Pith. sign in

REVIEW 4 major objections 8 minor 84 references

Efficient Federated Learning with Timely Update Dissemination

T0 review · 4 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read By pushing fresh global models to devices mid-training over idle downlink bandwidth, FedASMU and FedSSMU claim accuracy gains up to 145.87% and training-time cuts up to 97.59% in heterogeneous federated learning.

desk verdict Solid systems idea and broad empirical validation, but the convergence proof is a bounded-α abstraction and the adaptive machinery is unanalyzed. read the letter →

arxiv 2507.06031 v1 pith:J4NS7CHI submitted 2025-07-08 cs.DC cs.AIcs.LG

Juncheng Jia , Ji Liu , Chao Huo , Yihui Shen , Yang Zhou , Huaiyu Dai , Dejing Dou This is my paper
classification cs.DCcs.AIcs.LG
keywords federatedlearningasynchronoussynchronousstaleness-awareaggregationnon-IIDdatasystemheterogeneitydownlinkbandwidthtimelyupdatedissemination
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

This paper sets out to show that federated learning overlooks a cheap resource: downlink bandwidth is often broader and less busy than uplink, yet standard synchronous and asynchronous protocols wait until a full round finishes before sending the latest global model to devices. The proposed algorithms, FedASMU and FedSSMU, use that idle downlink capacity to push fresh global models to devices while they are still training locally, and pair the fresh models with dynamically learned aggregation weights on both server and device. The paper claims this shortens training time by up to 97.59% and raises accuracy by up to 145.87% relative to nine baselines across six models and five datasets, and that a convergence theorem guarantees the averaged gradient norm goes to zero under standard assumptions. A sympathetic reader would take the core truth to be that timely update dissemination, not just staleness weighting or data regularization, is a first-order lever on heterogeneous federated learning performance.

What carries the argument

The load-bearing objects are two parametrized aggregation rules whose weights are tuned online rather than fixed by a static staleness formula. On the server, the importance of an uploaded local model is $\alpha_t^i = \mu_\alpha \xi_t^i(o)/(1+\mu_\alpha \xi_t^i(o))$ with $\xi_t^i(o) = \lambda_t^i / (\sqrt{t}\,(t-o+1)^{\sigma_t^i}) + \iota_t^i$, where $t-o+1$ is staleness; the control parameters $\lambda_t^i, \sigma_t^i, \iota_t^i$ are updated by gradient descent on the global loss, with the global gradient approximated by a single device's local gradient. On the device, a fresh global model $w_g$ is merged into the local model with weight $\beta$ built from $\gamma$ and $\upsilon$ control parameters, and an RL-based selector (an LSTM meta-model plus Q-learning) picks the local epoch at which the device asks for the fresh model. Together these mechanisms update model importance in real time from staleness and loss signals, which is what the paper argues lets the global model absorb fresh information without being reverted by stale updates.

What would settle it

A direct check is to run FedASMU with the device-side request slot fixed to the final local epoch, so the fresh global model arrives too late to affect local training; if the reported accuracy and time gains over FedAsync persist, timely mid-training dissemination is not the operative mechanism.

Watch

Extended reading notes

Core claim

The central claim is that delayed update dissemination is a correctable bottleneck in federated learning: devices finish local training on stale global models, and the freshest information sits unused on the server. FedASMU addresses the asynchronous setting by having the server periodically trigger devices, discard updates beyond a staleness threshold $\tau$, aggregate accepted local models with a weight $\alpha_i^t$ that is learned online from a polynomial function of staleness, and let each device request the newest global model once per local training and merge it into its local model with a learned weight $\beta$. FedSSMU transfers the same two mechanisms into the synchronous round structure, treating the synchronous case as a special asynchronous one. The paper's evidence for the claim is a comparison in which both algorithms beat nine baselines on final accuracy and time-to-target accuracy, with relative accuracy gains up to 145.87% and training-time reductions up to 97.59%, plus Theorem 1 showing convergence of the averaged squared gradient norm to zero.

Load-bearing premise

The load-bearing premise is that a single device's local gradient, estimated as $(w_i^o - w_o)/(\eta_i L)$, is a faithful stand-in for the global loss gradient when the server tunes its aggregation weights; if that proxy is poor, the learned weights can be mistuned and the convergence theorem, which treats the weights only as bounded, will not by itself guarantee the reported accuracy gains.

Editorial extensions

If this is right

  • In deployments where downlink capacity is plentiful, federated training time to a target accuracy can be cut substantially (claimed up to 97.59%) without changing the local datasets or privacy posture.
  • Devices that finish early no longer have to sit idle: their fresh models are disseminated to slower devices mid-round, so the global model incorporates new information sooner in both synchronous and asynchronous settings.
  • The device-side merge should reduce client drift on non-IID data, because local models track the latest global model instead of training for many epochs on a stale copy.
  • Synchronous FL systems that cannot switch to asynchronous aggregation can still obtain most of the benefit through FedSSMU, which applies the same server- and device-side mechanisms inside the round structure.
  • Because the request timing is learned by reinforcement learning rather than fixed, the method does not require a hand-tuned schedule of when devices should fetch fresh models.

