acceptodds
Withdrawn from ICLR 2027

TriFed:Trustworthy Personalized Federated Generalization for Unseen Clients under Open Distribution Shifts

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated learning is valuable when data cannot be shared, yet most existing methods implicitly assume a relatively closed set of training clients and largely stationary local distributions, and often interpret deviations from the global direction uniformly as personalization signals. In open environments, the arrival of unseen clients, distribution drift, and fluctuations in update quality may occur simultaneously, coupling shared-knowledge transfer, local personalization, and the suppression of unreliable effects. To address this setting, we propose TriFed, a trustworthy personalized federated generalization framework. Based on cross-client evidence and local utility, TriFed organizes training contributions into three functional roles—shared retention, personalized reinforcement, and harmful suppression. Locally useful knowledge can therefore participate in shared learning while strengthening local discrimination, whereas sample contributions supported by negative-utility evidence are filtered through cross-round confirmation. On the server, global aggregation and cluster-level collaboration form two channels with isolated training and deployment states; at deployment, unseen clients use shared-model initialization and lightweight support-set adaptation. We evaluate seen and tail client performance, unseen-client zero-shot and adapted accuracy, open-perturbation responses, and overhead, with structural ablations of tri-decomposition and dual-channel collaboration. Ablation and gating diagnostics indicate distinct roles: tri-decomposition mainly affects robustness under open shifts and contamination, the private head primarily supports seen-client performance, and the measurable benefit of cluster-level collaboration is concentrated in adaptation initialization for unseen clients. On FEMNIST with 20% label noise, TriFed maintains both the average and bottom-10% performance of seen clients, while ranking within the top two entries of the same table for unseen-client zero-shot prediction and adaptation.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
federated learning, personalized federated learning, federated generalization, unseen clients, distribution shift, robust federated learning, few-shot adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_PcsMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TriFed, a trustworthy personalized federated generalization framework to address unseen clients in FL. The framework has three core components: shared FL part, private head with clustering FL part, and rejection.

Critical strengths
  • this paper focuses on an important problem in FL: unseen clients, and proposes TriFed, a trustworthy personalized federated generalization framework to address this problem.
  • the author has done a very clear literature review with recent works.
Critical weaknesses
  • I am not quite sure what the main contributions are. The authors summarize three bullet points in the Introduction. Regarding the first point, this reads to me more as a problem formulation and a description of the protocols rather than a concrete contribution. Regarding the second point, the proposed mechanism appears to consist of three common components: shared FL, private FL, and filtering (please see my further comments below regarding these components). I assume the authors consider this combination to be the core contribution, but I am not convinced that the combination itself is sufficiently strong to constitute the main contribution of a paper. I would encourage the authors to explain more explicitly what is novel beyond simply combining these components.

  • The filtering mechanism is defined based on “harmfulness,” which I find confusing. “Harmfulness” is a strong term in the FL literature and is typically used in adversarial settings. In this context, I assume the authors are instead referring to cross-client inconsistency. However, this raises another concern: if the underlying issue is merely cross-client inconsistency, why is filtering the most appropriate solution?

  • The paper also positions sample-level reliability control as a key distinction of TriFed, but it is not sufficiently established why such a fine-grained reliability mechanism is necessary or particularly well suited to the considered open federated setting.

  • Overall, I would conclude that the paper proposes a framework to address the unseen-client problem, but it does not clearly identify the central research problem or sufficiently define the scope. The current presentation leaves me with several open questions and would benefit from further refinement and additional supporting evidence.

Questions for authors

Please see Critical Weaknesses.

Reviewer_M3CeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TriFed, a personalized federated learning framework for unseen-client generalization under noisy contributions and distribution shifts. TriFed uses cross-client and cross-round evidence to suppress suspected harmful samples, while reinforcing private heads based on their utility on local probes. The server aggregates a shared model and separately clusters private-head offsets to construct deployment-only initialization candidates for unseen clients. At deployment, each unseen client selects an initialization using a labeled support set and adapts its classifier head before evaluation on a disjoint query set. Experiments on FEMNIST and CIFAR-10 examine seen-client average and tail performance, unseen-client zero-shot and few-shot accuracy, robustness to label noise and distribution drift, ablations, and runtime.

