acceptodds
Under review as a conference paper at ICLR 2027

Federated Graph Fraud Detection with Spectral Adaptation and Confidence-Aware Aggregation

Abstract

Graph-based Fraud Detection (GFD), which aims to identify fraud patterns within interaction networks, has attracted significant interest in trustworthy machine learning. However, real-world privacy regulations distribute graph data across decentralized clients, where fraudsters exhibit severe spectral heterogeneity both within and across these silos. Consequently, existing centralized fraud detectors relying on static filter banks fail to capture intricate patterns. More critically, fraudsters deliberately isolate themselves to evade detection, rendering any graph filter that operates via message-passing schemes completely ineffective. To address these challenges, we propose a federated learning framework featuring two major techniques: (1) , which employs dynamic spectral channel attention to handle intra-client heterogeneity, and performs per-frequency confidence-aware aggregation on filter coefficients to tackle inter-client heterogeneity. (2) , which extends FedSACA with a feature-induced structure augmentation subject to the frequency confidence evaluation, thereby preventing the introduction of potential noise. Extensive experiments on multiple datasets demonstrate the superior performance of our proposed techniques.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Fraud Detection

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_jmy6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores the challenges in graph-based federated fraud detection, including spectral heterogeneity on the client side and the identification of isolated fraudulent nodes. The authors propose a framework called FedSACA, which achieves local adaptation by learning spectral gates and uses the resulting gate values ​​to weight the aggregated client filtering coefficients. Building on this, the authors also introduce FedSACA+; this version constructs neighborhood relationships based on node features and suppresses potential noise introduced by new connections. Experiments conducted on four datasets demonstrate that FedSACA outperforms baseline methods in fraud detection benchmarks.

Critical strengths
  1. This paper explores the heterogeneity of fraud spectra both within and across clients. By visualizing the spectral energy distribution, the authors highlight the disparities among different clients, providing a clear motivation for employing adaptive spectral learning within a federated learning environment.
  2. FedSACA features a streamlined and intuitive design that combines local spectral gating with weighted aggregation at the server side. Building upon this, FedSACA+ introduces a feature-based nearest-neighbor graph as a distinct relational structure; it employs a conservative initial gating mechanism to regulate the influence of these added connections, thereby achieving structural enhancement within a unified framework.
  3. The experimental evaluation encompasses four datasets and a variety of baseline methods. This framework performs better than other baselines.
Critical weaknesses
  1. The motivation for this study stems primarily from the heterogeneity in the spectral patterns of fraudulent behavior within real-world institutions. However, the graph datasets used in the experiments are partitioned among clients artificially based on a Dirichlet distribution. It remains unclear to what extent this partitioning accurately represents clients in real-world institutional settings, nor is it clear whether the observed differences in spectral patterns arise from distinct fraudulent mechanisms or from topological changes induced by the graph partitioning process itself. Please provide the full details of the data partitioning scheme and evaluate the proposed method under varying degrees of heterogeneity.
  2. In Section 4.1, the authors state that fraudsters manipulate local connectivity to blend in with legitimate nodes. Please clarify which relationship types or operations in the dataset correspond to this behavior, and provide relevant examples, statistical data, or references.
  3. In Equation (7), si,ks_{i,k} is a gating weight learned jointly with the filter coefficients; in Equation (14), this weight is used to determine the contribution of each client during the aggregation process. Could the authors explain why a larger gating weight implies that the knowledge provided by a client is more reliable? To isolate the contribution of confidence-aware weighting, could the authors conduct an experiment where only Equation (14) is modified? For instance, replace it with uniform weighting.
Questions for authors

Please refer to my weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer jmy6 (1/3)

We thank the reviewer for their careful reading and constructive suggestions, which helped us strengthen the paper.

W1: Spectral heterogeneity and Dirichlet distribution.

We thank the reviewer for this insightful comment. The Dirichlet-based partitioning is a standard protocol for simulating non-IID data in real-world federated learning (Hsu et al., 2019 [1]; He et al., 2021 [2]). To address whether the observed heterogeneity originates from partitioning, we have added an analysis of the original centralized graph to section E.1 of the revised appendix. As shown in the new Figure 6, the fraud signal energy on the whole graph of YelpChi already exhibits a diverse, non-contiguous spectral profile (i.e., heterogeneity) without any partitioning. The Dirichlet partitioning merely simulates how fraud patterns are distributed naturally within and across clients, thereby exposing inter and intra spectral heterogeneity as exhibited in Figure 1. Furthermore, we have conducted additional experiments with varying degrees of heterogeneity (different Dirichlet α\alpha values) in C.3 of the revised appendix, demonstrating that FedSACA+ maintains robust performance.

W2: Fraudster manipulation.

We appreciate the reviewer’s constructive feedback. Our statement in Section 4.1 regarding fraudulent camouflaging is grounded in well-documented empirical findings in real-world fraud detection literature. Taking the YelpChi dataset as an example (e.g., Dou et al., 2020 [3]; Peng et al., 2021 [4]), fraudsters frequently manipulate local connectivity (such as densely reviewing each other via the R-S-R relation) to artificially blend in with benign nodes, forming tight local clusters with elevated connectivity and deceptive homophily at the spatial level. Consequently, relying solely on local spatial aggregations causes GNNs to aggregate massive noise from these camouflaged connections. This renders even camouflage-aware spatial methods (e.g., CARE-GNN) inadequate, as also evidenced by the degraded performance of in Table 1. This critical limitation directly motivates our core argument: to bypass spatial camouflage, we need to capture the distinct fraud patterns from a spectral perspective.

[1] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335, 2019.

[2] Chaoyang He, Keshav Balasubramanian, Emir Ceyani, Carl Yang, Han Xie, Lichao Sun, Lifang He, Liangwei Yang, Philip S Yu, Yu Rong, et al. Fedgraphnn: A federated learning system and benchmark for graph neural networks. In ICLR, 2021a.

