satisfy CN(v,t)=0 and Exit(v,t)=0, so the reported results do not establish that it reduces the structural footprints defined in Section 3.2.
  2. Missing baseline ASR and CA in Tables 3 and 6, unclear handling of trigger sizes and label permissions, and unspecified MTFD thresholds and loss-weight updates make the comparisons difficult to assess, especially given the drop in ASR from 94.55% to 48.83% when p_t changes in Table 10.
Questions for authors
  1. What fraction of FRGBA triggers is detected at each stage of MTFD, given that they still satisfy CN=0 and Exit=0?
  2. How do the baseline attacks compare under the defenses in Table 3 when trigger sizes, poisoning budgets, and label permissions are explicitly accounted for?
  3. Which data and rules were used to set MTFD's thresholds, p_v, p_t, and loss weights, and what are the standard deviations of the main results?
  4. How does removing victim-feature fusion alone affect ASR and detection when normalization and both similarity constraints are retained?
6 comments

We thank the reviewer for this question. We agree that the single-node leaf triggers adopted by FRGBA still satisfy both CN = 0 and Exit = 0, and therefore cannot evade the first-stage structural screening of MTFD. It is important to note that the first stage primarily identifies structural candidates and does not imply that these candidates will ultimately be confirmed as triggers.

One possible way to alter these structural conditions would be to introduce additional edges between trigger nodes and other nodes in the original graph. However, such modifications would increase the number of injected edges and introduce greater perturbations to the original graph topology. Therefore, FRGBA adopts a minimal single-node trigger to limit structural modifications, rather than introducing additional connections solely to evade the first-stage screening.

By introducing a cross-trigger similarity constraint, FRGBA reduces abnormally high feature similarities between corresponding nodes across different triggers. Consequently, although FRGBA triggers enter the first-stage candidate set, they are less likely to be confirmed during the second-stage similarity verification. Thus, FRGBA's resistance to MTFD primarily comes from weakening similarity-based detection in the second stage, rather than evading structural screening in the first stage. We have revised the manuscript to clarify that the single-node trigger design aims to limit structural perturbations rather than evade structural detection.

To quantitatively demonstrate this mechanism, we provide additional stage-wise detection results for FRGBA across four datasets. Specifically, we report the proportion of true triggers included in the first-stage candidate set and the proportion finally identified after second-stage verification. Both proportions use the total number of injected triggers as the denominator.

Table: Stage-wise detection of FRGBA triggers by MTFD (identified triggers / total triggers).

DatasetSplitStage 1: Structural CandidatesStage 2: Confirmed Triggers
CoraTraining21/210/21
Testing292/2920/292
CiteSeerTraining26/260/26
Testing359/3590/359
PubMedTraining157/1570/157
Testing2,125/2,1250/2,125
ogbn-arxivTraining1,354/1,35440/1,354
Testing18,248/18,2480/18,248

We thank the reviewer for raising this concern regarding the fairness of the experimental comparisons. To provide a more comprehensive evaluation of FRGBA's robustness against existing defenses, we conducted additional experiments with GTA, UGBA, DPGBA, and CNGBA under five defense methods: Prune, Prune+LD, OD, RIGBD, and LoSplit.

Regarding the experimental settings, to control for the effects of trigger size and poisoning budget, we followed the same setup as Table 2, using single-node triggers and a uniform poisoning budget of 1% for all attack methods. Regarding label-modification permissions, we retained the original label settings of each attack method to preserve their respective attack configurations and avoid altering the characteristics of the baseline methods through additional label constraints.

Table: Attack performance under five existing defenses on Cora (ASR (%) / CA (%)).

AttackPrunePrune+LDODRIGBDLoSplit
GTA25.20 / 72.1539.52 / 57.4376.35 / 81.9757.81 / 82.8436.80 / 83.21
UGBA86.92 / 84.4488.89 / 76.3064.71 / 85.9375.86 / 82.9658.33 / 84.07
DPGBA45.31 / 82.2232.29 / 70.0681.08 / 83.4642.70 / 81.6039.93 / 83.33
CNGBA89.76 / 86.0589.05 / 86.0576.78 / 83.02100.00 / 85.801.79 / 84.94
FRGBA96.53 / 82.1096.15 / 79.1491.32 / 83.0896.18 / 82.5994.45 / 83.70

The additional results on Cora demonstrate that FRGBA maintains strong attack effectiveness under existing defenses. Across the 20 attack-defense combinations involving the four baseline attacks, FRGBA achieves a higher ASR in 19 cases. Specifically, under Prune, Prune+LD, OD, and LoSplit, FRGBA outperforms the best-performing baseline by 6.77, 7.10, 10.24, and 36.12 percentage points, respectively. The only exception is under RIGBD, where FRGBA achieves a slightly lower ASR than CNGBA. These additional results further support the relative robustness of FRGBA under different defense settings.

We thank the reviewer for the questions regarding parameter selection. All similarity thresholds in our method are computed based on the corresponding reference feature similarity distributions, while the associated percentile parameters are selected based on preliminary experimental observations and empirical analysis.

