acceptodds
Under review as a conference paper at ICLR 2027

FedGME: Federated Gaussian Mixture-based Learning for Clinical EEG Classification

Abstract

Federated Learning (FL) enables collaborative model training across hospitals without sharing sensitive local patient data. In clinical EEG, however, label distributions differ strongly across sites, and rare classes, such as uncommon seizure types, may be missing from a hospital altogether. Weight-sharing methods such as FedAvg are costly, scale with model size and are biased toward majority classes under label skew, while prototype-sharing methods such as FedPCL summarise each class by a single point, which collapses the within-class variability intrinsic to EEG and carries little about classes a site has never seen. We propose FedGME (Federated Gaussian Mixture-based Learning for EEG), in which clients exchange low-rank Gaussian components per class instead of weights or prototypes. The server keeps the components of all sites side by side, so each class is shared as a mixture that preserves its spread and can be sampled, allowing clients to learn classes absent from their own data. Built on a pretrained EEG model and evaluated on three corpora from the TUH EEG Corpus, FedGME lets clients recognise classes they never observed, raising their balanced accuracy to 0.15–0.30, where prototype methods stay near zero and FedAvg reaches at most 0.10. On the two multiclass tasks, this improves overall balanced accuracy by 15.5% on average relative to FedAvg (up to +0.060 absolute), while transmitting only 0.2%–1.3% of FedAvg's payload and no model weights. Replacing FedPCL's prototype-wise contrastive loss with an established spread-preserving contrastive loss, weighted by patient identity, prevents the within-class collapse of prototype training and raises effective rank by up to 4.2× over FedPCL at the same payload. On the balanced binary task, where no class is absent, FedGME performs on par with FedAvg. Overall, sharing class distributions instead of model weights offers a communication-efficient strategy for mitigating label skew in federated biomedical signal analysis.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
federated learning, representation learning, seizure classification, EEG, contrastive learning, deep learning

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

The paper proposes a promising FL framework to deal with the substantial label distributions differerence between each client. Clients exchange low-rank Gaussian components per class instead of weights or prototypes. The idea is interesting, however, the experiments are not enough to make this paper a solid one.

Critical strengths
  1. The idea is interesting, sharing GMM rather than weights or prototypes. However, it has also been discussed in general areas [1].
  2. Label absent is a very important question to be solved.

[1] Wu Y, Zhang S, Yu W, et al. Personalized federated learning under mixture of distributions[C]//International conference on machine learning. PMLR, 2023: 37860-37879.

Critical weaknesses
  1. The experiments are not enough, only main results are presented, no ablation makes the paper not that solid.
  2. The experiment settings is not clear. How to define the absent labels for each client?
  3. Less critical related work: Personalized Federated Learning under Mixture of Distributions.
Questions for authors
  1. How does the client seperation of the experiment settings? How do you define the absent labels for each client? If a client has some absent labels, how do the metrices calculate?
1 comment

We thank the reviewer for recognising the problem of locally absent labels and the idea of sharing class mixtures instead of weights or prototypes, and for pointing us to FedGMM. Line numbers refer to the revised PDF.

S1 / W3. Relation to FedGMM.

FedGMM (Wu et al., 2023) is now discussed in Sec. 2.3 (l. 146–151) and contrasted with FedGME in the first contribution (l. 80–82).

What the two methods share. Both describe data with Gaussian mixtures and exchange Gaussian parameters instead of raw data.

How they differ. FedGMM models the input distribution, pairing a mixture over inputs or pretrained embeddings with predictors specific to each component, to handle covariate and concept shift. Its server averages both the Gaussian parameters and the predictor weights. Its Gaussians describe inputs rather than classes, so they cannot supply labelled features for a class a site lacks. FedGME instead models class-conditional features with one component per client and class, never averages these components and exchanges no model weights.