[3] Yingtong Dou, Zhiwei Liu, Li Sun, Yutong Deng, Hao Peng, and Philip S Yu. Enhancing graph neural network-based fraud detectors against camouflaged fraudsters. In Proceedings of the 29th ACM international conference on information & knowledge management, pp. 315–324, 2020.

[4] Hao Peng, Ruitong Zhang, Yingtong Dou, Renyu Yang, Jingyi Zhang, and Philip S Yu. Reinforced neighborhood selection guided multi-relational graph neural networks. ACM Transactions on Information Systems (TOIS), 40(4):1–46, 2021.

Thank you for your clear explanation. I have decided to maintain my score.

AuthorsOpenReview ↗

Response to Reviewer jmy6 (2/3)

W3: Justification of si,ks_{i,k}

We thank the reviewer for raising this important question. We clarify that si,ks_{i,k} is a task-driven spectral confidence score, rather than a manually assigned weight. Its learning mechanism and its role in aggregation can be understood from two complementary perspectives.

1. Gradient-based learning of task-dependent spectral confidence.

Recall that the filtered representation is

$ H_i=\sum_{k=0}^{K-1}s_{i,k}\theta_{i,k}V_{i,k}, \quad V_{i,k}=T_k(\hat A_i)X_i. $

Let

$ G_{i,k}= \left\langle \nabla_{H_i}\mathcal L_{\mathrm{CE}}^{(i)}, V_{i,k} \right\rangle_F. $

By the chain rule, the gradient of the local objective with respect to the confidence mask mi,km_{i,k} is

∂Li∂mi,k=si,k(1−si,k)(θi,kGi,k+λ)+μ(mi,k−mkt).\frac{\partial\mathcal L_i}{\partial m_{i,k}}=s_{i,k}(1-s_{i,k})(\theta_{i,k}G_{i,k}+\lambda)+\mu(m_{i,k}-m_k^t).

Ignoring the proximal term momentarily, gradient descent increases si,ks_{i,k} when

$ -\theta_{i,k}G_{i,k}>\lambda. $

This condition indicates that the corresponding spectral component provides sufficient reduction in the local fraud detection loss to overcome the sparsity penalty. Conversely, components that provide insufficient predictive utility are progressively suppressed.

Therefore, si,ks_{i,k} is learned according to the task-dependent contribution of each spectral basis, rather than being assigned independently of the downstream objective. Importantly, this mechanism also accounts for the sign of θi,k\theta_{i,k}, since the gradient depends on the actual contribution of the basis to the prediction loss.

2. Theoretical motivation for confidence-aware aggregation.

The use of confidence scores in aggregation is also motivated by inverse-variance weighting.

Consider a fixed spectral coefficient kk, and suppose each client provides an unbiased estimate

$ \theta_{i,k}=\theta_k^*+\epsilon_{i,k}, $

where the estimation errors are independent, zero-mean, and have variances σi,k2\sigma_{i,k}^2.

For a linear unbiased estimator

$ \hat\theta_k=\sum_i\alpha_{i,k}\theta_{i,k}, \qquad \sum_i\alpha_{i,k}=1, $

its variance is

$ \operatorname{Var}(\hat\theta_k) =\sum_i\alpha_{i,k}^2\sigma_{i,k}^2. $

Minimizing this variance under the unbiasedness constraint yields

αi,k∗=σi,k−2∑jσj,k−2.\alpha_{i,k}^*=\frac{\sigma_{i,k}^{-2}}{\sum_j\sigma_{j,k}^{-2}}.

Thus, clients with more reliable coefficient estimates (lower estimation variance) should receive larger aggregation weights.

FedSACA adopts the corresponding confidence-weighted form:

αi,k=si,k∑jsj,k.\alpha_{i,k}=\frac{s_{i,k}}{\sum_j s_{j,k}}.

Under the condition that the learned confidence scores are positively associated with coefficient estimation precision, i.e., si,k∝σi,k−2s_{i,k}\propto\sigma_{i,k}^{-2}, Eq. (14) recovers the minimum-variance aggregation weights.

We emphasize that this result provides a theoretical justification for precision-aware weighting, while the correspondence between the learned gates and actual client reliability is an empirical property that we further investigate below.

AuthorsOpenReview ↗

Response to Reviewer jmy6 (3/3)

3. Empirical validation of learned confidence.

We conduct two complementary experiments.

First, we measure the independent predictive utility of each spectral basis through leave-one-basis-out retraining. Specifically, we define

Ui,k=Lval,i(fi−k)−Lval,i(fi),U_{i,k}=\mathcal L_{\mathrm{val},i}(f_i^{-k})-\mathcal L_{\mathrm{val},i}(f_i),

where fi−kf_i^{-k} is retrained with the kk-th basis disabled. A positive value indicates that the basis contributes to better validation performance. We then examine the rank correlation between the learned confidence si,ks_{i,k} and its independently measured utility, controlling for the magnitude of θi,k\theta_{i,k} and the basis representation norm.

DatasetSpearman ρ(s,U) ↑Partial ρ ↑Top-2 Utility-positive Rate ↑
YelpChi0.740.6186.7%
Amazon0.690.5583.3%
FDCompCN0.780.6590.0%
Elliptic0.720.5886.7%
Average0.730.6086.7%

Table 1: Faithfulness of Learned Confidence to Spectral Basis Utility

As shown in Table 1, the learned confidence scores exhibit consistently strong positive correlations with independently measured basis utility across all four datasets (Spearman ρ=0.69–0.78\rho=0.69–0.78, averaging 0.730.73). Importantly, these correlations remain positive after controlling for the filter coefficient magnitude and basis representation norm (partial ρ=0.55–0.65\rho=0.55–0.65, averaging 0.600.60. Moreover, an average of 86.786.7% of the top-2 highest-confidence bases exhibit positive predictive utility. These findings support that the learned confidence scores reflect task-relevant spectral basis importance, rather than merely their multiplicative contribution to the filtering operation.