Specifically, for MTFD, our analysis in Section 3.3 reveals abnormally high feature similarities among trigger nodes across different triggers, which differ noticeably from the similarity distribution of clean-node pairs. Based on this observation, MTFD adopts the p-th percentile of the clean-node pairwise similarity distribution as the similarity detection threshold. In addition, each candidate node is required to have at least M high-similarity matches to reduce false positives caused by incidental similarities among benign nodes.

To further examine cross-trigger similarity across datasets, we report three statistics: CC 99th, the 99th percentile of pairwise cosine similarities between clean-node pairs; TT 1st, the 1st percentile of similarities between trigger-node pairs at the same structural positions across different triggers; and TT above CC 99th, the proportion of such pairs exceeding CC 99th.

Table 1: Cross-trigger feature similarity statistics across four datasets.

DatasetAttackCC 99thTT 1stTT above CC 99th
CoraGTA0.59680.97991.0000
UGBA0.46070.8286
DPGBA0.99991.0000
CNGBA0.91301.0000
FRGBA0.19100.0000
CiteSeerGTA0.67650.96861.0000
UGBA0.99991.0000
DPGBA0.98191.0000
CNGBA0.85711.0000
FRGBA0.27920.0000
PubMedGTA0.71960.98631.0000
UGBA0.30210.7173
DPGBA1.00001.0000
CNGBA0.93371.0000
FRGBA0.04480.0000
ogbn-arxivGTA0.94980.99961.0000
UGBA0.92950.9815
DPGBA1.00001.0000
CNGBA0.99831.0000
FRGBA0.68060.0013

For FRGBA, the victim–trigger similarity lower bound τv\tau_v and the cross-trigger similarity upper bound τt\tau_t are computed based on the feature similarity distributions of real edges between clean nodes and pairwise similarities among clean reference nodes, respectively. The percentile parameters pvp_v and ptp_t are empirically selected through preliminary experiments to balance attack effectiveness and the strength of the similarity constraints. The attack-loss weight is kept fixed, while the two similarity-constraint weights are dynamically adjusted according to their constraint violations to balance attack effectiveness and trigger stealthiness.

We thank the reviewer for the concern regarding the reliability of the experimental results. Table 2 reports the average attack performance of FRGBA across three target GNN architectures (GCN, GraphSAGE, and GAT), while Table 5 in Appendix H.1 provides the corresponding results for each architecture to illustrate performance variations across different models. The original experiments were conducted over five independent runs. To further evaluate the stability of the results, we conducted five new independent runs under the same experimental settings and report the mean and standard deviation of ASR and CA for FRGBA on each of the three target architectures. These results provide further insight into FRGBA's performance and run-to-run variability across different model architectures.

Table 2: FRGBA performance over five independent runs (mean ± standard deviation; ASR (%) and CA (%)).

DatasetDefenseMetricGCNGraphSAGEGAT
CoraNoneASR99.65 ± 0.3089.58 ± 1.0498.96 ± 0.52
CA83.45 ± 0.4382.34 ± 0.8584.20 ± 1.75
MTFDASR99.48 ± 0.0087.85 ± 1.5097.57 ± 1.31
CA83.21 ± 0.2183.21 ± 0.7783.82 ± 0.43
CiteSeerNoneASR99.75 ± 0.2289.48 ± 1.5499.87 ± 0.22
CA74.80 ± 0.3571.39 ± 0.3174.60 ± 1.06
MTFDASR99.62 ± 0.0089.22 ± 1.16100.00 ± 0.00
CA74.90 ± 0.3571.19 ± 0.6374.10 ± 0.30
PubMedNoneASR97.94 ± 0.5689.51 ± 0.4696.85 ± 0.27
CA85.20 ± 0.0885.91 ± 0.1184.02 ± 0.13
MTFDASR97.00 ± 0.5288.43 ± 0.5796.27 ± 0.65
CA84.91 ± 0.0885.80 ± 0.3183.80 ± 0.21
ogbn-arxivNoneASR90.41 ± 0.4680.14 ± 0.6999.51 ± 0.07
CA65.03 ± 0.3765.91 ± 0.4264.85 ± 0.09
MTFDASR90.28 ± 0.0380.24 ± 0.4299.57 ± 0.04
CA65.53 ± 0.0865.84 ± 0.0965.03 ± 0.07
Reviewer_sgUJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FRGBA, a graph backdoor attack that aims to reduce detectable footprints and constrain excessive feature similarity across different triggers. FRGBA consists of three main components: trigger generation, trigger stealthiness enhancement, and bilevel optimization. Extensive experiments across multiple datasets and GNN architectures demonstrate that the proposed attack achieves high attack success rates.