Why we compare against CCVR instead (l. 150–156). CCVR is the closest method to FedGME in both problem and mechanism: it targets label skew, estimates class-conditional feature Gaussians at each client and samples synthetic features from them to train the classifier. The two methods differ in what is shared and when. CCVR merges the statistics into one Gaussian per class, uses them after training and relies on FedAvg's shared model. FedGME keeps one component per client, uses the components during training and only for absent classes, and exchanges no weights. Applied on top of FedGME, CCVR's merging lowers absent-class accuracy on TUEV from 0.264 to 0.088 at K = 5 (Table 1; Sec. 4, l. 424–428).

W1. Not enough experiments; no ablation.

We added a leave-one-out ablation and three baselines.

Ablation (Table 7, App. C.1, l. 1117–1159; summary in Sec. 4, l. 473–476). Removing synthetic replay drops absent-class accuracy to zero on both corpora. Removing absent-subspace anchoring (L_freeze) reduces it by about 60%. Removing pretrained anchoring reduces it to 0.011 on TUEV and drops TUEV BAcc from 0.413 to 0.340. The two bank terms change every metric by at most 0.018, so we read them as close to neutral.

Baselines (Table 1; Sec. 4, l. 390–428). We added FedProto, FedSA, and CCVR on top of both FedAvg and FedGME. Among the methods that share no weights, FedGME has the highest BAcc on both multiclass corpora at both K, by 0.060–0.072 over the strongest such baseline. It is the only one above 0.012 on absent classes. FedAvg + CCVR, which shares all model weights every round, is stronger on TUEV, on TUSZ at K = 10 and on TUAB, and we report this openly (l. 420). FedGME and FedGME + CCVR lead on TUSZ at K = 5.

W2 / Q1. How are clients separated, how are absent labels defined, and how are metrics computed?

Clients. For each seed, every training patient is assigned to exactly one client, so clients hold disjoint sets of patients (App. B.3, l. 1021). Validation and test patients are never used for training, and all splits are patient-disjoint (App. B.1, l. 964–966; Tables 4–6).

Absent labels (Sec. 3.6, l. 369–377). After assigning patients, we remove selected classes from selected clients, while every class remains at two or more clients. A class is locally absent at a client if that client holds no training window of it.

  • On TUEV, clients keep three or four of the six classes (background at every client) and lack 2.6 on average at both K.
  • On TUSZ, clients keep two or three of the four classes and lack 1.6 on average at K = 5 and 1.7 at K = 10.
  • TUAB has no absent classes; the share of abnormal windows ranges from 30% to 73% across clients.

The number of clients per class is fixed for each corpus and K, and roughly doubles from K = 5 to K = 10; the seed decides which clients these are. Classes are removed from training only.

Metrics for a client with absent labels (Sec. 3.6, l. 378–386; Eq. 15).

  • Each client's linear probe has one output for every class in the full label set, including classes it never trained on, and every client is scored on the same full official test set.
  • BAcc is the mean recall over all classes.
  • Abs. is the mean recall over only the classes that client lacks. It is averaged over the clients that lack at least one class, and then over seeds.
  • A test window of an absent class counts as correct only if the probe predicts exactly that class. A method that never predicts absent classes therefore scores 0, which is exactly the failure the protocol is designed to detect.
  • All other metrics are computed per client and then averaged over clients and seeds.

We hope this clarifies the setting and the relation to FedGMM, and we would be glad to answer any further questions.

Reviewer_38zcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes FedGME for federated clinical EEG classification under label skew. Clients share low-rank Gaussian components per class instead of model weights, and they sample from the server's component bank to learn classes missing from local data. Experiment results on three TUH EEG corpora show that FedGME improves balanced accuracy over FedAvg on different tasks, especially on local absent classes while maintain the small fraction of communication cost compared to FedAvg.

Critical strengths
  • The method is well-motivated. Using Gaussian components keeps the within-class variability of EEG data that single-point prototypes lose, and lets clients learn classes missing from their local data.

