Pith. sign in

REVIEW 5 major objections 5 minor 46 references

Edge Association Strategies for Synthetic Data Empowered Hierarchical Federated Learning with Non-IID Data

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 This paper claims that a small dose of synthetic data, generated at edge servers and given to workers, materially improves federated model accuracy when local data is highly non-IID.

desk verdict A practically motivated integration of synthetic data and edge-association games for hierarchical FL; the empirical story is plausible, but the uniqueness/stability proof is not. read the letter →

arxiv 2506.18259 v1 pith:2MWJHPIW submitted 2025-06-23 cs.DC

Jer Shyuan Ng , Aditya Pribadi Kalapaaking , Xiaoyu Xia , Dusit Niyato , Ibrahim Khalil , Iqbal Gondal This is my paper · ORCID
classification cs.DC
keywords hierarchicalfederatedlearningsyntheticdatanon-IIDevolutionarygametheoryedgeassociationreplicatordynamicsincentivemechanism
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 claims that a hierarchical federated learning network can overcome two practical obstacles at once: the accuracy loss caused by non-IID data and the reluctance of workers to participate. It proposes that edge servers generate small synthetic datasets related to the learning task and distribute them to workers, while each worker chooses which edge server to join based on reward, data quantity, and computation cost. The paper models this choice as an evolutionary game and shows, analytically and by simulation, that the resulting cluster formation reaches a unique, stable equilibrium. Its headline experimental result is that adding just 5% synthetic data raises MNIST test accuracy from 0.8923 to 0.9316 at iteration 500 under the most extreme non-IID setting (each worker holds a single class), and that accuracy gains from larger synthetic fractions are small.

What carries the argument

The load-bearing mechanism is the replicator dynamics equation (5), $\dot{x}^{(z)}_n(t)=\delta x^{(z)}_n(t)(u^{(z)}_n(t)-\bar{u}^{(z)}(t))$, which describes how the population share of workers in group $z$ choosing edge server $n$ evolves according to whether that choice beats the group's average utility. Utility (equation (2)) is a linear reward pool $\gamma_n$ divided in proportion to data contribution, minus computation and communication costs, including $s_n$, the extra computation required to train on the edge-generated synthetic data. This means the same equation that settles workers into stable clusters also accounts for the cost of synthetic-data training. Stability is proved with the Lyapunov function $G(x^{(z)}(t))=(x^{(z)*}_n(t)-x^{(z)}_n(t))^2$, whose derivative is shown to be non-positive.

What would settle it

Run the same HFL setup at a 5% synthetic-data budget with a generator deliberately trained on a shifted distribution (for example, a different digit style or label-inverted CIFAKE) and check whether test accuracy still improves over the 0% baseline; if it does not, the accuracy claim fails.

Watch

Extended reading notes

Core claim

The central discovery is that a small, task-specific synthetic dataset generated at the edge and added to each worker's local data re-aligns heterogeneous local distributions enough to recover most of the accuracy lost to non-IID skew, while remaining compatible with incentive-driven edge association. Concretely, in the hardest tested scenario (one class per worker, non-IID edge servers) MNIST accuracy at iteration 500 rises from 0.8923 with 0% synthetic data to 0.9316 with 5%; on CIFAR-10, the two-class IID-edge scenario improves from 0.7062 to 0.8416 when 25% synthetic data is used, with the largest single gain again coming from the first 5%. The paper also establishes that the replicator dynamics governing workers' edge-server choices have a unique evolutionary equilibrium that is globally asymptotically stable, so no worker can raise its utility by switching servers once the equilibrium is reached.

Load-bearing premise

The load-bearing assumption is that an edge server can produce synthetic data whose distribution is close enough to the true task data that a small fraction of it makes non-IID local datasets more compatible; in the experiments this is ensured by construction because the synthetic MNIST and CIFAKE datasets closely mirror the real ones.

Editorial extensions

If this is right

  • Under the paper's assumptions, a 5% synthetic-data budget is sufficient to recover most of the accuracy lost to extreme non-IID data, and larger budgets yield diminishing returns.
  • Worker clusters self-organize into a unique, stable configuration through local utility comparisons, without a central coordinator deciding assignments.
  • The accuracy benefit persists whether edge servers aggregate IID or non-IID worker data, and on both digit (MNIST) and object (CIFAR-10) classification tasks.
  • The equilibrium is reached despite the added computation cost of training on synthetic data, so incentives and data augmentation do not conflict.
  • For a fixed cloud aggregation interval, more local updates per worker improve accuracy, so the framework preserves the usual HFL communication-computation tradeoff.