Critical strengths
  1. This paper explores the trigger footprint, which provides a useful perspective for studying the detectability of graph backdoor attacks.
  2. The experiments are extensive, encompassing across three GNN architectures (i.e., GCN, GraphSAGE, and GAT) and four citation-network datasets.
Critical weaknesses
  1. The three main components proposed in this paper are very similar to those in existing papers (e.g., GTA), and I am concerned that the paper lacks sufficient novelty.
  2. The proposed method cannot achieve satisfactory attack performance. As shown in Table 2, without defense, the proposed method achieves lower ASRs than existing methods across all datasets, which achieve nearly 100% ASRs.
  3. The proposed attack remains effective against several existing defense methods, but its performance degrades significantly under some defenses. As shown in Table 3, for example, its attack success rate drops from 90.46% to 76.48% under OD on ogbn-arxiv.
  4. The current experimental results do not sufficiently support the claim of stealth. It is recommended that the authors include a more direct assessment of trigger detectability to verify whether FRGBA actually reduces detectable traces.
Questions for authors

See weaknesses.

4 comments

We appreciate the reviewer's comment and welcome the opportunity to clarify the contributions of our work. We acknowledge that FRGBA shares certain similarities in its overall framework with existing graph backdoor attacks, such as GTA. Indeed, trigger generation, stealthiness enhancement, and alternating optimization of the trigger generator and surrogate model are common components of existing generative graph backdoor attacks, including GTA, UGBA, DPGBA, and CNGBA. Therefore, the novelty of our work does not lie in proposing an entirely new general framework, but rather in addressing an overlooked problem through specifically designed trigger construction and stealthiness enhancement mechanisms.

Specifically, existing methods primarily focus on the stealthiness of individual triggers. For example, UGBA employs local feature-similarity constraints, DPGBA adopts distribution-preserving generation, and CNGBA combines single-node triggers with conditional diffusion. However, we find that even with these adaptive generation strategies, nodes occupying the same structural positions across different triggers may still exhibit abnormally high feature similarities, resulting in detectable recurring cross-trigger node footprints.

To address this issue, FRGBA reconsiders trigger stealthiness from a cross-trigger perspective. On the one hand, we directly use victim node features as the basis for trigger construction and adopt different generation strategies for different feature types to reduce feature anomalies and enhance the diversity of generated triggers. On the other hand, we introduce a cross-trigger similarity constraint guided by the feature similarity distribution of clean nodes to explicitly suppress abnormally high feature similarities among different triggers. We further use MTFD to examine the detectability of these recurring cross-trigger patterns, and conduct ablation and cross-dataset experiments to evaluate the effectiveness and generalizability of FRGBA.

Therefore, the primary contribution of this work lies in identifying previously underexplored recurring cross-trigger feature footprints in existing generative graph backdoor attacks and proposing targeted mechanisms to suppress them, rather than introducing an entirely new general-purpose attack framework.

We thank the reviewer for the comments on the attack performance. We acknowledge that some existing methods can achieve ASRs close to 100% on certain datasets. However, such performance is not consistent across all datasets. For example, on ogbn-arxiv, DPGBA and CNGBA achieve ASRs of only 30.64% and 69.29%, respectively, whereas FRGBA achieves 90.46%. Moreover, FRGBA achieves ASRs of 96.70%, 96.96%, and 95.10% on Cora, CiteSeer, and PubMed, respectively, maintaining an ASR above 90% across all four datasets. These results demonstrate its relatively stable attack performance across datasets.

It is worth noting that a high ASR in the absence of defenses primarily reflects attack effectiveness and does not, by itself, establish the stealthiness of the generated triggers. As analyzed in Section 3, although some existing methods achieve high attack success rates, their generated triggers may still exhibit recurring cross-trigger feature patterns, leaving detectable feature footprints.

Therefore, FRGBA does not simply aim to maximize the ASR without defense, but rather seeks to maintain high attack effectiveness while further reducing detectable trigger footprints. To this end, we employ feature-aware trigger generation strategies tailored to different feature types and introduce a cross-trigger similarity constraint to balance attack effectiveness and stealthiness.

Although FRGBA achieves slightly lower ASRs than the best-performing baselines on some datasets, it maintains high attack success rates across all four datasets while effectively suppressing the recurring cross-trigger feature patterns identified in our analysis. These results demonstrate a balance between attack effectiveness and trigger stealthiness.

We thank the reviewer for the comments on FRGBA's robustness against existing defenses. We acknowledge that different defense mechanisms have varying impacts on FRGBA, and some defenses can indeed lead to a decrease in attack success rate. For example, on ogbn-arxiv, OD reduces FRGBA's ASR from 90.46% to 76.48%.

However, we believe that FRGBA's robustness should also be evaluated in comparison with existing attacks under the same defense settings. To this end, we conducted additional experiments evaluating GTA, UGBA, DPGBA, and CNGBA under five existing defenses.

Table: Attack performance under five existing defenses on Cora (ASR (%) / CA (%)).

