Pith. sign in

REVIEW 5 major objections 5 minor 22 references

Robust Federated Learning on Edge Devices with Domain Heterogeneity

T0 review · 5 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read FedAPC, by aligning local features with global prototypes built from augmented views, claims higher average accuracy under domain heterogeneity than FedAvg, MOON, and FedProto, with 80.57% on Digits and 66.39% on Office-10.

desk verdict FedAPC's reported gains come from data augmentation that the baselines never got, not from the prototype-contrastive alignment the paper proposes—a fixable flaw, but the central SOTA claim is not supported. read the letter →

arxiv 2505.10128 v1 pith:G5ZC5QSA submitted 2025-05-15 cs.LG cs.AI

Huy Q. Le , Latif U. Khan , Choong Seon Hong This is my paper · ORCID
classification cs.LGcs.AI
keywords federatedlearningdomainheterogeneityprototypecontrastivedataaugmentationshiftnon-IIDedgedevices
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes FedAPC, a federated learning method that targets the case where different edge devices hold data from different visual domains but the same label set. FedAPC claims that by computing local class prototypes from augmented views and training a contrastive loss that pulls local features toward same-class global prototypes, the global model becomes more robust to domain shift. On two four-domain benchmarks, Digits and Office-10, FedAPC reports higher average accuracy than FedAvg, MOON, and FedProto, 80.57% versus 78.90% and 66.39% versus 64.41%, together with faster and more stable convergence. A sympathetic reader would care because domain heterogeneity is a practical obstacle for federated systems deployed across real-world devices, and prototype-based regularization is communication-cheap.

What carries the argument

The load-bearing object is the augmented prototype, defined as the mean feature $\bar{z} = \frac{1}{N}\sum_n z_n$ over augmented views of an input, aggregated per class to form $\bar{p}^m_k$ and then averaged at the server into global prototypes $G^m$. The argument runs through the augmented prototype contrastive loss $\mathcal{L}_{APC} = -\log \frac{\sum_{g^+ \in G^+}\exp(s(h(x), g^+)/\tau)}{\sum_{g_k \in G}\exp(s(h(x), g_k)/\tau)}$, which is added to cross-entropy. This machinery carries the claim by coupling local feature learning to global semantic anchors that are derived from a deliberately diversified feature set, so that no single client's domain dominates the representation.

What would settle it

Run FedAPC against FedProto and MOON with identical data augmentation, the same backbone, the same seeds, and repeated trials; if the average accuracy advantage over the best baseline drops below run-to-run noise, such as overlapping within one standard deviation on both Digits and Office-10, the central claim is not supported. Additionally, checking whether removing the contrastive term while keeping augmentation reproduces the same accuracy would test whether the contrastive alignment matters at all.

Watch

Extended reading notes

Core claim

FedAPC establishes that prototype augmentation, building class prototypes from the mean of augmented feature views rather than from raw features, is sufficient to improve generalization across client domains. On each client, the method computes $\bar{p}^m_k = (1/|S^m_k|)\sum_{i \in S^m_k} \bar{z}_i$ from mean augmented features, uploads these local prototypes, and the server forms global prototypes $G^m = \sum_k \bar{p}^m_k$. During local training, the cross-entropy loss is supplemented by a contrastive term that maximizes the cosine similarity between each local feature and the global prototype of its class while suppressing similarity to other classes. The reported result is that this alignment reduces overfitting to a client's own domain and yields better average accuracy, especially on the hardest domains, SYN and DSLR, and smoother convergence over 100 communication rounds.

Load-bearing premise

The central claim assumes that the higher reported accuracy comes from the augmented-prototype contrastive mechanism rather than from unequal training conditions, because baselines are not given the same augmentation pipeline and the reported averages carry no variance.

