Abstract

Federated Learning (FL) enables collaborative model training across institutions without sharing raw data. Performance degrades, however, when label distributions are imbalanced across sites, a common scenario in medical settings where disease prevalence could vary drastically across institutions. Existing distribution-aware methods rely on label distribution statistics that clients cannot share in real-world federated settings due to privacy and regulatory constraints. We propose FedSwitch, a model-agnostic FL framework that estimates the global label histogram, under client-level epsilon-differential privacy, and uses the estimate to switch from FedAvg to distribution-aware aggregation. Each client adds partial discrete Laplace noise to its bounded label counts and secret-shares the result via Shamir’s scheme, designated decryptors reconstruct the noisy aggregate without ever observing individual contributions. The optimal switching round can be estimated via a closed-form variance decomposition that accounts for bounded contributions, client sampling, and differential-privacy noise components. We evaluate FedSwitch on medical imaging benchmarks under realistic heterogeneous settings, demonstrating consistent improvements over standard FL aggregation with negligible communication overhead.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6676_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Oxcart99-edu/MICCAI_FedSwitch

Link to the Dataset(s)

BloodMNIST dataset: https://medmnist.com OrganMNIST dataset: https://medmnist.com HAM1000 dataset: https://www.kaggle.com/datasets/kmader/skin-cancer-mnist-ham10000

BibTex

@InProceedings{CacPao_FedSwitch_MICCAI2026,
        author = { Cacace, Paolo AND Taiello, Riccardo AND Protani, Andrea AND Molina Van den Bosch, Marc AND Santos, Diogo Reis AND Brutti, Pierpaolo AND Serio, Luigi},
        title = { { FedSwitch: From Standard to Balanced Aggregation via Private Histogram Estimation in Federated Medical Image Classification } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16885},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduces FedSwitch, a model-agnostic federated learning (FL) framework designed to address label distribution imbalance across clients. The method estimates a global label distribution by leveraging client-level ε-differential privacy. Specifically, each client perturbs its local label counts using partial discrete Laplace noise and applies a secure sharing mechanism before sending the information to the server. A set of decryptors then reconstructs the aggregated noisy histogram without accessing individual client contributions, thereby preserving privacy. The authors evaluate FedSwitch on multiple medical imaging benchmarks under heterogeneous data settings and demonstrate consistent improvements over standard FL aggregation methods.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper addresses a timely and practically significant challenge in federated learning, namely, handling label distribution imbalance across clients. This issue is particularly critical in medical imaging applications, where data is inherently heterogeneous due to variations in acquisition protocols, patient populations, and institutional practices. By explicitly targeting distribution-aware aggregation, the work tackles a problem that directly impacts the reliability and fairness of FL models in real-world clinical settings.

    • A key strength of the paper is that it not only addresses label imbalance but does so while providing client-level differential privacy guarantees for shared label statistics. The proposed mechanism, perturbing label counts with discrete Laplace noise followed by a secret-sharing scheme, is both conceptually elegant and practically relevant. Notably, the design avoids additional privacy cost during aggregation and does not incur privacy budget composition across rounds, which is a non-trivial and valuable property in federated settings.

    • The method is evaluated on multiple medical imaging benchmarks, including HAM10000, BloodMNIST, and OrganMNIST3D, under heterogeneous data settings. The inclusion of diverse datasets and scenarios strengthens the empirical validation. The reported consistent improvements over standard FL aggregation methods further support the effectiveness of the proposed approach.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The paper does not convincingly justify the need for a new aggregation method, especially given the substantial progress in server-side aggregation strategies in federated learning over recent years. For instance, prior work such as FedMix [1] introduces adaptive client weighting mechanisms and operates in more complex settings like medical image segmentation with mixed supervision (pixel-level, bounding box, and image-level labels). In contrast, the current work is limited to medical image classification tasks, which represent a comparatively simpler setting. This gap weakens the motivation for proposing a new aggregation framework and makes it difficult to assess its practical significance. Furthermore, the references section is limited and does not adequately reflect recent advances (particularly post-2023 with citation of only one paper from 2024, 2025 and only one MICCAI 2023 paper) in federated learning and non-IID data handling. The lack of engagement with more recent literature reduces clarity on how the proposed method advances the state of the art.