AttackPrunePrune+LDODRIGBDLoSplit
GTA25.20 / 72.1539.52 / 57.4376.35 / 81.9757.81 / 82.8436.80 / 83.21
UGBA86.92 / 84.4488.89 / 76.3064.71 / 85.9375.86 / 82.9658.33 / 84.07
DPGBA45.31 / 82.2232.29 / 70.0681.08 / 83.4642.70 / 81.6039.93 / 83.33
CNGBA89.76 / 86.0589.05 / 86.0576.78 / 83.02100.00 / 85.801.79 / 84.94
FRGBA96.53 / 82.1096.15 / 79.1491.32 / 83.0896.18 / 82.5994.45 / 83.70

Taking Cora as an example, across the 20 attack-defense combinations involving the four baseline attacks, FRGBA achieves a higher ASR in 19 cases. Specifically, under Prune, Prune+LD, OD, and LoSplit, FRGBA outperforms the best-performing baseline by 6.77, 7.10, 10.24, and 36.12 percentage points, respectively. The only exception is under RIGBD, where FRGBA achieves a slightly lower ASR than CNGBA.

These results demonstrate that although some defenses can reduce FRGBA's attack effectiveness, it generally maintains stronger attack performance than the existing baseline methods under the evaluated defense settings on Cora.

We thank the reviewer for the suggestion regarding trigger stealthiness evaluation. To more directly assess FRGBA's ability to suppress detectable trigger footprints, we conducted additional cross-dataset analyses of trigger feature similarities.

Specifically, CC 99th denotes the 99th percentile of pairwise cosine similarities between clean-node pairs and serves as the reference threshold. TT 1st denotes the 1st percentile of cosine similarities between node pairs occupying the same structural positions across different triggers. TT above CC 99th represents the proportion of such cross-trigger node pairs whose similarities exceed the clean-node reference threshold. These metrics allow us to quantitatively examine abnormally high cross-trigger similarity and evaluate how effectively FRGBA suppresses recurring feature patterns.

Table 1: Cross-trigger feature similarity statistics across four datasets.

DatasetAttackCC 99thTT 1stTT above CC 99th
CoraGTA0.59680.97991.0000
UGBA0.46070.8286
DPGBA0.99991.0000
CNGBA0.91301.0000
FRGBA0.19100.0000
CiteSeerGTA0.67650.96861.0000
UGBA0.99991.0000
DPGBA0.98191.0000
CNGBA0.85711.0000
FRGBA0.27920.0000
PubMedGTA0.71960.98631.0000
UGBA0.30210.7173
DPGBA1.00001.0000
CNGBA0.93371.0000
FRGBA0.04480.0000
ogbn-arxivGTA0.94980.99961.0000
UGBA0.92950.9815
DPGBA1.00001.0000
CNGBA0.99831.0000
FRGBA0.68060.0013

The results show that existing attacks generally exhibit high cross-trigger similarities across datasets. In contrast, the proportion of cross-trigger pairs exceeding CC 99th is only 0%, 0%, 0%, and 0.13% for FRGBA on Cora, CiteSeer, PubMed, and ogbn-arxiv, respectively. These results indicate that FRGBA effectively suppresses recurring cross-trigger feature similarities beyond Cora.

Furthermore, we report the detection rates of FRGBA at the two stages of MTFD, including the proportion of true triggers entering the first-stage candidate set and the proportion finally identified after second-stage verification.

Table 2: Stage-wise detection of FRGBA triggers by MTFD (identified triggers / total triggers).

DatasetSplitStage 1: Structural CandidatesStage 2: Confirmed Triggers
CoraTraining21/210/21
Testing292/2920/292
CiteSeerTraining26/260/26
Testing359/3590/359
PubMedTraining157/1570/157
Testing2,125/2,1250/2,125
ogbn-arxivTraining1,354/1,35440/1,354
Testing18,248/18,2480/18,248

The results show that all evaluated FRGBA triggers enter the first-stage structural candidate set. However, none are confirmed in the second stage on Cora, CiteSeer, or PubMed, and only 40 out of 1,354 training triggers are confirmed on ogbn-arxiv, with none identified in its testing split.

These additional experiments provide a more direct assessment of FRGBA's stealthiness from both cross-trigger feature similarity and actual detection performance perspectives, supporting our explanation that FRGBA reduces recurring node footprints rather than evading structural screening.

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

The paper identifies three major footprints across triggers in existing graph backdoor attacks. It then introduces MTFD to exploit structural and recurring feature footprints and FRGBA to generate single-node triggers from victim features. Experiments cover four citation datasets and three GNN architectures.

Critical strengths
  1. The cross-trigger similarity analysis identifies a useful detection signal, and MTFD demonstrates its effectiveness against the evaluated attacks.
  2. FRGBA injects a single-node trigger while demonstrating strong empirical attack performance.
Critical weaknesses
  1. Method and implementation: There is a gap between the structural analysis and the method design: a single attached trigger node still satisfies the identified structural conditions, i.e., CN = Exit = 0. Moreover, the implemented losses also differ from Eqs. (5)–(6), while additional clean-model filtering and outer-loop sampling are insufficiently described.