Editorial extensions

If this is right

  • Adding augmentation-based prototypes to existing prototype federated learning yields accuracy gains on domain-heterogeneous benchmarks without changing the model architecture.
  • The method converges faster and more stably, so fewer communication rounds may be needed to reach the same average accuracy.
  • The largest gains appear on the hardest, least-represented domains, SYN and DSLR, suggesting the mechanism helps poorly represented clients most.
  • Because only prototypes and model parameters are exchanged, the approach fits communication-limited edge-device federated learning settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The claimed gap might shrink or disappear when baselines receive the same augmentation pipeline; a controlled comparison with matched augmentation would isolate the contribution of prototype contrastive alignment.
  • The closest clustering-based prototype method cited in the paper is not compared, so evaluating FedAPC against it would test whether the gain comes from augmented prototypes specifically or from contrastive prototype alignment generally.
  • A useful stress test is measuring per-client feature drift before and after alignment; if prototypes reduce inter-client distance in feature space, that would mechanistically explain the accuracy gain.
  • The method could be extended to partial label overlap or open-set domains, where global prototypes of classes absent on a client might mislead the contrastive term.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes FedAPC, a federated learning method for domain heterogeneity in which clients first compute local augmented prototypes as the mean feature of multiple augmented views (Eqs. 2 and 3), the server aggregates these into global prototypes (Eq. 4), and local training aligns local features with the corresponding global prototypes via a contrastive loss (Eq. 5). The total local objective is the sum of cross-entropy and the proposed contrastive loss (Eq. 6). The method is evaluated on Digits and Office-10 with eight clients, comparing against FedAvg, MOON, and FedProto, and the paper reports higher average accuracy and faster, more stable convergence, together with an ablation study on the effect of augmentation.

Significance. If the empirical claims were properly supported, FedAPC would be a simple and reasonably motivated extension of prototype-based FL, and the paper identifies a real limitation of existing prototype methods (dependence on unaugmented local features). The manuscript deserves credit for including an ablation study and for using external benchmark evaluation, so the reported gains are not obtained by fitting a model to the benchmark. However, the central claim of consistent superiority over existing methods currently rests on an unfair comparison: the baselines do not receive the same augmentation pipeline, and the paper's own ablation shows that the margin over the best baseline disappears when augmentation is removed. The omission of the closest prior prototype-contrastive method and the absence of variance reporting further weaken the empirical core. The idea is plausible and worth publishing if the comparison is put on equal terms, but the paper in its current form does not support the claim that FedAPC's prototype-contrastive alignment is the source of the reported gains.

major comments (5)
  1. [Section III.B, Tables I and II] The attribution of the reported improvement to FedAPC's prototype-contrastive learning is undermined by the ablation study. In Table II, removing augmentation from the local prototypes drops average accuracy to 78.22 on Digits, below MOON (78.90) and FedProto (78.70), and to 61.93 on Office-10, below FedProto (64.41) and MOON (62.13). Since the baselines in Table I are not trained with FedAPC's augmentation pipeline, the entire average gap in Table I can be explained by augmentation rather than by the contrastive alignment in Eq. 5. The authors should either run all baselines with the same augmentation pipeline or substantially temper the claim that FedAPC outperforms existing methods.
  2. [Section III.A, Implementation Details] The manuscript states 'We fix the seed to ensure the reproducibility' but also reports average accuracy over three runs. With a fixed seed, the three runs are not independent, and no standard deviation or per-run results are provided. Given that the reported improvements are 1.67 and 1.98 percentage points over the second-best method, run-to-run variance could be of the same magnitude. The paper should report mean and standard deviation over multiple independently seeded runs, or clearly state that a single seed was used and only one run's trajectory is shown.
  3. [Section III.A, Baselines] The closest prior prototype-contrastive method cited in the paper, FedCCL [14], is not included as a baseline, although it also addresses domain heterogeneity in federated learning with clustering-based prototypes and contrastive feature alignment. Without a comparison to this method, the claim of outperforming 'SOTA baselines' is not established. At minimum, the authors should compare against FedCCL or justify its exclusion with concrete differences in problem setting and evaluation.
  4. [Section III.B, Table I] The claim that 'FedAPC consistently surpasses the other methods' is not supported even by the reported numbers: on SVHN in Digits, FedAPC (81.98) is below MOON (82.18), and on Amazon in Office-10, FedAPC (78.42) is below FedAvg (79.79) and FedProto (83.89). The average-accuracy advantage is real in the table, but the per-domain results show that the method does not uniformly dominate. The text should be revised to describe the result as an average improvement with specific weaker domains noted.
  5. [Section II.B, Eq. (4)] Equation (4) writes the global prototype as a sum over client prototypes, G_m = sum_k \bar{p}^m_k, which conflicts with the text's statement that 'prototype aggregation is performed' by averaging. An unweighted sum over clients is not an average, and it ignores differing client dataset sizes, which are substantial here (e.g., different sample counts from different domains). If the intended operation is a weighted average (as in FedProto), the formula and the aggregation step in Algorithm 1 should be corrected and the weights specified.
minor comments (5)
  1. [Abstract] The phrase 'the global mode's convergence' should read 'the global model's convergence', and there are several typographical issues throughout the introduction, including 'boomming era' and 'paramaters'.
  2. [Section II.B, Eq. (5)] The notation in Eq. (5) is incomplete: the positive set G^+ is not defined, and the expression '(s(h(x),g +)' contains a misplaced parenthesis. It should be written as \exp(s(h(x), g^+)/\tau) with a clear definition such as G^+ = {g \in G : label(g) = y}.
  3. [Section II.B, Eq. (2)] The symbol N is described as 'the total number of augmented versions for each input', which is ambiguous; it should be clarified that N is the number of augmented views generated per sample, and the summation index should be written consistently over n = 1, ..., N.
  4. [Algorithm 1] The parallel loop in Algorithm 1 runs 'for k = 0, 1, ..., K', but the paper defines K clients indexed from 1 to K elsewhere; also, the server aggregation in Line 9 omits the prototype aggregation step that is described in the text and shown in Eq. (4).
  5. [Section III.A, Datasets] The paper should state the training/communication rounds T, the number of local epochs, and the optimizer details; these hyperparameters are needed for reproducibility and for judging whether all methods are trained under equal compute budgets.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: FedAPC is an externally benchmarked empirical method; the only self-citations are contextual and not load-bearing.