Reading between the lines

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

  • If the synthetic generator does not match the true task distribution, the 5% gain is not guaranteed; a direct test would compare a well-matched generator against a deliberately mismatched one at the same 5% budget.
  • The uniqueness and stability results rest on risk-neutral utilities and single-resource (data quantity) reward sharing; allowing risk-averse workers or multi-resource contributions could produce multiple equilibria or different cluster splits.
  • A practical extension is to make the synthetic fraction adaptive: since gains saturate quickly, edge servers could estimate the minimum fraction needed and stop generating synthetic samples, saving computation.
  • The same evolutionary-game machinery could be applied where workers choose among edge servers offering different generative models, turning generator fidelity into part of the selection criterion.
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 a hierarchical federated learning (HFL) system in which edge servers distribute synthetic data to FL workers and offer reward pools to incentivize participation. Worker association to edge servers is modeled as an evolutionary game with replicator dynamics, and the paper claims that a unique, globally asymptotically stable evolutionary equilibrium is reached. Simulations on MNIST and CIFAR-10 report that adding 5% synthetic data improves FL model accuracy under one-class and two-class non-IID data distributions. The paper's main contributions are the synthetic-data-empowered HFL framework, the evolutionary game formulation, and the claimed uniqueness and stability of the resulting equilibrium.

Significance. The topic is timely, and the combination of incentive design with synthetic-data augmentation in HFL is a plausible practical direction. The empirical experiments are clearly described, and the reported accuracy gains from 5% synthetic data, if confirmed, are practically interesting. However, the central theoretical claim of a unique and stable evolutionary equilibrium is not established: the proofs of Theorems 2 and 3 are invalid as written, and the utility model in Eq. (2) does not express the per-worker congestion effect that the paper invokes. The contribution therefore currently rests on simulation evidence alone, which is insufficient for the paper's stated claims.

