acceptodds
Under review as a conference paper at ICLR 2027

Mitigating Sparsity and Heterogeneity in Unsupervised Federated Graph Anomaly Detection

Abstract

Federated graph anomaly detection shows promising potential in academia and industry since it solves the limitations of traditional centralized GAD. However, existing methods suffer from local sparsity, intrinsic distribution shifts, and label reliance issues, resulting in limited semantic information in local clients, strong heterogeneity across clients, and vulnerability to label quality. To address these issues, we propose an unsupervised federated graph anomaly detection (UnFedGAD) framework, which enhances local semantic information through cross-client topology suture and addresses heterogeneity via personalized bidirectional parameter updating. Specifically, UnFedGAD discovers adequate semantics via cross-client communication with privacy protection, identifying the nearest neighboring client based on prototype similarity for each client. In addition, UnFedGAD employs prototype similarities to compute adaptive weights for global consensus generation and introduces a novel client updating mechanism to strike a dynamic balance between local and global models, effectively conquering client heterogeneity. To eliminate the dependence on labels, UnFedGAD generates pseudo-labels based on local-global consistency estimation for unsupervised training. Extensive experiments on eight public datasets demonstrate the effectiveness and advancement of UnFedGAD compared with nine baselines.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph anomaly detection, graph federated learning, unsupervised learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_M41pOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper studies node-level federated graph anomaly detection: a complete graph is split into several subgraphs held by different clients, and the clients cannot share raw data. The authors identify three problems with existing methods: edges cut across clients cause local sparsity, differing client distributions cause heterogeneity, and training depends on labels. They propose UnFedGAD, which has three main components:

  1. Cross-client topology suture. Clients upload node embeddings with Gaussian noise added. The server uses inner products of client prototypes to select a nearest client for each client, and then connects the nodes of that client to the local graph according to an affinity threshold, supplementing the missing neighborhood semantics.
  2. Bidirectional parameter updating based on prototype similarity. In server aggregation, weights are computed from client sample sizes and prototype similarity. In client updating, the global parameters, the nearest client's parameters, and the client's own parameters are mixed.
  3. Pseudo-label training. The sum of a node's inner product with its aggregated neighbor representation and its inner product with the client prototype is used as the score, and nodes scoring below a threshold are treated as pseudo-anomalies for training. The paper also gives an information-bottleneck-based upper bound on privacy leakage, and the appendix adds privacy attack experiments and a differential privacy extension. Experiments on 8 datasets compare against 9 baselines and report AUROC and AUPRC. The proposed method is best on all datasets, with an average AUROC gain of 5.84 points over the strongest baseline. There are also ablation, hyperparameter sensitivity, non-IID, noise robustness, and noise propagation analyses.
Critical strengths
  1. The problem setting is well targeted and the experiments are broad. Missing cross-client edges caused by subgraph partitioning is a real and meaningful problem in node-level federated graph anomaly detection. The paper puts suturing, personalized updating, and pseudo-labeling into one framework and provides fairly complete experiments: 8 datasets, 9 baselines, component ablations, comparisons against several federated aggregation strategies, and a non-IID setting built from 5 heterogeneous datasets.
  2. The authors proactively probe their own weaknesses. The appendix adds three kinds of privacy attacks, namely representation reconstruction, attribute inference, and edge reconstruction, as well as a differential privacy extension, and explicitly states that IB with added noise does not constitute formal differential privacy. The noise propagation experiment also acknowledges that suturing introduces heterophilic noise.
Critical weaknesses
  1. Several definitions in the method section do not match the paper's own data. Eq. (6) applies a softmax to the affinities and then binarizes with K^=0.5\hat{K}=0.5. After a softmax, at most two entries in the same group can be ≥0.5\ge 0.5, so I do not see how the suture could produce what Table 15 reports, where Cora client 1 has over fourteen hundred stitched edges by a lower-bound count. In Eq. (16), wˉmg+wˉmm′=1\bar{w}_{mg}+\bar{w}_{mm'}=1 and the term α Θlm(t)\alpha\,\Theta_l^m(t) is added on top, so with α=0.5\alpha=0.5 the coefficients sum to 1.5 and the parameters would grow with every update, which may be a typo. In addition, KK reads like a score threshold in Section 3.4 but is interpreted as a pseudo-anomaly ratio in Appendix H. Under the ratio reading, 20%–30% of nodes become pseudo-anomalies while the true anomaly rate is only 3%–7%. Section 3.1 also says a larger score means more anomalous, whereas the consistency scores later are lower for anomalies. Table 1 depends on these formulas, so without clarification it is hard to tell where the improvement comes from.
  2. The evidence given does not support the paper's privacy claims. Theorem 1 is presented as a strict upper bound on leakage, but under the paper's own setting it has no practical meaning. The encoder H=MLP(X)\mathbf{H}=\mathrm{MLP}(\mathbf{X}) is deterministic, so p(h∣x)p(h\mid x) is a point mass and the KL in M1M_1 is infinite. The theorem also bounds the training objective LIB\mathcal{L}_{\mathrm{IB}}, which contains a −0.01-0.01 weighted utility term, rather than the leakage I(X;H^)I(X;\hat{H}). In the derivation of M2M_2, XX and X^\hat{X} are used interchangeably, and I could not follow this step. The noise scale σ\sigma is learnable, which also means there is no guarantee in advance. Appendix C itself admits that IB with noise is not differential privacy, which is inconsistent with the main text. The supplementary DP experiments are also unconvincing: by Eq. (21)–(23), at ε=0.5\varepsilon=0.5 the per-dimension noise is about 19 times the clipping norm CC, yet AUROC only drops from 77.85 to 70.12, still far above the variant without suturing, and I cannot explain this result. Eq. (23) only applies to ε<1\varepsilon<1, but Table 6 goes up to ε=8\varepsilon=8. The protocol also uploads and distributes noisy embeddings of every node, and Eq. (16) obtains other clients' local parameters, and none of these exposure surfaces is discussed in the privacy analysis.
Questions for authors
  1. In Eq. (6), along which dimension is the softmax normalized? As written, at most two entries per group survive K^=0.5\hat{K}=0.5, so how are the many nodes with more than 10 stitched neighbors in Table 15 obtained? Is a weight sum of 1+α1+\alpha in Eq. (16) intentional, and how is the parameter scale kept stable in the actual implementation?
  2. Is the threshold KK a score threshold or a pseudo-anomaly ratio? If it is a ratio, K=0.2K=0.2 and 0.30.3 are far above the true anomaly rates, so what is the actual precision of the pseudo-anomalies? Which direction of the consistency score is used in the final evaluation? Eq. (11) requires s∈(0,1)s\in(0,1), but s=sG+sLs=s_G+s_L is a sum of inner products, so how is it restricted to this interval?
  3. How does M1M_1 stay finite under a deterministic encoder? What value does σ\sigma take after training? How is the noise calibrated when ε>1\varepsilon>1 in the DP experiments, and which datasets and ε\varepsilon does Table 5 use? At ε=0.5\varepsilon=0.5 the noise is far larger than the signal, so why does performance remain far above the variant without suturing?
  4. Please describe the evaluation protocol and check the tables. How are training and test nodes split? The five datasets in Table 4 have different feature dimensions, so how are the encoder parameters aggregated? How are the baselines adapted to federated anomaly detection, and were they tuned separately? In the ACM row at epoch 100 of Table 16, Homo plus Heter is 1253 while After minus Before is 3775, so is Homo incorrect?
Reviewer_o487OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes UnFedGAD for unsupervised node-level federated graph anomaly detection, targeting missing cross-client structural information and client heterogeneity. The framework shares noisy latent representations, uses client prototypes to identify a neighboring client, and augments local graphs with cross-client connections selected through an affinity threshold. Local and global consistency scores generate pseudo-labels for training, while prototype-dependent weights combine global, neighboring-client, and local model parameters. Experiments on eight public graph datasets report improvements over nine baselines, supported by component ablations, alternative federated updating strategies, heterogeneity experiments, and privacy-related analyses.

