Pith. sign in

REVIEW 6 major objections 4 minor 38 references

AFBS:Buffer Gradient Selection in Semi-asynchronous Federated Learning

T0 review · 6 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read AFBS prunes low-value buffered gradients in semi-asynchronous federated learning.

desk verdict A simple, plausible server-side idea for pruning stale gradients in semi-asynchronous FL, but the empirical case rests on single runs and a disabled baseline; worth reviewing, not yet worth believing. read the letter →

arxiv 2506.12754 v2 pith:EVF7Y4AJ submitted 2025-06-15 cs.LG cs.AI

classification cs.LGcs.AI
keywords federatedlearningasynchronousaggregationgradientselectionbufferclientclusteringrandomprojectionencryptionstalenessdataheterogeneity
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 AFBS, a server-side algorithm for semi-asynchronous federated learning that scores each gradient sitting in the aggregation buffer and discards the low-value ones before aggregation. The score is $\text{Score}(x)=V_x/(\tau_x+1)^2$, combining the source client's dataset size $V_x$ with the gradient's staleness $\tau_x$, and selection is done per cluster of clients that share a similar label distribution. The authors claim this is the first buffer-internal gradient-selection method that preserves privacy, and they report accuracy gains of up to 4.8% over the strongest prior method on CIFAR-100 together with a 75% reduction in virtual time to reach target accuracy. A reader should care because it suggests that stale updates should be filtered out rather than merely down-weighted, and that the server can do this without seeing raw client data.

What carries the argument

The carrying mechanism is a per-cluster gradient filter paired with a privacy-preserving client clustering step. The filter, described in Algorithm 2, scores each buffered gradient by $\text{Score}(x)=V_x/(\tau_x+1)^2$ and compares it with the highest-scoring gradient in the same cluster, keeping it only if its dataset size is no smaller or its staleness is no larger. The clustering step sends a label-distribution matrix that has been replicated, perturbed with Gaussian noise, and reduced by a random projection, so K-Means can group similar clients while the Johnson-Lindenstrauss lemma guarantees that pairwise distances are approximately preserved; the Gaussian noise makes the matrix full rank with probability 1, which the paper uses to argue decryption is impossible.

What would settle it

Run AFBS on CIFAR-100 and replace Algorithm 2 with random pruning that discards exactly the same number of gradients per buffer; if accuracy and time-to-target match AFBS, the score is not responsible for the gains. A sharper test is to measure each discarded gradient's actual effect by re-adding it to the aggregate and checking whether the global validation loss drops; if discarded gradients improve the loss more often than kept ones, the heuristic is inverted.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a semi-asynchronous federated learning buffer should be treated as a selection problem, not a weighting problem. Buffered gradients that are both older and sourced from smaller datasets than the cluster's best gradient are taken to be low-value, so AFBS keeps a gradient only when $V_x \ge V_m$ or $\tau_x \le \tau_m$, where $V_m$ and $\tau_m$ come from the highest-scoring gradient in the same cluster. Comparing only within clusters, built from random-projection-encrypted label distributions, prevents the filter from silently dropping data classes that happen to live on slow clients. On MNIST, CIFAR-10, CIFAR-100, and SST2, the paper reports that AFBS matches or beats six baselines, with the largest margins on the hardest vision tasks. The paper also proves convergence of the update rule and shows that the encrypted label matrix is full rank with probability 1, which it uses to argue the original distribution cannot be recovered.

Load-bearing premise

The load-bearing premise is that a gradient's usefulness is captured by its staleness and its source dataset size, so the score $V_x/(\tau_x+1)^2$ and the per-cluster filter keep the informative gradients and remove only harmful ones; if an old gradient from a small dataset can still be valuable, pruning will delete useful updates.

Editorial extensions

If this is right

  • Semi-asynchronous federated learning can gain both accuracy and speed by discarding, rather than down-weighting, stale gradients, because every averaged gradient that is removed also removes its share of the update step.
  • Server-side gradient summation becomes cheaper: the paper reports a 10%-30% reduction in per-round handling time over FedBuff, since fewer gradients reach aggregation.
  • Client clustering based on encrypted label distributions keeps rare or slow-client data in training, which the paper shows is necessary for the gains on heterogeneous data.
  • The same random-projection encryption is claimed to be a general matrix-encryption building block for other settings that need to hide a data matrix from a server.