major comments (5)
  1. [Section IV.A, Eq. (2)] The utility in Eq. (2) is not the per-worker payoff described in the text. The reward term γ_n d_z x_n^(z) / Σ_{z'=1}^Z d_{z'} x_n^(z') is the fraction of edge n's reward pool earned by population z as a whole, not the reward received by an individual FL worker. For a single population (Z=1), this term is identically γ_n, so utility is independent of x_n^(z) and no congestion effect exists; for Z>1, the term is increasing in x_n^(z) when other populations are held fixed, which is the opposite of the crowding behavior invoked in Section V.A. A per-worker formulation would divide d_z by the total data at edge n, not multiply it by x_n^(z). Since Eqs. (4)-(6) and all simulations inherit this payoff structure, the evolutionary game as specified does not model the association incentives the paper claims.
  2. [Section IV.B.1, Theorem 2] Theorem 2 proves, via a Lipschitz argument, that the initial-value problem in Eq. (5) has a unique solution for each initial state x0. This establishes uniqueness of trajectories, not uniqueness of a stationary equilibrium. The sentence immediately after the proof, which concludes that the evolutionary equilibrium 'not only exists but is also unique', does not follow; multiple rest points with separate basins of attraction are entirely compatible with the proved bound. No argument is given that a rest point of the replicator dynamics exists at all.
  3. [Section IV.B.2, Theorem 3] The Lyapunov analysis of the interior equilibrium is invalid. The function G(x^(z)(t)) = (x_n^(z)*(t) - x_n^(z)(t))^2 in Eq. (10) is not positive definite on the simplex: it vanishes whenever x_n^(z) = x_n^(z)* regardless of the other N-1 coordinates, and it cannot tend to infinity as the state grows because the state space X is compact. The sign argument in Eq. (12) relies on the unproved implications u_n > \bar u ⇒ x_n* > x_n and u_n < \bar u ⇒ x_n* < x_n; these are generally false because the average utility \bar u depends on the whole state. Moreover, the proof states that \partial x_n/\partial t < 0 when u_n > \bar u, whereas the replicator dynamics in Eq. (5) give \dot x_n > 0 in that case. Global asymptotic stability is therefore not established.
  4. [Section IV.B.2 and Section V.A] The paper assumes rather than proves the existence of an interior equilibrium and asserts without proof that boundary equilibria are unstable. Theorem 3 is stated only for an interior equilibrium, but the existence of such a point in the parameter ranges of Table II is never demonstrated analytically; Figs. 2-6 provide only selected initial conditions and cannot establish uniqueness or global stability. Because contribution 3 explicitly promises a unique and stable equilibrium, these gaps are load-bearing for the paper's central claim.
  5. [Section V.B] The synthetic-data experiments use MNIST-cGAN and CIFAKE as proxies, both of which are generated from models trained on the same task distribution as the real data, so the synthetic and real distributions nearly coincide. The experiments therefore do not test the framework's key assumption from Section III step 2 that edge servers can generate task-specific synthetic data of sufficient fidelity in real deployments where the generative model will be imperfect. The reported 5%-synthetic accuracy gain should be presented as an upper-bound result until experiments with distribution-mismatched generators are provided.
minor comments (5)
  1. [Section III, Eq. (1)] The notation 'k|κ1 ≠ 0' and 'k|κ1κ2 ≠ 0' is mathematically ambiguous; it should read 'k mod κ1 ≠ 0' and 'k mod (κ1κ2) ≠ 0' respectively, and the phrase 'three caseas' is a typo for 'three cases'.
  2. [Section IV.B, proof of Theorem 1] The proof equates |\dot x_n| with |\partial u_n / \partial x_v|, which is not correct: \dot x_n is the replicator dynamics value, not the partial derivative of utility. A Lipschitz bound can be obtained from compactness and continuity of the dynamics, but the derivation as written is flawed.
  3. [Algorithm 1] The termination condition on line 4 uses x^(z)*(t) = x^(z)(t), but x^(z)*(t) is not defined before the loop and exact equality is unlikely with floating-point updates; a tolerance-based stopping rule would be more appropriate.
  4. [Section V.A, Fig. 2] The notation x^(1,2)_1 = (0.1, 0.1) is confusing: it denotes the two-dimensional state (x_1^(1), x_1^(2)), but the vector parentheses and superscript are easy to misread as a single coordinate; a clearer notation such as (x_1^(1), x_1^(2)) = (0.1, 0.1) would help.
  5. [Section V.B, Figs. 7-9] The accuracy curves appear to be single runs without error bars or multiple seeds. Given the small number of workers (50) and the randomness in non-IID data assignment, reporting mean and standard deviation over several runs would materially strengthen the empirical claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the accuracy results are empirical measurements and the equilibrium claim is a model simulation, not a fitted prediction.

full rationale

The paper's central accuracy result is an empirical comparison between a 0% synthetic-data baseline and 5-25% synthetic-data conditions on MNIST and CIFAR-10; no accuracy value is fitted from the game-theoretic model or reused as a prediction, so the reported 5% gain is not circular. The evolutionary-game equilibrium is obtained by simulating the replicator dynamics that the paper defines, and the observation that workers have no incentive to switch at a rest point is definitional to that equilibrium concept rather than an independent empirical prediction. The self-citations [6]-[8] are background references on HFL resource allocation and edge association, and they are not load-bearing for the synthetic-data contribution or for the uniqueness/stability claims; no uniqueness theorem is imported from the authors' prior work. The main substantive weaknesses are correctness concerns rather than circularity: the proof of Theorem 3 uses a Lyapunov function that is not positive definite on the population simplex and asserts, without derivation, the sign relation between x_n^* - x_n and u_n - \bar u, while Theorem 2 proves uniqueness of the ODE trajectory for a fixed initial condition rather than uniqueness of the equilibrium rest point. Separately, the synthetic datasets used (MNIST-cGAN and CIFAKE) are near-distribution surrogates of the test distributions, which limits external validity but does not make the measured improvement an identity. These issues do not reduce the derived results to their inputs, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on behavioral assumptions about worker rationality, the realism of replicator dynamics, and the quality of arbitrary edge-generated synthetic data. None of these is independently validated, and several simulation parameters are chosen by hand without sensitivity analysis.