Critical strengths

1-Missing cross-client context and heterogeneous client distributions are relevant challenges for node-level federated graph anomaly detection. Combining semantic neighbor supplementation with personalized parameter updates is a reasonable approach to these challenges, especially when anomaly labels are unavailable. The proposed components have a clear connection to the practical limitations identified in the motivation.

2-Table 1 evaluates eight datasets against nine baselines using both AUROC and AUPRC, reporting means and standard deviations over five runs. UnFedGAD achieves the highest reported mean performance in all dataset–metric combinations. The component ablations and comparisons with alternative federated updating strategies provide additional evidence about the usefulness of the proposed design. This breadth of evaluation is a positive aspect of the paper, although the methodological inconsistencies discussed below currently limit how confidently I can interpret the results.

Critical weaknesses

1-In Section 3.3, the information-bottleneck objective is defined as LIB=I(X;H^)−0.01I(X^;H^)\mathcal{L}{\mathrm{IB}} = I(X;\widehat{H}) - 0.01 I(\widehat{X};\widehat{H}). However, the final conclusion of Appendix B is only LIB≤M1−0.01M2\mathcal{L}{\mathrm{IB}} \leq M_1 - 0.01M_2, whereas Theorem 1 presents the same expression as an upper bound on privacy leakage. Even if the intermediate inequalities are accepted, an upper bound on this difference of mutual informations does not establish the corresponding bound on I(X;H^)I(X;\widehat{H}), the quantity that Section 3.3 seeks to minimize for privacy protection. Subtracting a utility-related mutual-information term changes the quantity being bounded.

The definition of M1M_1 also differs between the main-text theorem and the statement in Appendix B, using pre-noise and post-noise representations, respectively. These definitions and the assumptions behind the bound need to be made consistent.