Reading between the lines

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

  • The paper leaves open whether the device-side merge still pays off under compressed or partial model updates; the overhead argument assumes full-model downlink transfers, so the trade-off could look different under compression.
  • The convergence theorem bounds the aggregation weight $\alpha$ but does not prove that the learned weight tracks the true global gradient, so the accuracy gains should be read as empirically established rather than guaranteed by the proof.
  • In networks where downlink and uplink bandwidth are comparable rather than asymmetric, the extra transfers could outweigh the freshness benefit; a direct comparison in symmetric-bandwidth conditions would delineate when the method helps.
  • The ablation study attributes most of the gain to the server-side dynamic weights; applying only the device-side merge would test whether freshness alone, without reweighting, is enough.
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

4 major / 8 minor

Summary. The paper proposes two federated learning algorithms, FedASMU and FedSSMU, that exploit additional downlink bandwidth to disseminate fresh global models to devices during local training. On the server side, FedASMU uses a dynamic, staleness-aware aggregation weight α_i^t that is tuned by gradient-based updates of control parameters λ, σ, ι; on the device side, an RL-based selector decides when to request a fresh global model, and the device merges it with the local model using an adaptively tuned weight β. The synchronous variant FedSSMU applies the same ideas in a synchronized setting. The paper claims a convergence theorem for the proposed methods and reports experiments over six models and five datasets showing accuracy improvements up to 145.87% and training-time reductions up to 97.59% relative to nine baselines.

Significance. If the empirical gains hold under repeated runs, the timely-dissemination idea is a useful contribution to heterogeneous FL, and the breadth of the evaluation (six models, five datasets, nine baselines, ablations, and scalability/bandwidth studies) is a clear strength. The extension to the synchronous setting and the ablation isolating server-side and device-side components are also valuable. However, the paper's central theoretical claim is not supported as stated: the convergence proof applies only to a generic bounded-weight aggregation and never analyzes the adaptive mechanisms that define the proposed methods. The empirical section also lacks multiple-seed uncertainty quantification and reports per-method optimal hyperparameter tuning, which makes the magnitude of the claimed gains hard to assess. The paper does not provide code or a detailed simulation-to-device mapping, further limiting reproducibility.

major comments (4)
  1. [Section 4.4, Theorem 1 and Appendix A] Theorem 1 as stated does not establish convergence of the proposed FedASMU/FedSSMU mechanisms. The proof treats α_i^t and β_i^{t_i−1} as arbitrary quantities satisfying α_min ≤ α_i^t ≤ 1 and 0 ≤ β ≤ 1 (Appendix A, final paragraph), and it never uses the functional forms in Eqs. (2) and (10), the control-parameter updates in Eqs. (3)–(5) and (11), or the RL request-time selector in Section 4.3.1. Consequently the theorem is compatible with any bounded, even adversarial, weights and does not substantiate the abstract claim that 'theoretical analyses substantiate the convergence of our proposed methodologies.' Please either extend the analysis to the actual adaptive rules or explicitly reframe the theoretical contribution as convergence of the generic staleness-bounded aggregation with exogenous weights.
  2. [Section 4.2, Eqs. (3)–(5)] The server-side control-parameter update rests on the approximation ∇_λ F(w_o) ≈ [(w_i^o − w_o)/(η_i L)]^T ∂w_o/∂λ, i.e., replacing the global loss gradient by one device's average local gradient. Under non-IID data no bound is stated on the error of this proxy; Eq. (4) can even have the opposite sign from the true gradient, turning the parameter step into ascent. Moreover, λ, σ, and ι are updated without projection, so ξ_i^t can become negative or make the denominator in Eq. (2) vanish, and α_i^t is not guaranteed to lie in (0,1). The analogue holds for γ, υ, and β in Eqs. (10)–(11). Thus the key conditions α_min > 0 and β ∈ [0,1] used in Theorem 1 can be violated by the algorithm itself. Please add projection steps, or prove that the unprojected updates preserve the needed bounds, and provide a quantitative bound on the proxy-gradient error.
  3. [Section 6.1.4 and Table 4] The empirical comparison reports a single run per setting and states that hyperparameters were 'meticulously fine-tuned ... reporting the optimal settings' for every method. Because FedASMU and FedSSMU introduce many additional hyperparameters (η_λ, η_σ, η_ι, η_γ, η_υ, μ_α, μ_β, RL learning rates, etc.), tuning each method to its optimum without reporting search ranges or repeated-seed variability makes it difficult to assess whether the accuracy/efficiency gains in Table 4 are significant or an artifact of tuning effort. Please report means and standard deviations over at least three seeds, and specify the hyperparameter search protocol (grid/random, ranges) applied uniformly to all methods.
  4. [Section 2 vs Section 3.2] The motivating premise is stated inconsistently. Section 2, Observation 2 says 'The downlink bandwidths of edge devices are typically much smaller than their uplink bandwidths,' while Section 3.2 says the approach 'assumes downlink bandwidth is typically larger and underutilized than uplink.' These statements are contradictory, and the paper's central idea of exploiting 'additional downlink bandwidth' depends on which direction is actually the abundant one. Please reconcile the two statements and, if the premise is that downlink is underutilized, provide a concrete system model of the bandwidth asymmetry used in the experiments.
minor comments (8)
  1. [Section 6.1.2 and Table 4] Fashion-MNIST is spelled 'FMNSIT' in Section 6.1.2 and 'FMNIST' in Table 4; please standardize the spelling.
  2. [Section 4.3.1] In the Q-learning action description, 'stay means staying with the same epoch (l* = l* + 1)' contains a typo; the formula should leave l* unchanged, not increment it.
  3. [Section 6.2.5] The sentence beginning 'The advantages of FedASMU becomes even more significant...' contains a stray ')' and is confusing; please rewrite the bandwidth-reduction comparison for clarity.
  4. [Table 3] The hyper-parameter table is visually garbled, with column boundaries misaligned; please reformat so each model's values are readable.
  5. [Eq. (10)] The formula for ϕ_i^{t_i−1}(g,o) has ambiguous parentheses: it should be clarified whether the factor (1 − υ_i sqrt(g − o + 1)) multiplies sqrt(g) or is inside the square root.
  6. [Assumption 5] The bounded-variance assumption writes E ∥∇F_i(w_{o,l};ζ_l) − F(w_{o,l})∥²; the second term should presumably be ∇F(w_{o,l}), not F(w_{o,l}).
  7. [Theorem 1 and Appendix A] The O(Lτ²G²L_max²/L_min³) term appears twice in the theorem statement; one is a typo.
  8. [Section 6.1.4] Please clarify how 100 heterogeneous devices are simulated on 44 Tesla V100 GPUs, including the mapping of devices to GPUs and how communication delays are emulated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the adaptive control parameters are self-tuned rather than predicted from themselves, the convergence theorem is an abstraction that does not smuggle in the paper's conclusion, and the empirical claims are benchmarked against external baselines.