  2. Evaluation: Tables 3 and 6 evaluate only FRGBA under defenses, without corresponding attack-baseline results. Table 7 describes node detection, but the code computes edge-level metrics for MTFD and node-level metrics for LoSplit. These issues make the comparative robustness difficult to assess.

Questions for authors
  1. In roleDDD.py, both regularizers sum the mean and maximum of squared violations; (L_{vt}) also uses threshold-based normalization. Which formulation produced the results? Please reconcile these differences with Eqs. (5)–(6) and clarify the outer-loop sampling.

  2. How were Table 7’s node-detection metrics calculated? rolecos_defense.py counts edges, whereas LoSplit counts nodes. Please clarify the denominators, report consistently defined metrics, and provide attack-baseline comparisons under existing defenses.

3 comments

First, regarding the fact that a single-node trigger still satisfies CN = 0 and Exit = 0, we agree with the reviewer's observation. A single-node leaf trigger indeed satisfies both structural conditions and is therefore included in the candidate set of MTFD's first-stage screening. Similarly, subgraph triggers connected to victim nodes through a single edge may also satisfy the same structural conditions. Altering these patterns by introducing additional connections between triggers and the original graph would introduce more edge perturbations. Therefore, FRGBA does not aim to evade the first-stage structural screening. Instead, it employs cross-trigger similarity constraints to reduce the likelihood that candidate triggers are confirmed during the second-stage verification. We have revised the manuscript to clarify that the single-node trigger design aims to limit structural perturbations rather than evade structural detection.

To further examine this mechanism, we report the proportion of true triggers included in MTFD's first-stage candidate set and the proportion finally confirmed after second-stage verification.

Table: Stage-wise detection of FRGBA triggers by MTFD (identified triggers / total triggers).

DatasetSplitStage 1: Structural CandidatesStage 2: Confirmed Triggers
CoraTraining21/210/21
Testing292/2920/292
CiteSeerTraining26/260/26
Testing359/3590/359
PubMedTraining157/1570/157
Testing2,125/2,1250/2,125
ogbn-arxivTraining1,354/1,35440/1,354
Testing18,248/18,2480/18,248

The results confirm that all evaluated triggers enter the first-stage candidate set, whereas very few are finally identified in the second stage. This distinction supports our explanation that FRGBA reduces recurring node footprints rather than avoiding structural screening.

Second, regarding the discrepancies between Eqs. (5)–(6) in the original manuscript and the implementation, we clarify that the implementation in roleDDD.py was correct and was used to obtain all reported experimental results. The discrepancies arose from the mathematical formulations in the original manuscript, which did not fully reflect the implemented loss computations. We have therefore revised the corresponding equations and descriptions without modifying the code.

Specifically, both regularizers combine the mean and maximum of squared constraint violations. The mean term captures overall violations, while the maximum term emphasizes the most severe individual violation. The victim–trigger similarity loss additionally normalizes its constraint violations by the larger of the absolute similarity threshold and a small positive constant (epsilon), reducing the influence of threshold scale and ensuring numerical stability. In contrast, the cross-trigger similarity loss does not involve additional normalization. The revised equations thus accurately reflect the implemented loss functions.

Finally, regarding clean-model filtering and additional node sampling in the outer optimization, we have clarified these procedures in the revised manuscript. We first train a reference GNN on the clean training graph and use it to predict the labels of unlabeled nodes. Only nodes predicted to belong to non-target classes are retained as candidates.

At each outer iteration, a subset of these candidate nodes is randomly sampled and combined with the fixed poisoned node set to form the outer optimization set. These additional sampled nodes are used exclusively for the outer optimization of the trigger generator; they are neither included in the poisoned training set nor involved in the inner optimization of the surrogate model. By evaluating the attack loss on these nodes, the generator is exposed to a more diverse range of victim-node features, improving its adaptability.

These revisions clarify FRGBA's relationship with MTFD's two-stage detection mechanism, align the mathematical formulations with the implementation, and detail the reference-model filtering and node-sampling procedures.

We thank the reviewer for pointing out the issues in the experimental evaluation. We agree that describing Precision, Recall, and FPR for different detection targets uniformly as node-level detection metrics in the original manuscript was inaccurate and could lead to the misconception that results obtained using different evaluation units are directly comparable.

In fact, we calculate these metrics according to the objects ultimately detected or removed by each defense method. For edge-pruning methods, including Prune, Prune+LD, and MTFD, we adopt edge-level evaluation, treating injected malicious edges (including victim–trigger attachment edges and edges within triggers) as positive samples and benign edges as negative samples. Accordingly, correctly removed malicious edges are counted as true positives (TP), incorrectly removed benign edges as false positives (FP), retained malicious edges as false negatives (FN), and retained benign edges as true negatives (TN). For node-detection methods, including OD, RIGBD, and LoSplit, we adopt node-level evaluation, treating poisoned nodes as positive samples and benign nodes as negative samples, and calculate TP, FP, FN, and TN based on whether the nodes are correctly identified.