Critical strengths
  1. The paper addresses a practically relevant setting by jointly considering personalization, unreliable training contributions, distribution shifts, and generalization to unseen clients.

  2. TriFed provides a coherent design that combines cross-client evidence, cross-round confirmation, and local utility, while isolating cluster-based deployment candidates from the global training state.

  3. The evaluation is relatively comprehensive and transparent, covering multiple datasets, four random seeds, label noise, distribution shifts, few-shot budgets, tail-client performance, and structural ablations.

  4. The results under label noise are particularly encouraging. On FEMNIST with 20% noise, TriFed outperforms the next-best method by 4.77 and 9.00 percentage points in average and bottom-10% seen-client accuracy, respectively.

Critical weaknesses
  1. The dynamic-shift evaluation lacks matched baseline comparisons, so TriFed’s comparative robustness remains unclear. This is especially important under concept drift, where affected clients suffer a 20.1-point accuracy drop. In addition, class-prior drift is simulated by rewriting labels rather than resampling class frequencies, potentially conflating prior and concept shifts.
  2. The harmful-sample detector needs more direct validation. The rejection rate for flipped-label samples decreases from 36.4% at 20% noise to 25.5% at 40% noise, yet precision, recall, false-positive rates, and class-conditional behavior are not reported. The current gating statistics do not establish whether rejected samples are genuinely harmful, particularly under legitimate client-specific shifts.
  3. Cluster-based initialization may overfit extremely small validation sets. At the 5% support budget, some clients have only 6-9 samples, leaving roughly 1-2 labels for selecting among multiple cluster candidates and mixing coefficients. Moreover, it is unclear whether baselines receive an equivalent initialization-selection budget. Selection rates, paired gains, and comparisons with simpler initialization strategies are needed.
  4. TriFed incurs additional computation, communication, storage, and privacy costs. Its sample-wise trial updates and cross-client neighbor retrieval may be expensive, while uploading plaintext probe features and labels limits applicability in privacy-sensitive settings. Beyond end-to-end runtime, the paper should report component-wise costs, communication volume, server memory, and scalability with the number of clients and probes.
  5. The empirical advantage is mixed outside noisy FEMNIST. Several baselines perform better on clean seen-client metrics, while some CIFAR-10 adaptation improvements are smaller than the reported marginal standard deviations. Paired statistical tests are needed to determine whether these gains are significant and whether they justify the additional complexity.
Questions for authors
  1. Could the authors provide matched baseline comparisons under feature, prior, and concept drift? Please also clarify why class-prior drift is simulated by relabeling rather than resampling class frequencies, as this may introduce concept drift.

  2. Could the authors report the precision, recall, and false-positive rate of harmful-sample detection across noise levels and drift types? How does the method avoid rejecting valid client-specific samples that disagree with cross-client consensus?

  3. How reliable is cluster-candidate selection when only one or two support-validation samples are available? Do all baselines receive the same validation and initialization-selection budget, and how does TriFed compare with simpler initialization strategies?

  4. Could the authors provide component-wise computation, communication, and memory costs, including neighbor retrieval, trial updates, and probe storage? How would privacy-preserving probe sharing affect efficiency and accuracy?

  5. Given the mixed clean-setting results and small CIFAR-10 adaptation margins, could the authors report paired statistical tests or confidence intervals to establish whether the improvements are significant and justify the added complexity?

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

The paper proposes TriFed, a personalized federated learning framework for open settings with unseen clients, distribution drift, and noisy updates. On each client, training samples are routed into three roles. They are retained for joint training (α), used to reinforce a private head according to local probe utility (β), or rejected when cross-client neighbors, a trial-update loss test, and cross-round repetition all flag them as harmful (γ). The server aggregates the shared model with FedAvg and separately clusters private-head offsets to provide optional initializations for unseen clients, which then adapt only the classifier head on a small support set. Experiments use FEMNIST as the main benchmark and CIFAR-10 as a secondary one, with most detailed results in the supplementary material.