full rationale

The paper's central empirical claims are evaluated against nine external baselines (FedAvg, FedProx, MOON, FedDyn, FedLWS, FedAsync, PORT, ASO-Fed, FedBuff, FedSA) on five public datasets, so they do not reduce to the paper's own fitted parameters. The server-side control parameters λ, σ, ι in Eq. (3) and device-side γ, υ in Eq. (11) are updated by gradient descent on the training loss; this is self-tuning of aggregation weights, not a prediction derived from the quantity it claims to produce. Theorem 1 bounds the averaged squared gradient for the generic aggregation w_{t+1}=(1−α_i^t)w_t+α_i^t w_i^o under Assumptions 1–5 and the additional bounds α_min≤α_i^t≤1, 0≤β≤1 (Appendix A, final paragraph). The proof never uses the functional forms in Eqs. (2), (10) or the update rules in Eqs. (3)–(5), (11); hence convergence is not assumed as an input. The gap is that these boundedness conditions are not enforced by the unprojected adaptive updates, so the theorem may not substantiate the actual FedASMU/FedSSMU mechanisms; this is an omitted-support/rigor concern, not a circular derivation. The only self-citation [48] identifies the prior conference version of FedASMU and is not load-bearing; no uniqueness theorem is imported. No equation in the derivation chain reduces to its own input, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The ledger shows the method's empirical gains rest on several learned parameter sets (control parameters, RL models) whose fitted values are not reported, plus hand-chosen hyperparameters (μ_α, μ_β, target accuracies). The convergence theory assumes away the adaptive mechanisms, and the core efficiency premise depends on the downlink-bandwidth assumption. No new physical entities are introduced.

free parameters (5)
  • server-side control parameters λ_i, σ_i, ι_i = unknown; learned via Eq. (3)
    These parameters set the aggregation weight α_i^t (Eq. 2) and are updated by gradient descent against an approximation of the global loss. The claimed accuracy improvements depend on their fitted values.
  • device-side control parameters γ_i, υ_i = unknown; learned via Eq. (11)
    These set the device-side merge weight β (Eq. 10) and are fit to the local loss. The device-side adaptation and its benefits rely on them.
  • RL meta-model parameters θ_t and device Q-tables H_i = unknown; learned from request reward
    The time-slot selector is trained online from rewards (Eq. 7,8); its choices are not analyzed in the convergence proof and their optimality is only empirically supported.
  • sigmoid scaling hyperparameters μ_α and μ_β = not specified in the text
    These constants scale the control-parameter functions in Eqs. (2) and (10). They are chosen by hand but their values are not reported in Table 3.
  • target accuracy thresholds for training-time comparison = e.g., 0.40 for AlexNet/CIFAR-10, 0.12 for AlexNet/CIFAR-100, etc. (Table 4)
    The reported 'Time' values are the duration to reach a hand-chosen target accuracy per model/dataset. These thresholds are arbitrary and directly determine the efficiency comparisons.
assumptions (5)
  • domain assumption Loss functions are L-smooth, μ-strongly convex, with unbiased sampling and bounded gradients/variance (Assumptions 1-5)
    Standard FL convergence conditions invoked in Section 4.4 and Appendix A; not verified for the evaluated deep networks (e.g., VGG, ResNet, Transformers), which are non-convex.
  • domain assumption Downlink bandwidth is typically larger and underutilized than uplink
    Stated in Section 3.2 as the premise for the timely-dissemination idea; if false, the extra global-model transfers could cancel the efficiency gains.
  • ad hoc to paper A single device's local gradient (via (w_i^o - w_o)/(η_i L)) approximates the global loss gradient for server-side parameter updates
    Used in Section 4.2 to derive Eqs. (4)-(5); this is a coarse proxy, especially after a mid-training merge, and is not justified in the text.
  • ad hoc to paper Only one fresh global model is requested during a local training session
    Assumed in Section 4.3.1 to simplify the RL action; the actual implementation might benefit or require multiple requests, which is not analyzed.
  • standard math The global loss F is convex when proving the device-side and server-side aggregation bounds
    The proof in Appendix A uses convexity of F; this follows from Assumption 2 (μ-strong convexity of each F_i) since F is a weighted average, but convexity is used explicitly rather than strong convexity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Federated Learning with Timely Update Dissemination." pith.science (2026). https://pith.science/paper/J4NS7CHI

@misc{pith2026250706031,
  author       = {Pith},
  title        = {Pith review of: Efficient Federated Learning with Timely Update Dissemination},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J4NS7CHI}},
  note         = {Machine review of arXiv:2507.06031}
}
read the original abstract

Federated Learning (FL) has emerged as a compelling methodology for the management of distributed data, marked by significant advancements in recent years. In this paper, we propose an efficient FL approach that capitalizes on additional downlink bandwidth resources to ensure timely update dissemination. Initially, we implement this strategy within an asynchronous framework, introducing the Asynchronous Staleness-aware Model Update (FedASMU), which integrates both server-side and device-side methodologies. On the server side, we present an asynchronous FL system model that employs a dynamic model aggregation technique, which harmonizes local model updates with the global model to enhance both accuracy and efficiency. Concurrently, on the device side, we propose an adaptive model adjustment mechanism that integrates the latest global model with local models during training to further elevate accuracy. Subsequently, we extend this approach to a synchronous context, referred to as FedSSMU. Theoretical analyses substantiate the convergence of our proposed methodologies. Extensive experiments, encompassing six models and five public datasets, demonstrate that FedASMU and FedSSMU significantly surpass baseline methods in terms of both accuracy (up to 145.87%) and efficiency (up to 97.59%).

