acceptodds
Under review as a conference paper at ICLR 2027

Hiding in Plain Sight: Reliable Backdoor Attack Against Vertical Federated Graph Learning via Reconstruction Consistency

Abstract

Vertical federated graph neural networks (VFGNNs) enable collaborative learning from distributed graph data, but remain vulnerable to backdoor attacks. Unlike Euclidean data, graph data contain interconnected nodes, and message propagation may dilute trigger features, making backdoor attack methods developed for conventional vertical federated learning less effective when applied directly. In this paper, we propose Graph Camouflage (GraphCamo), a backdoor attack that combines structure-guided trigger learning and masked reconstruction-based backdoor consolidation. Its core idea is to select trigger support nodes based on graph topology and use masked feature reconstruction to preserve relevant information during graph aggregation. Experiments on four datasets and three GNN models show that, with only four anchor nodes, GraphCamo achieves an average attack success rate of 99.24% and an average standard deviation across configurations of 0.75%, while maintaining competitive main-task accuracy. Further experiments demonstrate its resistance to different defense mechanisms, while analyses of embeddings and gradients indicate that GraphCamo is less detectable.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Vertical Federated Graph Learning, Backdoor Attack, Node Classification

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 3 reviews, above 27% of the others.
Reviewer_JzsgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GraphCamo studies a malicious participant in vertical federated graph learning who knows four target-class training nodes and controls its own features and local model. It uses personalized PageRank to select training nodes for perturbation and trains with reconstruction losses on clean and triggered features while leaving labels unchanged. At test time, it applies the learned perturbations throughout each victim's two-hop neighborhood. Experiments on four citation graphs and three encoders report an average attack success rate of 99.24%, with comparisons against seven attacks and four defenses.

Critical strengths
  1. Table 1 reports high attack success across twelve dataset and encoder combinations over five runs, with less variation than BVG in the reported results.
  2. Figure 6 and the tests across target classes support the contribution of the reconstruction losses and show that success is not confined to one target label.
Critical weaknesses
  1. The test attack in Appendix C.3 perturbs every node within two hops, whereas training selects a subset, but the numerical perturbation bound and affected node counts are not reported, making attack budgets difficult to compare.
  2. Table 1 lacks a no-attack accuracy baseline, and smaller changes between consecutive epochs do not establish that the attack is difficult to detect, leaving the accuracy-preservation and stealth claims insufficiently supported.
Questions for authors
  1. What perturbation bound and node counts are used during training and testing, and are these budgets matched across attacks?
  2. How much does attack success change if test-time perturbations use the same PageRank-based selection as training?
  3. What are the no-attack accuracy and the detection rates at a fixed false-positive rate for GraphCamo and the baselines?
  4. How do attack success and main-task accuracy change across defense strengths beyond the single settings in Appendix C.6?
3 comments
AuthorsOpenReview ↗

Response to Reviewer Jzsg

Thank you for your careful reading and suggestions. Your questions helped us clarify the attack settings and identify where further evidence was needed.

Perturbation Budgets During Training and Testing

GraphCamo uses the same coordinate-wise L∞ bound during training and testing: ε=0.1 for GCN/Cora and GCN/CiteSeer, and 0.08 otherwise. Training uses four known target-class anchors and the middle 50% of PPR-ranked neighbors at each hop, without neighbor labels. Testing modifies one target and its full one- and two-hop neighborhoods. Across 12 equally weighted configurations, the mean numbers of nodes modified during training and testing are 24.83 and 16.37, respectively, or 0.528% and 0.421% of the graph. Observed maxima are 62 and 313. Four anchors thus describe label knowledge, not all modifications. These small modified-node fractions provide further evidence of GraphCamo's effectiveness with limited node modifications.

Consistency of PPR Selection Between Training and Testing

Keeping the model and trigger fixed, we apply the training-time middle-50% rule at test time. Mean ASR is 95.64%, versus the paper's full-neighborhood reference of 99.24%, while 9 of 12 configurations exceed 97%. The mean number of modified nodes falls from 16.38 to 9.11, a 44.35% reduction. Most configurations thus retain high attack effectiveness with fewer modifications. Training selection limits interference around fixed anchors. Full-neighborhood testing aims to cover message-passing paths around diverse targets and reduce dependence on their neighbor rankings.