  • FedGME illustrates the effectiveness against other baselines, improving accuracy on absent classes and introduce less communication cost than FedAvg.

Critical weaknesses
  • The contribution are not clearly presented. The listed contributions (L77-92) mixed mthoed components, results and design choices. This makes it is difficult to tell what is technically new in the paper.

  • The presentation of the paper is not well-organized. Many key details are moved to the appendix. For example, Section 3 states that “definitions, estimation procedures and edge cases are in the matching parts of Appendix A.” As a result, the exact formulations of several core components appear only in Appendices A.4–A.6: the patient-aware local term (L_sup, L_cnce and their weights), pretrained anchoring, absent-subspace anchoring (L_freeze), and synthetic replay.

  • Figure 1 is hard to understand. Readers cannot see how each component works or how they interact, especially how the bank is leveraged for contrastive learning and synthetic replay. A cleaner figure would help to strengthen the visualization of the method.

  • Some recent baselines are missing in the experimental results. Please compare against more recent methods such as prototype-based FL: FedProto (AAAI 2022), FedTGP (AAAI 2024), feature-distribution sharing: CCVR (NeurIPS 2021).

  • The evaluation metrics are not clearly defined. Balanced accuracy, absent-class balanced accuracy (Abs.), effective rank are only cited in main text without definitions.

Questions for authors
  • The shared Gaussian components of a rare class may come from only one or two patients. Have the authors evaluated whether these components can leak patient-identifying information?
  • How sensitive is FedGME to the covariance rank r, absent-class weight ωabs\omega_{\mathrm{abs}}? Moreover, how do the performance and communication scale with more clients (e.g., K=50 or 100)?
4 comments

(1/3) Thank you for the concrete suggestions on presentation and evaluation. Each one led to a specific change in the revision. The new baselines are in Table 1 and the leave-one-out ablation is in Table 7 (App. C.1). Line numbers refer to the revised PDF.

W1. The contributions mix method components, results and design choices.

We agree. The revised list (l. 69–100) now separates the technical contributions from the results, and l. 73–74 states where the novelty lies: in what clients share and in how a client learns classes it lacks. The local objective builds on established contrastive losses.

  1. Per-client Gaussian components as the only shared state (l. 76–82; Sec. 3.2–3.3). The server keeps one low-rank component per client and class and never averages them. Both sending a component and scoring a representation against it cost O(rd) (Eq. 1, l. 223; Eq. 3, l. 256–260), and communication does not depend on model size (Eq. 17–18). Evidence: merging the components into one Gaussian per class, as CCVR does, lowers absent-class accuracy by 0.09–0.18, e.g. from 0.264 to 0.088 on TUEV (Table 1; Sec. 4, l. 424–428).

  2. Learning locally absent classes without a shared model (l. 83–88; Sec. 3.4–3.5). This contribution has three parts:

    • Synthetic replay trains the probe on features sampled from other clients' components of absent classes.
    • Absent-subspace anchoring protects the encoder directions that carry those classes.
    • Pretrained anchoring keeps components from independently trained encoders comparable across sites.

    Evidence: without replay, Abs. drops to zero; without L_freeze, it drops by about 60%; without pretrained anchoring, it drops to 0.011 on TUEV (Table 7).

  3. A patient-aware local objective (l. 89–92; Sec. 3.4, Eq. 5–6). We show that FedPCL's prototype-wise loss collapses classes within each client. We replace it with supervised contrastive and spreading losses weighted by patient identity. Evidence: at the same payload, effective rank on TUAB rises from 41.3/32.1 (FedPCL) to 142.5/134.4 (FedPCL-S), up to 4.2× (Sec. 4, l. 466–468; Table 11). On its own, this improves representation geometry, not accuracy (l. 468–471).

The results now form a separate bullet (l. 94–100). Among the methods that share no weights, FedGME is the only one that recognises absent classes (0.26–0.30 Abs., against ≤ 0.012), at 0.2–1.3% of FedAvg's traffic.