Figures

Figures reproduced from arXiv: 2507.06031 by the authors.

Figure 1
Figure 1. The system model of FedASMU. 3.2 Motivation Based on the observations at the end of Section 2, in order to address the problem defined in Equation (P1), we propose a timely update dissemination approach for both synchronous and asynchronous FL. The main idea is to enable the server to promptly disseminate fresh model updates from devices that have completed one round of local training to other devices that are still… view at source ↗
Figure 2
Figure 2. The accuracy and training time with AlexNet & CIFAR-10, AlexNet & [PITH_FULL_IMAGE:figures/full_fig_p023_2.png] view at source ↗
Figure 3
Figure 3. The accuracy and training time with VGG & CIFAR-100, TextCNN & [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The accuracy and training time with LeNet & CIFAR-10, LeNet & CIFAR [PITH_FULL_IMAGE:figures/full_fig_p023_4.png]
Figure 5
Figure 5. Figure 5: The accuracy and training time with CNN & CIFAR-100, ResNet & CIFAR [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: Impact of device numbers. slowest devices is 110, while the ratio is 440 with the highly heterogeneous level. The local training times for other devices randomly sampled within the maximum ratio for both levels. As depicted in [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: Impact of device heterogeneity. 6.2.5 Communication Overhead Analysis FedASMU and FedSSMU increase the communication overhead due to the addi￾tional transmissions of global models from server to devices. Since the downlink channel has high bandwidth, which incurs accep…
Figure 8
Figure 8. Figure 8: Impact of network bandwidth. refers to FedASMU lacking adaptive model update on devices. FedASMU-0 de￾notes the version of FedASMU that excludes both methods, effectively making it comparable to FedAsync with a staleness bound. The dynamic weight adjust￾ment feature si…
Figure 9
Figure 9. Figure 9: The accuracy and training time with FedASMU, FedASMU-DA, FedASMU [PITH_FULL_IMAGE:figures/full_fig_p028_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

84 extracted references · 72 canonical work pages

  1. [1]

    European Union’s General Data Protection Regulation (GDPR),

    EU, “European Union’s General Data Protection Regulation (GDPR),” 2018, https:// eugdpr.org/, accessed 2018-1

    2018

  2. [2]

    Privacy and big data,

    B. M. Gaff, H. E. Sussman, and J. Geetter, “Privacy and big data,” Computer, vol. 47, no. 6, pp. 7–9, 2014

    2014

  3. [3]

    Cybersecurity law of the people’s republic of china,

    Standing Committee of the National People’s Congress, “Cybersecurity law of the people’s republic of china,” https://www.newamerica.org/cybersecurity-initiative/digichina/blog/ translation-cybersecurity-law-peoples-republic-china/, 2021, online; accessed 22/02/2021. Efficient Federated Learning with Timely Update Dissemination 29

    2021

  4. [4]

    Advances and open problems in federated learn- ing,

    P. Kairouz, H. B. McMahan, and et al., “Advances and open problems in federated learn- ing,” Foundations and Trends® in Machine Learning , vol. 14, no. 1, 2021

    2021

  5. [5]

    From distributed machine learning to federated learning: a survey,

    J. Liu, J. Huang, Y. Zhou, X. Li, S. Ji, H. Xiong, and D. Dou, “From distributed machine learning to federated learning: a survey,” Knowledge and Information Systems (KAIS) , vol. 64, no. 4, pp. 885–917, 2022

    work page 2022

  6. [6]

    Fedvision: An online visual object detection platform powered by federated learning,

    Y. Liu, A. Huang, Y. Luo, H. Huang, Y. Liu, Y. Chen, L. Feng, T. Chen, H. Yu, and Q. Yang, “Fedvision: An online visual object detection platform powered by federated learning,” in AAAI Conf. on Artificial Intelligence (AAAI) , 2020, pp. 13 172–13 179

    work page 2020

  7. [7]

    Federated learning meets natural language processing: A survey,

    M. Liu, S. Ho, M. Wang, L. Gao, Y. Jin, and H. Zhang, “Federated learning meets natural language processing: A survey,” arXiv preprint arXiv:2107.12603 , 2021

    arXiv 2021

  8. [8]

    Fl-qsar: a federated learning-based qsar prototype for collaborative drug discovery,

    S. Chen, D. Xue, G. Chuai, Q. Yang, and Q. Liu, “Fl-qsar: a federated learning-based qsar prototype for collaborative drug discovery,” Bioinformatics, vol. 36, no. 22-23, pp. 5492–5498, 2021

    work page 2021

Show all 84 references
  1. [9]

    Fed-biomed: A general open-source frontend framework for federated learning in healthcare,

    S. Silva, A. Altmann, B. Gutman, and M. Lorenzi, “Fed-biomed: A general open-source frontend framework for federated learning in healthcare,” in Domain Adaptation and Rep- resentation Transfer, and Distributed and Collaborative Learning . Springer, 2020, pp. 201–210

    2020

  2. [10]