No-Attack Accuracy and Localization at Controlled Empirical FPR

To assess normal-task utility, we compare MTA from supplementary no-attack training with the paper's untriggered GraphCamo MTA across four datasets and three models. Equally weighting the 12 configurations gives 69.04% for the no-attack models and 68.07% for GraphCamo, a mean decrease of only 0.97 percentage points. These close values indicate that GraphCamo retains most normal classification capability after learning the attack, with a small overall utility cost. At the configuration level, GraphCamo has higher MTA than the no-attack reference in four cases, showing that normal accuracy does not decline in every setting. The added no-attack reference directly quantifies this performance cost and supports GraphCamo's preservation of main-task utility within the evaluated scope.

For localization, we score uploaded representations using kNN distance, class-based Mahalanobis distance and Isolation Forest. Without an extra split, thresholds from untriggered scores cap empirical node FPR at 1% or 5% on that same set. Success requires nodes directly modified in that attack to constitute over half the post-attack alarm set. The three detectors' mean success rates are 7.93%, 10.86% and 8.57% at 1%, and 3.03%, 4.98% and 0.68% at 5%, respectively. They usually fail to produce alarms dominated by modified nodes, supporting limited localizability in this offline setting. This is not conventional target-node detection. The added detector results cover GraphCamo.

Sensitivity to Defense Strength

We test four strengths per defense, averaging equally across 12 configurations.

For ISO noise ratios of 0.05, 0.1, 0.2 and 0.4, ASR is 96.46%, 97.89%, 93.62% and 92.83%. MTA is 67.06%, 66.00%, 65.74% and 61.94%, respectively. Stronger noise weakens the attack but costs utility. ASR remains above 92% at the strongest setting.

For DP-SGD at the same noise ratios, ASR is 97.33%, 96.48%, 95.73% and 90.88%, while MTA is 65.25%, 63.50%, 62.08% and 60.54%. Both decline as noise increases, but the attack persists.

For GC retention fractions of 0.5, 0.2, 0.1 and 0.05, ASR is 97.65%, 95.83%, 96.81% and 92.66%. MTA is 67.23%, 65.98%, 64.83% and 64.50%. Stronger compression reduces utility without consistently blocking the attack.

For ANP thresholds of 0.0025, 0.005, 0.01 and 0.02, ASR is 99.76%, 93.60%, 99.91% and 99.65%. MTA is 46.55%, 44.86%, 34.16% and 24.52%. Pruning substantially harms utility without a consistent ASR reduction.

Results at the default settings are taken from the paper: 0.1 for ISO, DP-SGD and GC, and 0.01 for ANP.

Thank you for focusing the discussion on budget clarity, consistent node selection and the effects of stronger defenses. The added experiments make the associated performance and utility trade-offs more concrete. If this evidence addresses your experimental concerns, we hope it will strengthen your confidence in the work and support a more favorable assessment. We remain available to clarify any protocol details.

Thank you for the detailed response and the additional experiments. I appreciate the clearer description of the experimental settings and the effort to quantify the performance and utility trade-offs. These additions provide useful context for the concerns raised in my review.

AuthorsOpenReview ↗

Response to Reviewer Jzsg

Thank you for your encouraging response and for the thoughtful feedback you have provided throughout this discussion. We will incorporate the clarifications and additional experimental results guided by your suggestions into the final version of the paper. Your insightful questions and constructive suggestions clearly reflect your deep expertise in this field. This paper represents a summer of sustained effort, and your constructive engagement has been a great encouragement to us.

If our responses and additional evidence have adequately addressed your concerns, we would sincerely appreciate your consideration of a more favorable assessment. This is very important for us. Thank you again for your time.

Reviewer_M1s6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes GraphCamo, a backdoor attack by a malicious passive party in vertical federated graph learning. It selects training trigger-support nodes using local personalized PageRank, learns shared hop-specific feature perturbations, and uses separate masked reconstruction branches for clean and triggered features. An auxiliary-gradient projection reduces conflict with the main training update. The paper provides restricted analyses of graph-induced attenuation and feature-direction recoverability. Experiments across four citation datasets and three GNN architectures report high attack success, low variability, clean-task accuracy, defense results, and additional participant-count and target-label analyses.

