acceptodds
Under review as a conference paper at ICLR 2027

Fed-DREAM: Bridging Fragmented Views in Hybrid Federated Learning via Relational Learning

Abstract

Hybrid federated learning (HBFL) addresses scenarios in which sample and feature spaces are fragmented across clients, so that each client holds only a partial view with limited predictive capacity. Existing HBFL methods aggregate views through feature or prediction sharing with static fusion strategies, missing the inter-client relations essential for a holistic representation. We propose **Fed-DREAM** (Federated Divisive RElational Aggregation and Mixup), a framework for collaborative relational learning in HBFL built on a three-stage pipeline: (1) *federated divisive clustering*, which constructs structurally aligned global clusters without exposing raw data, (2) *multi-sample representation-level mixup* for privacy-preserving synthetic feature sharing, and (3) an*edge-conditioned gated graph neural network* that fuses fragmented views by conditioning message passing on inter-client feature-membership relations. Experiments across tabular, image, and text modalities show that Fed-DREAM consistently outperforms state-of-the-art benchmarks while exhibiting lower membership-inference leakage than the highest-utility baseline.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
hybrid federated learning, multi-view learning, graph neural networks, privacy-preserving representation sharing

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_pC8oOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses hybrid federated learning (HBFL), where both sample and feature spaces are fragmented across clients. It proposes Fed-DREAM, a framework that integrates federated divisive clustering, multi-sample representation-level mixup, and an edge-conditioned gated graph neural network to enable cross-client representation fusion. Experiments on FashionMNIST, CIFAR-100, and Multi-News report improved predictive performance over existing methods, while membership inference attacks are conducted to evaluate its privacy-preserving capabilities.

Critical strengths
  1. The paper addresses the dual heterogeneity of sample and feature spaces in hybrid federated learning, a problem with research value and practical relevance.

  2. The proposed method combines clustering, representation-level mixup, and graph neural networks to facilitate cross-client collaborative learning without directly centralizing raw data.

  3. The experiments cover various heterogeneous configurations and include ablation studies, privacy evaluations, and system cost analyses.

Critical weaknesses
  1. Hybrid FedGraph (2024) has already proposed using a server-side GNN to aggregate client representations, modeling inter-client interactions based on feature-sharing relationships, and protecting representation privacy through clustering. It also introduces an encoder privacy score based on preimage diameter and bilevel optimization, which is highly similar to the approach adopted in this paper. However, this directly relevant work is neither cited nor discussed. The authors should clearly clarify the substantive novelty of Fed-DREAM over Hybrid FedGraph.

  2. The proposed framework combines federated clustering, Dirichlet Mixup, edge-conditioned message passing, and gated attention into a three-stage pipeline. However, the paper does not sufficiently explain what fundamental challenges this combination addresses that existing methods cannot handle.

  3. Each client independently samples Dirichlet weights within the same global cluster, and the resulting mixed representations with the same index are combined into a single graph sample. However, these representations do not necessarily correspond to the same original instances or mixing coefficients, so instance-level semantic consistency across client nodes is not guaranteed. This may also introduce a substantial distribution shift between training and inference.

  4. Fed-DREAM additionally trains a server-side ECG-GNN on top of the local encoders, whereas some baselines primarily rely on local model optimization, parameter aggregation, or prediction fusion. Therefore, the reported improvements may partly result from the additional server-side model capacity and training procedure, raising concerns about the fairness of the comparisons.

  5. The authors convert the original Multi-News summarization task into a multi-label classification task using labels automatically generated by gpt-4o-mini. This is not a standard Multi-News evaluation setting, and the paper lacks validation on established multi-label classification datasets with ground-truth annotations.

  6. The 12 HBFL configurations evaluated in this paper are primarily constructed by the authors rather than directly following the original experimental protocols of HyFEM, FedHD, or FedLabSync. Although evaluating all methods under the same partitions is reasonable, the current results only demonstrate the effectiveness of Fed-DREAM under these selected synthetic settings and are insufficient to establish its general applicability to broader HBFL scenarios.

  7. Privacy preservation is presented as a major contribution, but the privacy evaluation mainly relies on black-box, loss-based membership inference attacks against the deployed system. It does not directly assess the training-time privacy risks arising from the mixed representations, cluster memberships, and aggregated distances accessible to the server. In particular, when a client holds only one sample within a global cluster, Mixup degenerates into directly releasing the original embedding.