Reading between the lines

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

  • The score function uses only staleness and dataset volume; a natural stress test is whether adding gradient norm, loss value, or client update frequency would rank gradients better, and whether the 4.8% margin survives.
  • Clustering is performed once before training; under non-stationary or drifting label distributions, a fixed clustering could become stale, so an adaptive re-clustering variant is a natural follow-up.
  • The reported 75% time reduction is measured under a uniform latency distribution; under a long-tail distribution, where most clients are fast, the relative speed gain might be smaller.
  • The privacy argument shows algebraic recovery is hard because the projected matrix is full rank, but it does not test reconstruction or membership-inference attacks; that is an open question.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

6 major / 4 minor

Summary. The paper proposes AFBS, a server-side gradient-selection algorithm for semi-asynchronous federated learning. Clients first send a randomly projected, noise-perturbed label-distribution matrix, the server clusters clients by this information, and then, when the buffer is full, the server scores buffered gradients by Eq. (11), Score(x)=V_x/(tau_x+1)^2, and filters them within each cluster via Algorithm 2 before aggregation. The paper claims that AFBS is the first buffer-selection method in this setting, that it improves accuracy by up to 4.8% over the previous best method on CIFAR-100, and that it reduces time-to-target-accuracy by 75%. Supporting material includes a convergence proof and a privacy argument in the Supplementary, plus experiments on MNIST, CIFAR-10, CIFAR-100, and SST2 against FedAvg, FedAsync, FedBuff, CA2FL, FedFa, and FedDyn.

Significance. If the reported gains replicate, AFBS would be a useful, low-overhead addition to semi-asynchronous federated learning: it is a simple server-side heuristic, evaluated on four public datasets against six baselines, and Table 5 indicates that pruning actually reduces per-round server handle time relative to CA2FL and FedBuff. The paper is also candid in Section 4.2 about the FedFa configuration mismatch, which is a point in its favor even though it invalidates that particular baseline comparison. However, the absence of seed variance reporting, the unsupported 75% speedup number, the lack of an ablation isolating Eq. (11), and the fact that the convergence proof does not model the selection rule mean the current evidence establishes only a promising empirical heuristic, not a proven state-of-the-art method.

