Pith. sign in

REVIEW 5 major objections 7 minor 37 references

PNCS:Power-Norm Cosine Similarity for Diverse Client Selection in Federated Learning

T0 review · 5 major / 7 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 This paper argues that federated learning should select clients with the most complementary gradients, measured by L4-norm cosine similarity, and that this choice speeds convergence and raises accuracy on non-IID data.

desk verdict A simple and honest client-selection heuristic with consistent small gains, but the missing communication-cost accounting and same-benchmark metric tuning keep it from being convincing as-is. read the letter →

arxiv 2506.15923 v1 pith:E337S5VJ submitted 2025-06-18 cs.LG cs.AIcs.DC

Liangyan Li , Yangyi Liu , Yimo Ning , Stefano Rini , Jun Chen This is my paper · ORCID
classification cs.LGcs.AIcs.DC
keywords federatedlearningclientselectionnon-IIDdatagradientdiversitypower-normcosinesimilarityL4normconvergencespeedheterogeneity
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

Federated learning trains on many private datasets, but when those datasets are non-IID, choosing which clients update the model each round matters. This paper argues that the best choices are not the clients whose gradients point in the same direction, but those whose gradients are most complementary, because divergent updates correct each other. It proposes measuring complementarity with a power-norm cosine similarity, specifically the L4-norm version, and combines it with a rotation queue so no client is selected too often. The authors show on CIFAR-10 and Fashion-MNIST with a frozen VGG16 that this selection rule converges faster and to better final accuracy than three existing selection strategies under several heterogeneity levels.

What carries the argument

The central object is the power-norm cosine similarity $\cos_p(g_k,g_{k'}) = \langle g_k,g_{k'}\rangle_p / (\|g_k\|_p \|g_{k'}\|_p)$, where the generalized inner product is $\langle u,v\rangle_p = (\|u+v\|_p - \|u-v\|_p)/4$. The paper uses $p=4$, so the selected subset $S$ is chosen to minimize the average $\cos_4$ over all pairs in $S$. Around this sits the Age-of-Update Queue: selected clients are barred from selection for roughly $L/S$ rounds, forcing rotation. The machinery converts client selection into a diversity-ranking problem and then enforces temporal fairness.

What would settle it

Run PNCS against L2-cosine selection and random selection on CIFAR-10 with the full VGG16 fine-tuned rather than frozen, for 100 rounds, while counting the summary-transmission round in the total communication budget and recording the number of rounds to reach 40% and 60% test accuracy; if the L4 advantage disappears or the extra summary round eats the gain, the central claim is not general.

Watch

Extended reading notes

Core claim

The paper's central claim is that in heterogeneous federated learning, the server should select clients whose local gradients are maximally complementary, not maximally aligned, and that the right way to measure complementarity is the cosine similarity taken under the L4 norm. Concretely, the proposed PNCS policy computes, for every candidate subset of clients, the average of $\cos_4(g_k,g_{k'})$ over all pairs in the subset, where $\cos_p$ is defined through the generalized inner product $\langle u,v\rangle_p = (\|u+v\|_p-\|u-v\|_p)/4$. The server picks the subset with the smallest average value, i.e., the most negative alignment, and an Age-of-Update Queue then keeps selected clients out of contention for a few rounds to force rotation. On CIFAR-10 and Fashion-MNIST with a frozen VGG16 feature extractor and non-IID shard partitions, the paper reports that this policy reaches target accuracy in fewer rounds and settles at higher final accuracy than FedCor, AFL, and Power-of-Choice baselines.

Load-bearing premise

The load-bearing premise is that the L4-norm cosine similarity, chosen from a feature-importance study on CIFAR-10 with a frozen VGG16, one layer, and 15 iterations, remains the best selection metric in other settings, and that the extra summary-transmission round is cheap enough not to cancel the convergence gains.

Editorial extensions

If this is right

  • Under the paper's settings, PNCS reaches 40% test accuracy on CIFAR-10 with 4 selected clients by round 3, while all three baselines need more than 5 rounds.
  • Using a queue length of about 4 improves convergence under high heterogeneity, while length 6 hurts, showing that rotation must be balanced against gradient-based selection.
  • Negative PNCS alignment, meaning choosing the most complementary pair, is consistently the best single predictor of pair performance in the logistic-regression analysis, over cosine similarities with p = 1, 2, and 4.
  • The two-layer variant, computing similarity over layers 3 and 6, maintains or improves accuracy relative to the one-layer variant, while the baselines often degrade when moved to two layers.