  8. The experiments compare Fed-DREAM with only three methods published in or before 2024, without including more recent approaches. The lack of comparisons with newer methods raises concerns about whether the reported improvements adequately demonstrate the novelty and contribution of the proposed framework.

Questions for authors
  1. The paper introduces three parameter-sharing configurations: no sharing, selective sharing, and global sharing. However, these strategies are applied to different datasets and experimental settings without a systematic comparison under the same data conditions. What are the respective applicability and limitations of these strategies?

  2. In Tables 13–14, removing edge features results in only a minor performance degradation, whereas replacing ECG-GNN with a simple Mean-pool MLP leads to a much larger drop. Can the authors clarify whether the improvements primarily come from modeling feature relationships or from the additional server-side model capacity and nonlinear fusion capability?

  3. When a client holds only one sample within a global cluster, Algorithm 5 directly releases the original embedding. Could the authors report how frequently this occurs across the experimental settings and explain whether the proposed privacy protection remains effective in such cases?

  4. Table 10 reports three independent runs for each of the two Multi-News configurations, with the results pooled for a one-sided paired t-test. However, based on the six paired observations described in the paper, the reported means and standard deviations appear inconsistent with the p-values of 0.018, 0.016, and 0.006. Could the authors clarify how these p-values were computed?

9 comments
AuthorsOpenReview ↗

Response to Reviewer pC8o (1/5)

We thank the reviewer for a thorough review. We take the first weakness especially seriously and answer it first, because it concerns what the paper contributes beyond prior work. For the remaining points we give the evidence already in the paper where it exists and report the additional experiments we ran: Hybrid FedGraph as a baseline, a reconstruction attack by the server, an analysis of single-member releases, and FM3–FM5 without parameter sharing.

W1 — Relation to Hybrid FedGraph (arXiv 2404.09443, 2024)

The reviewer is right that Hybrid FedGraph should have been cited and discussed; it is a directly related prior work, and omitting it was an error that we correct in the revision. We state what the two works share and what Fed-DREAM adds.

Both works fuse client representations at the server with a graph model whose edges encode feature sharing, and both cluster samples so that the server does not receive individual embeddings; the privacy score of Appendix A.2 (preimage diameter, bi-level estimation) is Hybrid FedGraph's construction, which we use unchanged and now credit. Fed-DREAM differs in the three stages that follow local training. Hybrid FedGraph clusters at the server, grouping samples at random within each class because following the data geometry there would require disclosing the embeddings; Fed-DREAM clusters in the shared distance geometry under threshold homomorphic encryption, so the server sees only encrypted pairwise distances, medoid indices and memberships (Sec. 3.3). Hybrid FedGraph releases one plain mean per cluster; Fed-DREAM releases Dirichlet mixtures with server-blind weights, which obfuscates and augments at once (Sec. 3.4). Hybrid FedGraph's GCN averages neighbor messages, with the edge vector entering only the first layer; ECG-GNN conditions message generation and gating on the edge attribute at every layer and is trained with client dropout for partial participation (Sec. 3.5). Hybrid FedGraph was evaluated in one vertical setting against clustering variants only.

Results. Hybrid FedGraph is now a baseline on all twelve settings, run under the same partitions, local encoders and server-side hyperparameters as Fed-DREAM (Tables 1, 2, 9 and 10). Fed-DREAM is higher in every FashionMNIST and CIFAR-100 setting at both participation rates: on average by 5.7/5.2 pp on FashionMNIST (P = 1.0/0.3) and by 10.7/9.3 pp on CIFAR-100. On Multi-News, Micro-F1 is higher by 3.5–7.1 pp and mAP by 4.2–6.0 pp; Hybrid FedGraph has the lower Hamming Loss in both settings, and Sec. 4.3 explains why we regard Micro-F1 and mAP as the representative metrics for this task.

Revision. Sec. 2; Appendix A.2; Tables 1, 2, 9 and 10; Sec. 4.3; Appendix D.