W2. Key formulations appear only in the appendix.

Sec. 3 now contains every equation that defines the method:

  • the low-rank component (Eq. 1, l. 223);
  • the bank score and its O(rd) form (Eq. 2–3, l. 247–260);
  • the cross-client term (Eq. 4, l. 280–288);
  • the patient-aware local terms (Eq. 5–6, l. 305–315);
  • pretrained anchoring and the full objective (l. 318; Eq. 7, l. 323);
  • absent-subspace anchoring (Eq. 8, l. 337);
  • the sampling rule (Eq. 9, l. 348);
  • the replay loss (Eq. 10, l. 351).

Each subsection gives the concept and its rationale (l. 199–201). Appendix A now mirrors the main-text subsections one to one (l. 758–760) and holds only estimation details, derivations and edge cases:

  • the estimator and shrinkage rule (A.2, Eq. 11–12);
  • the inverse covariance (A.3, Eq. 13);
  • width calibration (A.6, Eq. 14);
  • the metrics (A.7, Eq. 15–16);
  • communication accounting (A.8, Eq. 17–18).

A notation table (Table 2, l. 766–796) is referenced from Sec. 3.1 (l. 215).

W3. Figure 1 is hard to follow.

Figure 1 (p. 4, l. 162–192) has been redrawn as one communication round, with the client on the left and the server on the right, in six labelled panels:

  • (a) a client computes its components;
  • (b) the server keeps the bank;
  • (c)–(e) the client trains against the bank: the cross-client term, the local term and the two anchoring terms;
  • (f) synthetic replay trains the probe.

The bank is drawn as a client × class grid. Empty cells mark classes a client does not observe, and dash-dotted columns mark the classes absent at client k. A legend distinguishes trainable and frozen modules, losses, stop-gradients, communication and replay, and the caption walks through the panels in order. Sec. 3 follows the same panels (l. 196–199), and each subsection points to its panel, e.g. Fig. 1(c) at l. 279 and Fig. 1(f) at l. 346.

(2/3) W4. Missing baselines.

We added FedProto (Tan et al., 2022a), FedSA (Zhou et al., 2025) and CCVR (Luo et al., 2021), the latter on top of both FedAvg and FedGME (Table 1; Sec. 4, l. 390–399; communication in Table 3).

Methods that share no weights. FedGME has the highest BAcc on both multiclass corpora at both K and exceeds the strongest such baseline by 0.060–0.072 (l. 404–407). Local and all prototype methods stay at or below 0.012 Abs. (l. 413–415). They represent each class by one point and train the probe on local labels only, so the probe never sees an absent class.

CCVR. FedAvg + CCVR scores highest on TUEV (0.510, 0.514), on TUSZ at K = 10 and on TUAB (l. 420). Its gain lies in retraining the probe on FedAvg's shared encoder, which requires exchanging all 6.5M parameters every round, 77 to 423 times FedGME's traffic (l. 421–424). Applied on top of FedGME, where encoders are trained locally, CCVR changes BAcc by at most 0.011 and lowers Abs. by 0.09–0.18 (l. 424–428). We report FedAvg + CCVR as the reference for settings where weights can be shared.

W5. Metrics are not defined.

The metrics are now defined in Sec. 3.6 (l. 378–386), with formal definitions in App. A.7 (Eq. 15–16, l. 910–929) and the readout protocol in App. B.5 (l. 1106–1112).

Readout. Each client's linear probe has one output for every class in C, including classes the client never trained on. Every client is scored on the same full official test set, and metrics are averaged over clients and then over five seeds.