free parameters (5)
  • Reward pools gamma_n = gamma1=100, gamma2=300, gamma3=500; swept to 900 in Fig. 5
    Reward pool sizes are chosen by hand and directly shape the equilibrium split; no derivation or measured calibration is given.
  • Unit costs alpha, beta = alpha=beta=0.001
    Unit computation and communication costs are set to convenient simulation values with no empirical basis.
  • Resource ranges c_z, m_z, s_n = c_z and m_z in [10,50], s_n in [2,6]
    Resource heterogeneity values are selected to create the plotted behaviors and have no empirical basis.
  • Adaptation rate delta = 0.001-0.1
    Controls replicator-dynamics convergence speed; the paper shows speed varies but the equilibrium point is unchanged.
  • Synthetic data fractions = 0%, 5%, 10%, 15%, 20%, 25%
    The 5% suffices claim depends on this sweep and on the quality of the particular synthetic datasets used.
assumptions (6)
  • domain assumption Each FL worker associates with exactly one edge server at a time.
    Section III states this explicitly; it makes the evolutionary game single-choice and rules out multi-homing.
  • domain assumption FL workers are risk-neutral, so the utility function U is linear.
    Section IV.A says without loss of generality, but it is a behavioral assumption, not a mathematical triviality.
  • domain assumption Workers' edge-switching follows replicator dynamics with full information about other workers' utilities.
    Equation (5) is the model's engine; no empirical evidence is given that real IoT workers update this way.
  • ad hoc to paper Boundary equilibria are unstable.
    Stated without proof in Section IV.B.2; this claim is needed for the interior-equilibrium stability story.
  • domain assumption Edge servers can generate useful task-specific synthetic data from pre-trained models.
    Section III step 2 and Section V.B assume cGAN and CIFAKE are adequate for MNIST and CIFAR-10 tasks.
  • standard math ODE uniqueness via Lipschitz continuity and Lyapunov stability theory are applicable.
    Invoked in Theorems 2 and 3, but applied incorrectly, so the standard background does not rescue the conclusion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Edge Association Strategies for Synthetic Data Empowered Hierarchical Federated Learning with Non-IID Data." pith.science (2026). https://pith.science/paper/2MWJHPIW

@misc{pith2026250618259,
  author       = {Pith},
  title        = {Pith review of: Edge Association Strategies for Synthetic Data Empowered Hierarchical Federated Learning with Non-IID Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2MWJHPIW}},
  note         = {Machine review of arXiv:2506.18259}
}
read the original abstract

In recent years, Federated Learning (FL) has emerged as a widely adopted privacy-preserving distributed training approach, attracting significant interest from both academia and industry. Research efforts have been dedicated to improving different aspects of FL, such as algorithm improvement, resource allocation, and client selection, to enable its deployment in distributed edge networks for practical applications. One of the reasons for the poor FL model performance is due to the worker dropout during training as the FL server may be located far away from the FL workers. To address this issue, an Hierarchical Federated Learning (HFL) framework has been introduced, incorporating an additional layer of edge servers to relay communication between the FL server and workers. While the HFL framework improves the communication between the FL server and workers, large number of communication rounds may still be required for model convergence, particularly when FL workers have non-independent and identically distributed (non-IID) data. Moreover, the FL workers are assumed to fully cooperate in the FL training process, which may not always be true in practical situations. To overcome these challenges, we propose a synthetic-data-empowered HFL framework that mitigates the statistical issues arising from non-IID local datasets while also incentivizing FL worker participation. In our proposed framework, the edge servers reward the FL workers in their clusters for facilitating the FL training process. To improve the performance of the FL model given the non-IID local datasets of the FL workers, the edge servers generate and distribute synthetic datasets to FL workers within their clusters. FL workers determine which edge server to associate with, considering the computational resources required to train on both their local datasets and the synthetic datasets.

Figures

Figures reproduced from arXiv: 2506.18259 by the authors.