major comments (6)
  1. [§4.1, Tables 2-4] The experimental core rests on single-run comparisons with no seeds or standard deviations. The headline CIFAR-100 accuracy gap of 4.8 percentage points and the speedup claims in Table 4 can easily be within seed noise for deep models, so multiple runs with mean and standard deviation (or a small number of seeds with individual values) are needed before the central accuracy claim is established. Table 2 also shows that AFBS is not the best method on every setting (CA2FL wins on MNIST at K=5 and FedDyn wins on both SST2 settings), so the abstract's general statement of superior performance should be qualified.
  2. [§4.2, Tables 2 and 4] The FedFa baseline is evaluated outside its intended regime. The paper itself explains in the last paragraph of Section 4.2 that FedFa was designed for a long-tail response-time distribution while the experiments use a uniform distribution, which is why FedFa reports only 1.15% on CIFAR-100 and fails to reach any target accuracy in Table 4. A baseline configured outside the conditions of its original paper is effectively disabled rather than fairly compared, so the FedFa rows should not be used to support the state-of-the-art comparison.
  3. [Abstract, §4.3, Table 4] The claimed 75% reduction in time to reach target accuracy on CIFAR-100 is not supported by Table 4. For target accuracy 0.15 at K=1, AFBS takes 138,274 virtual seconds, while the previous best competitor by accuracy (CA2FL) takes 300,248 seconds, which is a reduction of about 54%; compared with FedBuff (399,040 seconds) the reduction is about 65%. The abstract and conclusion should either state the exact baseline that yields 75% or correct the number.
  4. [§3.6 Eq. (11), §4.4-4.5] No experiment isolates the contribution of the scoring function. The ablations in Table 3 and Figure 8 vary clustering and buffer size, but the paper never compares Eq. (11) with alternative scores or with a random-pruning control at the same retention rate. The AFBS-versus-FedBuff contrast shows only that pruning at the server helps, not that the particular score is responsible. In addition, Section 3.6 describes selection as probabilistic via Eq. (12), whereas Algorithm 2 implements a deterministic threshold rule at line 9; the implemented rule needs to be specified unambiguously.
  5. [§6, Eqs. (22) and (34), Assumption 5] The convergence proof in the Supplementary does not prove convergence of AFBS. It analyzes a generic weighted average of client updates in Eq. (22) under Assumption 5, which simply assumes the post-selection buffer length is bounded between Cmin and Cmax. The score in Eq. (11), the threshold filter in Algorithm 2, and the staleness-dependent weight lambda are not present in the derivation. Thus the bound in Eq. (34) and the final inequality in Eq. (40) apply to any bounded-buffer semi-asynchronous method, not to the proposed selection mechanism; the proof should either model the selection operation or be restated as a convergence result for the base FedBuff-style aggregation.
  6. [§3.5, Eq. (10)] The privacy argument for the encryption scheme is incomplete. Eq. (10) establishes only that the noise-perturbed matrix A+G is full rank with probability 1 when G is Gaussian. Full rank does not imply that the original label distribution cannot be recovered from A_new, and the paper gives no information-theoretic or computational leakage bound, no differential-privacy guarantee, and no discussion of what an adversary can infer from the publicly sent label-distribution matrix. The claim that encryption 'effectively prevents' decryption should be softened or supported with a formal analysis.
minor comments (4)
  1. [Table 3] The table is typeset incorrectly: entries such as '98.2057.21 50.2721.32 20.24w/ clustering' run together and are unreadable; the rows and columns should be separated properly.
  2. [§3.5] The dimensions of the random projection matrix R and the output matrix A_new are not specified. The text says A is in R^{1×d} and is replicated to a d×d matrix, but the dimensions of R^T and the resulting A_new are needed for reproducibility.
  3. [§3.6, Algorithm 1] Algorithm 1 uses a single lambda = 1/sqrt(tau_min+1) for all selected gradients, whereas Eq. (7) and the surrounding discussion describe per-gradient staleness weights lambda_i; the relationship between these two formulations should be clarified.
  4. [Throughout] There are several typographical issues, including 'A synchronous FL Buffer Selection' in the abstract (presumably 'Asynchronous'), 'FedA VG' with an erroneous space in Table 2, and 'isolate single-client training' in Section 3.2, which should read 'isolated single-client training'.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: AFBS's claimed gains come from external benchmark comparisons, not from fitting or self-referential equations; the only self-citation is minor and non-load-bearing.

full rationale

The central performance claim (up to 4.8% accuracy gain and 75% time reduction to target accuracy on CIFAR-100) is supported by Table 2 and Table 4, which compare AFBS against externally published algorithms (FedAvg, FedAsync, FedBuff, CA2FL, FedFa, FedDyn) on public datasets under a fixed virtual runtime. These numbers are measured outcomes, not identities forced by the method's construction. The gradient scoring rule in Eq. (11) and the filter in Algorithm 2 are hand-chosen heuristics (the paper explicitly says 'we hypothesize' in Section 3.2) rather than parameters fitted to the test set; a heuristic design choice is not circular. The convergence proof in the Supplementary proves boundedness of the aggregate update under standard assumptions plus Assumption 5 (bounded buffer length), and it does not secretly re-import the target accuracy improvement; if anything it is more general than the selection rule, which is a weakness of scope rather than circularity. The only self-citation is reference [5] (Lu et al., 2025), used in the straggler context in Section 1; it is not load-bearing for the AFBS algorithm, the scoring rule, or the experimental claims. The admitted issue that FedFa is evaluated outside its long-tail response-time regime (end of Section 4.2) is a fairness/correctness concern about a baseline, not a circular reduction of AFBS's claim to its own inputs. Overall, no prediction in the paper is equivalent by construction to an input, and no load-bearing conclusion depends on a self-citation chain.