Critical strengths
  1. The method addresses a graph-specific obstacle rather than only transferring a tabular attack. Combining neighborhood-aware support selection with clean and triggered reconstruction gives a coherent mechanism for maintaining trigger information during message passing while limiting interference with useful representations.
  2. The reported attack success is strong across multiple configurations. Table 1 reports near-saturated ASR with substantially lower variability than several baselines. The ablations, participant-count study, and all-target-label evaluations provide useful evidence that the result is not confined to one GNN or one target class. The theoretical analyses are explicitly restricted and do not claim a general ASR guarantee.
Critical weaknesses
  1. The effective attack budget and inference setting are not adequately characterized. Four labeled anchors are not equivalent to four modified nodes: training perturbs selected neighbors, and Appendix C.3 states that inference evaluates each attacked node separately while modifying every node in its local zero-, one-, and two-hop neighborhoods. This is a materially larger, query-specific intervention than the anchor count alone suggests. The paper does not report the distribution of modified-node counts or sufficiently document matched perturbation budgets across baselines. I also could not locate a numerical value for the perturbation bound ϵ\epsilon introduced in Eq. (7) in the supplied manuscript. Without this information, the practical significance and comparability of near-100% ASR remain uncertain.
  2. The evidence does not yet establish the stronger clean-utility and stealthiness claims. Table 1 compares main-task accuracy among attacks but does not provide matched no-attack MTA values. Moreover, MTA on an untriggered graph does not measure collateral errors on unselected nodes in the graph modified for an actual attack. The consecutive-epoch embedding/gradient MSE in Figures 4–5 and Appendix D.1 measures temporal smoothness, not detectability by a calibrated detector. The four defense configurations are useful stress tests, but do not substitute for clean-utility comparisons or detection measurements. These gaps directly affect the paper’s central positioning around reliability, preservation, and camouflage.
Questions for authors
  1. Budget clarification: Please report the exact perturbation bound, feature-validity constraints, and the mean, median, and upper-tail number of modified nodes at training and inference. Are baseline budgets matched on those quantities, and are the reported ASRs all measured using the same single-target graph-perturbation protocol?

  2. Evidence affecting the recommendation: Can you provide paired no-attack MTA and accuracy on unselected nodes within the same attacked graph? Is there any detector evaluation at a controlled false-positive rate supporting the camouflage claim? Without such evidence, please distinguish temporal stability from demonstrated low detectability.

2 comments
AuthorsOpenReview ↗

Response to Reviewer M1s6

Thank you for examining the assumptions and evidence behind our claims so closely. Your comments helped us distinguish label knowledge from actual modifications and temporal stability from localization.

Perturbation Constraints, Feature Validity, Modification Counts, and Evaluation Protocol

GraphCamo adds continuous perturbations to the malicious party's local features with a coordinate-wise L∞ bound: ε=0.1 for GCN/Cora and GCN/CiteSeer, and 0.08 otherwise. Training and testing use the same bound. All 60 audited runs satisfy it within floating-point tolerance. Graph structure and other parties' inputs are unchanged. Features are not clipped to their original range or projected to binary values, so negative or nonbinary values can occur.

Four known target-class anchors define the required label knowledge. Training also modifies selected neighbors without requiring their labels. Testing modifies one target and its full one- and two-hop neighborhoods. The mean, median and P95 counts of nodes modified during training average 24.83, 24.67 and 32.02, respectively. The corresponding test-query statistics average 16.37, 7.75 and 70.45. Observed maxima are 62 and 313. These summaries equally weight 12 configurations. The reported medians and P95 values are averages of within-group statistics, not pooled quantiles. On average, modified nodes account for 0.528% of graph nodes during training and 0.421% per test query. These small fractions provide further evidence of GraphCamo's effectiveness with limited node modifications.

ASR is the fraction of non-target-class test nodes classified as the target class after triggering. Input-space attacks evaluate one target per query and restore the original input between queries. GraphCamo modifies the neighborhood but scores success only on that target. GF and VILLAIN likewise evaluate individual targets, perturbing uploaded representations instead. Thus, the ASR definition and evaluation unit are shared, while perturbation spaces differ.