Figure 1
Figure 1. Synthetic-data-empowered Hierarchical Federated [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. shows the evolution of the population states of the two populations in joining edge server 1 given varying 0.0 0.2 0.4 0.6 0.8 1.0 Population share of Population 1, x (1) 1 0.0 0.2 0.4 0.6 0.8 1.0 P o p ulatio n sh are of P o p ulatio n 2, x (2) 1 Trajectories and direction fields x (1, 2) 1 =(0.1, 0.1) x (1, 2) 1 =(0.3, 0.4) x (1, 2) 1 =(0.4, 0.7) x (1, 2) 1 =(0.6, 0.9) [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Population shares of different populations. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Data quantities given different reward pools. 10 15 20 25 30 35 40 45 50 Computation resource of FL workers in population 1, c1 0.84 0.86 0.88 0.90 0.92 0.94 0.96 D a t a q u a n tit y, D n n = 1 n = 2 n = 3 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: FL model accuracy given various non-IID scenarios. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: FL model accuracy on MNIST dataset given varying amount of synthetic data samples under various non-IID scenarios. per training iteration is employed for the local computation of FL training of each FL worker. When the FL workers no longer able to increase their utilit…
Figure 9
Figure 9. Figure 9: FL model accuracy on CIFAR-10 dataset with varying amount of synthetic data samples when the edge servers are IID, each FL worker has two classes of data samples.. 0 100 200 300 400 500 Training Iteration 0.800 0.825 0.850 0.875 0.900 0.925 0.950 0.975 1.000 Test Accur…
Figure 10
Figure 10. Figure 10: FL model accuracy on MNIST dataset given different values of κ1 and κ2 for fixed cloud interval, κ1κ2. data is particularly pronounced when each FL worker has only a single class of data, i.e., when the data distribution is more non-IID, as observed in Figures 8(a) an…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 44 canonical work pages

  1. [1]

    Automatic Healthcare Diagnosis and Prediction Assessment based on AI Multi-Classification Algorithm,

    S. Endargiri and K. Laabidi, “Automatic Healthcare Diagnosis and Prediction Assessment based on AI Multi-Classification Algorithm,” in 12 2020 16th International Computer Engineering Conference (ICENCO) , pp. 112–117, 2020

    work page 2020

  2. [2]

    Federated Learning for Object Detection in Autonomous Ve- hicles,

    D. Jallepalli, N. C. Ravikumar, P. V . Badarinath, S. Uchil, and M. A. Suresh, “Federated Learning for Object Detection in Autonomous Ve- hicles,” in 2021 IEEE Seventh International Conference on Big Data Computing Service and Applications (BigDataService) , pp. 107–114, 2021

    work page 2021

  3. [3]

    Privacy-Preserving Traffic Flow Prediction: A Federated Learning Approach,

    Y . Liu, J. J. Q. Yu, J. Kang, D. Niyato, and S. Zhang, “Privacy-Preserving Traffic Flow Prediction: A Federated Learning Approach,”IEEE Internet of Things Journal , vol. 7, no. 8, pp. 7751–7763, 2020

    work page 2020

  4. [4]

    FELIDS: Federated learning-based intrusion detection sys- tem for agricultural Internet of Things,

    O. Friha, M. A. Ferrag, L. Shu, L. Maglaras, K.-K. R. Choo, and M. Nafaa, “FELIDS: Federated learning-based intrusion detection sys- tem for agricultural Internet of Things,” Journal of Parallel and Dis- tributed Computing, vol. 165, pp. 17–31, 2022

    work page 2022

  5. [5]

    Communication-efficient Learning of Deep Networks from Decentral- ized Data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y. Arcas, “Communication-efficient Learning of Deep Networks from Decentral- ized Data,” in Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (A. Singh and J. Zhu, eds.), vol. 54 of Proceedings of Machine Learning Research , pp. 1273–1282, PMLR, 20–22 Apr 2017

    work page 2017

  6. [6]

    Decentralized Edge Intelligence: A Dynamic Resource Allocation Framework for Hierarchical Federated Learning,

    W. Y . B. Lim, J. S. Ng, Z. Xiong, J. Jin, Y . Zhang, D. Niyato, C. Leung, and C. Miao, “Decentralized Edge Intelligence: A Dynamic Resource Allocation Framework for Hierarchical Federated Learning,” IEEE Transactions on Parallel and Distributed Systems , vol. 33, no. 3, pp. 536–550, 2022

    work page 2022

  7. [7]

    Dynamic Edge Association and Resource Allocation in Self-Organizing Hierarchical Federated Learning Networks,

    W. Y . B. Lim, J. S. Ng, Z. Xiong, D. Niyato, C. Miao, and D. I. Kim, “Dynamic Edge Association and Resource Allocation in Self-Organizing Hierarchical Federated Learning Networks,” IEEE Journal on Selected Areas in Communications , vol. 39, no. 12, pp. 3640–3653, 2021

    work page 2021

  8. [8]

    Reputation-Aware Hedonic Coalition Formation for Ef- ficient Serverless Hierarchical Federated Learning,

    J. S. Ng, W. Y . B. Lim, Z. Xiong, X. Cao, J. Jin, D. Niyato, C. Leung, and C. Miao, “Reputation-Aware Hedonic Coalition Formation for Ef- ficient Serverless Hierarchical Federated Learning,” IEEE Transactions on Parallel and Distributed Systems , vol. 33, no. 11, pp. 2675–2686, 2022

    work page 2022

Show all 46 references
  1. [9]