Assumptions & free parameters 7 free parameters · 7 assumptions · 0 invented entities

The method's gains rest on hand-picked scoring and filtering rules (exponent 2, OR filter), an underspecified random projection dimension, and a noise scale chosen without privacy analysis. The convergence proof adds an extra assumption (bounded buffer length after selection) that is not guaranteed by the algorithm. No genuinely new entities are introduced; the 'encrypted label distribution matrix' is a data transformation, not a new physical or mathematical object.

free parameters (7)
  • Score exponent = 2
    Hand-chosen in Score(x)=Vx/(tau_x+1)^2 (Eq. 11); no derivation or sensitivity analysis for other exponents.
  • Selection filter = Vx >= Vm OR tau_x <= tau_m
    Hand-chosen deterministic rule in Algorithm 2, lines 7-9; differs from the probabilistic rule in Eq. (12) with no explanation.
  • Gaussian noise scale sigma = 1e-3
    Chosen as 'a specified small constant' in Eq. (9); sets the privacy-utility trade-off but no privacy analysis or sensitivity study.
  • Random projection dimension k = not specified
    Eq. (9) projects to R^T with unspecified dimension; essential for JL distance preservation and for the rank and decryption argument.
  • Obsolescence decay exponent = 1/2
    lambda = 1/sqrt(tau_min+1) in Section 4.1; the exponent is hand-picked, no ablation.
  • Number of clusters K = 1, 3, 5 in experiments
    K is a hyperparameter of K-means clustering in Algorithm 1; not tuned per dataset except by the experimenter.
  • Buffer size C = 10 (default)
    Buffer capacity hyperparameter; Figure 8 tests robustness, showing modest effect for C in 8-20.
assumptions (7)
  • standard math Johnson-Lindenstrauss lemma: random projection approximately preserves pairwise distances (Eq. 8).
    Invoked in Section 3.5 to justify clustering on projected label distributions; standard result.
  • standard math Gaussian noise matrix makes A+G full rank with probability 1 (Eq. 10).
    Proof in Supplementary uses continuity of Gaussian vectors; standard result.
  • domain assumption Assumptions 1-4: unbiased gradients, bounded variance, bounded gradients, L-smoothness.
    Standard stochastic optimization assumptions in the Supplementary convergence proof; not specific to AFBS.
  • ad hoc to paper Assumption 5: buffer length after selection satisfies Cmin <= Ct <= Cmax.
    This assumes the selection algorithm keeps the buffer in a bounded interval; the actual AFBS algorithm provides no such guarantee and the proof never models the selection rule.
  • domain assumption Clients' label distributions can be clustered from random-projection-encrypted versions.
    Section 3.5 assumes the server can group clients by the projected noisy label matrix, and that this preserves the similarity structure needed for fair gradient comparison.
  • ad hoc to paper Gradient informational value is captured by staleness and dataset size only.
    Section 3.6 defines Score(x)=Vx/(tau_x+1)^2 with no derivation; the claim that dropping low-score gradients helps training is the core heuristic premise.
  • domain assumption All clients use the same random projection matrix R.
    Eq. (9) and the clustering step imply a shared R; the paper never specifies how R is generated or distributed, nor what happens if clients use different R.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AFBS:Buffer Gradient Selection in Semi-asynchronous Federated Learning." pith.science (2026). https://pith.science/paper/EVF7Y4AJ

@misc{pith2026250612754,
  author       = {Pith},
  title        = {Pith review of: AFBS:Buffer Gradient Selection in Semi-asynchronous Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EVF7Y4AJ}},
  note         = {Machine review of arXiv:2506.12754}
}
read the original abstract