Under both evaluation settings, Precision, Recall, and FPR are calculated using the same formulas: TP/(TP+FP), TP/(TP+FN), and FP/(FP+TN), respectively, where all counts are based on the corresponding evaluation units (edges or nodes). To eliminate the ambiguity in the original manuscript, we have revised the text preceding Table 7 to explicitly distinguish between edge-level and node-level evaluations and clarify the corresponding positive and negative sample definitions and evaluation protocols. The results in Table 7 remain unchanged, as the revision only clarifies the evaluation targets and metric calculation procedures.

Furthermore, to address the lack of attack-baseline comparisons under existing defenses, we conducted additional experiments evaluating GTA, UGBA, DPGBA, and CNGBA against five defense methods: Prune, Prune+LD, OD, RIGBD, and LoSplit.

Table: Attack performance under five existing defenses on Cora (ASR (%) / CA (%)).

AttackPrunePrune+LDODRIGBDLoSplit
GTA25.20 / 72.1539.52 / 57.4376.35 / 81.9757.81 / 82.8436.80 / 83.21
UGBA86.92 / 84.4488.89 / 76.3064.71 / 85.9375.86 / 82.9658.33 / 84.07
DPGBA45.31 / 82.2232.29 / 70.0681.08 / 83.4642.70 / 81.6039.93 / 83.33
CNGBA89.76 / 86.0589.05 / 86.0576.78 / 83.02100.00 / 85.801.79 / 84.94
FRGBA96.53 / 82.1096.15 / 79.1491.32 / 83.0896.18 / 82.5994.45 / 83.70

Across the 20 attack-defense combinations involving the four baseline attacks, FRGBA achieves a higher ASR in 19 cases. Specifically, under Prune, Prune+LD, OD, and LoSplit, FRGBA outperforms the best-performing baseline by 6.77, 7.10, 10.24, and 36.12 percentage points, respectively. The only exception is under RIGBD, where FRGBA achieves a slightly lower ASR than CNGBA. These additional results enable a more comprehensive comparison of FRGBA's robustness against existing attacks under the same defense settings.

Thank you for the clarifications. Q1 is addressed, but the structural concern in W1 remains: single-node triggers still satisfy CN=Exit=0. W2 and Q2 are only partly addressed, since Table 7 retains different evaluation units, and the added baseline comparisons cover only Cora. I will therefore keep my score unchanged.

Reviewer_8jZuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies detectable patterns left by graph backdoor triggers in node classification. It identifies unusual feature values, recurring attachment structures, and high similarity between triggers generated for different nodes. Based on this, it proposes FRGBA, which attaches one victim-dependent trigger node and constrains feature similarity across triggers. The attack performs well, particularly against the paper’s proposed defense, MTFD.

Critical strengths
  • Looking across multiple triggers, rather than judging each trigger alone, is a useful angle.
  • The attack maintains high success rates under several defenses while generally preserving clean accuracy.
Critical weaknesses
  • The feature anomaly in Table 1 seems largely due to baseline triggers not following the datasets’ row-normalization preprocessing. This makes that finding less convincing as a fundamental weakness of existing attacks. It would be useful to see what happens when the baseline triggers are normalized.
  • The largest improvement is against MTFD, a defense designed around the footprints identified in this paper. For the other defenses, Table 3 reports only FRGBA, so it is hard to tell whether it is actually more robust than the competing attacks.
  • I am not convinced the single-node design reduces the structural footprint described in Section 3. A single attached leaf still satisfies both CN = 0 and Exit = 0, which are the paper’s own structural screening conditions. Also, all attacks use single-node triggers in Table 2, so this contribution is not isolated there.
  • Most of the footprint analysis and the main ablation are on Cora. The evaluation covers four datasets, so the broader claim about graph backdoors feels a bit ahead of the evidence.
Questions for authors
  1. If the baseline trigger features are passed through the same row-normalization step as clean nodes, how much of the feature-footprint result remains?
  2. How do the competing attacks perform under the five existing defenses in Table 3, using the same experimental setup?
  3. What evidence shows that the single-node design reduces structural detectability under the proposed CN and Exit criteria?
  4. Do the cross-trigger similarity patterns and ablation trends also hold on CiteSeer, PubMed, and ogbn-arxiv?
4 comments

We thank the reviewer for this helpful suggestion. We further applied post-hoc row-wise normalization to the trigger features generated by UGBA, DPGBA, and CNGBA, and re-evaluated their attack performance using three target GNN architectures on Cora. GTA was excluded because its triggers contain numerous negative feature values, which cannot be corrected by the same row-wise scaling applied to nonnegative clean features.

Table: Attack performance after trigger feature row-wise normalization on Cora (ASR (%) / CA (%)).