Critical strengths
  • The paper studies a realistic setting in which unseen clients, drift, and noisy updates occur together. Its protocol evaluates seen-client average and tail accuracy alongside zero-shot and few-shot accuracy on unseen clients, using nested support budgets and a fixed query set.
  • On FEMNIST with 20% label noise (Table A9), TriFed clearly improves seen-client accuracy, by 4.77 points in R-tr and 9.00 points in bottom-10% accuracy over the best competing method. The supplement documents hyperparameters, split sizes, and paired statistics in detail. The authors are also explicit that "trustworthy" does not imply a privacy or Byzantine-robustness guarantee.
Critical weaknesses
  • The technical novelty is limited. Each building block has a close precedent:

    • Shared and private classification heads are a standard design in personalized FL (e.g., FedPer, FedRep, FedBABU).
    • Filtering samples by cross-client neighbor agreement and loss changes is common in noisy-label FL.
    • The single-sample trial update in Eq. (6) is essentially a first-order influence estimate.
    • Cosine clustering of head offsets with a coordinate-wise median follows standard clustered FL and robust aggregation practice.
    • Unseen-client adaptation is head-only fine-tuning on a support set, and server aggregation in Eq. (8) is plain FedAvg.

    The paper itself positions TriFed as combining "three concerns that are usually treated separately." The combination relies on many hand-set constants listed in Appendix A.3, A.4, and Table A2. These include a 0.8 confidence threshold, top-3 neighbors covering at least two source clients, a 64ε(1+R) loss tolerance, two-round confirmation, at most four clusters with a minimum size of 10, a window of recent offsets, mixing magnitudes of 0.2/0.1/0.05, and a 10⁻⁷ acceptance margin. The paper offers no analysis of why this design should work and no sensitivity study for any of these choices.

  • TriFed is not state of the art in most of the reported settings:

    • Clean FEMNIST (Table 1): TriFed is best on only one of six accuracy metrics (ZS-Std, by 1.54 points). It ranks fourth on R-tr (91.74 vs. 92.92 for PGFedMo), third on Worst (75.52 vs. 78.12), and fourth on Adapt15-Std, where even FedAvg is higher (84.97 vs. 84.80). Ditto, a 2021 method, is higher than TriFed on four of the six metrics.
    • Clean CIFAR-10 (Table A11): TriFed trails PGFed, PGFedMo, and Ditto on both seen-client metrics (e.g., Worst 56.29 vs. 59.69 for Ditto). Its zero-shot accuracy is about 1.9 points below FedAvg on both unseen-client groups and lower than every baseline except FedRG. This contradicts the claim that tri-decomposition helps unseen-client generalization, yet the conclusion calls CIFAR-10 "complementary cross-environment evidence."
    • CIFAR-10 with 20% label noise (Table A12): TriFed ranks fourth on both R-tr and Worst, and its zero-shot accuracy is again lower than every baseline except FedRG.
    • FEMNIST with 40% label noise (Table A10): FedRG is higher than TriFed on all six metrics, including by 10.19 points on ZS-Std and 6.74 points on ZS-Shift. GAS+Median also exceeds TriFed on both zero-shot metrics.
    • Even on FEMNIST with 20% label noise (Table A9), FedRG is higher on three of the four unseen-client metrics.

    TriFed is also about 1.7 times slower than FedAvg on clean FEMNIST (1.29 vs. 0.77 hours).

  • Where TriFed does lead, the margins are small relative to the variance. On clean FEMNIST, its margins over plain FedAvg are +0.10 on ZS-Shift and +0.62 on Adapt15-Shift, both within the seed-to-seed standard deviations in Table 1. In Table A4, the few-shot results that Section 4.4 describes as "highest" lead the next-best method by only 0.21 to 0.71 points, against standard deviations of up to 1.04. On clean CIFAR-10, TriFed's lead at Adapt15 is 0.51 and 0.23 points, while its own standard deviations are 2.91 and 2.76. With only four seeds and no significance tests for cross-method comparisons, these differences cannot be called reliable. Figures 3 and 4 also report means without any uncertainty.