AuthorsOpenReview ↗

Response to Reviewer pC8o (2/5)

W2 — What fundamental challenge the combination addresses

The challenge is stated in the introduction but, we agree, not sharply: when a client's view covers only a small part of the feature space, static fusion of the clients' outputs fails, because neither averaged parameters (HyFEM, FedHD) nor averaged predictions (FedLabSync) can express which client is informative for which sample. Table 13 gives the clearest evidence: on FM1, where each client holds 2.3% of the features, fusing the clients' embeddings without relational structure (Mean-pool MLP, 0.6641) is worse than simply averaging their independent predictions (w/o node connection, 0.6940), because embeddings from clients that see almost disjoint features blur when averaged; only when the fusion is conditioned on which clients overlap with which (ECG-GNN, 0.7493) does early fusion become a gain. On the high-overlap settings the order is the expected one (late fusion < plain early fusion < relational early fusion, Table 14). The three stages thus serve one purpose: the clustering and mixup make a server-side relational fuser trainable without releasing individual embeddings, and the relational fuser is what the low-overlap regime needs.

W3 — Instance-level consistency across client nodes and train/inference shift

This is a correct description of the design, and it is deliberate. The mixing weights must be sampled independently on each client, because if the server generated or observed them it could solve the mixtures for the underlying embeddings exactly (Appendix A.4, "Server-Blind Weight Generation"); independence is the price of keeping the release one-way. Consistency is therefore approximate and enforced at the cluster level: all nodes of synthetic sample i are mixtures of samples from the same cluster GkG_k.

How much this costs at inference, where the server sees real embeddings of one sample, is measured. The Raw-embedding reference of Sec. 4.5 trains the same ECG-GNN on the original embeddings and labels; Fed-DREAM is within 2.9 pp of it in every setting at P = 1.0 and above it in six of twelve, and FashionMNIST accuracy is flat up to C = 20 (Appendix F.3), so the shift is small where clusters are tight and the mixup regularization compensates for it, as in centralized mixup. Sec. 3.4 states both facts, and Sec. 4.5 measures the resulting shift with the Raw-embedding reference.

W4 / Q2 — Server-side capacity versus relational modeling

Collaboration in the benchmarks. None of the benchmarks is local-only: FedLabSync aligns the predictions of label-sharing clients over communication rounds, FedLabSync-WA also averages the shared parameters, HyFEM and FedHD couple the local models through a matched server model and global-gradient tracking, and Hybrid FedGraph trains a server-side graph model. Fed-DREAM uses the same local architectures, training budget and communication rounds as the benchmarks (Appendix C.5) and the same parameter sharing as FedLabSync-WA where available (removing that sharing on FM3–FM5 costs 0.5–1.5 pp, Appendix E.2); its collaborative stage is the ECG-GNN. The comparison is between collaborative procedures, not between a local model and one with an added server.

Capacity versus relational modeling. Three comparisons separate the two factors. (i) The Mean-pool MLP keeps the node-encoding module, width (256), dropout and schedule of ECG-GNN and removes only the single message-passing layer (edge-conditioned messages and gating); Fed-DREAM exceeds it in all ten settings, by 2.2 pp on average and by 8.5/7.3 pp on FM1/FM2 (Tables 13–14), where plain pooling even falls below averaging the clients' independent predictions (see W2). (ii) Hybrid FedGraph, a server-side graph model on the same encoders with the same width (see W1), trails Fed-DREAM in all ten settings, by 0.5–14.4 pp and 8.8–15.1 pp at P = 1.0. (iii) Fed-DREAM exceeds Oracle feature union, a centralized model with full feature access, in all five FashionMNIST settings. Server-side capacity alone does not produce the gain.

Q2, answered directly. The reviewer reads Tables 13–14 correctly. The large drop from replacing ECG-GNN by the Mean-pool MLP is the contribution of relational message passing and gating as a whole; the edge feature is one input to that mechanism, with the smallest marginal value where overlaps are large and uniform (CIFAR crops, FM3–FM5) and the largest on FM1–FM2. The smaller effects are consistent and significant: removing the edge feature lowers accuracy in nine of the ten settings (paired one-sided t-tests per dataset, p = 0.001 on FashionMNIST and 0.018 on CIFAR-100), and removing the gate lowers it in all ten (p = 0.013 and 0.002). Sec. 4.4 states this reading: relational fusion is the dominant factor, and edge features and gating add smaller but consistent and significant gains.