AttackGCNGraphSAGEGAT
UGBA27.37 / 84.6724.02 / 83.1120.11 / 85.70
DPGBA41.67 / 80.0030.73 / 83.3324.48 / 85.19
CNGBA9.42 / 86.328.31 / 86.8811.36 / 86.14

The results show that all three attacks experience substantial reductions in ASR after row-wise normalization compared with their respective original results. Among them, DPGBA exhibits the smallest performance drop, followed by UGBA, while CNGBA experiences the largest reduction. This trend is consistent with the magnitude of row-sum deviations reported in the original Table 1, with DPGBA exhibiting the smallest deviations, followed by UGBA and CNGBA.

We acknowledge that the row-sum anomalies are partly attributable to inconsistencies in feature preprocessing and should not be regarded as an inherent weakness of existing attacks. However, row-wise normalization does not eliminate all feature footprints. For any nonzero, nonnegative feature vector, row-wise normalization is equivalent to positive scalar scaling, which preserves cosine similarity. Consequently, pairwise cosine similarities between trigger nodes, and thus the cross-trigger similarity patterns measured by this metric, remain unchanged after normalization.

We thank the reviewer for this helpful suggestion. To provide a more comprehensive evaluation of FRGBA's robustness against existing defenses, we conducted additional experiments with GTA, UGBA, DPGBA, and CNGBA under five defense methods: Prune, Prune+LD, OD, RIGBD, and LoSplit. All experiments followed the same experimental setup as Table 3.

Table: Attack performance under five existing defenses on Cora (ASR (%) / CA (%)).

AttackPrunePrune+LDODRIGBDLoSplit
GTA25.20 / 72.1539.52 / 57.4376.35 / 81.9757.81 / 82.8436.80 / 83.21
UGBA86.92 / 84.4488.89 / 76.3064.71 / 85.9375.86 / 82.9658.33 / 84.07
DPGBA45.31 / 82.2232.29 / 70.0681.08 / 83.4642.70 / 81.6039.93 / 83.33
CNGBA89.76 / 86.0589.05 / 86.0576.78 / 83.02100.00 / 85.801.79 / 84.94
FRGBA96.53 / 82.1096.15 / 79.1491.32 / 83.0896.18 / 82.5994.45 / 83.70

The results show that FRGBA maintains high attack effectiveness under all five defenses on the Cora dataset. Across the 20 attack-defense combinations involving the four baseline attacks, FRGBA achieves a higher ASR in 19 cases. Specifically, under Prune, Prune+LD, OD, and LoSplit, FRGBA outperforms the best-performing baseline by 6.77, 7.10, 10.24, and 36.12 percentage points, respectively. The only exception occurs under RIGBD, where CNGBA achieves a slightly higher ASR than FRGBA. These results demonstrate that FRGBA generally maintains stronger attack effectiveness than existing baseline attacks under the five defense methods.

We thank the reviewer for raising this important concern. We agree that a single trigger node attached as a leaf to a victim node satisfies both CN(v,t)=0CN(v,t)=0 and Exit(v,t)=0Exit(v,t)=0, and therefore enters the candidate set in the first-stage structural screening of MTFD.

Moreover, this structural pattern is not unique to single-node triggers. Under the subgraph-trigger setting considered in this work, a trigger subgraph is connected to the victim node through a single edge. Such an attachment pattern can also satisfy the CN and Exit screening conditions. Altering these structural patterns would require introducing additional connections between the trigger subgraph and the original graph, resulting in more edge modifications and greater perturbation to the original graph topology.

We therefore clarify that the primary purpose of FRGBA's single-node trigger design is to reduce the number of injected nodes and edges, thereby limiting structural modifications to the original graph, rather than to evade the structural screening criteria defined by CN and Exit. We have revised the manuscript to distinguish between reducing the magnitude of structural modifications and evading structural detection, as these are two different aspects. Regarding Table 2, all attack methods adopt single-node triggers to ensure a controlled comparison under the same trigger-size constraint.

More importantly, FRGBA's resistance to MTFD primarily stems from its cross-trigger similarity constraint. By reducing feature similarity among different triggers, FRGBA weakens the recurring similarity patterns exploited in MTFD's second-stage verification. Thus, FRGBA does not evade the first-stage structural screening but reduces the likelihood of trigger confirmation in the subsequent stage.

To further investigate this mechanism, we provide additional statistics on the structural and node-level footprints of FRGBA across the training and testing splits of four datasets.

Table: Stage-wise detection of FRGBA triggers by MTFD (identified triggers / total triggers).

DatasetSplitStage 1: Structural CandidatesStage 2: Confirmed Triggers
CoraTraining21/210/21
Testing292/2920/292
CiteSeerTraining26/260/26
Testing359/3590/359
PubMedTraining157/1570/157
Testing2,125/2,1250/2,125
ogbn-arxivTraining1,354/1,35440/1,354
Testing18,248/18,2480/18,248

The results show that all evaluated trigger instances satisfy the structural screening conditions. In contrast, no triggers are confirmed in the second stage on Cora, CiteSeer, or PubMed in either split. On ogbn-arxiv, only 40 out of 1,354 training triggers are confirmed, while none are identified in the testing split. These results further illustrate the distinction between structural candidate screening and subsequent node-level footprint identification, supporting our explanation of how FRGBA reduces trigger detectability during the second-stage verification.