Reading between the lines

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

  • Beyond the paper, a direct extension would be to test whether an even higher-order norm, such as p = 6 or 8, improves selection on larger models, since the feature study only compares p = 1, 2, and 4.
  • Beyond the paper, counting the summary-transmission round in total communication could shrink the reported gains, because the GSCCS setting adds one extra transmission before selected clients send full gradients; the paper assumes summaries are cheap but does not report end-to-end communication cost.
  • Beyond the paper, the diversity principle connects to coreset selection and submodular maximization, where the goal is also to pick a subset that represents the whole; average pairwise L4 similarity is one concrete, computable proxy for that goal.
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

5 major / 7 minor

Summary. The paper proposes PNCS, a federated-learning client-selection method in which the server first collects compressed gradient summaries from all clients, then selects a subset that minimizes the average L4-norm cosine similarity (interpreted as maximizing gradient diversity), while an Age-of-Update Queue prevents immediate reselection. The method is motivated by a logistic-regression analysis of pairwise gradient features on a small CIFAR-10 task, and the main experiments compare PNCS against AFL, FedCor, and Power-of-Choice on CIFAR-10 and Fashion-MNIST with a frozen VGG16 feature extractor. The authors claim that PNCS improves both convergence speed and final accuracy, and they frame the work as addressing communication cost.

Significance. If the central claims were fully supported, the contribution would be useful: a simple, interpretable selection heuristic with a plausible diversity criterion and a queue-based rotation mechanism, backed by comparisons to three established baselines on two datasets. The paper's strengths include a systematic feature-importance exploration, a clearly specified selection protocol, and an honest statement of some limitations in Section IV-A. However, the validation currently has load-bearing gaps: the communication cost of the summary phase is never accounted for, the choice of L4 is made on the same benchmark used for evaluation, and the queue length is tuned post hoc on those same experiments. These issues prevent the manuscript from supporting its claimed communication-efficiency advantage and 'consistent improvements' with current evidence.

major comments (5)
  1. [Section IV and Section VI-B-1] The GSCCS setting adds a summary-transmission round that is never included in any comparison. The research question in Section I explicitly asks about communication cost, and Section VI-B-1 is headed 'Test accuracy versus communication,' but the reported results (Table II and Figures 2-3) plot accuracy against iteration number, not against transmitted bytes. Section IV-A states that the sketch size r is left for future work and that the framework 'assumes summaries are accurate.' As written, the evidence does not rule out that PNCS is strictly worse than baselines when measured as accuracy per total transmitted bits, since baselines transmit only J full gradients while PNCS transmits K summaries plus J full gradients per round. This is an internally acknowledged open issue and is load-bearing for the paper's communication-efficiency claim.
  2. [Section V-B and Section VI] The choice of L4-norm cosine similarity as the best selection feature is derived from an analysis on CIFAR-10 with a frozen VGG16, a single classifier layer, and only 15 iterations (Section V-B). The main evaluation then uses the same dataset and the same frozen feature extractor to demonstrate gains. This creates a circularity: the feature is selected on the same benchmark used for the headline result. The Fashion-MNIST experiment (Figure 3) provides partial independent support, but only for one heterogeneity level and one selection size. The paper needs either a theoretical justification for L4 or a held-out validation (different dataset/architecture/training horizon) before claiming that PNCS 'consistently' improves over baselines.
  3. [Section VI-C and Figure 1] The queue length L=4 is selected by inspecting Figure 1, which uses the same CIFAR-10 experiments that later produce the main results in Table II and Figure 2. This post hoc tuning can inflate the apparent benefit of the queue. The paper should report queue-length sensitivity on a separate validation split or otherwise demonstrate that the improvement is not an artifact of this particular choice. In addition, the notation is confusing: the queue parameter is called L in Section V-C and Q in Figure 1 and Section VI-C, and the formula t' > t + L/S mixes the queue length with the selected subset size S, which is itself also used for the heterogeneity shard count.
  4. [Section V-A, Equation (10)] For p ≠ 2, the quantity ⟨u, v⟩_p = ( ||u+v||_p − ||u−v||_p ) / 4 is not a bilinear inner product, so cosp is not a cosine similarity in the standard sense. The paper does not characterize the range of cosp, its behavior under scaling, or whether the 'negative alignment' interpretation remains valid for p=4. Since PNCS is built on this quantity, the authors should either justify the terminology, prove relevant properties (e.g., boundedness, monotonicity), or provide an empirical sanity check such as the distribution of cosp values and a comparison with the standard L2 cosine similarity.
  5. [Section VI-A and Table II] The text states that each experiment is repeated with 10 random seeds, but the reported numbers are point estimates without error bars or confidence intervals. The 'Std' row in Table II appears to be the standard deviation across shard/iteration settings, not across random seeds. The claim that PNCS 'consistently outperforms' the baselines requires variance information across seeds, especially for the small margins observed in some settings (e.g., Shard 5 in Table II).
minor comments (7)
  1. [Section VI-B-1] The text refers to 'Table ??' as a placeholder; the actual table is Table II, and the header 'Test accuracy versus communication' is misleading because no communication-cost plot is provided.
  2. [Table II and Section VI-A] Table II's caption says 'Layers 1/2' while Section VI-A describes one-layer (layer 6) and two-layer (layers 3 and 6) configurations; these references should be reconciled.
  3. [Section III-A and Section V-C] The symbol S is overloaded: it denotes both the set of selected clients and its cardinality, and later the number of shards (heterogeneity level). This makes statements such as '|S| = S' and Figure 2's 'S=4' confusing; please use distinct symbols.
  4. [Section V-A] The symbol J is used both for the number of shards in the feature-analysis setup (Section V-B: J = {1,2,5}) and for the number of selected clients in Eq. (5); this reuse should be avoided.
  5. [References] References [19] and [27] appear irrelevant to the topics they are cited for: [19] concerns head-and-neck cancer treatment outcomes, and [27] concerns COVID-19 research synthesis. These are likely citation errors and should be replaced with the intended Hungarian-matching and reputation-based client-selection works.
  6. [Section V-B] The sentence 'The features we consider are quite extensive, such as: The features we consider include:' contains a repeated phrase and should be rewritten.
  7. [Section II] The method is called 'POWD' in related work but 'PoC' (Power-of-Choice) in the experiments and Figure captions; please standardize the name.

