acceptodds
Withdrawn from ICLR 2027

FedPGRL: Prototype-Guided Refinement for Noisy Long-Tailed Federated Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated Learning (FL) under label noise and long-tailed class distributions poses a fundamental challenge, as biased local supervision and skewed data jointly distort global optimization. This issue is especially pronounced when noisy annotations disproportionately affect rare classes, amplifying representation collapse across clients. We propose Federated Learning with Prototype-Guided Refinement Learning (FedPGRL), a unified framework built on CLIP2FL that addresses this coupling through prototype-driven sample refinement. FedPGRL leverages class prototypes as stable semantic anchors to guide three refinement stages: Prototypical Distance Selection (PDS), Prototypes Contrast Refinement (PCR), and Multi-Prototypical Mask Refinement (MPMR), which collectively generate soft labels and suppress noise while preserving minority-class structure. We further introduce Class Prototypical Regularization (CPR) to explicitly encourage inter-class separation and intra-class compactness during federated optimization. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet validate that FedPGRL consistently improves convergence stability and generalization under both noisy and imbalanced FL settings.

Primary area:
reinforcement learning
Keywords:
Robust Federated Learning, Federated Long Tail Learning, Federated Noisy Label Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ae4iOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents FedPGRL for federated learning scenarios plagued by both long‑tailed class distributions and heterogeneous label noise. Built upon CLIP2FL, FedPGRL incorporates three prototype‑guided refinement modules. These modules include Prototypical Distance Selection (PDS), Prototypes Contrast Refinement (PCR), and Multi‑Prototypical Mask Refinement (MPMR). PDS leverages prototype distances alongside a Gaussian mixture model to distinguish clean samples from noisy ones. PCR generates prototype‑aware soft labels for potentially corrupted samples. MPMR further derives corrected representations to support class‑gradient training. The work also introduces Class Prototypical Regularization (CPR) while preserving the original CLIP‑driven knowledge regularization. Evaluations are carried out on CIFAR‑10‑LT, CIFAR‑100‑LT and Tiny‑ImageNet‑LT under simulated label noise and long‑tailed distribution settings.

Critical strengths

--Simultaneously addressing class imbalance and heterogeneous annotation noise is practically meaningful, and the experiments cover multiple imbalance and noise configurations.

--FedPGRL substantially improves over CLIP2FL under severe noise and imbalance on both CIFAR-10-LT and CIFAR-100-LT, and the component ablations indicate that PDS, PCR, and MPMR each contribute to performance.

Critical weaknesses

--The paper clearly defines an input sample x and its learned feature z=F_\theta(x), but subsequent prototype operations repeatedly use x directly in feature-space computations, including PDV calculation, cosine similarity with class prototypes, MPMR interpolation, and CPR regularization. For example, Eq. (7) combines x with a weighted sum of feature prototypes, which is dimensionally inconsistent if x denotes the raw image. Similar ambiguity appears in Eq. (9). These issues affect the correctness and reproducibility of the central PDS–PCR–MPMR–CPR pipeline, rather than being merely cosmetic notation problems.

--CPR is repeatedly described as encouraging intra-class compactness while explicitly pushing different classes apart, yet its loss is only an MSE attraction to the assigned class prototype, with no inter-class repulsion, margin, or contrastive term. In addition, the paper claims that FedPGRL requires “no explicit server-side classifier retraining,” while Section 2.5 explicitly introduces a classifier retraining stage. The manuscript also states that weighted gradient aggregation preserves privacy, despite providing no formal privacy mechanism or leakage analysis. These inconsistencies weaken the technical credibility of the method description.

--PDS is motivated as a way to avoid unreliable prototypes caused by noisy labels, but the cluster initialization and class alignment still rely on prototypes computed from those same noisy labels. The resulting process effectively uses noisy labels to construct noisy prototypes, uses them to align clusters, and then uses the aligned clusters to infer which labels are noisy. This is particularly concerning under severe long-tail conditions, where minority classes may contain very few samples and their prototypes can be highly unstable. The paper reports only final classification accuracy and does not directly evaluate clean/noisy detection quality, relabeling accuracy, or performance separately for head and tail classes.