We thank the reviewer for pointing out that the footprint analysis was primarily conducted on Cora, which may limit the generalizability of our findings. We agree with this concern. To examine whether the cross-trigger similarity patterns and the effectiveness of FRGBA's key components generalize across datasets, we conducted additional similarity analyses and ablation studies on CiteSeer, PubMed, and ogbn-arxiv.

Table 1: Cross-trigger feature similarity statistics. CC 99th is the 99th percentile of clean-node pairwise cosine similarity; TT 1st is the 1st percentile of same-position node similarities across different triggers; TT above CC 99th is the fraction of such pairs exceeding CC 99th.

DatasetAttackCC 99thTT 1stTT above CC 99th
CoraGTA0.59680.97991.0000
UGBA0.46070.8286
DPGBA0.99991.0000
CNGBA0.91301.0000
FRGBA0.19100.0000
CiteSeerGTA0.67650.96861.0000
UGBA0.99991.0000
DPGBA0.98191.0000
CNGBA0.85711.0000
FRGBA0.27920.0000
PubMedGTA0.71960.98631.0000
UGBA0.30210.7173
DPGBA1.00001.0000
CNGBA0.93371.0000
FRGBA0.04480.0000
ogbn-arxivGTA0.94980.99961.0000
UGBA0.92950.9815
DPGBA1.00001.0000
CNGBA0.99831.0000
FRGBA0.68060.0013

The results show that baseline attacks exhibit consistently high cross-trigger similarities across datasets. For most baselines, all such pairs exceed the clean-node reference threshold. Even for UGBA, the proportions are 82.86%, 71.73%, and 98.15% on Cora, PubMed, and ogbn-arxiv, respectively. In contrast, FRGBA reduces these proportions to 0%, 0%, 0%, and 0.13% on Cora, CiteSeer, PubMed, and ogbn-arxiv. Thus, excessive cross-trigger similarity is not specific to Cora.

Table 2: Cross-dataset FRGBA ablations (ASR (%) / CA (%)). FRGBA\L: remove cross-trigger similarity constraint; FRGBA\S: remove sparse normalized generation; FRGBA\B: remove both; FRGBA\V: remove victim-feature fusion; FRGBA\LV: remove fusion and the cross-trigger constraint. -- denotes not evaluated.

DatasetAttackNoneMTFDNormalization
CoraFRGBA96.70 / 83.2196.01 / 83.8295.49 / 84.20
FRGBA\L85.07 / 83.585.04 / 84.4483.86 / 83.33
FRGBA\S100.00 / 83.33100.00 / 84.205.04 / 84.19
FRGBA\B99.48 / 84.195.21 / 84.075.04 / 83.46
CiteSeerFRGBA96.96 / 74.1096.70 / 73.6094.92 / 73.19
FRGBA\L90.11 / 73.403.55 / 74.0089.19 / 73.19
FRGBA\S99.87 / 74.5099.62 / 74.103.25 / 74.20
FRGBA\B99.24 / 74.103.55 / 73.503.52 / 73.50
PubMedFRGBA95.10 / 85.0594.55 / 84.9794.13 / 84.87
FRGBA\L91.35 / 84.937.79 / 85.1990.99 / 84.73
FRGBA\S99.00 / 84.787.82 / 85.087.51 / 84.69
FRGBA\B97.58 / 84.787.71 / 84.867.54 / 84.92
ogbn-arxivFRGBA90.46 / 65.1090.06 / 65.51--
FRGBA\L88.92 / 65.5583.09 / 65.61--
FRGBA\V87.12 / 64.3880.15 / 64.60--
FRGBA\LV82.30 / 65.2426.23 / 65.41--

The ablations on CiteSeer and PubMed show trends largely consistent with Cora. Removing the cross-trigger similarity constraint reduces ASR under MTFD from 96.70% and 94.55% to 3.55% and 7.79%, respectively. Removing the sparse normalized design reduces ASR under Normalization from 94.92% and 94.13% to 3.25% and 7.51%.

On ogbn-arxiv, removing the cross-trigger constraint causes a smaller MTFD ASR decrease (90.06% to 83.09%). In the sparse-feature setting, the learnable fusion weight may allow generator outputs to dominate victim features, increasing the risk of recurring trigger similarities when the constraint is removed. In contrast, continuous-feature generation directly preserves victim features and adds bounded perturbations, retaining victim-specific information even without the constraint. Removing victim-feature fusion alone yields 80.15% ASR under MTFD, whereas removing both fusion and the constraint reduces ASR to 26.23%, highlighting their complementary roles. Normalization is not evaluated on ogbn-arxiv because it uses continuous rather than sparse row-normalized features. Overall, the results support the effectiveness of FRGBA's key components beyond Cora while revealing dataset-dependent differences.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.