    Adaptive Hierar- chical Federated Learning Over Wireless Networks,

    B. Xu, W. Xia, W. Wen, P. Liu, H. Zhao, and H. Zhu, “Adaptive Hierar- chical Federated Learning Over Wireless Networks,” IEEE Transactions on Vehicular Technology, vol. 71, no. 2, pp. 2070–2083, 2022

    2022

  2. [10]

    Client-Edge-Cloud Hi- erarchical Federated Learning,

    L. Liu, J. Zhang, S. Song, and K. B. Letaief, “Client-Edge-Cloud Hi- erarchical Federated Learning,” in ICC 2020 - 2020 IEEE International Conference on Communications (ICC) , pp. 1–6, 2020

    2020

  3. [11]

    Bilateral pricing for dynamic association in federated edge learning,

    B. Pan, J. Lu, S. Cao, J. Liu, W. Tian, and M. Li, “Bilateral pricing for dynamic association in federated edge learning,” IEEE Transactions on Mobile Computing, pp. 1–14, 2025

    2025

  4. [12]

    Federated Learning With Non-IID Data: A Survey,

    Z. Lu, H. Pan, Y . Dai, X. Si, and Y . Zhang, “Federated Learning With Non-IID Data: A Survey,” IEEE Internet of Things Journal , vol. 11, no. 11, pp. 19188–19209, 2024

    2024

  5. [13]

    IOFL: Intelligent-Optimization-Based Federated Learning for Non-IID Data,

    X. Li, H. Zhao, and W. Deng, “IOFL: Intelligent-Optimization-Based Federated Learning for Non-IID Data,” IEEE Internet of Things Journal, vol. 11, no. 9, pp. 16693–16699, 2024

    2024

  6. [14]

    Delayed Gradient Aver- aging: Tolerate the Communication Latency for Federated Learning,

    L. Zhu, H. Lin, Y . Lu, Y . Lin, and S. Han, “Delayed Gradient Aver- aging: Tolerate the Communication Latency for Federated Learning,” in Advances in Neural Information Processing Systems (M. Ranzato, A. Beygelzimer, Y . Dauphin, P. Liang, and J. W. Vaughan, eds.), vol. 34, p...

    2021

  7. [15]

    Deep Gradient Compression: Reducing the Communication Bandwidth for Distributed Training,

    Y . Lin, S. Han, H. Mao, Y . Wang, and W. J. Dally, “Deep Gradient Compression: Reducing the Communication Bandwidth for Distributed Training,” arXiv preprint arXiv:1712.01887 , 2020

    2020 arXiv

  8. [16]

    QSGD: Communication-Efficient SGD via Gradient Quantization and Encod- ing,

    D. Alistarh, D. Grubic, J. Li, R. Tomioka, and M. V ojnovic, “QSGD: Communication-Efficient SGD via Gradient Quantization and Encod- ing,” in Advances in Neural Information Processing Systems (I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. ...

    2017

  9. [17]

    A Joint Learning and Communications Framework for Federated Learning Over Wireless Networks,

    M. Chen, Z. Yang, W. Saad, C. Yin, H. V . Poor, and S. Cui, “A Joint Learning and Communications Framework for Federated Learning Over Wireless Networks,” IEEE Transactions on Wireless Communications , vol. 20, no. 1, pp. 269–283, 2021

    2021

  10. [18]

    Federated Learning Over Wireless IoT Networks With Optimized Communication and Resources,

    H. Chen, S. Huang, D. Zhang, M. Xiao, M. Skoglund, and H. V . Poor, “Federated Learning Over Wireless IoT Networks With Optimized Communication and Resources,” IEEE Internet of Things Journal, vol. 9, no. 17, pp. 16592–16605, 2022

    2022

  11. [19]

    Resource Allocation for Multi-Task Federated Learning Algorithm over Wireless Communication Networks,

    B. Cao, M. Chen, Y . Ben, Z. Yang, Y . Hu, C. Huang, and Y . Cang, “Resource Allocation for Multi-Task Federated Learning Algorithm over Wireless Communication Networks,” in 2022 IEEE Wireless Communi- cations and Networking Conference (WCNC) , pp. 590–595, 2022