Circularity Check

2 steps flagged · score 6.0 of 10

L4 cosine metric and queue length are selected using CIFAR-10 and then used to claim CIFAR-10 gains, making part of the central improvement fitted rather than independently predicted.

  1. fitted input called prediction [Section V-B (Numerical Findings) and Section VI-B-1 (CIFAR-10 results)]
    "After careful experimentation, we conclude that the cosp(g, g′) for p ∈ [4] provides the most accurate and robust user selection performance. ... Table ?? demonstrates that our PNCS method consistently outperforms the baseline methods on CIFAR-10 under different data heterogeneity (Shard = 1, 2, 5)."

    The PNCS metric (L4-norm cosine similarity) is selected by fitting a logistic regression to accuracy labels from CIFAR-10/VGG training runs (Section V-B), and then the same CIFAR-10 benchmark is used to demonstrate that PNCS outperforms baselines. The reported CIFAR-10 gain therefore partially inherits the feature-selection fit on the same dataset; it is not an out-of-sample prediction for CIFAR-10. The Fashion-MNIST result is independent, but the headline CIFAR-10 claim is statistically forced to some degree.

  2. fitted input called prediction [Section V-C (AoU-Queue) and Section VI-C (Ablation Study on Queue)]
    "Figure 1 shows that a queue length of Q=4 optimally balances stability and adaptability, achieving faster convergence in heterogeneous data (small shards) and slightly improving early convergence in homogeneous data (large shards), while longer queues (Q=6) underperform due to diminishing returns."

    The queue length L is tuned via an ablation study on CIFAR-10 (Fig. 1) and then the CIFAR-10 experiments in Table II and Figures 2 use this tuned value. Thus the hyperparameter is selected on the same benchmark that is later reported as evidence of PNCS's effectiveness. This tuning-on-test-set step makes part of the CIFAR-10 advantage a self-fulfilling choice rather than an independent validation.

full rationale

The paper's derivation chain is empirical rather than theoretical: Section V-B performs a feature-importance analysis on CIFAR-10 to identify L4-norm cosine similarity as the best single feature, and Section V-C introduces an Age-of-Update queue whose length is later chosen by a CIFAR-10 ablation. The main CIFAR-10 comparisons with FedCor, AFL, and PoC are then presented as evidence that PNCS 'significantly improves convergence speed and final model accuracy.' Because the metric and queue length were both selected using CIFAR-10, the CIFAR-10 portion of the headline claim is partly fitted, not independently predicted. This is a genuine circularity of the fitted-input-called-prediction type. However, the paper does include an external benchmark (Fashion-MNIST, Figure 3) where PNCS also outperforms the baselines, and the comparisons against multiple baselines provide some independent content. There is no self-citation chain, no imported uniqueness theorem, and no formal equation reducing the result to its inputs. The communication-cost gap (summaries' dimension and extra transmission round are not counted) is a reporting/completeness concern, not a circularity, and is excluded from this score. Overall partial circularity: score 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method relies on the choice of norm order and queue length as fitted hyperparameters, and on several assumptions about the transferability of the feature analysis and the reliability of gradient sketches. No new physical entities are introduced.