Asynchronous federated learning (AFL) accelerates training by eliminating the need to wait for stragglers, but its asynchronous nature introduces gradient staleness, where outdated gradients degrade performance. Existing solutions address this issue with gradient buffers, forming a semi-asynchronous framework. However, this approach struggles when buffers accumulate numerous stale gradients, as blindly aggregating all gradients can harm training. To address this, we propose AFBS (Asynchronous FL Buffer Selection), the first algorithm to perform gradient selection within buffers while ensuring privacy protection. Specifically, the client sends the random projection encrypted label distribution matrix before training, and the server performs client clustering based on it. During training, server scores and selects gradients within each cluster based on their informational value, discarding low-value gradients to enhance semi-asynchronous federated learning. Extensive experiments in highly heterogeneous system and data environments demonstrate AFBS's superior performance compared to state-of-the-art methods. Notably, on the most challenging task, CIFAR-100, AFBS improves accuracy by up to 4.8% over the previous best algorithm and reduces the time to reach target accuracy by 75%.

Figures

Figures reproduced from arXiv: 2506.12754 by the authors.

Figure 1
Figure 1. Schematic diagram of buffer gradient selection. According to the gradient [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The illustration of the proposed AFBS framework. Clients first need to encrypt the label distribution using random projection before training. The server will cluster the clients based on this information. During training, the server collects gradients sent by clients until the buffer is full. Once the buffer is full, the server employs the AFBS algorithm to perform Gradient Selection on the gradients within each cl… view at source ↗
Figure 3
Figure 3. Comparison of Aggregating Stale Gradients, Fresh Gradients, and Individual [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The data distribution of pre-clustering versus non-pre-clustering strategies par [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Changes in cosine similarity of models with consistent and inconsistent label [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Comparison of different algorithms on different datasets when K = 1. [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Comparison results of clustering or not. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Impact of buffer size under different K. [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 24 canonical work pages

  1. [1]

    Federated machine learning: Con- cept and applications,

    Q. Yang, Y . Liu, T. Chen, and Y . Tong, “Federated machine learning: Con- cept and applications,”ACM Transactions on Intelligent Systems and Tech- nology, vol. 10, no. 2, 2019

  2. [2]

    Federated visual classification with real-world data distribution,

    T.-M. H. Hsu, H. Qi, and M. Brown, “Federated visual classification with real-world data distribution,” inComputer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X. Berlin, Heidelberg: Springer-Verlag, 2020, p. 76–92. [Online]. Available: https://doi.org/10.1007/978-3-030-58607-2_5

  3. [3]

    Straggler-resilient federated learning: Leveraging the interplay between statistical accuracy and system heterogeneity,

    A. Reisizadeh, I. Tziotis, H. Hassani, A. Mokhtari, and R. Pedarsani, “Straggler-resilient federated learning: Leveraging the interplay between statistical accuracy and system heterogeneity,”IEEE Journal on Selected Areas in Information Theory, vol. 3, no. 2, pp. 197–205, 2022

  4. [4]

    Breaking barriers of system heterogeneity: Straggler-tolerant multimodal federated learning via knowledge distillation,

    J. Chen, H. Tang, J. Cheng, M. Yan, J. Zhang, M. Xu, Y . Hu, and L. Nie, “Breaking barriers of system heterogeneity: Straggler-tolerant multimodal federated learning via knowledge distillation,” inProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, K. Larson, Ed. International Joint Conferences on Artificia...

  5. [5]

    Corrected with the Latest Version: Make Robust Asynchronous Federated Learning Possible

    C. Lu, Y . Sun, P. Li, and Z. Yang, “Corrected with the latest version: Make robust asynchronous federated learning possible,” 2025. [Online]. Available: https://arxiv.org/abs/2504.04081

  6. [6]

    Advances and open problems in federated learning,

    P. e. a. Kairouz, “Advances and open problems in federated learning,” 2021. [Online]. Available: https://arxiv.org/abs/1912.04977

  7. [7]

    Fedsa: A staleness-aware asynchronous federated learning algorithm with non-iid data,

    M. Chen, B. Mao, and T. Ma, “Fedsa: A staleness-aware asynchronous federated learning algorithm with non-iid data,”Future Generation Computer Systems, vol. 120, pp. 1–12, 2021. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0167739X21000649

  8. [8]