full rationale

The paper makes no theoretical prediction and fits no constant to the benchmark claim. FedAPC's components are stated as training objectives: augmented local prototypes (Eqs. 2-3), global prototype aggregation (Eq. 4), and a prototype contrastive loss (Eq. 5). Nothing in these equations is defined in terms of the reported accuracies, and the Table I numbers are measured against external datasets (Digits, Office-10) and external baselines (FedAvg, MOON, FedProto), so the central empirical claim does not reduce to its inputs. The references to the authors' prior work ([12], [14]) are contextual survey citations about complete prototypes and clustering-based unbiased prototypes; they are not used as a uniqueness theorem or to forbid alternative methods, so they are not load-bearing. The ablation in Table II does show that removing augmentation drops FedAPC below MOON and FedProto on both averages, and the omission of FedCCL [14] plus the absence of variance makes the SOTA comparison weaker; these are fairness and reproducibility concerns, not circularity. Accordingly the reader's score of 1.0 is appropriate: at most there are minor non-load-bearing self-citations, with the derivation itself self-contained.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities or mathematical constants. Its load-bearing commitments are modeling choices: augmented-average prototypes, unnormalized global aggregation, and contrastive alignment, together with standard hyperparameters whose values are not reported.

free parameters (4)
  • temperature tau in Eq. 5
    Controls contrastive sharpness; no value is reported and the paper does not state how it was chosen.
  • number of augmented views N per sample in Eq. 2
    Defines the augmented mean feature; never specified.
  • augmentation operations and strengths
    Erasing, cropping, and flipping are listed without magnitudes or probabilities; results likely depend on them.
  • local training rounds or epochs
    Algorithm 1 says 'for each local round' but does not give the number; learning rate 0.01 and batch size 32 are given.
assumptions (4)
  • domain assumption All clients share the same label space (P_i(y) = P_j(y))
    Used to define class-matched global prototypes and positive pairs in Eq. 5; fails under label heterogeneity.
  • ad hoc to paper Mean feature of augmented views is a valid class prototype
    Eqs. 2 and 3 define prototypes as averages; the paper offers no argument that this average is stable under domain shift.
  • domain assumption Data augmentation preserves semantic class
    Contrastive alignment assumes augmented views of x are positive examples of the same class; heavy augmentation could violate this.
  • ad hoc to paper Averaging local prototypes at the server yields a useful global prototype
    Eq. 4 sums local augmented prototypes without normalizing by client sample counts; the paper does not justify this aggregation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Federated Learning on Edge Devices with Domain Heterogeneity." pith.science (2026). https://pith.science/paper/G5ZC5QSA