AuthorsOpenReview ↗

Response to Reviewer pC8o (3/5)

W5 — Multi-News labels generated by gpt-4o-mini

We agree that this is not the standard Multi-News evaluation, and we explain why we built the task this way. The hybrid setting requires a text dataset in which one sample is covered by several parties that hold different content and share one target. To our knowledge, no public text dataset with human annotations meets this condition. Multi-News supplies the structure naturally, several articles on the same event, one per client, but its original task, abstractive summarization, cannot serve as the shared target: the compared methods fuse the clients' outputs by averaging logits or by a graph-level prediction, which is defined for a label vector, not for generated text. We therefore derived a 54-label vocabulary from the reference summaries and assigned labels with gpt-4o-mini, using the prompts given in Fig. 1 (Appendix B.1). On the comparison itself: every method is trained and evaluated on exactly the same label set, so any noise in the generated labels affects all methods alike and cannot favor one of them; the prompts are provided so that the labeling can be reproduced and inspected.

W6 — Synthetic HBFL configurations

To our knowledge there is no public real-world HBFL dataset, i.e., one that records which parties hold which samples and which features. Every HBFL study we are aware of, including HyFEM, FedHD, Hybrid FedGraph and FedLabSync, therefore partitions standard datasets synthetically, and their protocols differ from each other (feature groups in HyFEM, random feature subsets in FedHD, a vertical FashionMNIST split in Hybrid FedGraph), so there is no single original protocol under which all methods could be compared. Evaluating all methods under one controlled family of partitions is the alternative, and the family is broad: sample ratios from 6% to 40%, feature ratios from 2.3% to full-size crops, 3 to 50 clients, three sharing regimes, and full and partial participation (Tables 5 and 9); FM3–FM5 follow the grouped-feature protocol of Zhang et al. (2020b) directly, as Appendix B.2 states.

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

This paper proposes Fed-DREAM, a framework for hybrid federated learning (HBFL), where both samples and features are fragmented across clients. The motivation is that existing HBFL methods fuse client views through static parameter aggregation or prediction sharing, which ignores inter-client relations and exchanges unobfuscated predictions or representations during training. The contributions are (1) a privacy-preserving federated divisive clustering that aggregates pairwise embedding distances under threshold homomorphic encryption and forms global clusters of size [CC,2$C$-1], (2) a multi-sample representation-level mixup in which each client releases Dirichlet-interpolated synthetic embeddings and logits within each cluster, and (3) an edge-conditioned gated GNN (ECG-GNN) at the server that conditions both message generation and gating on inter-client feature overlap. Experiments cover FashionMNIST, CIFAR-100, and Multi-News under 12 synthetic partitions, compared against HyFEM, FedHD, and FedLabSync. Fed-DREAM achieves the best result in every setting, with gains of 0.01 to 0.73 percentage point (pp) on FashionMNIST and 1.59 to 3.99 pp on CIFAR-100 (In Table 1).

Critical strengths
  • This paper identifies a problem and a potential gap in hybrid federated learning (HBFL) fusion, where each client observes only a subset of features together with a subset of samples. Existing HBFL methods focus on static model alignment through fixed, sample-agnostic aggregation weights or simple prediction sharing, which discards the relational context essential for reconciling fragmented views.

  • The experiments are comprehensive, including 3 modalities (tabular, text, and image modalities) across 12 settings (FM1–FM5, CI1–CI5, MN1–MN2), three reference conditions (local-only, oracle feature union, and zero-filled union), paired one-sided t-tests over runs with re-randomized partitions.

Critical weaknesses

The paper presents a privacy-preserving representation sharing mechanism combining federated divisive clustering and multi-sample representation-level mixup as one of its main contributions. However, the privacy evaluation does not test the server as an adversary. The membership inference audit only tests an external adversary querying the deployed model.

Questions for authors
  • The definition of the edge attribute amja_{mj} is somewhat confusing, since the overlap between clients appears to vary across samples on CIFAR-100 and Multi-News. Can you explain how amja_{mj} is computed for a synthetic graph sample?
3 comments
AuthorsOpenReview ↗

Response to Reviewer aFTR

We thank the reviewer for the careful reading and for the positive assessment of the problem framing and the experimental breadth. We address the weakness and the question below and summarize the revisions at the end.

W1 — The privacy evaluation does not test the server as an adversary

We agree that this is the right adversary to test, and we have now done so.

Which server-side threat is relevant. The server knows the index allocations (Nm,Fm)(N_m, F_m) by construction (Sec. 3.1), so it knows which samples each client holds; membership is not a secret with respect to the server, and a membership-inference audit against it would be vacuous. What the clustering and mixup stages protect from an honest-but-curious server is the content of the training-time releases: the individual local embeddings behind each mixture and, through them, the raw features xm,nx_{m,n}. The relevant server-side attack is thus reconstruction, the attack class we discussed but did not measure in Sec. 3.4 and Appendix A.4.