  • BAcc is the mean recall over all classes, our primary metric because plain accuracy hides rare classes.
  • Abs. is the mean recall over only the classes a client lacks, averaged over the clients that lack at least one class. A test window of an absent class counts as correct only if the probe predicts exactly that class, so a probe that never predicts absent classes scores 0.
  • eRank (effective rank) and Unif. (uniformity; lower means more uniform) are computed on each client's ℓ₂-normalised encoder embeddings.
  • Communication counts the floats a client sends and receives per round (App. A.8).
  • Corpus-specific metrics follow Yang et al. (2023) and are reported in App. C.2: Cohen's κ, weighted F1, AUROC and AUPRC.

Q1. Can the components of a rare class from one or two patients leak patient identity?

They could. The revision now names exactly this risk in Sec. 5 (l. 518–520): a rare class may come from one or two patients, and EEG representations encode patient identity. We do not claim formal privacy, and the same limitation applies to every method that exchanges class-level statistics, including FedProto, FedPCL and FedSA.

What is sent. For each class, in each of the two spaces, a client sends a count, a unit-norm centroid, r = 7 eigenvectors with their eigenvalues and one residual variance. All of these are pooled over every window of the class and computed on ℓ₂-normalised embeddings. No patient identifiers (Table 2, l. 770), per-window features or model weights leave a site.

The rarest classes send less. Below a minimum number of windows, a class sends only its centroid, or its centroid and eigenvalues, with no eigenvectors (App. A.2, l. 821–827). The TUSZ absence seizures (60 training windows), for example, fall back to this reduced form (Sec. 5, l. 514–515). This reduces the structure that is shared, but a centroid computed from one patient can still carry identity information.

Compared with other methods. FedAvg sends all 6.5M trainable parameters every round, and shared weights are known targets of inversion and membership-inference attacks. CCVR sends a full 256 × 256 covariance per class (33,153 values, against 2,057 for a FedGME component in the same space; Table 3).

Mitigations. Because FedGME shares only a few pooled statistics, standard mechanisms apply directly. A client can withhold a component until a minimum number of patients contribute to it, and a Gaussian mechanism with per-patient clipping can be applied to the centroids and eigenvalues. A formal privacy analysis is future work (Sec. 6, l. 532).

(3/3)Q2. Sensitivity to r and ω_abs, and scaling to K = 50 or 100.

We agree that these analyses are important. The revision already notes that we do not study sensitivity to r, ω_abs or ζ (Sec. 5, l. 513). The experiments are in progress, but we could not finish them before the end of the discussion period. Each setting must be trained on several corpora and values of K with five seeds, and the compute available during the rebuttal went to the new baselines and the ablation. We will add the sweeps over r and ω_abs and the runs at K ∈ {50, 100} to the supplementary material of the final version. In the meantime, the paper already supports the following:

Communication grows linearly in K and r. A component costs (r + 1)d + r + 2 values per space (Eq. 17), and a client's traffic per round is at most (K + 1)·|C|·2,578 floats (Eq. 18). At K = 50 and K = 100, this gives 0.79M and 1.56M floats on TUEV, i.e. 6.0% and 11.9% of FedAvg's 13.1M, and 0.53M and 1.04M on TUSZ (4.0% and 8.0%). Scoring against the bank costs O(B·|B|·r·d), also linear in the number of components (App. A.3, l. 838–840).

Trend from K = 5 to K = 10. Abs. rises (TUEV 0.264 → 0.295; TUSZ 0.257 → 0.300), plausibly because the bank holds more components per absent class. TUEV BAcc decreases slightly (0.418 → 0.397) as the number of patients per client halves (Sec. 4, l. 408–418).

What we expect at K = 50–100. Each client will hold fewer windows, so more classes will fall back to centroid-only components (App. A.2). Communication will also grow linearly, which motivates merging components on the server at scale (Sec. 5, l. 517–518; Sec. 6, l. 533). On TUSZ, the absence class has only 60 training windows (Table 5), which limits how many clients can meaningfully hold it.

We thank the reviewer again and are happy to clarify any point further.