    Client Selection in Federated Learning: Principles, Challenges, and Opportunities

    L. Fu, H. Zhang, G. Gao, M. Zhang, and X. Liu, “Client selection in federated learning: Principles, challenges, and opportunities,” 2023. [Online]. Available: https://arxiv.org/abs/2211.01549

Show all 38 references
  1. [9]

    Fedmccs: Multicriteria client selection model for optimal iot federated learning,

    S. Abdulrahman, H. Tout, A. Mourad, and C. Talhi, “Fedmccs: Multicriteria client selection model for optimal iot federated learning,”IEEE Internet of Things Journal, vol. 8, no. 6, pp. 4723–4735, 2021

  2. [10]

    A sur- vey on federated learning systems: Vision, hype and reality for data privacy and protection,

    Q. Li, Z. Wen, Z. Wu, S. Hu, N. Wang, Y . Li, X. Liu, and B. He, “A sur- vey on federated learning systems: Vision, hype and reality for data privacy and protection,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 4, pp. 3347–3366, 2023

  3. [11]

    Asynchronous federated optimization,

    C. Xie, S. Koyejo, and I. Gupta, “Asynchronous federated optimization,”

  4. [12]

    Communication-efficient federated deep learn- ing with layerwise asynchronous model update and temporally weighted ag- gregation,

    Y . Chen, X. Sun, and Y . Jin, “Communication-efficient federated deep learn- ing with layerwise asynchronous model update and temporally weighted ag- gregation,”IEEE Transactions on Neural Networks and Learning Systems, vol. 31, no. 10, pp. 4229–4238, 2020

  5. [13]

    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,” inProceedings of the 20th International Conference on Artificial Intelligence and Statistics, ser. Proceedings of Machine Learning Rese...

  6. [14]

    Adaptive federated optimization,

    S. Reddi, Z. Charles, M. Zaheer, Z. Garrett, K. Rush, J. Kone ˇcný, S. Kumar, and H. B. McMahan, “Adaptive federated optimization,” 2021. [Online]. Available: https://arxiv.org/abs/2003.00295

  7. [15]

    Federated optimization in heterogeneous networks,

    T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V . Smith, “Federated optimization in heterogeneous networks,” 2020. [Online]. Available: https://arxiv.org/abs/1812.06127

  8. [16]

    Fedbn: Federated learning on non-iid features via local batch normalization,

    X. Li, M. Jiang, X. Zhang, M. Kamp, and Q. Dou, “Fedbn: Federated learning on non-iid features via local batch normalization,” 2021. [Online]. Available: https://arxiv.org/abs/2102.07623

  9. [17]

    Model-contrastive federated learning,

    Q. Li, B. He, and D. Song, “Model-contrastive federated learning,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 10 708–10 717

  10. [18]

    Fedasmu: Efficient asynchronous federated learning with dynamic staleness-aware model update,

    J. Liu, J. Jia, T. Che, C. Huo, J. Ren, Y . Zhou, H. Dai, and D. Dou, “Fedasmu: Efficient asynchronous federated learning with dynamic staleness-aware model update,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 12, pp. 13 900–13 908, Mar. 2024. [O...

  11. [19]

    Tackling the data heterogeneity in asynchronous federated learning with cached update calibration,

    Y . Wang, Y . Cao, J. Wu, R. Chen, and J. Chen, “Tackling the data heterogeneity in asynchronous federated learning with cached update calibration,” inThe Twelfth International Conference on Learning Representations, 2024. [Online]. Available: https://openreview.net/forum? id=...

  12. [20]

    Federated learning with buffered asynchronous aggregation,

    J. Nguyen, K. Malik, H. Zhan, A. Yousefpour, M. Rabbat, M. Malek, and D. Huba, “Federated learning with buffered asynchronous aggregation,”

  13. [21]

    Fedfa: a fully asynchronous training paradigm for federated learning,

    H. Xu, Z. Zhang, S. Di, B. Liu, K. A. Alharthi, and J. Cao, “Fedfa: a fully asynchronous training paradigm for federated learning,” inProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, ser. IJCAI ’24, 2025. [Online]. Available: https://d...