Second, to verify whether the learned confidence scores provide useful information for cross-client aggregation, we compare the original Eq. (14) against uniform aggregation and randomly permuted confidence weights. The permutation preserves the distribution of confidence scores while breaking their correspondence with the respective client coefficients. We keep the remaining aggregation procedure, including Eq. (13), unchanged.

DatasetLearned s (AUC/F1) ↑Uniform (AUC/F1) ↑Permuted s (AUC/F1) ↑ΔAUC vs. Permuted ↑
YelpChi89.8 / 74.484.1 / 70.380.9 / 66.0+8.9
Amazon99.3 / 92.295.4 / 89.891.2 / 85.4+8.1
FDCompCN76.2 / 68.070.0 / 62.061.2 / 57.6+15.0
Elliptic99.2 / 93.693.5 / 88.188.4 / 82.9+10.8

Table 2: Effectiveness of Learned Confidence in Federated Aggregation

As shown in the above Table 2, confidence-aware aggregation consistently outperforms uniform averaging by 3.93.9–6.26.2 AUC points (averaging 5.45.4) across all four datasets. More importantly, randomly permuting the learned confidence scores leads to substantially larger performance degradation, with AUC dropping by 8.18.1–15.015.0 points (averaging 10.710.7). Since permutation preserves the confidence distribution but disrupts its correspondence with client coefficients, these results suggest that the learned scores encode useful client-specific information for knowledge aggregation. Together with Table 1, these findings support interpreting si,ks_{i,k} as a task-conditioned confidence proxy that reflects spectral basis relevance and provides informative weights for federated aggregation.

As shown in our newly added ablation study in the revised appendix (Variant B in Table 4), this uniform aggregation yields an AUC of 81.67%81.67\%%, which is significantly lower than the 87.71%87.71\%% achieved by our full confidence-aware aggregation (Variant C/FedSACA). This large performance gap (drop of ∼6\sim 6%%) strongly validates that weighting by confidence effectively prevents the global filter from being polluted by clients' conflicting and unreliable spectral preferences, proving the necessity of our design.

Thank you again. Please let us know if our answers address your concerns, or whether we can provide further clarifications. We hope the new results and clarifications support a stronger acceptance rating.

Reviewer_kzr5OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies federated graph fraud detection with heterogeneous spectral patterns and isolated nodes. FedSACA places learnable sigmoid gates on Chebyshev filter coefficients and uses the gate values to weight coefficient aggregation across clients. FedSACA+ adds a local feature-based kNN relation whose gates are initialized with a negative bias. The evaluation uses YelpChi, Amazon, FDCompCN, and Elliptic with five clients and Dirichlet partitions, reporting AUC and F1-Macro in the main paper and AP, G-Mean, and Recall in the appendix.

Critical strengths
  1. Federated graph fraud detection combines privacy constraints, graph fragmentation, class imbalance, and heterophilic fraud patterns. The paper identifies a useful interaction between spectral heterogeneity and fragmented graph structure rather than considering these issues independently.

  2. FedSACA+ consistently improves over FedSACA and the listed federated baselines on four datasets. The additional AP, G-Mean, and Recall results in the appendix also provide useful evidence beyond AUC and F1-Macro. The method is computationally relatively simple compared with more complicated path-based approaches.

Critical weaknesses
  1. Section 4.1, pp.7, Eqs. (14), does not justify interpreting the gates as confidence in distinct frequency bands. The index kk denotes Chebyshev polynomial order, while Tk(1−λ)T_k(1-\lambda) responds across the spectrum rather than within a localized band. Two equally confident clients with coefficients of equal magnitude and opposite signs cancel exactly under Eq. (14). Thus, confidence-weighted averaging alone does not establish the claimed frequency alignment.

  2. The proposed method lacks sufficient methodological novelty. FedSACA adds sigmoid gates to Chebyshev filter coefficients and aggregates those coefficients through a normalized weighted average. FedSACA+ adds a feature-based kNN relation and initializes its gates with a negative bias. The paper does not establish a substantial methodological advance beyond this combination of adaptive coefficient weighting and graph augmentation.

Questions for authors
  1. For Section 4.1, pp. 5 to 7, what mathematical property makes si,ks_{i,k} a confidence score for a frequency band? Please show the effective response gi(λ)=∑k=0K−1si,kθi,kTk(1−λ)g_i(\lambda)=\sum_{k=0}^{K-1}s_{i,k}\theta_{i,k}T_k(1-\lambda) before and after aggregation, including clients whose coefficients disagree in sign.
  2. For Section 5.2, pp. 8 and 9, can you compare the proposed aggregation with FedAvg on the same Chebyshev model, both with and without the same kNN relation? Can you also report matched comparisons with the related methods cited above under compatible settings?
4 comments
AuthorsOpenReview ↗

Response to Reviewer kzr5 (1/4)

We sincerely thank the reviewer for the careful evaluation and constructive suggestions. We have substantially strengthened the revised manuscript with theoretical clarifications, effective spectral response analysis, controlled ablations, and additional experimental evidence. We address the concerns as follows.

W1 & Q1: Chebyshev-Basis Confidence, Spectral Conflicts, and Frequency Responses

1. Clarification of spectral bases and confidence scores.

We thank the reviewer for identifying this theoretical imprecision. We agree that individual Chebyshev polynomial orders are not isolated frequency bands. Accordingly, we have revised the terminology throughout the manuscript, replacing frequency bands with Chebyshev bases and per-frequency aggregation with per-basis aggregation.

Importantly, although an individual Chebyshev basis Tk(1−λ)T_k(1-\lambda) responds across the spectrum, their weighted combination can approximate complex spectral responses, including non-contiguous frequency patterns:

$ g_i(\lambda)=\sum_{k=0}^{K-1}s_{i,k}\theta_{i,k}T_k(1-\lambda). $

Thus, our spectral adaptation does not require individual bases to correspond to isolated frequency intervals. Instead, the learned coefficients and gates jointly construct an adaptive filter over the complete graph spectrum.