    Federated learning for smart healthcare: A survey,

    D. C. Nguyen, Q.-V. Pham, P. N. Pathirana, M. Ding, A. Seneviratne, Z. Lin, O. Dobre, and W.-J. Hwang, “Federated learning for smart healthcare: A survey,” ACM Computing Surveys (CSUR) , vol. 55, no. 3, pp. 1–37, 2022

    2022

  3. [11]

    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 (AISTATS), 2017, pp. 1273–1282

    2017

  4. [12]

    Distributed and deep vertical federated learning with big data,

    J. Liu, X. Zhou, L. Mo, S. Ji, Y. Liao, Z. Li, Q. Gu, and D. Dou, “Distributed and deep vertical federated learning with big data,” Concurrency and Computation: Practice and Experience, p. e7697, 2023

    2023

  5. [13]

    Efficient asynchronous federated learning with sparsification and quantization,

    J. Jia, J. Liu, C. Zhou, H. Tian, M. Dong, and D. Dou, “Efficient asynchronous federated learning with sparsification and quantization,” Concurrency and Computation: Practice and Experience, 2023, to appear

    2023

  6. [14]

    Joint device scheduling and resource allo- cation for latency constrained wireless federated learning,

    W. Shi, S. Zhou, Z. Niu, M. Jiang, and L. Geng, “Joint device scheduling and resource allo- cation for latency constrained wireless federated learning,”IEEE Transactions on Wireless Communications, vol. 20, no. 1, pp. 453–467, 2020

    2020

  7. [15]

    Oort: Efficient federated learning via guided participant selection,

    F. Lai, X. Zhu, H. V. Madhyastha, and M. Chowdhury, “Oort: Efficient federated learning via guided participant selection,” in USENIX Symposium on Operating Systems Design and Implementation (OSDI) , 2021, pp. 19–35

    2021

  8. [16]

    Client selection for federated learning with heterogeneous resources in mobile edge,

    T. Nishio and R. Yonetani, “Client selection for federated learning with heterogeneous resources in mobile edge,” in IEEE Int. Conf. on communications (ICC) , 2019, pp. 1–7

    2019

  9. [17]

    Federated fingerprint learning with heterogeneous architectures,

    T. Che, Z. Zhang, Y. Zhou, X. Zhao, J. Liu, Z. Jiang, D. Yan, R. Jin, and D. Dou, “Federated fingerprint learning with heterogeneous architectures,” in IEEE Int. Conf. on Data Mining (ICDM) . IEEE, 2022, pp. 31–40

    2022

  10. [18]

    Fast federated machine unlearning with nonlinear functional theory,

    T. Che, Y. Zhou, Z. Zhang, L. Lyu, J. Liu, D. Yan, D. Dou, and J. Huan, “Fast federated machine unlearning with nonlinear functional theory,” in Int. Conf. on Machine Learning (ICML). PMLR, 2023, pp. 4241–4268

    2023

  11. [19]

    Tackling the objective inconsistency problem in heterogeneous federated optimization,

    J. Wang, Q. Liu, H. Liang, G. Joshi, and H. V. Poor, “Tackling the objective inconsistency problem in heterogeneous federated optimization,” in Advances in Neural Information Processing Systems (NeurIPS), vol. 33, 2020, pp. 7611–7623

    2020

  12. [20]

    Federated learning of large language models with parameter-efficient prompt tuning and adaptive optimization,

    T. Che, J. Liu, Y. Zhou, J. Ren, J. Zhou, V. S. Sheng, H. Dai, and D. Dou, “Federated learning of large language models with parameter-efficient prompt tuning and adaptive optimization,” in Conf. on Empirical Methods in Natural Language Processing (EMNLP). Singapore: Associati...

    2023

  13. [21]

    SCAFFOLD: Stochastic controlled averaging for federated learning,

    S. P. Karimireddy, S. Kale, M. Mohri, S. Reddi, S. Stich, and A. T. Suresh, “SCAFFOLD: Stochastic controlled averaging for federated learning,” in Int. Conf. on Machine Learning (ICML), vol. 119, 2020, pp. 5132–5143

    2020

  14. [22]

    Measuring the effects of non-identical data distri- bution for federated visual classification,

    T.-M. H. Hsu, H. Qi, and M. Brown, “Measuring the effects of non-identical data distri- bution for federated visual classification,” arXiv preprint arXiv:1909.06335 , 2019

    1909 arXiv

  15. [23]

    Asynchronous federated learning on heterogeneous devices: A survey,

    C. Xu, Y. Qu, Y. Xiang, and L. Gao, “Asynchronous federated learning on heterogeneous devices: A survey,” arXiv preprint arXiv:2109.04269 , 2021

    2021 arXiv

  16. [24]

    SAF A: A semi-asynchronous protocol for fast federated learning with low overhead,

    W. Wu, L. He, W. Lin, R. Mao, C. Maple, and S. Jarvis, “SAF A: A semi-asynchronous protocol for fast federated learning with low overhead,”IEEE Transactions on Computers, vol. 70, no. 5, pp. 655–668, 2020. 30 Jia and Liu et al

    2020

  17. [25]

    How asynchronous can federated learning be?