Reviewer_38zcOpenReview ↗

Response to Authors' Rebuttal

I would like to thank the authors for their detailed rebuttal and revisions, which have addressed several of my initial concerns. The improved presentation, clearer contributions, and additional baseline comparisons are helpful. However, concerns remain regarding the robustness and practical applicability of FedGME, such as limited evaluation, scalability and potential privacy risks. Overall, I believe the paper still requires substantial revisions that may be difficult to address within the current submission cycle. Therefore, I will maintain my score at 2.

Reviewer_Bju3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedGME, a communication-efficient federated learning framework for clinical EEG classification under label skew and locally absent classes. Instead of exchanging model parameters, clients share low-rank Gaussian descriptors of their class representations, which the server retains as client-specific components of a class-wise mixture. Experiments on three TUH-derived datasets show improved balanced accuracy on the multiclass tasks.

Critical strengths
  • The paper addresses a practically relevant problem.
  • The overall design is coherent and intuitively motivated.
  • The paper is clearly structured and easy to follow.
Critical weaknesses
  • The evidence for cross-institutional applicability remains limited to simulated clients from a single data source.
  • The current partitions ensure that each class is available at multiple clients. If a class exists at only one site, that site has no remote same-class component for the cross-client positive term, while all other clients must rely on a single source distribution. It is unclear how well the method would perform under this condition.
  • Avoiding raw-data and model-weight exchange does not guarantee that the transmitted means and covariance information are private, especially when a rare class is represented by only one or two patients.
  • The method’s dependence on the pretrained representation and the effectiveness of cross-client alignment require further validation. A shared pretrained reference provides a useful starting point, but it may not adequately represent the rare classes of interest.
  • The comparison with relevant competing methods is insufficient. The experimental comparison mainly includes Local, FedAvg, FedPCL and its variants. Several approaches discussed in the paper, such as FedVLS, FedSA, CCVR, are not included.
  • The ablation study does not sufficiently isolate the contributions of the main components.
Questions for authors

Please see the weakness part.

3 comments

(1/3) We thank the reviewer for a careful review that goes straight to the assumptions FedGME rests on: one data source, classes held by one site, privacy, and dependence on the pretrained reference. In response, the revision adds three baselines (Table 1; Sec. 4, l. 390–428), a leave-one-out ablation (Table 7, App. C.1, l. 1117–1159, summarised in Sec. 4, l. 473–476) and a longer limitations paragraph (Sec. 5, l. 512–520).

W1. All evidence comes from simulated clients drawn from one data source.

We agree that this is the main limitation of our evaluation. The revision now states it directly in Sec. 5 (l. 512–513): our partitions reproduce label skew and differences between patients, but not differences in hardware or preprocessing between institutions. Sec. 3.6 (l. 369) now describes the clients as simulated hospitals from one archive, and Sec. 6 (l. 532) lists evaluation on data from several institutions as future work. Two aspects of the setup reduce the concern without removing it:

  • The shared reference comes from outside TUH. The BIOT checkpoint was pretrained without labels on PREST, a clinical sleep-lab dataset with no TUH recordings (Sec. 3.1, l. 209–210; App. B.2, l. 970–1015). Every client's encoder is anchored to this external model, and the anchoring is necessary: without it, TUEV BAcc drops from 0.413 to 0.340 (Table 7, l. 1149).
  • TUH itself is heterogeneous. It spans about 14 years of routine clinical EEG, originally recorded at four sampling rates (250, 256, 400 and 512 Hz), before our resampling to 200 Hz (Obeid & Picone, 2016; App. B.4, l. 1071). Clients with disjoint patients inherit this variability.

W2. A class held by only one client.