2. Why is the learned si,ks_{i,k} meaningful?

The confidence score si,k=σ(mi,k)s_{i,k}=\sigma(m_{i,k}) is a task-driven Chebyshev-basis selection weight, rather than a manually assigned or statistically calibrated probability. Its learning mechanism directly connects basis selection to fraud detection performance.

For clarity, omitting the relation index and letting Vi,k=Tk(A^i)XiV_{i,k}=T_k(\hat A_i)X_i, the gradient of the local objective with respect to mi,km_{i,k} is:

$s_{i,k}(1-s_{i,k}) \theta_{i,k} \langle \nabla_{H_i}\mathcal L_{\mathrm{CE}},V_{i,k} \rangle+\lambda_{\mathrm{reg}} +\mu(m_{i,k}-m_k^{global}). $

This reveals that basis selection is driven by three complementary factors: its influence on the supervised fraud detection objective, the sparsity constraint suppressing unnecessary bases, and the proximal term promoting stable federated optimization. Therefore, si,ks_{i,k} reflects the preference learned by client ii for retaining basis kk, providing a task-dependent signal for coefficient aggregation.

AuthorsOpenReview ↗

Response to Reviewer kzr5 (2/4)

We further conducted two complementary empirical validations, reported in our response to Reviewer jmy6:

  • Basis-utility faithfulness: Learned confidence scores correlate positively with independently measured leave-one-basis-out retraining utility, achieving an average Spearman correlation of 0.73, partial Kendall correlation of 0.60, and 86.7% positive predictive utility among the two highest-confidence bases.
  • Aggregation utility: We compare learned confidence weighting against uniform weighting and randomly permuted confidence weights, keeping the remaining aggregation procedure unchanged.
DatasetLearned s (AUC/F1)Uniform (AUC/F1)Permuted s (AUC/F1)ΔAUC vs. Permuted ↑
YelpChi89.8 / 74.484.1 / 70.380.9 / 66.0+8.9
Amazon99.3 / 92.295.4 / 89.891.2 / 85.4+8.1
FDCompCN76.2 / 68.070.0 / 62.061.2 / 57.6+15.0
Elliptic99.2 / 93.693.5 / 88.188.4 / 82.9+10.8

Table 1: Effectiveness of Learned Confidence in Federated Aggregation.

The learned confidence consistently outperforms uniform weighting by 3.9–6.2 AUC points and permuted confidence by 8.1–15.0 points. Since permutation preserves the confidence values but disrupts their correspondence with client coefficients, the substantial performance degradation indicates that the learned confidence–coefficient associations contain useful information for federated aggregation, rather than merely acting as arbitrary weights.

Together, the gradient analysis and empirical evidence support interpreting si,ks_{i,k} as a task-driven basis-importance signal that is useful for aggregation.

3. Why do opposite-signed coefficients not necessarily imply harmful spectral conflicts?

We appreciate the reviewer's mathematical counterexample. Indeed, opposite-signed coefficients with equal weights can cancel under Eq. (14). However, coefficient cancellation is not mathematically equivalent to harmful interference in the final spectral response or downstream detection performance.

First, Eq. (13) establishes a consensus over learned basis-selection preferences, rather than averaging the signed filter coefficients. The confidence-aware aggregation in Eq. (14) admits an exact weighted least-squares interpretation:

θkglobal=∑i=1Nsi,kθi,k∑i=1Nsi,k=arg⁡min⁡θ∈R∑i=1Nsi,k(θ−θi,k)2.\theta_k^{\mathrm{global}}=\frac{\sum_{i=1}^{N}s_{i,k}\theta_{i,k}}{\sum_{i=1}^{N}s_{i,k}}=\underset{\theta\in\mathbb{R}}{\arg\min}\sum_{i=1}^{N}s_{i,k}(\theta-\theta_{i,k})^2.

Since si,k=σ(mi,k)∈(0,1)s_{i,k}=\sigma(m_{i,k})\in(0,1), the objective is strictly convex and admits a unique global minimizer.

Thus, Eq. (14) determines the optimal confidence-weighted consensus in the coefficient space. Clients assigning higher confidence to a given Chebyshev basis exert greater influence on the corresponding global filter coefficient.

Importantly, even when equally confident clients have opposite coefficients, their cancellation represents the optimal weighted least-squares consensus, rather than an arbitrary numerical artifact.

More generally, the contribution of client ii relative to client jj on basis kk is:

$\frac{s_{i,k}}{s_{j,k}}, \qquad w_{i,k}=\frac{s_{i,k}}{\sum_j s_{j,k}}.$

Therefore, when clients assign different importance to a basis, the more confident client's coefficient receives greater weight instead of being averaged indiscriminately.

Second, a sign disagreement at one Chebyshev order does not necessarily imply substantially different complete frequency responses. The effective filter is a combination of all bases:

$\sum_{k=0}^{K-1} s_{i,k}\theta_{i,k}T_k(1-\lambda). $

For two clients (i,j)(i,j), since (∣Tk(1−λ)∣≤1|T_k(1-\lambda)|\leq1) for (λ∈0,2\lambda\in0,2),

$ |g_i(\lambda)-g_j(\lambda)| \leq \sum_{k=0}^{K-1} |s_{i,k}\theta_{i,k}-s_{j,k}\theta_{j,k}|. $

Hence, the complete response difference depends on the effective coefficients of all bases, rather than the sign of any single coefficient. Opposite signs at one order alone are insufficient to establish that the overall responses are severely conflicting.

Effective response visualization. In Appendix H, Figure 8, we visualize the effective responses of individual clients before aggregation and the resulting global response after aggregation. To directly examine the reviewer's concern, we also identify a client pair exhibiting opposite-signed coefficients at a particular Chebyshev order and report the corresponding coefficient values and confidence scores (on basis k=2, client 0 with 1.2 and client 1 with -1.0). This analysis distinguishes local coefficient sign disagreement from its actual effect on the complete spectral response. Figure 8 also shows that the clients' spectral preferences (a, b) prior to aggregation are not obscured after aggregation (c).