free parameters (3)
  • Norm order p = 4
    Chosen as the most predictive feature in the logistic regression analysis on CIFAR-10 (Section V-B); not derived from first principles.
  • Queue length L = 4
    Selected via ablation on CIFAR-10 (Figure 1) as the best performing length; tuned on the same dataset used for evaluation.
  • Sketch dimension r
    Assumed large enough to accurately represent gradients, but not specified or tuned (Section IV-A).
assumptions (4)
  • domain assumption The polarization identity in Eq. (10) defines a meaningful similarity measure for p=4 even though it may not be positive semidefinite for p>2.
    The method relies on cos4 as a signed similarity; the paper does not analyze its properties.
  • ad hoc to paper Selecting clients with highly negative PNCS (complementary gradients) improves global model updates.
    This is an empirical observation from the feature analysis (Section V-B), not a proven result.
  • domain assumption Gradient summaries transmitted by clients are accurate enough for meaningful selection.
    The paper assumes this without analyzing the effect of sketch size (Section IV-A).
  • ad hoc to paper Logistic regression labels based on 15-iteration, single-layer fine-tuning on CIFAR-10 predict which features are best for the full multi-layer training.
    The paper extrapolates from a short single-layer experiment to the full setting without evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PNCS:Power-Norm Cosine Similarity for Diverse Client Selection in Federated Learning." pith.science (2026). https://pith.science/paper/E337S5VJ

@misc{pith2026250615923,
  author       = {Pith},
  title        = {Pith review of: PNCS:Power-Norm Cosine Similarity for Diverse Client Selection in Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E337S5VJ}},
  note         = {Machine review of arXiv:2506.15923}
}
read the original abstract

Federated Learning (FL) has emerged as a powerful paradigm for leveraging diverse datasets from multiple sources while preserving data privacy by avoiding centralized storage. However, many existing approaches fail to account for the intricate gradient correlations between remote clients, a limitation that becomes especially problematic in data heterogeneity scenarios. In this work, we propose a novel FL framework utilizing Power-Norm Cosine Similarity (PNCS) to improve client selection for model aggregation. By capturing higher-order gradient moments, PNCS addresses non-IID data challenges, enhancing convergence speed and accuracy. Additionally, we introduce a simple algorithm ensuring diverse client selection through a selection history queue. Experiments with a VGG16 model across varied data partitions demonstrate consistent improvements over state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2506.15923 by the authors.

Figure 1
Figure 1. Comparison of PNCS with different queue lengths [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Comparison with baselines on the CIFAR-10 dataset [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparison with baselines on the Fashion-MNIST [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 28 canonical work pages

  1. [19]

    Pre-treatment quality of life as a predictor of distant metastasis- free survival and overall survival in patients with head and neck cancer who underwent free flap reconstruction,

    M. Chen, K. Ho, Y . Hung, C. Su, C. Kuan, H. Tai, N. Cheng, and C. Lin, “Pre-treatment quality of life as a predictor of distant metastasis- free survival and overall survival in patients with head and neck cancer who underwent free flap reconstruction,” European Journal of Oncology Nursing, vol. 41, pp. 1–6, 2019

    work page 2019

  2. [1]

    Federated learning: Collaborative ma- chine learning without centralized training data,

    B. McMahan and D. Ramage, “Federated learning: Collaborative ma- chine learning without centralized training data,” Google Research Blog, vol. 3, 2017

    work page 2017

  3. [2]

    On the convergence of fedavg on non-iid data,

    X. Li, K. Huang, W. Yang, S. Wang, and Z. Zhang, “On the convergence of fedavg on non-iid data,” 2020. [Online]. Available: https://arxiv.org/abs/1907.02189

    arXiv 2020

  4. [3]

    Federated learning: Strategies for improving communica- tion efficiency,

    J. Kone ˇcn`y, “Federated learning: Strategies for improving communica- tion efficiency,” arXiv preprint arXiv:1610.05492 , 2016

    arXiv 2016

  5. [4]

    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 . PMLR, 2017, pp. 1273– 1282

    2017

  6. [5]

    A state-of-the-art survey on solving non-iid data in federated learning,

    X. Ma, J. Zhu, Z. Lin, S. Chen, and Y . Qin, “A state-of-the-art survey on solving non-iid data in federated learning,” Future Generation Computer Systems, vol. 135, pp. 244–258, 2022

    work page 2022

  7. [6]