--Although many combinations of imbalance factors and synthetic noise rates are considered, the experiments are still mainly conducted on CIFAR-10-LT, CIFAR-100-LT, and Tiny-ImageNet-LT with simulated imbalance and label corruption. This does not substantiate the manuscript’s claim of effectiveness under “real-world FL settings.” Moreover, for a long-tailed learning paper, overall accuracy is insufficient because it may be dominated by head classes; detailed Many/Medium/Few results are only provided for Tiny-ImageNet in the appendix. For a noisy-label method, the paper should also directly report noise-detection precision/recall, corrected-label accuracy, and tail-class relabeling quality.

--FedPGRL is explicitly built on CLIP2FL, and its main additions consist of prototype-distance filtering, GMM-based clean/noisy separation, prototype-based soft relabeling, hard pseudo-label conversion, prototype interpolation, and MSE prototype regularization. These are individually well-established techniques in noisy-label learning and prototype-based representation learning. The current ablation shows that the proposed components improve the CLIP2FL backbone, but does not demonstrate that the particular design is superior to simpler alternatives such as confidence-based relabeling, loss/GMM filtering, nearest-prototype relabeling, class-balanced loss, or standard supervised contrastive regularization.

--FedPGRL introduces multiple client-side operations, including feature extraction, K-means clustering, GMM fitting, prototype construction, PCR, MPMR, CPR, and class-gradient computation, together with server-side gradient matching, federated-feature optimization, and classifier retraining. However, the paper does not report client runtime, server runtime, memory overhead, communication per round, total communication, or scaling with the number of classes and clients. This is inconsistent with the manuscript’s claims regarding improved scalability and communication efficiency, especially since the authors themselves acknowledge that runtime and memory scaling remain future work.

--The paper includes several grammatical errors, awkward expressions, inconsistent terminology, and typographical mistakes. For example, “Prototypes Contrast Refinement” and “Prototypical Contrast Refinement” are used inconsistently; “EXPERIENCE RESULTS” should be “EXPERIMENTAL RESULTS”; “distibution” is misspelled; “the left part of 3” should be revised to “the left part of Fig. 3”; “AdaM algorithm” should be written as “Adam optimizer”; and “FedPRGL” appears instead of “FedPGRL” in the appendix. In addition, the figure formatting is inconsistent: the caption of Fig. 1 is placed above the figure, whereas the captions of the other figures are placed below their corresponding figures. The manuscript would benefit from a thorough proofreading and a careful check of formatting consistency throughout the paper.

Questions for authors

--How should the prototype-based operations in PDS, PCR, MPMR, and CPR be interpreted consistently?

--CPR is described as improving both intra-class compactness and inter-class separation, but the presented objective mainly minimizes the distance between samples and their assigned class prototypes. Could the authors explain where the explicit inter-class separation is enforced?

--Since PDS relies on noisy-label-derived class information to initialize or align clusters, how robust is the clean/noisy sample selection when label noise is severe, especially for tail classes with very few samples?

Reviewer_zoJ5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to address the challenge of simultaneous label noise and long-tail data distribution in federated learning. It introduces a client-side "prototype-guided refinement" mechanism built upon the CLIP2FL framework. This mechanism consists of three stages: Prototype Distance Selection (PDS), which performs K-means clustering on local features and aligns them with the class means of noisy labels to derive "clean" class prototypes; it then uses a Gaussian Mixture Model to categorize samples; Prototype Contrastive Refinement (PCR) generates soft labels for suspicious samples by calculating the cosine similarity between these samples and the clean class prototypes, which are then used for local training. Multi-Prototype Masked Refinement (MPMR) constructs "masked prototypes" by taking a weighted sum of clean prototypes, interpolates these with the original features, and applies an argmax operation to derive hard labels, which are subsequently used to compute class-specific gradients for transmission to the server. Furthermore, the authors propose a Class Prototype Regularization (CPR) loss, which combines cross-entropy loss with the mean squared error between sample features and their corresponding clean class prototypes while retaining the original CLIP knowledge distillation component from CLIP2FL. On the server side, the approach follows the CReFF/CLIP2FL pipeline: aggregating gradients, training federated features via gradient matching, and retraining the classifier using class-balanced federated features. Experiments on CIFAR-10/100-LT and Tiny-ImageNet-LT demonstrate that FedPGRL outperforms FedAvg, FedProx, FedCorr, FedNed, CReFF, and CLIP2FL across all evaluated settings.

Critical strengths
  1. The problem addressed in this paper has practical significance. The co-occurrence of noise and long-tailed distributions is a realistic scenario; furthermore, Figure 1a clearly demonstrates the vulnerability of existing methods under these combined conditions, providing a strong motivation.