Main-Task Utility, Collateral Effects, and Node Localization

For normal-task utility, we compare MTA from supplementary no-attack training with the paper's untriggered GraphCamo MTA across four datasets and three models. Equally weighting the 12 configurations gives 69.04% for the no-attack models and 68.07% for GraphCamo, a mean difference of only 0.97 percentage points. Thus, in this comparison, the attack-trained models retain normal-input accuracy close to the no-attack reference, with an average performance cost below one percentage point. GraphCamo also has higher MTA in four configurations, showing that accuracy does not decline in every setting. This reference complements the paper's MTA comparisons across attack methods by directly quantifying the normal-task performance difference relative to no-attack models, providing further evidence for GraphCamo's utility preservation.

For collateral effects, we fix the attack-trained model, trigger one test target and compare predictions before and after, restoring the graph between queries. A new error occurs when a previously correct prediction becomes incorrect. The mean rate of new errors is 0.407% among other test nodes and 0.224% after also excluding all directly modified nodes. This indicates limited average interference, particularly on unmodified test nodes.

For localization, we score uploaded representations using kNN distance, class-based Mahalanobis distance and Isolation Forest. Without a separate split, thresholds from untriggered scores cap empirical node FPR at 1% or 5% on that same set. Success requires over half the post-attack alarm set to consist of nodes directly modified in that attack. The detectors' mean success rates are 7.93%, 10.86% and 8.57% at 1%, and 3.03%, 4.98% and 0.68% at 5%, respectively. These results support limited alarm-set localizability in the current offline setting.

The new experiments assess normal-task utility, collateral errors within the attacked graph and localization quality. Representation and gradient MSE values across consecutive epochs measure temporal stability.

Your scrutiny helped us separate label knowledge from actual modifications and temporal stability from localizability, and add direct utility and collateral-effect measurements. We appreciate your emphasis on verifiable evidence. If these clarifications ease your concerns about support for our claims, we would greatly appreciate your renewed consideration and support in light of the additional evidence.

Thank you for your efforts in addressing my questions; this has resolved most of my concerns. After reviewing the author's response and the comments from other reviewers, I am inclined to maintain my score.

Reviewer_w5LxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GraphCamo, a clean-label backdoor attack tailored to vertical federated graph neural networks (VFGNNs) that achieves high and stable attack success without degrading main-task performance. The attack couples (i) structure-guided trigger learning that uses local personalized PageRank (PPR) to select a small, multi-hop set of trigger-support nodes and learns shared perturbations via PGD, and (ii) masked reconstruction-based backdoor consolidation that jointly reconstructs clean and triggered features with a gradient-conflict correction step to preserve both clean-task and trigger information through message passing. Across four citation benchmarks and three GNN backbones, the method reports near-100% ASR with low variance while maintaining competitive main-task accuracy and shows robustness to several defenses

Critical strengths
  1. The attack integrates topology-aware trigger-support selection (via local PPR) with a dual-branch masked reconstruction objective that explicitly preserves both clean and triggered feature directions after aggregation

  2. Broad evaluation across four standard node-classification datasets (Cora, Cora_ml, CiteSeer, PubMed) and three local backbones (GCN, GAT, SGC) under a standard VFGNN protocol.

  3. The problem setting, adversary capabilities, and protocol details are clearly specified; Algorithm 1 and Figure 2 convey the attack pipeline well.

Critical weaknesses
  1. The attack assumes access to several target-class anchor nodes; while common in VFL backdoor work, the practicality and acquisition pathway deserve deeper discussion and quantification.

  2. Evaluation is limited to small citation benchmarks; no larger, heterophilous, or industrial-scale graphs, making generalization uncertain.

  3. Recent graph backdoor advancements (e.g., Ba-Logic’s clean-label logic poisoning) are not discussed; while not in VFGNN, they provide a relevant perspective on stabilizing targeted backdoors under propagation.

Questions for authors
  1. Why select the “middle 50%” of PPR-ranked neighbors at each hop instead of the top-k? Could you report a sensitivity analysis across the selection fraction ρ and “middle vs. top” choices, and provide intuition for this design?

  2. What are the precise perturbation bounds ε used per dataset/model, and how do ASR/MTA trade off with ε? Relatedly, how does the auxiliary edge-drop rate (and whether it is uniform across nodes/epochs) affect performance?

1 comment
AuthorsOpenReview ↗

Response to Reviewer w5Lx

Thank you for your thoughtful questions about our design choices and evaluation scope. They led us to examine node selection and perturbation settings and extend the experiments to a larger graph.