AuthorsOpenReview ↗

Response to Reviewer kzr5 (3/4)

Most importantly, the global model does not impose identical spectral preferences permanently: clients can further adapt their local coefficients and gates after receiving the shared model. Thus, global knowledge sharing and client-specific spectral adaptation remain complementary.

W2 & Q2: Methodological Novelty and Matched Experimental Comparisons

1. Novelty lies in addressing two coupled obstacles in federated graph fraud detection.

We respectfully clarify that FedSACA and FedSACA+ are not simply an arbitrary combination of sigmoid gating, weighted averaging, and kNN augmentation. Their designs arise from two distinct but tightly connected limitations of federated graph fraud detection.

First, Fraud Spectral Heterogeneity arises because fraud patterns exhibit diverse, non-contiguous frequency characteristics within clients and heterogeneous preferences across clients. A filter that cannot adapt to these patterns remains inadequate even when sufficient graph connectivity is available. FedSACA addresses this through task-driven Chebyshev-basis selection and confidence-aware cross-client coefficient aggregation.

Second, Fraudulent Node Isolation prevents neighborhood-dependent information from reaching disconnected fraudsters. Even a well-adapted spectral filter cannot recover neighbor information that is absent from the original graph. FedSACA+ therefore introduces an additional feature-induced kNN relation, while applying conservative, learnable spectral gating to limit the influence of potentially misleading augmented relationships.

Conversely, merely restoring connectivity does not resolve spectral heterogeneity: the augmented graph still requires a filter capable of adapting to the resulting fraud-related spectral patterns.

These two challenges are complementary: spectral adaptation determines how available graph information should be filtered, whereas structure augmentation determines what additional relational information becomes available. FedSACA+ integrates both capabilities in a unified federated framework while avoiding the exchange of raw graph data.

This problem-driven integration, together with the empirical investigation of intra-/inter-client fraud spectral heterogeneity and fraudulent-node isolation, constitutes the principal contribution of our work.

2. Matched comparisons with uniform coefficient aggregation, with and without kNN.

To directly address the reviewer's request, we isolate the coefficient aggregation mechanism while preserving the same local Chebyshev architecture and other applicable components.

Coefficient AggregationkNN RelationNegative InitializationYelpChi AUC ↑F1-Macro ↑
UniformNoN/A81.67 ± 0.3369.79 ± 0.76
Confidence-aware, Eq. (14)NoN/A87.71 ± 0.1873.60 ± 0.68
UniformYesYes84.12 ± 0.5370.35 ± 0.72
Confidence-aware, Eq. (14)YesNo88.99 ± 0.1573.87 ± 0.61
Confidence-aware, Eq. (14)YesYes89.84 ± 0.1274.38 ± 0.56

Table 2: Component-Level Comparisons of Coefficient Aggregation and kNN Structure Augmentation on YelpChi.

Without kNN augmentation, confidence-aware aggregation improves AUC from 81.67% to 87.71%, a gain of 6.04 percentage points. With the same kNN relation and negative initialization, the improvement is from 84.1% to 89.84%, a gain of 5.74 points.

These comparisons directly demonstrate that confidence-aware aggregation provides substantial benefits both with and without structure augmentation. Furthermore, introducing kNN augmentation improves confidence-aware aggregation from 87.71% to 89.84%, while negative initialization further improves performance from 88.99% to 89.84%.

The results support two important conclusions: (i) confidence-aware coefficient aggregation contributes independently of graph augmentation, and (ii) the two components address complementary challenges and can jointly improve detection performance.

AuthorsOpenReview ↗

Response to Reviewer kzr5 (4/4)

We also emphasize that harmful spectral conflicts should ultimately be evaluated through their effects on downstream predictions, rather than inferred solely from the existence of heterogeneous coefficient signs. The consistent gains from cross-client coefficient sharing indicate that, under the evaluated settings, the benefits of confidence-guided knowledge transfer outweigh potential destructive interference.

This is particularly relevant to federated fraud detection, where fraudulent samples are scarce (ranging from approximately 2.23% on Elliptic to 14.53% on YelpChi among our four datasets). Cross-client collaboration allows individual clients to benefit from complementary fraud-related knowledge that would otherwise be difficult to learn from their limited local supervision.

3. Comparisons with related methods under compatible settings.

We have also strengthened the revised experimental comparisons. Beyond the original baselines, we added FedSage+ to the performance tables, enabling a more direct comparison with an existing federated graph structure-augmentation method.

Our benchmark includes representative federated graph learning approaches such as FedPAM, FedGH, and Fed-Kalter, as well as spectral fraud detection methods adapted to the federated setting, including BWGNN, SplitGNN, and SEC-GFD.

The experiments follow the federated evaluation protocol described in Section 5.1, including five clients, consistent train/validation/test partitions, and the specified communication and local training budgets. We have also clarified the computational limitations of the reported OOT results: PMP exceeded the 72-hour execution limit before completing the required federated training. Its result is therefore reported as unavailable rather than as an inferior performance outcome.

We have revised the related references and experimental descriptions accordingly to improve the transparency and reproducibility of these comparisons.

Conclusion.

We sincerely appreciate the reviewer's comments, which motivated us to clarify the mathematical interpretation of confidence, distinguish basis-wise coefficient consensus from complete spectral-response alignment, and strengthen the experimental isolation of our contributions.

We believe the revised analyses establish that FedSACA provides task-driven spectral adaptation and empirically effective confidence-guided knowledge sharing, while FedSACA+ complements these capabilities by addressing missing neighborhood information for isolated fraudsters.

Rather than claiming that every possible coefficient conflict is eliminated, our contribution is to demonstrate that these mechanisms substantially improve federated fraud detection under heterogeneous and structurally incomplete graph data. We hope the additional theoretical explanations and controlled evidence address the remaining concerns and would greatly appreciate the reviewer reconsidering the assessment.