    [1] Wicaksana, Jeffry, Zengqiang Yan, Dong Zhang, Xijie Huang, Huimin Wu, Xin Yang, and Kwang-Ting Cheng. “Fedmix: Mixed supervised federated learning for medical image segmentation. “ IEEE Transactions on Medical Imaging 42, no. 7 (2022): 1955-1968.

    • The proposed framework combines multiple components, namely, client-level differential privacy, discrete Laplace noise injection, and Shamir secret sharing, resulting in a relatively complex design. However, the paper does not clearly justify the necessity of each component or their combined role. In particular, while the method is framed as privacy-preserving, there is no empirical evaluation under concrete privacy attacks (e. g. , membership inference or model inversion attacks) to demonstrate its effectiveness in mitigating such risks. Additionally, the absence of ablation studies makes it difficult to assess the contribution of individual components or understand whether the full complexity of the framework is required.

    • In page 3, although the paper describes a secure aggregation threat model involving honest-but-curious decryptors and collusion assumptions, it does not clearly define the attacker’s objectives, capabilities, or success criteria. More importantly, the work does not include experimental validation under this threat model. Merely stating that the method operates within a given threat model is insufficient without demonstrating robustness against realistic adversarial scenarios.

    • The experimental evaluation is restricted to a small set of baselines (FedAvg, FedProx, FedIIC, FedIC), despite the availability of widely used adaptive federated optimization methods such as FedAdagrad, FedYogi, and FedAdam [2]. The omission of these baselines limits the strength of the empirical comparison. Moreover, key experimental analyses are missing: 1.The rationale and intuition behind defining R* in page 5 is not clearly stated. No clear intuition or justification is provided for the choice of hyperparameter R, despite its dataset-specific values (in page 6, R is considered as 21, 21, 10 for HAM10000, BloodMNIST, and OrganMNIST3D datasets, respectively). 2.No sensitivity or ablation analysis is conducted for hyperparameters. 3.Client scalability is not evaluated (e. g. , varying number of clients N or participation rate m). 4.Computational and communication overheads, including runtime, GPU utilization, and efficiency metrics, are not reported. These gaps make it difficult to assess the robustness, scalability, and practical applicability of the proposed method.

    [2] Reddi, S. J. , Charles, Z. , Zaheer, M. , Garrett, Z. , Rush, K. , Konečný, J. , Kumar, S. , McMahan, H. B. : Adaptive federated optimization. In: International Conference on Learning Representations (2021)

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper addresses an important and timely problem in federated learning, particularly distribution-aware aggregation under heterogeneous medical data, and proposes a privacy-preserving framework combining differential privacy and secure aggregation. However, the motivation for introducing a new aggregation method is not fully convincing given the extensive prior work in federated aggregation. The related work is limited and does not adequately position the method against recent advances. Additionally, the framework introduces multiple components (DP noise, secret sharing) without sufficient justification or ablation, and there is no empirical evaluation under concrete privacy attack scenarios. The experimental evaluation, while covering multiple datasets, lacks comparisons with stronger baselines, as well as analysis of scalability, hyperparameter sensitivity, and computational/communication overhead. Overall, while the paper has merit, these limitations weaken its contribution, leading to a weak reject recommendation.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    While the rebuttal clarifies several points (e.g., aggregation terminology and positioning), it does not sufficiently address my core concerns regarding the motivation and justification for the proposed aggregation method. In particular, the need for introducing a new aggregation framework remains unclear given the substantial prior work in distribution-aware and adaptive aggregation strategies. The rebuttal reiterates the integration of differential privacy and secure aggregation, but does not convincingly explain why these components, in their current combination, are necessary or how they provide a clear advancement over existing approaches.

    Additionally, the framework combines multiple elements, but the rebuttal does not adequately justify their individual roles or the necessity of the overall design. The argument that each component is required, without further clarification of their contribution to the final performance or robustness, leaves the design difficult to assess.

    While some clarifications are helpful, they do not sufficiently strengthen the conceptual grounding or address the primary concerns about novelty, necessity, and positioning relative to prior work. Therefore, I maintain my original assessment.