    N. Su and B. Li, “How asynchronous can federated learning be?” in IEEE/ACM Int. Symposium on Quality of Service (IWQoS) , 2022, pp. 1–11

    2022

  18. [26]

    Asynchronous federated optimization,

    C. Xie, S. Koyejo, and I. Gupta, “Asynchronous federated optimization,” arXiv preprint arXiv:1903.03934, 2019

    1903 arXiv

  19. [27]

    Device scheduling with fast convergence for wireless feder- ated learning,

    W. Shi, S. Zhou, and Z. Niu, “Device scheduling with fast convergence for wireless feder- ated learning,” in IEEE Int. Conf. on Communications (ICC) , 2020, pp. 1–6

    2020

  20. [28]

    Efficient device scheduling with multi-job federated learning,

    C. Zhou, J. Liu, J. Jia, J. Zhou, Y. Zhou, H. Dai, and D. Dou, “Efficient device scheduling with multi-job federated learning,” inAAAI Conf. on Artificial Intelligence (AAAI), 2022, pp. 9971–9979

    2022

  21. [29]

    Multi-job intelli- gent scheduling with cross-device federated learning,

    J. Liu, J. Jia, B. Ma, C. Zhou, J. Zhou, Y. Zhou, H. Dai, and D. Dou, “Multi-job intelli- gent scheduling with cross-device federated learning,” IEEE Transactions on Parallel and Distributed Systems (TPDS) , vol. 34, no. 2, pp. 535–551, 2022

    2022

  22. [30]

    FedDUAP: Federated learning with dynamic update and adaptive pruning using shared data on the server,

    H. Zhang, J. Liu, J. Jia, Y. Zhou, and H. Dai, “FedDUAP: Federated learning with dynamic update and adaptive pruning using shared data on the server,” in Int. Joint Conf. on Artificial Intelligence (IJCAI) , 2022, pp. 2776–2782

    2022

  23. [31]

    Fjord: Fair and accurate federated learning under heterogeneous targets with ordered dropout,

    S. Horvath, S. Laskaridis, M. Almeida, I. Leontiadis, S. Venieris, and N. Lane, “Fjord: Fair and accurate federated learning under heterogeneous targets with ordered dropout,” Advances in Neural Information Processing Systems (NeurIPS), vol. 34, pp. 12 876–12 889, 2021

    2021

  24. [32]

    Fedhisyn: A hierarchical synchronous federated learning framework for resource and data heterogeneity,

    G. Li, Y. Hu, M. Zhang, J. Liu, Q. Yin, Y. Peng, and D. Dou, “Fedhisyn: A hierarchical synchronous federated learning framework for resource and data heterogeneity,” in Int. Conf. on Parallel Processing (ICPP) , 2022, pp. 1–10, to appear

    2022

  25. [33]

    Hierarchical federated learning across heterogeneous cellular networks,

    M. S. H. Abad, E. Ozfatura, D. Gunduz, and O. Ercetin, “Hierarchical federated learning across heterogeneous cellular networks,” in IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 8866–8870

    2020

  26. [34]

    Federated opti- mization in heterogeneous networks,

    T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V. Smith, “Federated opti- mization in heterogeneous networks,” in Machine Learning and Systems (MLSys) , vol. 2, 2020, pp. 429–450

    2020

  27. [35]

    Feder- ated learning based on dynamic regularization,

    D. A. E. Acar, Y. Zhao, R. Matas, M. Mattina, P. Whatmough, and V. Saligrama, “Feder- ated learning based on dynamic regularization,” inInt. Conf. on Learning Representations (ICLR), 2021, pp. 1–36

    2021

  28. [36]

    No fear of heterogeneity: Classifier calibration for federated learning with non-iid data,

    M. Luo, F. Chen, D. Hu, Y. Zhang, J. Liang, and J. Feng, “No fear of heterogeneity: Classifier calibration for federated learning with non-iid data,” in Advances in Neural Information Processing Systems (NeurIPS) , vol. 34, 2021, pp. 5972–5984

    2021

  29. [37]

    Adaptive federated optimization,

    S. J. Reddi, Z. Charles, M. Zaheer, Z. Garrett, K. Rush, J. Konevcn´ y, S. Kumar, and H. B. McMahan, “Adaptive federated optimization,” in Int. Conf. on Learning Representations (ICLR), 2021, pp. 1–38

    2021

  30. [38]

    Model-contrastive federated learning,

    Q. Li, B. He, and D. Song, “Model-contrastive federated learning,” in IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , 2021, pp. 10 713–10 722

    2021

  31. [39]

    Partialfed: Cross-domain personalized federated learning via partial initialization,

    B. Sun, H. Huo, Y. YANG, and B. Bai, “Partialfed: Cross-domain personalized federated learning via partial initialization,” in Advances in Neural Information Processing Systems (NeurIPS), vol. 34, 2021, pp. 23 309–23 320

    2021

  32. [40]

    Quped: Quantized personalization via distillation with applications to federated learning,

    K. Ozkara, N. Singh, D. Data, and S. Diggavi, “Quped: Quantized personalization via distillation with applications to federated learning,” in Advances in Neural Information Processing Systems (NeurIPS), vol. 34, 2021, pp. 3622–3634

    2021

  33. [41]

    Adaptive gradient-based meta-learning methods,

    M. Khodak, M.-F. F. Balcan, and A. S. Talwalkar, “Adaptive gradient-based meta-learning methods,” in Advances in Neural Information Processing Systems (NeurIPS) , vol. 32, 2019, pp. 1–12

    2019

  34. [42]

    Federated multi-task learning,

    V. Smith, C.-K. Chiang, M. Sanjabi, and A. S. Talwalkar, “Federated multi-task learning,” in Advances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017, pp. 1–11

    2017

  35. [43]

    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 (FGCS) , vol. 120, pp. 1–12, 2021

    2021

  36. [44]

    Asynchronous online federated learning for edge devices with non-iid data,

    Y. Chen, Y. Ning, M. Slawski, and H. Rangwala, “Asynchronous online federated learning for edge devices with non-iid data,” in IEEE Int. Conf. on Big Data (Big Data) , 2020, pp. 15–24

    2020

  37. [45]

    Attention-weighted federated deep reinforcement learning for device-to-device assisted heterogeneous collaborative edge caching,

    X. Wang, R. Li, C. Wang, X. Li, T. Taleb, and V. C. Leung, “Attention-weighted federated deep reinforcement learning for device-to-device assisted heterogeneous collaborative edge caching,” IEEE Journal on Selected Areas in Communications, vol. 39, no. 1, pp. 154–169, 2020. Ef...