Attack. An honest-but-curious server trains a decoder from a released embedding to the raw input of a victim client. We consider two adversaries: (i) a mixup-trained decoder fitted on Dirichlet-mixed releases, and (ii) a raw-trained decoder fitted on unmixed (embedding, input) pairs of the victim, a strictly stronger prior. Both are given raw inputs in the victim's feature view for decoder training, a supervision a real server does not have, so the results upper-bound what the server can learn. The target of each release is the raw input of the member with the largest Dirichlet weight. Decoder-training and evaluation pools are disjoint (80/20), with the clustering run independently on each. We attack FM1, FM4, CI1 and CI5, sweeping C ∈ {1, 2, 3, 4, 5, 7, 10, 15, 20} over five seeds; at C = 1 the releases are unmixed and the two adversaries coincide.

Results. Reconstruction quality degrades monotonically with C in all four settings and against both adversaries: from C = 1 to the deployed C, the RMSE on FM1 rises from about 0.07 to 0.14–0.18 and the PSNR on CI1 and CI5 falls from about 28–29 dB to 19–21 dB, and both keep worsening up to C = 20; cosine similarity, SSIM and LPIPS follow the same trend (Appendix F.2, Figures 5–6). For releases that mix several members, the protection against the server therefore comes from the mixing and grows with C, which is direct evidence that the clustering and mixup stages work against the server as intended.

Revision. Appendix F.2; Sec. 4.5.

Q1 — How is the edge attribute amja_{mj} computed for a synthetic graph sample?

amja_{mj} is a fixed attribute of the client pair (m, j), not of the sample: it is the indicator vector of Fm∩FjF_m \cap F_j, the raw features held by both clients, and it is identical for every original and every synthetic graph sample. We see why the submitted text suggested otherwise, and the revision clarifies the two datasets in question.

  • CIFAR-100. Each client holds one fixed r × r crop region applied to all of its images. "A random r × r crop of each image" in Sec. 4.1 read as a per-image crop; the revision says "a fixed r × r crop region (r ∈ [16, 32]), drawn at random once per client and applied to all of its images". The pixel overlap between two clients is thus the same for all samples.
  • Multi-News. A client's feature set is the token vocabulary of the articles it holds, and amja_{mj} is the intersection of the two clients' vocabularies; since both are fixed per client, the overlap is again the same for all samples.

Revision. Sec. 3.5.1.

Thanks for the reply. I will keep my orginal positive score.

Thank you for reading our reply and for confirming the score. We appreciate your constructive review, which helped us improve the paper.

Reviewer_FAC1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a federated learning method for scenarios where both features and samples are fragmented. First, representation models are trained. Clients aggregate and share pairwise sample distances in an encrypted manner, followed by a two-layer clustering procedure performed locally on each client and at the server side based on these distances. Next, each client generates synthetic embeddings and logits for every cluster and sends them to the server, which aggregates these outputs to build a synthetic training dataset for training the graph aggregation model. During inference, the encoded outputs from clients are aggregated into an inference graph at the server to yield global predictions.

Critical strengths
  1. Valuable research problem: This work targets federated learning with simultaneous sample and feature fragmentation. The research setting is practically relevant, and related literature remains relatively scarce.

  2. Well-designed pipeline: It integrates privacy-preserving clustering, representation-level Mixup, and relation-aware graph modeling into an end-to-end framework suited for the target scenario.

  3. Comprehensive experiments: Evaluations are conducted across tabular, image, and text modalities, including ablation studies, privacy attack assessments, and computational overhead analysis.

Critical weaknesses
  1. A critical assumption of this work is that all sample holders possess labels, which is not a standard assumption for general HBFL. In more prevalent real-world settings, labels can be viewed as a special type of feature that resides only on one or a subset of clients. Label information is often highly sensitive and far less likely to be shared than ordinary features. The requirement that every label is universally available restricts the method to a narrow cross-silo setting and substantially limits its practical applicability.

  2. Ambiguity regarding parameter sharing. The paper states parameter sharing is optional, yet this point raises confusion about whether the method imposes constraints on network architectures. Without architectural constraints, semantically compatible modules cannot be guaranteed under misaligned feature spaces. If constraints are enforced, partial parameter sharing becomes feasible but introduces a host of attendant issues. It remains unclear whether the networks adopted in this work match those used by state-of-the-art baselines, and whether semantic correspondences between features must be pre-known. This suggests the models are largely trained independently.