Review #2

  • Please describe the contribution of the paper

    FedSwitch is introduced in this paper as a federated learning mechanism that aims to alleviate the problem of label distribution skewness among the participating clients during medical imaging classification tasks. The basic principle behind the proposed approach involves computing the private label distribution by leveraging ε-differential privacy through the use of client-level privacy-preserving mechanisms for generating noise using the Pólya decomposition approach and then utilizing the Shamir secret sharing technique to reconstruct the label histogram, after which switching from conventional Federated Averaging (FedAvg) to inverse frequency weighting is performed by choosing the appropriate round based on a variance decomposition criterion.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.Distributed DP noise combined with Shamir secret sharing allows avoiding a trusted aggregator, which is a reasonable assumption in the context of multi-institutional healthcare environments. The choice of client-level ε-DP noise (and not sample-level one) seems appropriate given the threat model.

    2.The histogram is computed only once, and thus no composition happens across rounds. This is clearly a strong point of the work compared to methods querying the same sensitive statistics multiple times.

    3.Instead of treating the switching round as an empirical parameter, the authors propose a variance decomposition (equation 3), rendering R* both principled and understandable. Such a result can potentially be applied to other strategies for federated averaging.

    4.It is truly impressive how well FedSwitch behaves in the presence of severe heterogeneity (αℓ = 0.1), where baselines achieve macro-F1 ≤ 0.15, but FedSwitch scores 0.61.Analysis of the per-class performance for the least frequent label df (prevalence of 1.2%) and the t-SNE visualization prove its ability to give representations to underrepresented categories.

    5.Experiments on 2D dermoscopy, 2D microscopy, and 3D CT images with different degrees of intrinsic imbalance provide a fairly complete perspective on when the strengths of this technique shine.

    6.Secret sharing happens only once for each client at the initialization phase, and the extra communication cost per round is just one scalar (ϕᵢ).

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.While the paper benchmarks against FedAvg, FedProx, FedIIC, and FedLC, it does not consider other privacy-preserving methods such as SAAFL (referenced as [22]). This paper also uses cryptographic techniques in label-aware aggregation. It would have been helpful to compare against SAAFL in order to establish the extent to which the gain of FedSwitch is due to its distribution-aware aggregation or just using global histogram information that baselines do not have access to.

    2.The estimated global histogram is used only for once and will be used across all future rounds. There can be changes in the data distribution or addition of new users to the network in practical scenarios, but the paper leaves this as future work without discussing or quantifying its sensitivity to changes in data distribution over time.

    3.Using C = minᵢ nᵢ is appropriate in experimental scenarios to ensure no exclusion of client datasets. In practice, there may be heterogeneity in dataset sizes that are unknown, making it impractical to use one-shot DP query on minᵢ nᵢ.

    4.This paper focuses on histogram privacy, but fails to employ any type of end-to-end privacy guarantee via DP-SGD or even update-level privacy like SecAgg for the conducted experimental evaluations. Given that the research focuses on practical clinical implementation, it seems that not having an analysis with both histogram and model update privacy will limit real-world applicability of the results obtained.

