- 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:
- 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.
- 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.
- 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
- 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.
- 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
- 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 . After a softmax, at most two entries in the same group can be , 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), and the term is added on top, so with the coefficients sum to 1.5 and the parameters would grow with every update, which may be a typo. In addition, 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.
- 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 is deterministic, so is a point mass and the KL in is infinite. The theorem also bounds the training objective , which contains a weighted utility term, rather than the leakage . In the derivation of , and are used interchangeably, and I could not follow this step. The noise scale 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 the per-dimension noise is about 19 times the clipping norm , 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 , but Table 6 goes up to . 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
- In Eq. (6), along which dimension is the softmax normalized? As written, at most two entries per group survive , so how are the many nodes with more than 10 stitched neighbors in Table 15 obtained? Is a weight sum of in Eq. (16) intentional, and how is the parameter scale kept stable in the actual implementation?
- Is the threshold a score threshold or a pseudo-anomaly ratio? If it is a ratio, and 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 , but is a sum of inner products, so how is it restricted to this interval?
- How does stay finite under a deterministic encoder? What value does take after training? How is the noise calibrated when in the DP experiments, and which datasets and does Table 5 use? At the noise is far larger than the signal, so why does performance remain far above the variant without suturing?
- 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?
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