Reviewer_ocr7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedSACA and its extension FedSACA+ for federated graph fraud detection. To address distinct intra- and inter-client heterogeneitywithout accessing raw data or sensitive topology, the paper proposes Federated Graph Fraud Detection with Spectral Adaptation and Confidence-Aware Aggregation (FedSACA), which rethinks filter adaptation and fraud spectral-domain aggregation in the federated setting. To address fraudulent node isolation, the paper extends FedSACA to FedSACA+ by introducing a feature-induced structure augmentation.

Critical strengths

S1: The problem definitions which seperate Fraud Spectral Heterogeneity into two levels are clear.

S2: The framework is easy to follow. In this framework, each research problem has a related design. Like Intra-Client Spectral Heterogeneity corresponds to Spectral Chennel Attendtion.

S3: The baselines in experiment section is sufficient.

Critical weaknesses

W1: Some of the statement I think they are too absolute. For example, This deliberate isolation severely disrupts the GNN message-passing scheme, rendering the filters entirely ineffective on disconnected nodes as it severs signal propagation pathways. and Furthermore, FedSACA+ restores filter effectiveness on isolated fraudsters via noise-free structure augmentation.

W2: The experiment section does not have any ablation study to show the performance gains from each core components.

W3: The spectral heterogeneity and islated fraudsters are evaluated insufficiently.

Questions for authors

Q1: The paper argues that simply averaging parameters can cause problems when clients have different spectral preferences. But here, you're still averaging the masks across clients. Could you explain how this step actually helps address spectral heterogeneity?

Q2: In Equation 17, you initialize the confidence of the added kNN relation to a low value, but this confidence can increase during training. Also, the mechanism controls the entire relation rather than individual edges. So how can we be sure that noisy connections won't be reinforced as well?

Q3: I have two questions about Figure 5. First, the AUC and F1 scores in Figure 5 seem lower than those reported in Table 1, even though both use YelpChi. Could you clarify why the results are different?

Q4: In Figure 5(d), the performance remains quite stable even when (c) is extremely large. Since a large (c) makes the initial confidence of the kNN relation almost zero, wouldn't it also make it very difficult for the model to reactivate those connections during training? If the performance is still similar, does that mean the added kNN connections are not contributing much to the improvement?

5 comments
AuthorsOpenReview ↗

Response to Reviewer ocr7 (1/2)

We thank the reviewer for their careful reading and constructive suggestions.

W1: Clarification of isolated nodes and structure augmentation.

Thank you for your suggestion for improving the statement. An isolated node's self-feature information remains available through the zero-order Chebyshev term (but it can only fit to the constant filter, i.e., 1); however, the isolated node cannot receive neighborhood messages that higher-order terms (k>0k>0) rely on. Since k>0k>0 terms are important to effectively fit to the specified and intricate fraud frequency response, the isolation therefore significantly reduces detection capability. We have revised "filters entirely ineffective" to clarify that neighborhood-dependent message-passing propagation basic to filter capabilities becomes unavailable. We acknowledge that the method does not avoid introducing any noise at all. However, we restore filter effectiveness through the reconnection of knn and exert a strong negative constant bias on the new induced relation to limit noise to the greatest extent. Moreover, the self-adaptive spectral channel attention also adjusts masks to a state that is most helpful in discovering the fraudulent nodes, which filters out most noise. Our experiments in Table 1 demonstrate that FedSACA+ outperforms FedSACA, further validating that the possible induced noise during knn augmentation does not degrade the performance. We have replaced "noise-free structure augmentation" with "confidence-controlled structure augmentation".

W2: Component contributions.

We appreciate your advice on the ablation study. The following data is the ablation experiments conducted under YelpChi. This data indicates that the four modules are beneficial for fraud detection.

VariantLocal spectral adaptationConfidence aggregationkNN augmentationNegative initializationAUCF1-MacroAPG-MeanRecall
A✗✗✗✗73.48±0.8873.48_{\pm 0.88}63.13±1.1163.13_{\pm 1.11}56.18±1.2256.18_{\pm 1.22}68.66±1.4568.66_{\pm 1.45}69.27±1.3669.27_{\pm 1.36}
B✓✗✗✗81.67±0.3381.67_{\pm 0.33}69.79±0.7669.79_{\pm 0.76}60.78±0.8160.78_{\pm 0.81}78.76±1.1378.76_{\pm 1.13}78.84±1.1078.84_{\pm 1.10}
C (FedSACA)✓✓✗✗87.71±0.1887.71_{\pm 0.18}73.60±0.6873.60_{\pm 0.68}66.83±0.5566.83_{\pm 0.55}83.02±0.6183.02_{\pm 0.61}83.15±0.5883.15_{\pm 0.58}
D✓✓✓✗88.99±0.1588.99_{\pm 0.15}73.87±0.6173.87_{\pm 0.61}68.11±0.3268.11_{\pm 0.32}84.34±0.3584.34_{\pm 0.35}84.42±0.3184.42_{\pm 0.31}
E (FedSACA+)✓✓✓✓89.84±0.1289.84_{\pm 0.12}74.38±0.5674.38_{\pm 0.56}69.24±0.1269.24_{\pm 0.12}85.28±0.1585.28_{\pm 0.15}85.35±0.1485.35_{\pm 0.14}
AuthorsOpenReview ↗

Response to Reviewer ocr7 (2/2)

W3: Heterogeneity and isolated fraudster evaluation.

We appreciate this suggestion. Due to the issue with the OpenReview Markdown representation, we have to present our explanation in Appendix E and Appendix F. Please refer to these sections in the revised pdf version.

Q1: Why does FedSACA average the confidence masks?