    2020

  38. [46]

    Federated meta-learning with attention for diversity-aware human activity recognition,

    Q. Shen, H. Feng, R. Song, D. Song, and H. Xu, “Federated meta-learning with attention for diversity-aware human activity recognition,” Sensors, vol. 23, no. 3, p. 1083, 2023

    2023

  39. [47]

    Fedssa: Semantic similarity-based aggregation for efficient model-heterogeneous personalized federated learning,

    L. Yi, H. Yu, Z. Shi, G. Wang, X. Liu, L. Cui, and X. Li, “Fedssa: Semantic similarity-based aggregation for efficient model-heterogeneous personalized federated learning,” arXiv preprint arXiv:2312.09006, 2023

    2023 arXiv

  40. [48]

    Fedasmu: Effi- cient 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: Effi- cient asynchronous federated learning with dynamic staleness-aware model update,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 12, 2024, pp. 13 900–13 908

    2024

  41. [49]

    A comprehensive survey on knowledge- defined networking,

    P. A. D. S. N. Wijesekara and S. Gunawardena, “A comprehensive survey on knowledge- defined networking,” in Telecom, vol. 4, no. 3. MDPI, 2023, pp. 477–596

    2023

  42. [50]

    Fellas: Enhancing federated sequential recommendation with llm as external services,

    W. Yuan, C. Yang, G. Ye, T. Chen, N. Q. V. Hung, and H. Yin, “Fellas: Enhancing federated sequential recommendation with llm as external services,” ACM Transactions on Information Systems , 2024

    2024

  43. [51]

    Hetefedrec: Federated recom- mender systems with model heterogeneity,

    W. Yuan, L. Qu, L. Cui, Y. Tong, X. Zhou, and H. Yin, “Hetefedrec: Federated recom- mender systems with model heterogeneity,” in 2024 IEEE 40th International Conference on Data Engineering (ICDE) . IEEE, 2024, pp. 1324–1337

    2024

  44. [52]

    A systematic review on federated learn- ing system: a new paradigm to machine learning,

    R. K. Chaudhary, R. Kumar, and N. Saxena, “A systematic review on federated learn- ing system: a new paradigm to machine learning,” Knowledge and Information Systems , vol. 67, no. 2, pp. 1811–1914, 2025

    1914

  45. [53]

    A survey of federated learning for edge com- puting: Research problems and solutions,

    Q. Xia, W. Ye, Z. Tao, J. Wu, and Q. Li, “A survey of federated learning for edge com- puting: Research problems and solutions,” High-Confidence Computing , vol. 1, no. 1, p. 100008, 2021

    2021

  46. [54]

    Federated-learning-enabled intelligent fog radio access networks: Fundamental theory, key techniques, and future trends,

    Z. Zhao, C. Feng, H. H. Yang, and X. Luo, “Federated-learning-enabled intelligent fog radio access networks: Fundamental theory, key techniques, and future trends,” IEEE wireless communications, vol. 27, no. 2, pp. 22–28, 2020

    2020

  47. [55]

    In-edge ai: Intelligentizing mobile edge computing, caching and communication by federated learning,

    X. Wang, Y. Han, C. Wang, Q. Zhao, X. Chen, and M. Chen, “In-edge ai: Intelligentizing mobile edge computing, caching and communication by federated learning,” Ieee Network, vol. 33, no. 5, pp. 156–165, 2019

    2019

  48. [56]

    Federated learning on internet of things: Extensive and systematic review

    M. Aggarwal, V. Khullar, S. Rani, T. A. Prola, S. B. Bhattacharjee, S. M. Shawon, and N. Goyal, “Federated learning on internet of things: Extensive and systematic review.” Computers, Materials & Continua , vol. 79, no. 2, 2024

    2024

  49. [57]

    Fladen: Federated learning for anomaly detection in iot networks,

    F. Hendaoui, R. Meddeb, L. Trabelsi, A. Ferchichi, and R. Ahmed, “Fladen: Federated learning for anomaly detection in iot networks,”Computers & Security, vol. 155, p. 104446, 2025

    2025

  50. [58]

    Analyzing federated learning through an adversarial lens,

    A. N. Bhagoji, S. Chakraborty, P. Mittal, and S. Calo, “Analyzing federated learning through an adversarial lens,” in International conference on machine learning . PMLR, 2019, pp. 634–643

    2019

  51. [59]

    Exploiting unintended feature leakage in collaborative learning,

    L. Melis, C. Song, E. De Cristofaro, and V. Shmatikov, “Exploiting unintended feature leakage in collaborative learning,” in 2019 IEEE symposium on security and privacy (SP) . IEEE, 2019, pp. 691–706

    2019

  52. [60]

    Differential privacy,

    C. Dwork, “Differential privacy,” in International colloquium on automata, languages, and programming. Springer, 2006, pp. 1–12

    2006

  53. [61]

    Practical secure aggregation for federated learning on user-held data,

    K. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan, S. Patel, D. Ramage, A. Segal, and K. Seth, “Practical secure aggregation for federated learning on user-held data,” arXiv preprint arXiv:1611.04482 , 2016

    2016 arXiv

  54. [62]

    Towards federated learning at scale: System design,

    K. Bonawitz, H. Eichner, W. Grieskamp, D. Huba, A. Ingerman, V. Ivanov, C. Kiddon, J. Kone vcn` y, S. Mazzocchi, B. McMahan et al. , “Towards federated learning at scale: System design,” Machine Learning and Systems (MLSys) , vol. 1, pp. 374–388, 2019