  2. The performance gains observed in experiments under high-noise conditions are significant, indicating that the proposed method effectively mitigates the contamination of class-specific gradients caused by noise.

Critical weaknesses
  1. My primary concerns relate to the writing. Specifically:

. Throughout the paper, citations use the \citet style (e.g., "FedCorr Xu et al. (2022)"); these should be changed to \citep.

. There is no "Related Work" section in the main text; it has been placed entirely in the appendix. I suggest moving the "Related Work" section to the main text.

. There are errors in cross-references: the main text refers to Tiny-ImageNet results in "Table 5," but they appear in Table 4 in the appendix; additionally, Algorithm 1 references an "Eq. 16" that does not exist.

. The notation is inconsistent: in the main text (line 196), xx is defined as a sample, yet in several equations (such as Equations 5 and 6), xx appears to represent a feature—a definition necessary for operations involving prototypes. This is confusing.

. There are several spelling errors. For instance, the heading for Section 2.1 uses "notion" instead of "notation"; Section 3.2 contains "Experience results" and "Experience of α" (likely intended to be "Experiments" and "Effect of α," respectively); and there are typos such as "distibution" and "Refienment."

. The caption for Figure 1 appears above the figure, while captions for other figures appear below; the formatting should be consistent throughout the paper.

. The figures included in the manuscript are not in vector format.

  1. The method proposed in this paper appears to be an incremental combination of existing approaches. The server-side design is the same as that of CReFF/CLIP2FL. On the client side, the introduced components, including PDS (feature clustering combined with GMM-based separation of clean and noisy samples), PCR (soft labeling based on prototype similarity), and CPR (an MSE-based loss pulling samples toward prototypes), are standard techniques within the fields of noisy-label learning and prototype-based learning. The novelty, therefore, seems to lie primarily in the specific combination of these techniques and their adaptation to the federated learning setting.

  2. Regarding the experiments, the paper does not compare the proposed method against any Federated Learning approaches that simultaneously address both noise and class imbalance (such as FedNoRo). Instead, it only compares with methods designed exclusively for handling noise or long-tailed distributions as baselines; under these conditions, it is hardly surprising that the proposed method performs well.

  3. Furthermore, the paper does not report any computational cost comparison.

Questions for authors

Please see weaknesses.

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

This paper studies federated learning under both label noise and class imbalance. The proposed FedPGRL method uses prototype-based sample separation and label refinement to handle noisy training data. It also incorporates prototype regularization and server-side feature synthesis. Experiments are conducted on long-tailed versions of CIFAR-10, CIFAR-100, and Tiny-ImageNet under different noise settings.

Critical strengths

The problem considered in this paper is practically relevant. Label noise and class imbalance often occur together in federated learning, while many existing studies consider them separately. The proposed use of prototypes to identify and refine potentially noisy samples is reasonable. The experimental results also show noticeable improvements in some challenging settings, particularly when the noise level and imbalance ratio are both high (64.5 vs 51.8 on CIFAR-10-LT at IF=100 with (0.8, 0.5) noise).

Critical weaknesses

I have several concerns about the method and its experimental validation. First, the method description is not sufficiently clear. Some important variables and intermediate operations are not properly defined. For example, the construction of noisy prototypes is unclear. The description of the GMM-based sample separation is also difficult to follow. Second, the technical novelty is limited. The main components, including GMM-based noise separation, prototype-based label refinement, and federated feature synthesis, have been studied in previous work. The paper does not adequately explain how their combination specifically addresses the interaction between label noise and class imbalance. Third, the experimental analysis is insufficient. The authors report improvements in overall classification accuracy, but provide little analysis of noise detection performance for minority classes. This is particularly important because prototype estimation may be unreliable when only a few local samples are available. Additional experiments are needed to determine which components are responsible for the reported improvements. Moreover, the paper claims efficiency advantages without providing sufficient computational or communication analysis. The paper mentions that the experiment on efficiency is in the appendix, but it was not found. There are also several notation and presentation errors that should be corrected.

Questions for authors

What is the specific role of CLIP in the proposed framework? Which components of CLIP2FL are retained? How reliable is the noise separation procedure for minority classes with very few local samples? Per-class precision and recall would be useful. Were all baseline methods tuned under the same noise and imbalance settings? In particular, the relatively poor performance of FedNed requires some explanation.

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

Related papers

Discussion (0)

Sign in to comment.