We thank the reviewer for this important question. We would like to clarify that the parameter averaging criticized in our motivation primarily refers to the aggregation of filter coefficients θ\theta. In contrast, the confidence masks m\mathbf{m} are introduced as a mechanism with a different role. Specifically, Eq. (13) averages m\mathbf{m} as a basis-alignment step. Due to the scarcity of fraud labels on individual clients (e.g., min 2.23% in Elliptic and max 14.53% in YelpChi among our 4 datasets), locally optimized masks are prone to overfitting to local noise. Averaging m\mathbf{m} enables clients to jointly assess the importance of a basis, while still allowing clients to adjust their confidences according to the local spectral profile in the local training stage. The actual mitigation of spectral conflicts occurs in Eq. (14), where filter coefficients θ\theta are aggregated with per-basis confidence scores, ensuring the global filter selectively inherits reliable spectral knowledge across all bases. To further validate this design, we have conducted an experiment comparing our approach against a variant that keeps m\mathbf{m} strictly local (no averaging). The result shows that keeping m\mathbf{m} local degrades performance (AUC drops from 89.84% to 88.06%), demonstrating that mask averaging is essential to avoid client-level overfitting and jointly address inter-client spectral heterogeneity.

Q2: How can the proposed mechanism avoid reinforcing noisy kNN connections?

When optimizing the filter θ\theta, the objective is to ensure that each basis is well-behaved. In evaluating the spectra of these bases, it is not a single edge that reflects the spectral response; rather, it is the entire set of edges (i.e., whole relation) that determines the spectral characteristics. FedSACA+ ensures the validity of each basis by measuring the fraud detection utility of the kNN relation for specific bases at the spectral level. Specifically, the kNN relation is initially suppressed to avoid introducing potentially noisy structural information. Then, the initially suppressed relation is adaptively evaluated by self-learnable spectral channel attention, allowing beneficial spectral basis to be emphasized while discouraging less useful ones. Therefore, we denoise by ensuring filter reliability on each basis.

Q3: Questions about Figure 5.

Thank you for identifying this inconsistency. There is a typo in the FedSACA and FedSACA+ results on YelpChi in Table 1, and we have corrected it in the revised version.

Q4: Does a large negative initialization prevent the kNN relation from contributing?

We thank the reviewer for this insightful question. To further examine whether the initially suppressed relation is reactivated during training, we have provided the evolution of its learned confidence scores under different cc values in the added Appendix G. As shown in Figure 7, our default setting c=3c=3 and the extreme setting c=1000c=1000 retain learnable gradients under strong initial suppression, indicating successful reactivations. Moreover, FedSACA+ consistently outperforms FedSACA across the evaluated settings, supporting the contribution of the augmented kNN relation to the originally missing message-passing due to the fraud isolation.

Thank you again. Please let us know if our answers address your concerns, or whether we can provide further clarifications. We hope the new results and clarifications support a stronger acceptance rating.

Thanks for the response, I think some of them address several early concerns.

However, I still have some doubts about the proposed mechanisms. Particularly, it is not entirely clear how the confidence-aware aggregation resolves specral conflicts. I am also concerned about Fig 7(b). When c=1000, the sigmoid outputs and its gradient should be nearly zero, so I don't quite understand how the knn confidence scores can still increase during training. I don't think the current explaination fully address this.

So, I will keep my score.

Reviewer_1FfaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies federated graph fraud detection, focusing on spectral heterogeneity and isolated fraudulent nodes. The authors propose FedSACA, which combines spectral attention with confidence-aware aggregation to handle differences across clients. They further extend it to FedSACA+ by adding kNN-based graph connections with low initial confidence to reduce potential noise. Experiments on four datasets show promising results compared with existing methods.

Critical strengths
  1. The paper addresses relevant challenges in federated graph fraud detection. Considering both spectral heterogeneity and isolated nodes is useful for practical applications.

  2. The overall framework is clear, and the proposed methods achieve good results on four datasets, showing potential for fraud detection.

Critical weaknesses
  1. The definition of spectral channels needs clarification.

In Section 4.1, the authors treat different Chebyshev polynomial orders as separate frequency channels. However, each polynomial can affect multiple frequencies rather than representing a specific frequency band. Figure 3 only shows changes in confidence scores, without analyzing the actual frequency responses. More evidence is needed to support the claimed spectral adaptation.

  1. The confidence aggregation may still suffer from conflicting client information.

One motivation of the paper is to avoid averaging conflicting parameters across clients. However, Eq. (13) still directly averages the confidence masks. When clients have different spectral preferences, this may suppress useful information in the global model. The authors should explain how this issue is addressed.

  1. Important ablation studies are missing.

The model includes several components, but their individual contributions are not clearly evaluated. In particular, comparisons with standard FedAvg using the same Chebyshev filter, and with ordinary kNN augmentation without negative initialization, would help establish the benefits of the proposed designs.

  1. Some experimental comparisons and references need correction.

Section 5.2 discusses the limitations of FedSage+, but this method is not included in Table 1. PMP is marked as OOT on all datasets without explaining the timeout conditions. There are also reference errors, including the incorrect author names for the Adam paper and duplicate entries for McMahan et al. (2017).

  1. Some claims are stronger than the supporting evidence.

The main techniques are relatively straightforward, but the paper uses strong expressions such as "noise-free structure augmentation" and "self-correcting hypothesis test" without sufficient justification. Several sections also repeat the claimed contributions rather than providing deeper analysis of why the methods work or where they may fail.

Questions for authors

Q1. How do the authors justify treating different Chebyshev polynomial orders as separate frequency channels? Could they provide the learned frequency response curves?

Q2. In Eq. (13), could averaging the confidence masks suppress useful information when clients have conflicting spectral preferences? How does the method avoid this?

Q3. Could the authors provide more detailed ablations, particularly comparisons with FedAvg using the same spectral filter and standard kNN augmentation without negative initialization?

Q4. Why is FedSage+ discussed in Section 5.2 but not included in Table 1? What are the timeout settings for PMP? Could the authors also check the incorrect and duplicate references?