    2019

  55. [63]

    Timely communication in federated learning,

    B. Buyukates and S. Ulukus, “Timely communication in federated learning,” in IEEE IN- FOCOM 2021-IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS). IEEE, 2021, pp. 1–6

    2021

  56. [64]

    Hide your model: A param- eter transmission-free federated recommender system,

    W. Yuan, C. Yang, L. Qu, Q. V. H. Nguyen, J. Li, and H. Yin, “Hide your model: A param- eter transmission-free federated recommender system,” in 2024 IEEE 40th International Conference on Data Engineering (ICDE) . IEEE, 2024, pp. 611–624

    2024

  57. [65]

    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,” in Int. Conf. on Artificial Intelligence and Statistics (AISTATS) , vol. 151, 2022, pp. 3581–3607

    2022

  58. [66]

    Fedpa: An adaptively partial model aggregation strategy in federated learning,

    J. Liu, J. H. Wang, C. Rong, Y. Xu, T. Yu, and J. Wang, “Fedpa: An adaptively partial model aggregation strategy in federated learning,”Computer Networks, vol. 199, p. 108468, 2021. 32 Jia and Liu et al

    2021

  59. [67]

    Sageflow: Robust federated learning against both stragglers and adversaries,

    J. Park, D.-J. Han, M. Choi, and J. Moon, “Sageflow: Robust federated learning against both stragglers and adversaries,” in Advances in Neural Information Processing Systems (NeurIPS), vol. 34, 2021, pp. 840–851

    2021

  60. [68]

    More effective distributed ml via a stale synchronous parallel parameter server,

    Q. Ho, J. Cipar, H. Cui, S. Lee, J. K. Kim, P. B. Gibbons, G. A. Gibson, G. Ganger, and E. P. Xing, “More effective distributed ml via a stale synchronous parallel parameter server,” Advances in neural information processing systems (NeurIPS) , vol. 26, 2013

    2013

  61. [69]

    J. F. Bard, Practical Bilevel Optimization: Algorithms and Applications . Springer, 1998

    1998

  62. [70]

    A stochastic approximation method,

    H. Robbins and S. Monro, “A stochastic approximation method,” The annals of mathe- matical statistics, pp. 400–407, 1951

    1951

  63. [71]

    Technical note q-learning,

    C. J. C. H. Watkins and P. Dayan, “Technical note q-learning,” Machine Learning, vol. 8, pp. 279–292, 1992

    1992

  64. [72]

    Online reinforcement learning by bayesian inference,

    Z. Xia and D. Zhao, “Online reinforcement learning by bayesian inference,” in Int. Joint Conf. on Neural Networks (IJCNN) , 2015, pp. 1–6

    2015

  65. [73]

    Neural architecture search with reinforcement learning,

    B. Zoph and Q. V. Le, “Neural architecture search with reinforcement learning,” in Int. Conf. on Learning Representations (ICLR) , 2017

    2017

  66. [74]

    Hierarchical reinforcement learning with the maxq value function de- composition,

    T. G. Dietterich, “Hierarchical reinforcement learning with the maxq value function de- composition,” Journal of artificial intelligence research , vol. 13, pp. 227–303, 2000

    2000

  67. [75]

    FedL WS: Federated learning with adaptive layer-wise weight shrinking,

    C. Shi, J. Li, H. Zhao, D. dan Guo, and Y. Chang, “FedL WS: Federated learning with adaptive layer-wise weight shrinking,” in The Thirteenth International Conference on Learning Representations, 2025

    2025

  68. [76]

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

    H. Xiao, K. Rasul, and R. Vollgraf, “Fashion-mnist: a novel image dataset for benchmark- ing machine learning algorithms,” arXiv preprint arXiv:1708.07747 , 2017

    2017 arXiv

  69. [77]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hinton et al. , “Learning multiple layers of features from tiny images,” 2009

    2009

  70. [78]

    Distilled one-shot federated learning,

    Y. Zhou, G. Pu, X. Ma, X. Li, and D. Wu, “Distilled one-shot federated learning,” arXiv preprint arXiv:2009.07999, vol. abs/2009.07999, no. 1, pp. 1–16, 2021

    2009 arXiv

  71. [79]

    Tiny imagenet visual recognition challenge,

    Y. Le and X. Yang, “Tiny imagenet visual recognition challenge,” CS 231N, vol. 7, no. 7, p. 3, 2015

    2015

  72. [80]

    Federated learning on non-iid data silos: An exper- imental study,

    Q. Li, Y. Diao, Q. Chen, and B. He, “Federated learning on non-iid data silos: An exper- imental study,” arXiv preprint arXiv:2102.02079 , 2021

    2021 arXiv

  73. [81]

    Handwritten digit recognition with a back-propagation network,

    Y. LeCun, B. Boser, J. Denker, D. Henderson, R. Howard, W. Hubbard, and L. Jackel, “Handwritten digit recognition with a back-propagation network,” in Advances in Neural Information Processing Systems (NeurIPS) , vol. 2, 1989, pp. 1–9

    1989

  74. [82]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) . Las Vegas, USA: IEEE, 2016, pp. 770–778

    2016

  75. [83]

    Imagenet classification with deep con- volutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep con- volutional neural networks,” in Annual Conf. on Neural Information Processing Systems (NeurIPS), 2012, pp. 1106–1114

    2012

  76. [84]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” in Int. Conf. on Learning Representations (ICLR) , 2015. Efficient Federated Learning with Timely Update Dissemination 33 A Appendix A.1 Convergence Analysis Proof. First, we de...

    2015

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.