In every partition we evaluate, each class is held by two or more clients (Sec. 3.6, l. 372; App. B.3, l. 1023). The case you raise is therefore untested, and the revision lists it as a limitation (Sec. 5, l. 513–514). The method itself runs unchanged:

  • At the client that holds the class, its windows have no component of their class from another client. They therefore drop out of the cross-client term (the set I_B in Eq. 4, l. 284) and are trained by the local and anchoring terms only (Sec. 3.4).
  • At every other client, the class appears as a mixture with a single component. Synthetic replay samples from that component, with the sampling width matched to the spread of the client's own classes (Eq. 14, App. A.6, l. 896–908). Absent-subspace anchoring uses its centroid as the consensus centroid (App. A.6, l. 874–884). Neither step needs a second component.

The remaining risk is that this one component carries the bias of the only site that holds the class, with nothing to balance it. We therefore expect a lower absent-class score than when several sites hold the class. This is consistent with the rise in Abs. We see this from K = 5 to K = 10, when the bank holds more components per absent class (Sec. 4, l. 416–418). If that site also has very few windows of the class, its component reduces to a centroid only, or to a centroid with an isotropic spread (App. A.2, l. 821–827), so even less structure is shared.

(2/3) W3. The transmitted means and covariances are not guaranteed to be private.

We agree, and we do not claim formal privacy. Sec. 5 (l. 518–520) now states explicitly that components are not private by construction, and Sec. 6 (l. 532) lists formal guarantees as future work. The same limitation applies to every method that exchanges class-level statistics, including FedProto, FedPCL and FedSA.

What is sent. For each class, in each of the two spaces, a client sends a count, a unit-norm centroid, r = 7 eigenvectors with their eigenvalues and one residual variance. That is 2,578 values per class for both spaces together (Eq. 17, App. A.8, l. 933–939). All of these are pooled over every window of the class and computed on ℓ₂-normalised embeddings. Patient identifiers are never sent (Table 2, l. 770), and no per-window features or model weights leave a site (Fig. 1 caption, l. 191–192). Classes too small to support a covariance send only a centroid, or a centroid and eigenvalues (App. A.2, l. 821–827).

Exposure compared with other methods. FedGME sends more per class than the prototype methods, because it adds a spread to the centroid. It sends far less than the methods that share weights or full covariances. FedAvg sends all 6.5M trainable parameters every round (App. A.8, l. 951), and shared model weights are known targets of inversion and membership-inference attacks. CCVR sends a full 256 × 256 covariance per class: 33,153 values, compared with 2,057 for a FedGME component in the same space (Table 3, l. 973–1004).

Mitigations fit easily. FedGME shares only a few pooled statistics, so standard mechanisms apply directly. For example, a Gaussian mechanism with per-patient clipping could be applied to the centroids and eigenvalues, and a client could withhold a component when too few patients contribute to it. A formal privacy analysis is beyond the scope of this paper.

W4. Dependence on the pretrained representation and on alignment across clients.

We agree that FedGME depends on the shared reference, and the revision explains why the method relies on it (Sec. 3.4, l. 320–322; contribution 2, l. 86–88). The ablation confirms it. Without pretrained anchoring, TUEV BAcc drops from 0.413 to 0.340 and Abs. from 0.255 to 0.011, and TUSZ Abs. drops from 0.257 to 0.119 (Table 7, l. 1149 and 1157).

Why the method relies on it. This follows from exchanging statistics instead of weights. A class summary from one site is useful at another site only if both map data into compatible feature spaces. There are two ways to achieve this. All clients can share one encoder, as in FedAvg + CCVR. Or all clients can be tied to one fixed reference, as in the original FedPCL, which freezes the pretrained backbone (Sec. 2.3, l. 140). FedGME uses a weaker form of the second approach:

  • All four encoder blocks are trained (App. B.4, l. 1071).
  • L_anchor is a soft cosine penalty (l. 318).
  • L_freeze constrains at most ⌊0.1 d_h⌋ = 25 of the 256 directions. In our experiments it constrains at most 3, because a client lacks at most 3 classes (Sec. 3.6, l. 372–374; App. A.6, l. 876–877).