  • The ablations and diagnostics do not support the claimed division of roles:

    • In the clean setting, removing tri-decomposition raises R-tr by 1.42 points (to 93.16, above every baseline in Table 1), Worst by 1.01, and Adapt15-Std by 0.60 (Table A3). Even before any perturbation, the filter rejects 6.7% of sample visits (Table A6, rounds 1 to 50), which is consistent with this loss.
    • The filter catches only part of the injected noise. According to Figure A3, flipped-label samples are rejected in only 25.5% to 36.4% of their visits, and 39.6% to 49.5% of them are never rejected. Meanwhile, 13.6% to 15.2% of clean-label samples are rejected at least once.
    • In the clean setting, the cluster channel changes the adaptation metrics by only 0.19 and 0.61 points, within the standard deviations in Table A3. The supplement shows larger paired gains under label noise (Tables A13 and A14) but gives no paired analysis for the clean setting. The abstract's general claim of a "measurable benefit" therefore holds only under contamination.
    • The private head produces large seen-client gains (5.78 and 11.60 points), but this is the standard personalization component that the PFL baselines already provide. Removing it does not hurt unseen-client accuracy and even raises ZS-Shift (78.38 vs. 77.59) and Adapt15-Std (85.13 vs. 84.80).
  • The drift evaluation does not show that TriFed handles drift, and the comparison is incomplete and possibly uneven:

    • The drift experiments (Section 4.6, Figures 6 and 8, Tables A6 to A8) report only TriFed, with no baseline. Under concept drift, the rejection rate on affected clients rises to 19.9% during the ramp. Their sentinel accuracy under the new label rule then falls by 20.1 points, while controls improve by 13.4 points (Table A8). Under class-prior drift, affected clients gain only 1.1 points vs. 12.2 for controls. The filter thus appears to treat legitimate drift as noise. The paper does not make clear whether drift should be suppressed or adapted to, and these results suggest the method does not adapt to it.
    • None of the unseen-client or federated generalization methods the paper discusses (FLUX, the profile-mapping method of Li et al. (2026), GAINS, TGFed) is used as a baseline. Noisy-label FL methods cited in the related work (FedClean, FedELR, FedSIR) are also missing, as are head-based PFL methods such as FedRep or FedBABU, which are the most direct comparison for a private-head design with head-only adaptation.
    • Each baseline is run with a single fixed hyperparameter setting (Appendix B.4), and FedRG is described as using "its preset for the clean condition," so it is unclear what it uses under noise. The main text says tuning details are in Appendix B, but no tuning procedure is described there.
    • TriFed uses information the baselines do not have. Appendix A.1 states that probe features and observed labels are uploaded in plaintext, and the harmful-evidence step retrieves neighbors from other clients' labeled probes. This departs from the premise in the first sentence of the abstract, that FL is valuable "when data cannot be shared." It also makes the comparison with parameter-only baselines hard to interpret. Neither the main paper nor the supplement reports the communication cost of this exchange.
  • The manuscript is not ready in completeness or presentation:

    • The headline claims come from the one favorable setting. The abstract and the third contribution highlight FEMNIST with 20% label noise, which appears only in Table A9. The main text points to the 40% comparison (Table A10) without mentioning that FedRG outperforms TriFed on every metric there. It also presents CIFAR-10 as supporting evidence without noting that TriFed trails on most metrics.
    • The supplement contradicts itself. It states that Table A12 (CIFAR-10, 20% noise) is the source of the right half of Figure 3. However, the values in Figure 3 match the clean Table A11 (e.g., TriFed R-tr +15.1 = 74.74 − 59.60, whereas Table A12 would give +16.0). The caption of Table A12 says its TriFed row "comes from the same runs as Table A11," yet the two rows differ (72.05 vs. 74.74 R-tr). Table A10 also uses a different client set (140 clients without the held-out filter) from every other table, so its numbers are not directly comparable.
    • The title and abstract emphasize "trustworthy," but Section 2.4 restricts the term to the reliability of training contributions. This framing overstates what the method provides.
    • Section 3 is dense with notation, and the constants that define the routing rules appear only in the supplement.
    • Formatting errors appear throughout. Examples include "Eq. equation 4" and "Eq. equation 12" in Sections 3.3 and 3.5 (and similar errors in Algorithm A1), author-year citations without parentheses (e.g., "without centralizing raw data McMahan et al. (2017)"), and ambiguous citations where "Liu et al. (2025)" and "Wang et al. (2025)" each refer to two different papers. The reference list uses a numbered, citation-order style rather than the ICLR format.
Questions for authors
  • Beyond combining personalized heads, sample filtering, influence-style utility scores, and clustered head offsets, what is the new technical insight? Please provide a sensitivity analysis for the routing and clustering constants listed in Appendix A.3 and A.4.
  • TriFed is not the best method on clean FEMNIST, on clean or noisy CIFAR-10, or on FEMNIST with 40% noise. Beyond FEMNIST with 20% noise, in which concrete setting should a practitioner prefer TriFed over Ditto, PGFedMo, or FedRG? Please report significance tests for the cross-method comparisons.
  • Why does the robustness observed on FEMNIST with 20% noise not carry over to 40% noise or to CIFAR-10? Please report the precision and recall of the γ filter on flipped labels, and explain how a filter that rejects at most about a third of flipped-label visits produces the gains in Table A9.
  • Removing tri-decomposition improves clean seen-client accuracy. Can the authors show that the filter does not harm clean data, and provide a paired analysis of the cluster channel in the clean setting?
  • Under concept and class-prior drift, is the method meant to suppress or adapt to the new labels? Please report baseline results under all four drift scenarios.
  • Can the authors add the missing baselines (unseen-client methods such as FLUX, profile mapping, and TGFed, noisy-label methods such as FedClean and FedELR, and head-based PFL methods such as FedRep and FedBABU)? Please also describe how baseline hyperparameters were tuned, and quantify how much of TriFed's gain depends on sharing labeled probe features, including the communication cost.
  • Can the authors resolve the inconsistencies among Figure 3, Table A11, Table A12, and Table A10, and revise the abstract and contributions to reflect results across all settings rather than the single favorable one?
Reviewer_qd7YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes TriFed, a personalized federated learning framework targeting three issues at the same time: (i) unseen clients at deployment, (ii) distribution drift, and (iii) unreliable or noisy local training contributions. TriFed uses three sample-level variables, \alpha, \beta, and \gamma, to determine whether each local sample should contribute to global training, personalized training, or be excluded as potentially harmful. Each client maintains a shared backbone, a global classifier head, and a private head. After local training, clients estimate the offset between their global and private heads, which is then used for client clustering. At deployment, an unseen client initializes from the final shared model, optionally selects a cluster-specific head using its support and validation data, and adapts only the classifier head using a small labeled support set. The paper also provides a detailed analysis of TriFed on Fashion-MNIST using multiple evaluation metrics.

Critical strengths

S1. Considering both unseen clients and noisy local dataset is one of the main novelty of this paper. Although there's some paper consider one of the aspect but no paper tackle there issues at the same time.

S2. Mechanism for removing the harmful samples is interesting. Although it is simple but it is straightforward and the result in this points is strong. With 20% label-flip noise, the authors report that TriFed exceeds the best competing method by 4.77 % in average seen-client accuracy and 9.00% in bottom-10% accuracy.

Critical weaknesses

W1. Presentation and clarity. The paper is difficult to follow. It is relatively easy to understand what the method aims to achieve, but it is difficult to understand how the method achieves it. Figure 1 should be simplified because the current version contains too much information. Its image quality is also insufficient, making it difficult to read when zoomed in or printed. The authors should explain the method more clearly at a higher level of abstraction.

W2. Limited empirical significance. The results in Table 1 do not demonstrate significant improvements over existing methods. TriFed does not achieve the best performance on five of the six metrics, with ZS-Std being the only exception. The experimental scope is also narrow, covering only two relatively small datasets, FEMNIST and CIFAR-10. The evaluation does not consider different levels of non-IID data or numbers of clients, which are commonly examined in FL research. The authors should discuss and evaluate these settings.

W3. Incomplete coverage of related work. The paper compares TriFed with other methods for unseen clients, such as Ditto, but does not discuss the paper “Enhancing the Generalization of Personalized Federated Learning with Multi-head Model and Ensemble Voting.” The authors should expand the related-work discussion and provide an appropriate comparison.

Questions for authors

See above

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

Related papers

Discussion (0)

Sign in to comment.