    2022

  12. [20]

    Robust Federated Learning With Noisy Communication,

    F. Ang, L. Chen, N. Zhao, Y . Chen, W. Wang, and F. R. Yu, “Robust Federated Learning With Noisy Communication,” IEEE Transactions on Communications, vol. 68, no. 6, pp. 3452–3464, 2020

    2020

  13. [21]

    Federated Learning in Mobile Edge Networks: A Comprehensive Survey,

    W. Y . B. Lim, N. C. Luong, D. T. Hoang, Y . Jiao, Y .-C. Liang, Q. Yang, D. Niyato, and C. Miao, “Federated Learning in Mobile Edge Networks: A Comprehensive Survey,” IEEE Communications Surveys & Tutorials , vol. 22, no. 3, pp. 2031–2063, 2020

    2020

  14. [22]

    Robust Decentralized Federated Learning Using Collaborative Deci- sions,

    A. Gouissem, K. Abualsaud, E. Yaacoub, T. Khattab, and M. Guizani, “Robust Decentralized Federated Learning Using Collaborative Deci- sions,” in 2022 International Wireless Communications and Mobile Computing (IWCMC), pp. 254–258, 2022

    2022

  15. [23]

    Acceler- ating Federated Learning With Cluster Construction and Hierarchical Aggregation,

    Z. Wang, H. Xu, J. Liu, Y . Xu, H. Huang, and Y . Zhao, “Acceler- ating Federated Learning With Cluster Construction and Hierarchical Aggregation,” IEEE Transactions on Mobile Computing , vol. 22, no. 7, pp. 3805–3822, 2023

    2023

  16. [24]

    Resource Efficient Cluster-Based Federated Learning for D2D Communications,

    J.-P. Jung, Y .-B. Ko, and S.-H. Lim, “Resource Efficient Cluster-Based Federated Learning for D2D Communications,” in 2022 IEEE 95th Vehicular Technology Conference: (VTC2022-Spring), pp. 1–5, 2022

    2022

  17. [25]

    Communication-efficient Hierarchical Federated Learning for IoT Heterogeneous Systems with Imbalanced Data,

    A. A. Abdellatif, N. Mhaisen, A. Mohamed, A. Erbad, M. Guizani, Z. Dawy, and W. Nasreddine, “Communication-efficient Hierarchical Federated Learning for IoT Heterogeneous Systems with Imbalanced Data,” Future Generation Computer Systems , vol. 128, pp. 406–419, 2022

    2022

  18. [26]

    Sample-level Data Selection for Federated Learning,

    A. Li, L. Zhang, J. Tan, Y . Qin, J. Wang, and X.-Y . Li, “Sample-level Data Selection for Federated Learning,” in IEEE INFOCOM 2021 - IEEE Conference on Computer Communications , pp. 1–10, 2021

    2021

  19. [27]

    ISFL: Federated Learning for Non-i.i.d. Data with Local Importance Sampling,

    Z. Zhu, P. Fan, C. Peng, and K. B. Letaief, “ISFL: Federated Learning for Non-i.i.d. Data with Local Importance Sampling,” arXiv preprint arXiv:2210.02119, 2023

    2023 arXiv

  20. [28]

    FedDC: Federated Learning with Non-IID Data via Local Drift Decoupling and Correction,

    L. Gao, H. Fu, L. Li, Y . Chen, M. Xu, and C.-Z. Xu, “FedDC: Federated Learning with Non-IID Data via Local Drift Decoupling and Correction,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10102–10111, 2022

    2022

  21. [29]

    Personalized Federated Learning for Intelligent IoT Applications: A Cloud-Edge Based Framework,

    Q. Wu, K. He, and X. Chen, “Personalized Federated Learning for Intelligent IoT Applications: A Cloud-Edge Based Framework,” IEEE Open Journal of the Computer Society , vol. 1, pp. 35–44, 2020

    2020

  22. [30]

    Hybrid-FL for Wireless Networks: Cooperative Learning Mechanism Using Non-IID Data,

    N. Yoshida, T. Nishio, M. Morikura, K. Yamamoto, and R. Yonetani, “Hybrid-FL for Wireless Networks: Cooperative Learning Mechanism Using Non-IID Data,” in ICC 2020 - 2020 IEEE International Confer- ence on Communications (ICC) , pp. 1–7, 2020

    2020

  23. [31]

    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) , pp. 691–706, 2019

    2019

  24. [32]