    5.Although the variance breakdown of R* was justified, the paper lacks an ablation study between the two criteria used in switching namely between the criterion of R* and a simple criterion of either a fixed early round or late round of switching. This could have helped understand the sensitivity to the switching criterion used.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The problem of enabling distribution-aware federated learning while preserving client privacy addressed by FedSwitch is both justified and practical. The method itself is conceptually sound, relying on a reasonable combination of proven technologies (Shamir sharing and Pólya-decomposed Laplace noise). The experimentally validated switching criterion adds an element of mathematical rigor to the proposal. The experimental evidence on the HAM10000 dataset in the presence of extreme class imbalance is strong and well-separated from baselines. The most pressing issues are the lack of comparison with the nearest baseline which provides client privacy (SAAFL) and the lack of model updates’ confidentiality validation in the experiments. This does not detract from the primary contribution made by the paper, but raises doubts about the comprehensiveness of the experiment setup.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    The paper proposes a model-agnostic federated learning framework designed to address severe label imbalance in medical image classification under strict privacy constraints. Its core contribution is a privacy-preserving global label distribution estimation mechanism that combines client-level differential privacy with Shamir secret sharing, enabling distribution-aware aggregation without exposing sensitive label statistics. Additionally, the method introduces a closed-form variance-based criterion to determine the optimal round to switch from standard FedAvg to inverse-frequency weighting. This allows the framework to balance performance and privacy effectively, achieving improved recognition of rare classes while maintaining negligible communication overhead.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    This paper: 1.combines client-level differential privacy with Shamir secret sharing to estimate global label distribution without exposing sensitive statistics—addressing a key limitation of prior methods; 2.proposes a clean mechanism to switch from FedAvg to distribution-aware aggregation, improving stability and avoiding noisy early-stage estimates; 3.provides a closed-form variance decomposition to determine the optimal switching round, offering clear insight into DP noise, sampling, and heterogeneity; 4.specifically targets label skew and strict privacy constraints common in real-world medical scenarios; 5.demonstrates significant gains, especially on highly imbalanced datasets and rare classes; 6.requires only one-shot estimation with minimal additional communication, making it efficient and scalable.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The core idea of using label distribution–aware reweighting (e.g., inverse-frequency weighting) is not new. The main difference here lies in the privacy-preserving estimation, rather than the aggregation itself. 2.The combination of differential privacy and secure aggregation (e.g., Shamir secret sharing) follows a well-established paradigm in federated analytics. The proposed protocol is technically sound but largely an integration of existing components. 3.Experiments assume relatively synchronous training, stable client participation, and known hyperparameters. The robustness of the method under more practical conditions (e.g., partial participation variability, stragglers, or unknown client dataset sizes) is not explored.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The proposed framework is well-designed, combining differential privacy and secure aggregation to estimate global label distributions without exposing sensitive information, which is a meaningful contribution in realistic settings. The adaptive switching strategy and accompanying variance analysis further strengthen the methodological coherence. Moreover, the experimental results are convincing, showing clear improvements, particularly for rare classes under severe heterogeneity. However, the overall methodological novelty is moderate, as several components build on existing techniques. Despite this, the strong empirical performance and practical relevance makes this paper looks good overall.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The rebuttal reasonably addresses the main concerns. The paper’s main contribution is not the inverse-frequency weighting itself, but the integration of privacy-preserving global label histogram estimation with distribution-aware federated aggregation, which is meaningful and technically coherent. Some limitations remain, such as dynamic client participation, stragglers, unknown client sizes, model-update privacy, and comparison with stronger privacy-aware baselines. However, these are not fatal enough to overturn my original recommendation.