  14. [22]

    Robust federated learning in a heterogeneous environment,

    A. Ghosh, J. Hong, D. Yin, and K. Ramchandran, “Robust federated learning in a heterogeneous environment,” 2019. [Online]. Available: https://arxiv.org/abs/1906.06629 28

  15. [23]

    Pfa: Privacy-preserving federated adaptation for effective model personalization,

    B. Liu, Y . Guo, and X. Chen, “Pfa: Privacy-preserving federated adaptation for effective model personalization,” 2021. [Online]. Available: https://arxiv.org/abs/2103.01548

  16. [24]

    Clustered federated learn- ing: Model-agnostic distributed multitask optimization under privacy con- straints,

    F. Sattler, K.-R. Müller, and W. Samek, “Clustered federated learn- ing: Model-agnostic distributed multitask optimization under privacy con- straints,”IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 8, pp. 3710–3722, 2021

  17. [25]

    An efficient framework for clustered federated learning,

    A. Ghosh, J. Chung, D. Yin, and K. Ramchandran, “An efficient framework for clustered federated learning,” 2021. [Online]. Available: https://arxiv.org/abs/2006.04088

  18. [26]

    Casa: Clustered federated learning with asynchronous clients,

    B. Liu, Y . Ma, Z. Zhou, Y . Shi, S. Li, and Y . Tong, “Casa: Clustered federated learning with asynchronous clients,” inProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, ser. KDD ’24. New York, NY , USA: Association for Computing Machinery,...

  19. [27]

    Research on k-means clustering algorithm: An improved k-means clustering algorithm,

    S. Na, L. Xumin, and G. Yong, “Research on k-means clustering algorithm: An improved k-means clustering algorithm,” in2010 Third International Symposium on Intelligent Information Technology and Security Informatics, 2010, pp. 63–67

  20. [28]

    Random projection in dimensionality reduction: applications to image and text data,

    E. Bingham and H. Mannila, “Random projection in dimensionality reduction: applications to image and text data,” inProceedings of the Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ser. KDD ’01. New York, NY , USA: Association for Computing...

  21. [29]

    The johnson-lindenstrauss lemma is optimal for linear dimensionality reduction,

    K. G. Larsen and J. Nelson, “The johnson-lindenstrauss lemma is optimal for linear dimensionality reduction,” 2014. [Online]. Available: https://arxiv.org/abs/1411.2404

  22. [30]

    Flgo: A fully customizable federated learning platform,

    Z. Wang, X. Fan, Z. Peng, X. Li, Z. Yang, M. Feng, Z. Yang, X. Liu, and C. Wang, “Flgo: A fully customizable federated learning platform,” 2023. [Online]. Available: https://arxiv.org/abs/2306.12079 29

  23. [31]

    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

  24. [32]

    Mnist handwritten digit database,

    Y . LeCun, C. Cortes, and C. Burges, “Mnist handwritten digit database,”ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, vol. 2, 2010

  25. [33]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, “Learning multiple layers of features from tiny images,” Uni- versity of Toronto, Tech. Rep., 2009

  26. [34]

    Bag of tricks for efficient text classification,

    A. Joulin, E. Grave, P. Bojanowski, and T. Mikolov, “Bag of tricks for efficient text classification,” 2016. [Online]. Available: https: //arxiv.org/abs/1607.01759

  27. [35]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,

    H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,” 2017

  28. [36]

    Federated learning based on dynamic regularization,

    D. A. E. Acar, Y . Zhao, R. M. Navarro, M. Mattina, P. N. Whatmough, and V . Saligrama, “Federated learning based on dynamic regularization,” 2021. [Online]. Available: https://arxiv.org/abs/2111.04263 30

  29. [2020]

    Available: https://arxiv.org/abs/1903.03934

    [Online]. Available: https://arxiv.org/abs/1903.03934

  30. [2022]

    Available: https://arxiv.org/abs/2106.06639

    [Online]. Available: https://arxiv.org/abs/2106.06639

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.