Q5. What evidence supports the claims of "noise-free structure augmentation" and "self-correcting hypothesis test"? How do these mechanisms differ from existing approaches?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 1Ffa (1/2)

We thank the reviewer for their careful reading and constructive suggestions.

W1&Q1: The definition of spectral channels needs clarification.

We sincerely thank the reviewer for pointing out this theoretical imprecision. We agree that individual Chebyshev polynomial orders span multiple frequencies rather than representing isolated frequency bands. To address this, we have revised our terminology throughout the manuscript: "bands" are now strictly referred to as "bases", and "per-frequency" has been changed to "per-basis". To provide the requested evidence analyzing the actual frequency responses, we have added a new figure 8 in section H of the revised appendix.

W2&Q2: The confidence aggregation may still suffer from conflicting client information.

We thank the reviewer for this important question. We would like to clarify that the parameter averaging criticized in our motivation primarily refers to the aggregation of filter coefficients θ\theta. In contrast, the confidence masks m\mathbf{m} are introduced as a mechanism with a different role. Specifically, Eq. (13) averages m\mathbf{m} as a basis-alignment step. Due to the scarcity of fraud labels on individual clients (e.g., min 2.23% in Elliptic and max 14.53% in YelpChi among our 4 datasets), locally optimized masks are prone to overfitting to local noise. Averaging m\mathbf{m} enables clients to jointly assess the importance of a basis, while still allowing clients to adjust their confidences according to the local spectral profile in the local training stage. The actual mitigation of spectral conflicts occurs in Eq. (14), where filter coefficients θ\theta are aggregated with per-basis confidence scores, ensuring the global filter selectively inherits reliable spectral knowledge across all bases. To further validate this design, we have conducted an experiment comparing our approach against a variant that keeps m\mathbf{m} strictly local (no averaging). The result shows that keeping m\mathbf{m} local degrades performance (AUC drops from 89.84% to 88.06%), demonstrating that mask averaging is essential to avoid client-level overfitting and jointly address inter-client spectral heterogeneity.

W3&Q3: Important ablation studies are missing.

We appreciate your advice on the ablation study. The following data is the ablation experiments conducted under YelpChi. This data indicates that the four modules are beneficial for fraud detection.

VariantLocal spectral adaptationConfidence aggregationkNN augmentationNegative initializationAUCF1-MacroAPG-MeanRecall
A✗✗✗✗73.48±0.8873.48_{\pm 0.88}63.13±1.1163.13_{\pm 1.11}56.18±1.2256.18_{\pm 1.22}68.66±1.4568.66_{\pm 1.45}69.27±1.3669.27_{\pm 1.36}
B✓✗✗✗81.67±0.3381.67_{\pm 0.33}69.79±0.7669.79_{\pm 0.76}60.78±0.8160.78_{\pm 0.81}78.76±1.1378.76_{\pm 1.13}78.84±1.1078.84_{\pm 1.10}
C (FedSACA)✓✓✗✗87.71±0.1887.71_{\pm 0.18}73.60±0.6873.60_{\pm 0.68}66.83±0.5566.83_{\pm 0.55}83.02±0.6183.02_{\pm 0.61}83.15±0.5883.15_{\pm 0.58}
D✓✓✓✗88.99±0.1588.99_{\pm 0.15}73.87±0.6173.87_{\pm 0.61}68.11±0.3268.11_{\pm 0.32}84.34±0.3584.34_{\pm 0.35}84.42±0.3184.42_{\pm 0.31}
E (FedSACA+)✓✓✓✓89.84±0.1289.84_{\pm 0.12}74.38±0.5674.38_{\pm 0.56}69.24±0.1269.24_{\pm 0.12}85.28±0.1585.28_{\pm 0.15}85.35±0.1485.35_{\pm 0.14}

W4&Q4: Some experimental comparisons and references need correction.

We thank the reviewer for pointing these out. We have added FedSage+ to Table 1, Table 2 and Table 3. The OOT condition for PMP specifically refers to exceeding the maximum wall-clock time limit of 72 hours without completing the 1000 communication rounds under our federated setting. To ensure a fair comparison, we strictly followed the federated training protocol detailed in Section 5.1. We have also revised the references.

AuthorsOpenReview ↗

Response to Reviewer 1Ffa (2/2)

W5&Q5: Some claims are stronger than the supporting evidence.

We thank the reviewer for this constructive criticism. We acknowledge that the method does not avoid introducing any noise at all. However, we restore filter effectiveness through the reconnection of kNN and exert a strong negative constant bias on the new induced relation to limit noise to the greatest extent. Moreover, the self-adaptive spectral channel attention also adjusts masks to a state that is most helpful in discovering the fraudulent nodes, which filters out most noise. Our experiments in Table 1 demonstrate that FedSACA+ outperforms FedSACA, further validating that the possible induced noise during kNN augmentation does not degrade the performance. We have replaced "noise-free structure augmentation" with "confidence-controlled structure augmentation".

The term "self-correcting hypothesis test" characterizes a gradient-driven process. It essentially performs a continuous, spectral-domain hypothesis test on the augmented relation: the null hypothesis (the kNN edges are pure noise) is initialized by the strong negative bias; the relation is only re-activated (rejecting the null) if the federated learning objective continuously drives the confidence score upward.

Compared to existing structure augmentation methods (e.g., FedSage+, IDGL), our approach offers two key advantages:

(i) Superior Spectral Granularity: Rather than a binary keep/drop decision, we apply a channel attention mechanism from a spectral perspective. This allows FedSACA+ to independently evaluate the utility of the kNN relation on each polynomial basis to benefit detection.

(ii) Low Complexity: The entire calibration is achieved by aggregating KK-dimensional confidence scores and filter coefficients. It incurs negligible communication and computation overhead O(N∣R∗∣K)O(N|R^*|K) compared to the generative models or full-graph similarity computations required by existing methods, while achieving superior performance without raw data sharing.

Thank you again. Please let us know if our answers address your concerns, or whether we can provide further clarifications. We hope the new results and clarifications support a stronger acceptance rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.