  3. Scalability bottlenecks. The method computes distances for all shared sample pairs before encrypted aggregation, which incurs combinatorial complexity and easily leads to computational explosion. This issue is already reflected in the empirical estimates in the paper. It is necessary to propose potential mitigation strategies or directions for improvement.

  4. Averaging distances across different clients is not always justified, especially when feature partitions carry distinct semantic meanings. Consider two clients from a hospital and a bank respectively. Similarity between two samples in medical records does not imply similarity in their financial profiles. Two samples a and b may be close in the medical view but divergent in the financial view. The averaged distance may not correspond to any meaningful downstream task; it merely produces an arithmetic compromise between incommensurable views and may mislead both clients. A core missing assumption must be made explicit: local distance metrics from different clients contribute consistently to global sample similarity. Put simply, if samples a and b are close from one client’s perspective, they are expected to also be close from the perspectives of other clients.

  5. The encoder models are mostly trained independently on each client. The framework relies on a shared graph-level loss to pull embeddings into a common latent space. However, this indirect regularization is relatively weak compared with explicit constraints such as contrastive loss or mutual information objectives.

Questions for authors
  1. Can the method be extended to the more general HBFL setting where labels are available only on one or a subset of clients?

  2. What are the precise architectural and semantic alignment prerequisites for parameter sharing? Are comparable network architectures consistently used across different methods?

  3. For the high complexity introduced by pairwise distance computation over all samples, can optimizations such as neighbor sampling or post-clustering distance calculation be adopted?

  4. Does the paper account for the semantic nature of feature partitions across clients?

  5. When local encoders are trained independently, can the graph-level loss reliably align the semantics of embeddings produced by different clients?

6 comments
AuthorsOpenReview ↗

Response to Reviewer FAC1 (1/3)

We thank the reviewer for a detailed review that engages with the assumptions behind the method. W1, W2 and W4 concern assumptions that the submission left implicit; for these we state the assumption and where the revision now states it. W3 is addressed with a new experiment and W5 with the existing ablations. We follow the reviewer's numbering and pair each weakness with the corresponding question.

W1 / Q1 — Label availability

We agree the assumption should be visible. Fed-DREAM assumes that each client holds the label of every sample it holds, the cross-silo situation in which each institution records the outcome of its own patients or customers. This is the assumption made by most HBFL methods, including all the benchmarks we compare against: HyFEM, FedHD, FedLabSync, and FedGraph (Jang et al., 2024), which the revision adds as a baseline; it restricts the comparison to the setting these methods share rather than to a narrower one.

Extension to partially labeled clients (Q1). The clustering stage (Sec. 3.3) and the ECG-GNN (Sec. 3.5) do not use labels; labels enter only through the mixed logits of Sec. 3.4, which the server averages over contributing clients into the collective target. A client without labels can therefore still contribute embeddings to every cluster it participates in, and only the logit average is restricted to the label holders: in Algorithm 5, line 13 becomes an average over CCiCC_i ∩ {label holders} instead of CCiCC_i. The one condition is that every cluster has at least one label holder, which the size-constrained clustering can enforce as it enforces the minimum size.

Revision. Sec. 3.1; Sec. 5.

W2 / Q2 — Parameter sharing and architectural constraints

The reviewer's last sentence is correct: the local encoders are trained independently by default. Parameter sharing is not a contribution of this work; we adopt it as practiced in prior HBFL work (Zhang et al., 2020b; Gao et al., 2023) and only organize that practice into the three configurations of Sec. 3.2. Our contribution lies in the stages that follow local training (clustering, representation release, server-side fusion), which operate on the local encoders however they were trained.

Prerequisites (Q2). A module is shareable among clients only when its input has the same structure on each of them, i.e., the same raw features (the feature groups of FM3–FM5) or a common input format (the CIFAR-100 crops resized to 32 × 32), and this correspondence is assumed to be known in advance, as in the prior work above. Where no such module exists (FM1–FM2, Multi-News), nothing is shared.

Comparability (Q2). All baselines use the same local architectures, local epochs and communication rounds as Fed-DREAM (Appendix C.5). FedLabSync-WA, HyFEM and FedGraph assume exactly the same parameter sharing as Fed-DREAM wherever they are evaluated (FedGraph reuses the trained local encoders of Fed-DREAM), so the comparison with them isolates the stages after local training.