Rationale and Sensitivity Analysis of PPR-Based Neighbor Selection

The middle range aims to balance trigger propagation with normal representations. Low-PPR neighbors may transmit triggers weakly, while the highest-ranked neighbors may be important for normal prediction. We vary training selection, keeping other settings and full-neighborhood testing fixed. Means equally weight four datasets and three models.

For middle selections of 25%, 50% and 75%, mean ASR is 93.94%, 99.24% and 93.36%. Top selections of 25%, 50% and 75% yield 92.94%, 91.94% and 93.81%. Selecting all neighbors gives 98.00%. ASR values for middle-50% and all-neighbor selection are from the paper. Top-ranked selection has no consistent advantage. Expanding the middle selection from 25% to 75% increases the mean number of modified nodes from 17.95 to 38.40 without improving ASR. The default achieves 68.07% MTA, supporting an effective compromise rather than a universal optimum.

Sensitivity to the Perturbation Bound and Auxiliary Edge Dropping

The coordinate-wise L∞ bound on local input perturbations is ε=0.1 for GCN/Cora and GCN/CiteSeer, and 0.08 otherwise, unchanged at test time. At 0.25, 0.5, 1 and 2 times this budget, mean ASR is 76.16%, 92.16%, 99.24% and 99.19%. MTA is 65.38%, 66.21%, 68.07% and 68.04%, respectively. Halving ε retains over 90% ASR in 9 of 12 configurations. Smaller ε does not consistently improve MTA, while the default already achieves high ASR.

At auxiliary drop rates of 0, 0.05, 0.1, 0.2 and 0.4, mean ASR is 97.43%, 96.20%, 99.24%, 98.75% and 97.26%. MTA is 66.92%, 67.10%, 68.07%, 67.07% and 67.21%. Resampling at rate 0.1 every epoch gives 98.97% ASR and 67.30% MTA, with ASR above 96% in every configuration. Results at the default setting are from the paper. Effectiveness remains high across rates and with resampling.

The original rule uses a common drop threshold and edge-based deterministic scores fixed within a run. Both directions of an undirected edge share a decision. The rule is uniform across nodes, but the mask is fixed across epochs within a run. Only the auxiliary reconstruction view changes, not the main-task or inference graph.

Supplementary Discussion: Assumptions About Knowledge of Target-Class Anchors

GraphCamo assumes knowledge of four target-class training nodes' identities and labels. Neighbor selection uses only the local graph, without neighbor labels. The paper compares 4, 6, 8 and 10 anchors. Four already yield high ASR, with generally limited gains from more anchors. This quantifies a modest label-knowledge requirement. Obtaining labels remains a threat-model assumption.

Supplementary Evaluation: Attack Effectiveness and Main-Task Utility on a Larger Graph

On the full ogbn-arxiv graph of 169,343 nodes, we use official splits, anchors comprising about 0.83% of training nodes and ε=0.32. No-attack references use the same architecture and training budget.

For GCN, GAT and SGC, ASR is 92.32%, 98.05% and 94.16%, while untriggered MTA is 61.36%, 63.56% and 61.54%. No-attack MTA is 64.58%, 64.90% and 64.66%, giving differences of 3.21, 1.34 and 3.12 percentage points from unrounded means. GraphCamo retains high attack effectiveness and most normal-task utility on a larger citation graph, especially GAT with 98.05% ASR and a 1.34-point MTA difference.

Supplementary Discussion: Relationship to Ba-Logic

Ba-Logic uses a clean GCN to select target-class nodes with uncertain predictions, generates feature-conditioned trigger subgraphs with new nodes and edges, and alternates surrogate and generator optimization without changing labels. It encourages trigger nodes' total gradient importance to exceed that of clean nodes, alongside target-class loss and feature-similarity constraints, making triggers influential during clean-label learning.

Both methods seek to sustain trigger signals through learning and graph propagation. GraphCamo uses a malicious passive party's local graph, features, model and returned gradients to perturb existing nodes within a bound. PPR selection, clean and triggered reconstruction, and gradient-conflict handling coordinate attack success, utility and representation preservation. These distinctions clarify GraphCamo's contribution.

From the choice of the middle PPR range to parameter sensitivity, larger graphs and Ba-Logic, your questions prompted a fuller account of our design and its scope. We sincerely appreciate this guidance. We hope these analyses make the rationale and contribution clearer and provide a stronger basis for your recognition and support of this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.