@misc{pith2026250510128,
  author       = {Pith},
  title        = {Pith review of: Robust Federated Learning on Edge Devices with Domain Heterogeneity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G5ZC5QSA}},
  note         = {Machine review of arXiv:2505.10128}
}
read the original abstract

Federated Learning (FL) allows collaborative training while ensuring data privacy across distributed edge devices, making it a popular solution for privacy-sensitive applications. However, FL faces significant challenges due to statistical heterogeneity, particularly domain heterogeneity, which impedes the global mode's convergence. In this study, we introduce a new framework to address this challenge by improving the generalization ability of the FL global model under domain heterogeneity, using prototype augmentation. Specifically, we introduce FedAPC (Federated Augmented Prototype Contrastive Learning), a prototype-based FL framework designed to enhance feature diversity and model robustness. FedAPC leverages prototypes derived from the mean features of augmented data to capture richer representations. By aligning local features with global prototypes, we enable the model to learn meaningful semantic features while reducing overfitting to any specific domain. Experimental results on the Office-10 and Digits datasets illustrate that our framework outperforms SOTA baselines, demonstrating superior performance.

Figures

Figures reproduced from arXiv: 2505.10128 by the authors.

Figure 1
Figure 1. Problem illustration of federated learning under domain [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of FedAPC. where LCE is the cross-entropy loss calculated between the predicted output and its corresponding label. C. FedAPC Algorithm This section presents the FedAPC algorithm, as outlined in Alg. 1. In each communication round, server broadcast the global model θt and global prototypes G to all clients. Each client then conducts the local training process with its private data Dk, updating the local… view at source ↗
Figure 3
Figure 3. Visualization of average accuracy across with [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 5 canonical work pages

  1. [14]

    Fedccl: Federated dual-clustered feature contrast under domain heterogeneity,

    Y . Qiao, H. Q. Le, M. Zhang, A. Adhikary, C. Zhang, and C. S. Hong, “Fedccl: Federated dual-clustered feature contrast under domain heterogeneity,” Information Fusion, vol. 113, p. 102645, 2025

    work page 2025

  2. [1]

    Federated learning for healthcare: Systematic review and architecture proposal,

    R. S. Antunes, C. Andr ´e da Costa, A. K ¨uderle, I. A. Yari, and B. Eskofier, “Federated learning for healthcare: Systematic review and architecture proposal,” ACM Transactions on Intelligent Systems and Technology (TIST), vol. 13, no. 4, pp. 1–23, 2022

    2022

  3. [2]

    Federated rec- ommendation systems,

    L. Yang, B. Tan, V . W. Zheng, K. Chen, and Q. Yang, “Federated rec- ommendation systems,” in Federated Learning: Privacy and Incentive . Springer, 2020, pp. 225–239

    2020

  4. [3]

    Federated learning review: Fundamentals, enabling technologies, and future applications,

    S. Banabilah, M. Aloqaily, E. Alsayed, N. Malik, and Y . Jararweh, “Federated learning review: Fundamentals, enabling technologies, and future applications,” Information processing & management , vol. 59, no. 6, p. 103061, 2022

    2022

  5. [4]

    Deepseek-inspired exploration of rl-based llms and synergy with wireless networks: A survey,

    Y . Qiao, P.-N. Tran, J. S. Yoon, L. X. Nguyen, E.-N. Huh, D. Niyato, and C. S. Hong, “Deepseek-inspired exploration of rl-based llms and synergy with wireless networks: A survey,” arXiv preprint arXiv:2503.09956 , 2025

    arXiv 2025

  6. [5]

    Communication-efficient learning of deep networks from decentralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Artificial Intelligence and Statistics , 2017, pp. 1273–1282

    2017

  7. [6]