Attendant issues. We are aware of them: the architectural constraints needed to make a module shareable, the privacy regulations that may forbid it, and the risk of negative transfer among heterogeneous clients (Sec. 3.2). For this reason sharing is left as an option the participants set, and the stages that follow local training do not depend on it: on FM3–FM5 we reran Fed-DREAM with no sharing (new Appendix E.2, Table 19); accuracy falls by 0.54/1.46/1.14 pp (paired one-sided t-tests, p < 0.01), while Fed-DREAM without sharing still exceeds Local-only by 6.0–7.9 pp, so the stages after local training account for most of the gain.

Revision. Sec. 3.2; Appendix E.2 (new Table 19).

AuthorsOpenReview ↗

Response to Reviewer FAC1 (2/3)

W3 / Q3 — Scalability of pairwise distance computation

We agree, and neighbor sampling fits the protocol directly: we implemented it and evaluated it on three settings.

Neighbor sampling. Stage 1 currently aggregates every pair observed by at least one client, ∑m(∣Nm∣2)\sum_m \binom{|N_m|}{2} encrypted items (Table 22). In the variant, each client computes the distances among its own samples in plaintext, as it already does, keeps for every sample its k nearest neighbors, and encrypts only the distances of those pairs; the server aggregates each pair over the clients that submitted it, and a pair that no client submitted is treated as a maximum-distance pair, which Algorithm 4 already handles since it averages only over observed entries. The count becomes at most k∑m∣Nm∣k \sum_m |N_m|, linear in the sample counts, and a client that submits a pair releases exactly what it releases under full aggregation.

Result (k = 10 on FM1, CI1 and MN1). We reran Stages 1–4, the mixup and the server training on the trained local encoders of the main experiments with their seeds and compare with full aggregation (new Table 25). The encrypted items fall by 246×, 504× and 2,344× on FM1, CI1 and MN1, the estimated THE time without packing from 130–1,651 h to 38–55 minutes, and the local k-nearest-neighbor search takes 2–130 ms per client on a GPU. Accuracy changes by −0.14 pp on FM1 (p = 0.14, paired one-sided t-test), by −1.89 pp on CI1 (p < 0.01) and, on MN1, by −0.18 pp in Micro-F1 (p = 0.026) with mAP and Hamming loss within 0.03 pp. Against the best benchmark of Tables 1 and 2, listed in Table 25, the variant stays above it on CI1 and MN1 and at its level on FM1. Neighbor sampling thus removes the quadratic term at an accuracy change of 0.1–1.9 pp for k = 10.

Revision. Sec. 3.3 (Stage 1); Appendix G (new Table 25).

W4 / Q4 — Averaging distances across semantically different views

The reviewer is right that the submitted paper did not state the assumption under which averaging is sensible; the revision adds it. We also explain why a disagreement between views has a limited effect, given what the clusters are used for.

  • The assumption, stated. "Local distances are assumed to agree in expectation with similarity under the shared target: if two samples are close in one client's view, they are not systematically far in another's." The revision adds this to Sec. 3.3 together with the reason it is reasonable in HBFL: every client predicts the same label yny_n, so the views are not arbitrary pairs of attribute sets but partial descriptions of the same outcome. In the hospital/bank example, collaboration is only motivated when a common target (e.g., an insurance or credit event) exists, and under that target the two views are related through the label even if their raw attributes are incommensurable.
  • What the clusters are used for. A cluster is not a prediction and does not need to be "meaningful" as a task; it only decides which of a client's own samples are mixed together in Sec. 3.4. Mixing is done within the client, in its own embedding space, so a disagreement between views cannot produce an invalid representation; it can only make the mixed set less homogeneous in that client's view, which lowers the augmentation's benefit but not its correctness. Distances are also normalized to [0, 1] per client before averaging (Sec. 3.3), so scale mismatch between views is removed.

Revision. Sec. 3.3.

AuthorsOpenReview ↗

Response to Reviewer FAC1 (3/3)

W5 / Q5 — Alignment of independently trained encoders

Fed-DREAM does not require the clients' embeddings to lie in one shared latent space. The server maps each client's embedding through a client-specific node encoder ξτ(m)\xi_{\tau(m)} (Sec. 3.5.1), and ECG-GNN fuses the mapped nodes conditioned on the feature overlap between clients. The graph-level loss is therefore end-to-end supervision on the fused prediction, the quantity needed at inference, and the node encoders are alignment maps learned under it; alignment is learned to the extent that this objective requires.