Author Feedback

We thank R1, R2, R4 for recognizing label-skewed FL relevance, the one-shot DP query mechanism integration, and the significant gains achieved by our method. Novelty [R1W1, R2W1, R4W1-2]. To the best of our knowledge, FedSwitch is the first method to perform label-based inverse-weighted aggregation under formal client-level DP [15]. While inverse weighting itself is established, our contribution lies in efficiently integrating it with a secure federated analytics layer, enabling distribution-aware aggregation without requiring clients to disclose their label histograms. Importantly, the key point is not that FedSwitch directly competes with prior non-DP methods on privacy grounds, but that even under strict client-level DP(ε=1) FedSwitch still achieves stronger empirical results. As in Sec. 1, SAAFL [22], the closest privacy-related method, provides cryptographic protection (no DP) but assumes full participation. We will clarify this in the paper. Pipeline integrity [R1W2]. A component-removal ablation would be uninformative because each block enforces a distinct formal property. Distributed discrete-Laplace noise is required for ε-DP. Shamir is required for rolling accumulation across sampled rounds with a single final release. SecAgg [12] protects single-round aggregates, and we use it after R* to protect the final estimate. Baselines [R1W4]. FedAdam/Yogi/Adagrad are server-side optimizers orthogonal to label-aware aggregation; they could be combined with FedSwitch rather than compared to it. After comparison with the standard methods, we focused on client-side label-aware baselines (FedIIC [4], FedLC [7]) which target the same problem. Segmentation [R1W1]. While segmentation is arguably more sophisticated, it is addressed via different mechanisms; advances there do not remove the need for new image classification-specific aggregators, a distinct and clinically relevant setting. FedMix targets mixed-supervision segmentation, whereas we target classification leveraging private histogram estimation. Extending FedSwitch to segmentation by informing loss-level weighting (Dice + weighted CE) with the private class-frequency estimate is considered as a direction for future work. R* / ablation [R1W4, R2W5]. R* is derived from (Eq.3), not tuned: the earliest round reaching coverage ρ. We fix a high ρ value, assuming a known client population and uniform infrastructure, reasonable in medical imaging clinical federations that are often built on formal multi-institutional agreements. We preferred an ablation on strong ε values to validate robustness and performance consistency. Threat model [R1W3, R2W4]. Our contribution focuses on the protection over individual (Shamir [11]) and aggregated (DP) client histograms. The adversary is the server colluding with up to M−t−1 honest-but-curious decryptors. Model-update privacy is orthogonal and addressable by local DP [23], distributed DP [21], DP-SGD [20] and SecAgg (Sec. 2). Practicality and scalability [R1W4,R2W2-3,R4W3]. Our experiments span three datasets with different imbalance and modalities, client counts N∈{50,100}, and label skew α_l∈{0.1,0.5,1.0}. Consistent gains, especially at α_l=0.1, highlight robustness under severe label skew. The protocol overhead is lightweight (Sec. 3), runs on CPU, and is negligible relative to model updates. We will clarify in the paper. We set C = min nᵢ to avoid excluding clients; when sizes are unknown, dataset distributions can be estimated via a one-shot DP query [16], and min nᵢ is deducible from the lowest non-empty bin. Discussing client dataset variability over time or strong participation variability (stragglers etc.) would require a multi-release protocol with composed privacy budget, together with an extensive analysis of edge case scenarios that the paper page limit can hardly accommodate. We acknowledge this as future work and prioritized validating the one-shot mechanism, supported by per class analysis and t-SNE visualization.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The paper proposes a model-agnostic federated learning framework for medical image classification under client label-distribution imbalance. The method privately estimates the global label histogram using client-level differential privacy and Shamir secret sharing, then switches from standard FedAvg to inverse-frequency distribution-aware aggregation. The method is evaluated on HAM10000, BloodMNIST, and OrganMNIST3D.

    The reviews are mixed. The paper addresses an important and practical problem, the combination of client-level differential privacy and Shamir secret sharing is practical and meaningful, and the empirical results appear strong for severe heterogeneity. However, there are several major concerns need to be addressed: 1.Several components building on existing techniques and the need for new aggregation method needs to be further justified. 2.The experiments may lack comparison with certain baseline methods. 3.Ablation study may be insufficient. 4.The evaluation does not empirically test concrete privacy attacks or provide end-to-end privacy guarantees for model updates. 5.Concerns about stable client participation, unknown client dataset sizes, changing distributions over time, and missing scalability or overhead analyses.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The paper introduces FedSwitch, a model-agnostic federated learning framework that addresses severe label distribution skew across clients in medical image classification under strict privacy constraints. Based on the reviewers’ feedback and the authors’ rebuttal, I recommend accepting this manuscript. The rebuttal sufficiently clarified the main contribution as the integration of privacy-preserving global label histogram estimation with distribution-aware aggregation, which R4 found “meaningful and technically coherent”, and the reviewers acknowledged the strong empirical results, particularly under severe heterogeneity. The authors should carefully prepare the final version to resolve the motivation and justification concerns from R1.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    Though the reviews are mixed, the concerns raised by reviewer R1 appear to be valid. Specifically, it is not clear if the proposed distribution-aware aggregation is necessary and beneficial (even assuming there is no privacy constraint). The argument in the rebuttal about the infeasibility of component-wise ablation is not convincing. Given these major limitations, the paper should not be accepted in its current form.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The authors’ rebuttal adequately addresses the major concerns raised by the reviewers. While Reviewer 1 remains unconvinced about the novelty and justification, both other reviewers acknowledge that the rebuttal clarifies the core contribution—integrating client-level DP with secure aggregation for distribution-aware FL—and resolves key issues regarding baselines, threat modeling, and experimental design. The method demonstrates strong empirical gains under severe label skew, and the technical approach is sound.



back to top