    Federated Learning with Non-IID Data,

    Y . Zhao, M. Li, L. Lai, N. Suda, D. Civin, and V . Chandra, “Federated Learning with Non-IID Data,” arXiv preprint arXiv:1806.00582 , 2018

    2018 arXiv

  25. [33]

    Feature Matching Data Synthesis for Non-IID Federated Learning,

    Z. Li, Y . Sun, J. Shao, Y . Mao, J. H. Wang, and J. Zhang, “Feature Matching Data Synthesis for Non-IID Federated Learning,” IEEE Trans- actions on Mobile Computing , pp. 1–16, 2024

    2024

  26. [34]

    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 ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 8866–8870, 2020

    2020

  27. [35]

    The Non-IID Data Quagmire of Decentralized Machine Learning,

    K. Hsieh, A. Phanishayee, O. Mutlu, and P. Gibbons, “The Non-IID Data Quagmire of Decentralized Machine Learning,” in Proceedings of the 37th International Conference on Machine Learning (H. D. III and A. Singh, eds.), vol. 119 of Proceedings of Machine Learning Research, pp. ...

    2020

  28. [36]

    Social-Trust and Power- Efficient Relay Selection for Device-to-Device Underlaying Distributed Shared Network,

    M. Sun, N. Bao, J. Zuo, X. Sun, and S. Pan, “Social-Trust and Power- Efficient Relay Selection for Device-to-Device Underlaying Distributed Shared Network,” in 2020 International Conference on Computing, Networking and Communications (ICNC) , pp. 992–996, 2020

    2020

  29. [37]

    Energy-Efficiency Optimization- Based User Selection and Power Allocation for Uplink NOMA-Enabled IoT Networks,

    Y . Wu, S. Liu, X. Lin, and L. Sun, “Energy-Efficiency Optimization- Based User Selection and Power Allocation for Uplink NOMA-Enabled IoT Networks,” in 2023 IEEE 12th International Conference on Educa- tional and Information Technology (ICEIT) , pp. 321–325, 2023

    2023

  30. [38]

    Node Selection Strategy Design Based on Reputation Mechanism for Hierarchical Federated Learning,

    S. Xin, L. Zhuo, and C. Xin, “Node Selection Strategy Design Based on Reputation Mechanism for Hierarchical Federated Learning,” in 2022 18th International Conference on Mobility, Sensing and Networking (MSN), pp. 718–722, 2022

    2022

  31. [39]

    Pytorch Conditional CGAN

    M. Rahmdel, “Pytorch Conditional CGAN.” https://github.com/arturml/ mnist-cgan

  32. [40]

    CIFAKE: Image Classification and Explainable Identification of AI-Generated Synthetic Images,

    J. J. Bird and A. Lotfi, “CIFAKE: Image Classification and Explainable Identification of AI-Generated Synthetic Images,” IEEE Access, vol. 12, pp. 15642–15650, 2024

    2024

  33. [41]

    Risk Neutral is Best for Risky Decision Making,

    R. E. Marks, “Risk Neutral is Best for Risky Decision Making,” tech. rep., mimeo, 2016. https://www.agsm.edu.au/bobm/papers/rnisbest.pdf

    2016

  34. [42]

    Engwerda, LQ Dynamic Optimization and Differential Games

    J. Engwerda, LQ Dynamic Optimization and Differential Games . John Wiley & Sons, 2005

    2005

  35. [43]

    Lyapunov Stability Theory,

    S. Sastry, “Lyapunov Stability Theory,” in Nonlinear Systems: Analysis, Stability, and Control, pp. 182–234, Springer New York, 1999. 13

    1999

  36. [44]

    J.-J. E. Slotine and W. Li, Applied Nonlinear Control, vol. 199. Prentice Hall Englewood Cliffs, NJ, 1991

    1991

  37. [45]

    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 ICC 2019 - 2019 IEEE International Conference on Communications (ICC) , pp. 1–7, 2019

    2019

  38. [46]

    FedMCCS: Mul- ticriteria Client Selection Model for Optimal IoT Federated Learning,

    S. Abdulrahman, H. Tout, A. Mourad, and C. Talhi, “FedMCCS: Mul- ticriteria Client Selection Model for Optimal IoT Federated Learning,” IEEE Internet of Things Journal , vol. 8, no. 6, pp. 4723–4735, 2021

    2021

Pith tools

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