Evidence. Table 17 isolates the node encoder: with one encoder shared by all clients, FM1/FM2 accuracy is 0.6584/0.7140; with client-specific encoders it is 0.7493/0.7946. On CIFAR-100, where global parameter sharing already synchronizes the local encoders, the shared node encoder is the better choice (Table 18), so the alignment capacity matches how aligned the inputs already are. Table 13 gives the most direct evidence on FM1: Mean-pool MLP 0.6641 < w/o node connection 0.6940 < ECG-GNN 0.7493; only client-specific node encoding combined with relational fusion outperforms averaging the clients' independent predictions.

Explicit alignment objectives. A contrastive or mutual-information term needs, during training, representations of the same sample from different clients at the server, which is the information Fed-DREAM is designed not to release. The clustering does provide positive pairs at the cluster level, since all nodes of a synthetic sample come from the same cluster, so such a term could be added without further release. We agree that the alignment pressure of the graph-level loss is weaker than that of an explicit objective and do not claim that the local encoders share a latent space.

Revision. Appendix E.2.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision summary

We thank the reviewers for their careful and constructive reviews. We have uploaded a revised PDF. Below we summarize the main changes; detailed answers to each point are given in the replies to the individual reviews.

1. Hybrid FedGraph as a discussed and compared prior work (pC8o W1, W8; FAC1 W1)

Hybrid FedGraph (Jang et al., 2024) was not cited in the submission; this was an error. The revision discusses it in Sec. 2 and adds it as a baseline on all twelve settings under the same partitions, local encoders and server-side hyperparameters as Fed-DREAM (Tables 1, 2, 9 and 10; Sec. 4.3; Appendix D). Fed-DREAM is higher in every FashionMNIST and CIFAR-100 setting at both participation rates and, on both Multi-News settings, higher in the Micro-F1 and mAP metrics.

2. New experiments

  • Reconstruction attack by the server (aFTR W1; pC8o W7). An honest-but-curious server trains decoders from the released embeddings to a victim client's raw inputs, with a sweep over the cluster size CC (Appendix F.2, summarized in Sec. 4.5).
  • Single-member releases (pC8o Q3). The case in which a client holds one member of a cluster and releases an unmixed embedding is analyzed in the new Appendix F.4 (Table 21) and stated in Sec. 4.5 and Sec. 5.
  • Fed-DREAM without parameter sharing on FM3–FM5 (pC8o Q1; FAC1 W2). Appendix E.2, Table 19: accuracy falls by 0.54–1.46 pp without sharing, and Fed-DREAM without sharing still exceeds Local-only by 6.0–7.9 pp.
  • Neighbor-sampled Stage 1 (FAC1 W3). Each client restricts the pairs it encrypts to its local kk-nearest-neighbor lists, which makes the encrypted item count linear in the sample counts; evaluated on FM1, CI1 and MN1 (Appendix G, Table 25; stated in Sec. 3.3).

3. Assumptions now stated explicitly

  • Each client holds the labels of its own samples (Sec. 3.1), with the label-holders-only extension named as future work in Sec. 5 (FAC1 W1).
  • The condition under which a module is shareable, and that the correspondence is known in advance (Sec. 3.2) (FAC1 W2; pC8o Q1).
  • Local distances are assumed to agree in expectation with similarity under the shared target (Sec. 3.3) (FAC1 W4).
  • The edge attribute amja_{mj} is defined explicitly as the indicator vector of Fm∩FjF_m \cap F_j, fixed per client pair (Sec. 3.5.1) (aFTR Q1).
  • The node encoders are the only cross-client alignment, and an explicit same-cluster alignment term could be added without releasing further information (Appendix E.2) (FAC1 W5).

4. Corrections

  • The p-values of Table 10 were miscalculated; they are recomputed with paired one-sided t-tests (p < 0.01 for all three metrics against FedLabSync), and Improv. in Table 2 is now measured against the best benchmark per metric (pC8o Q4).
  • Sec. 3.4 now qualifies the privacy statement of the mixup as referring to membership inference, and the threat-model paragraph of Appendix F.2 lists the server's complete training-time view (pC8o W7).

5. Other changes

Sec. 5 (Limitation and Future Work) is extended with the Stage-1 scaling, the single-member case and the label-holder extension. Sections 1–2 were shortened to keep the main text within the page limit.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.