Statistics from other sites are used only where local data are missing. Classes a client holds are learned from its real windows. Components from other sites enter the probe only for absent classes, in the same softmax as the real data and with a matched width (Eq. 10, l. 351; Eq. 14). The clients' feature spaces must therefore be comparable, not identical.

The CCVR results agree with this. CCVR merges all clients' statistics into one Gaussian per class, which assumes a shared feature space. Applied on top of FedGME, it changes overall BAcc by at most 0.011 and lowers absent-class accuracy by 0.09–0.18, e.g. from 0.264 to 0.088 on TUEV and from 0.257 to 0.166 on TUSZ at K = 5 (Table 1; Sec. 4, l. 424–428). Per-client components transfer between sites better than merged ones, and this is why FedGME keeps them separate.

The reference provides the coordinate system, not the description of rare classes. It was pretrained without labels on sleep-lab EEG, so it was never trained to separate seizure types or epileptiform events. The components describing a rare class are estimated by the clients that hold it, with their own trained encoders. On top of this reference, FedGME raises Abs. on TUEV from 0.000 (Local) to 0.264 (Table 1).

Alignment comes mainly from the shared anchor, not from the cross-client term. Removing L_cross changes BAcc by less than 0.01 (Table 7, l. 1146 and 1154), whereas removing L_anchor causes the drops reported above.

We use a single backbone (Sec. 5, l. 513). If a reference already maps a rare class onto the same region as a common class, anchoring cannot correct this.

(3/3) W5. Missing competing methods.

We added FedProto (Tan et al., 2022a), FedSA (Zhou et al., 2025) and CCVR (Luo et al., 2021), the latter on top of both FedAvg and FedGME. The results are in Table 1, the baselines are described in Sec. 4 (l. 390–399), and their communication is broken down in Table 3.

Methods that share no weights. FedGME has the highest BAcc on both multiclass corpora at both K. It exceeds the strongest such baseline by 0.060–0.072 (Sec. 4, l. 404–407). Local and all prototype methods stay at or below 0.012 Abs. (l. 413–415). They represent each class by one point and train the probe on local labels only, so the probe never sees an absent class.

Methods that share weights. FedAvg + CCVR scores highest on TUEV (0.510 and 0.514), on TUSZ at K = 10 (0.397) and on TUAB. FedGME and FedGME + CCVR lead on TUSZ at K = 5 (l. 420). The gain of FedAvg + CCVR comes from retraining the probe on a shared encoder. That encoder requires exchanging all 6.5M parameters every round, 77 to 423 times FedGME's traffic (l. 421–424). We report FedAvg + CCVR as the reference for settings where weights can be shared.

W6. The ablation does not isolate the main components.

The new leave-one-out ablation (Table 7, App. C.1, l. 1117–1159; summary in Sec. 4, l. 473–476) removes one component at a time and keeps everything else fixed. All results are for K = 5. The full model was re-run alongside the variants, so its scores differ slightly from Table 1 (caption, l. 1136).

  • Without synthetic replay, Abs. drops to zero (TUEV 0.255 → 0.000; TUSZ 0.257 → 0.001), and BAcc falls by 0.072 and 0.055.
  • Without L_freeze, Abs. falls by about 60% (to 0.102 and 0.103).
  • Without pretrained anchoring, Abs. falls to 0.011 on TUEV and 0.119 on TUSZ, and TUEV BAcc falls to the level of the prototype baselines (0.340).
  • The two bank terms (L_cross and L^h_bank) change every metric by at most 0.018, so we read them as close to neutral.

The original ablation ladder changed several elements at each step. Its two intermediate variants are kept as the last rows of Table 7 (bank only, FedPCL-SM; prototypes only, FedPCL-S). Both stay near zero Abs. on both corpora, which shows that the gain comes from the mechanisms that use the bank, not from exchanging richer statistics alone.

We hope these changes address your concerns, and we would be glad to discuss any point further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.