    Advances and open problems in federated learning,

    P. Kairouz, H. B. McMahan, B. Avent, A. Bellet, M. Bennis, A. N. Bhagoji, K. Bonawitz, Z. Charles, G. Cormode, R. Cummings et al. , “Advances and open problems in federated learning,” Foundations and Trends® in Machine Learning , vol. 14, no. 1–2, pp. 1–210, 2021

    2021

  8. [7]

    Federated learning: Challenges, methods, and future directions,

    T. Li, A. K. Sahu, A. Talwalkar, and V . Smith, “Federated learning: Challenges, methods, and future directions,” IEEE signal processing magazine, vol. 37, no. 3, pp. 50–60, 2020

    2020

Show all 22 references
  1. [8]

    Heterogeneous feder- ated learning: State-of-the-art and research challenges,

    M. Ye, X. Fang, B. Du, P. C. Yuen, and D. Tao, “Heterogeneous feder- ated learning: State-of-the-art and research challenges,” ACM Computing Surveys, vol. 56, no. 3, pp. 1–44, 2023

    2023

  2. [9]

    Prototype completion with primitive knowledge for few-shot learning,

    B. Zhang, X. Li, Y . Ye, Z. Huang, and L. Zhang, “Prototype completion with primitive knowledge for few-shot learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 3754–3762

    2021

  3. [10]

    Morphological prototyping for unsupervised slide representation learning in computational pathology,

    A. H. Song, R. J. Chen, T. Ding, D. F. Williamson, G. Jaume, and F. Mahmood, “Morphological prototyping for unsupervised slide representation learning in computational pathology,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 1...

    2024

  4. [11]

    Fedproto: Federated prototype learning across heterogeneous clients,

    Y . Tan, G. Long, L. Liu, T. Zhou, Q. Lu, J. Jiang, and C. Zhang, “Fedproto: Federated prototype learning across heterogeneous clients,” in Proceedings of the AAAI conference on artificial intelligence , vol. 36, no. 8, 2022, pp. 8432–8440

    2022

  5. [12]

    Cross-modal prototype based multimodal federated learning under severely missing modality,

    H. Q. Le, C. M. Thwal, Y . Qiao, Y . L. Tun, M. N. Nguyen, and C. S. Hong, “Cross-modal prototype based multimodal federated learning under severely missing modality,” arXiv preprint arXiv:2401.13898 , 2024

    2024 arXiv

  6. [13]

    Rethinking federated learning with domain shift: A prototype view,

    W. Huang, M. Ye, Z. Shi, H. Li, and B. Du, “Rethinking federated learning with domain shift: A prototype view,” in 2023 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) . IEEE, 2023, pp. 16 312–16 322

    2023

  7. [15]

    Geodesic flow kernel for unsupervised domain adaptation,

    B. Gong, Y . Shi, F. Sha, and K. Grauman, “Geodesic flow kernel for unsupervised domain adaptation,” in 2012 IEEE conference on computer vision and pattern recognition . IEEE, 2012, pp. 2066–2073

    2012

  8. [16]

    A database for handwritten text recognition research,

    J. J. Hull, “A database for handwritten text recognition research,” IEEE Transactions on pattern analysis and machine intelligence, vol. 16, no. 5, pp. 550–554, 1994

    1994

  9. [17]

    Gradient-based learning applied to document recognition,

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE , vol. 86, no. 11, pp. 2278–2324, 1998

    1998

  10. [18]

    Reading digits in natural images with unsupervised feature learning,

    Y . Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, A. Y . Ng et al. , “Reading digits in natural images with unsupervised feature learning,” in NIPS workshop on deep learning and unsupervised feature learning , vol. 2011, no. 2. Granada, 2011, p. 4

    2011

  11. [19]

    Effects of degradations on deep neural network architectures. arxiv 2018,

    P. Roy, S. Ghosh, S. Bhattacharya, and U. Pal, “Effects of degradations on deep neural network architectures. arxiv 2018,” arXiv preprint arXiv:1807.10108, 1807

    2018 arXiv

  12. [20]

    Model-contrastive federated learning,

    Q. Li, B. He, and D. Song, “Model-contrastive federated learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 10 713–10 722

    2021

  13. [21]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al. , “Pytorch: An imperative style, high-performance deep learning library,” Advances in neural information processing systems , vol. 32, 2019

    2019

  14. [22]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.