I appreciate the empirical attacks and the explicitly scoped DP extension in Appendix C. Nevertheless, these do not resolve the mismatch between Theorem 1 and its proof. The appendix also distinguishes its fixed-encoder, fixed-topology representation-release guarantee from an end-to-end guarantee for the complete training protocol. Since sharing node representations is essential to the proposed topology-suture mechanism, a precise and valid privacy statement materially affects the assessment of the method.

2-Equation (6) applies softmax to the cross-client similarity matrix and retains entries satisfying s^ij≥K^\widehat{s}_{ij} \geq \widehat{K}, while Section 4.1 sets K^=0.5\widehat{K}=0.5. Under the standard row-wise interpretation of softmax, each row sums to one. Therefore, the number of retained neighbors for a node must satisfy niK^≤1n_i\widehat{K}\leq 1, giving an upper bound of two cross-client neighbors at the stated threshold.

However, Appendix L, Table 15 reports many nodes with 11–20 cross-client neighbors; for example, client 1 contains 83 nodes in this bin. This is incompatible with the above interpretation of Equation (6) under the default setting. Moreover, the prose immediately following Equation (6) describes thresholding the inner products, whereas the equation thresholds the normalized affinities. These operations can produce substantially different graphs.

The manuscript does not explain whether the discrepancy arises from a different normalization rule, a different thresholding operation, or statistics accumulated across rounds. Because topology suture is a central contribution and the ablations indicate that performance depends strongly on it, this ambiguity makes it difficult to verify which graph-construction procedure produced the reported gains.

Questions for authors

1-Please provide a corrected statement and derivation that distinguish a bound on I(X;H^)I(X;\widehat{H}) from a bound on LIB\mathcal{L}_{\mathrm{IB}}, and clarify the consistent definition of M1M_1. If the result is intended only as a variational bound on the training objective, how should the main-text privacy claims be revised? Please also clarify how these claims relate to the explicitly limited scope of the DP extension in Appendix C.

2-Please specify the softmax dimension and whether the threshold is applied to raw similarities or normalized affinities. How does the implemented rule, with K^=0.5\widehat{K}=0.5, produce the neighbor-count distribution in Table 15? Please clarify whether these counts describe a single constructed graph or accumulate connections across rounds, and confirm which implementation was used for Table 1 and the topology-suture ablation.

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

In this paper, the authors propose UnFedGAD for unsupervised node-level anomaly detection on distributed graphs, addressing missing cross-client neighborhood information and heterogeneous client distributions. The framework shares perturbed node representations, uses prototype similarity to select neighboring clients, and augments local graphs with cross-client connections. It generates pseudo-labels from local and global consistency scores and personalizes model updates by combining global, neighboring-client, and local parameters. An information-bottleneck objective is introduced to limit information leakage. Experiments on eight datasets containing synthetic or real anomalies demonstrate that UnFedGAD outperforms the compared federated node anomaly detection methods.

Critical strengths
  1. The proposed method is well motivated by the challenges introduced by graph partitioning, which addresses the coupled effects of missing cross-client context and heterogeneous client distributions.
  2. The detector combines neighborhood consistency with similarity to the client prototype, incorporating complementary views of node abnormality. This provides a clear rationale for integrating local and broader contextual information into pseudo-label generation.
  3. The evaluation covers both synthetic and real anomalies and includes directly relevant federated graph anomaly detection baselines. The consistent performance improvements in the experiment provide encouraging evidence for the overall framework.
Critical weaknesses
  1. The privacy analysis does not establish the claimed protection for the complete federated training process. The main proof bounds an information-bottleneck objective, and the additional differential-privacy analysis does not cover learned encoders or exchanged model updates.
  2. The reported cross-client neighborhood statistics are not explained by the stated topology-construction rule. More deeply, which graph augmentation procedure leads to the improvements?
  3. The aggregation ablations do not isolate the claimed benefit of explicitly retaining the previous global model. With normalized weights, the server update reduces to a weighted average of current local models, so removing individual terms does not establish a separate knowledge-preservation mechanism.
  4. The training protocol does not clearly substantiate the fully unsupervised setting. The description permits labeled warm-up during the early training stage, which is contradicted by the setting.
Questions for authors
  1. How does the information-bottleneck bound establish the stated privacy guarantee, and does this guarantee cover encoder training and exchanged model updates?
  2. What topology-construction procedure was actually used in the experiments, and how does it explain the cross-client neighborhood sizes described in the paper?
  3. Given that the normalized server update reduces to a weighted average of current local models, what distinct contribution of the previous global model do the aggregation ablations demonstrate?
  4. Were anomaly labels used during warm-up in any reported experiments? Please clarify the actual warm-up objective and the supervision it requires.

Comments on the paper

1 comment

Dear all,

We're the half-way through the discussion phase. The authors have provided their response to your comments. Some of you have started involving discussion, but all of us can take action now.

If the provided responses do not address (even part of) your concerns, this is your chance to raise it. Please engage with the authors as soon as possible, we need to keep the conversation going!

Thanks,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.