    A comprehensive survey on client selection strategies in federated learning,

    J. Li, T. Chen, and S. Teng, “A comprehensive survey on client selection strategies in federated learning,” Computer Networks, p. 110663, 2024

    2024

  8. [7]

    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 15th {USENIX} Symposium on Operating Systems Design and Implementation ( {OSDI} 21), 2021, pp. 19–35

    work page 2021

Show all 37 references
  1. [8]

    Eiffel: Efficient and fair scheduling in adaptive federated learning,

    A. Sultana, M. M. Haque, L. Chen, F. Xu, and X. Yuan, “Eiffel: Efficient and fair scheduling in adaptive federated learning,” IEEE Transactions on Parallel and Distributed Systems , vol. 33, no. 12, pp. 4282–4294, 2022

    2022

  2. [9]

    Diverse client selection for federated learning via submodular max- imization,

    R. Balakrishnan, T. Li, T. Zhou, N. Himayat, V . Smith, and J. Bilmes, “Diverse client selection for federated learning via submodular max- imization,” in International Conference on Learning Representations , 2022

    2022

  3. [10]

    Fast-convergent federated learning,

    H. T. Nguyen, V . Sehwag, S. Hosseinalipour, C. G. Brinton, M. Chiang, and H. V . Poor, “Fast-convergent federated learning,” IEEE Journal on Selected Areas in Communications , vol. 39, no. 1, pp. 201–218, 2020

    2020

  4. [11]

    A pre- liminary study of improving evolutionary multi-objective optimization via knowledge transfer from single-objective problems,

    L. Huang, L. Feng, H. Wang, Y . Hou, K. Liu, and C. Chen, “A pre- liminary study of improving evolutionary multi-objective optimization via knowledge transfer from single-objective problems,” in 2020 IEEE International Conference on Systems, Man, and Cybernetics (SMC) , 2020, ...

    2020

  5. [12]

    Fairness-aware client selection for federated learning,

    Y . Shi, Z. Liu, Z. Shi, and H. Yu, “Fairness-aware client selection for federated learning,” in 2023 IEEE International Conference on Multimedia and Expo (ICME) , 2023, pp. 324–329

    2023

  6. [13]

    Lyapunov- based optimization of edge resources for energy-efficient adaptive fed- erated learning,

    C. Battiloro, P. Di Lorenzo, M. Merluzzi, and S. Barbarossa, “Lyapunov- based optimization of edge resources for energy-efficient adaptive fed- erated learning,” IEEE Transactions on Green Communications and Networking, vol. 7, no. 1, pp. 265–280, 2022

    2022

  7. [14]

    Joint client selection and cpu frequency control in wireless federated learning networks with power constraints,

    Z. Zhou, S. Shi, F. Wang, Y . Zhang, and Y . Li, “Joint client selection and cpu frequency control in wireless federated learning networks with power constraints,” Entropy, vol. 25, no. 8, p. 1183, 2023

    2023

  8. [15]

    A greedy agglomerative framework for clus- tered federated learning,

    M. Mehta and C. Shao, “A greedy agglomerative framework for clus- tered federated learning,” IEEE Transactions on Industrial Informatics , vol. 19, no. 12, pp. 11 856–11 867, 2023

    2023

  9. [16]

    Dynamic federated learning for gmec with time-varying wireless link,

    S. Zhai, X. Jin, L. Wei, H. Luo, and M. Cao, “Dynamic federated learning for gmec with time-varying wireless link,” IEEE Access, vol. 9, pp. 10 400–10 412, 2021

    2021

  10. [17]

    Mflces: Multi-level federated edge learning algorithm based on client and edge server selection,

    Z. Liu, S. Duan, S. Wang, Y . Liu, and X. Li, “Mflces: Multi-level federated edge learning algorithm based on client and edge server selection,” Electronics, vol. 12, no. 12, p. 2689, 2023

    2023

  11. [18]

    Budgeted online selection of candidate iot clients to participate in federated learning,

    I. Mohammed, S. Tabatabai, A. Al-Fuqaha, F. El Bouanani, J. Qadir, B. Qolomany, and M. Guizani, “Budgeted online selection of candidate iot clients to participate in federated learning,” IEEE Internet of Things Journal, vol. 8, no. 7, pp. 5938–5952, 2020

    2020

  12. [20]

    Deep reinforcement learning assisted federated learning algorithm for data management of iiot,

    P. Zhang, C. Wang, C. Jiang, and Z. Han, “Deep reinforcement learning assisted federated learning algorithm for data management of iiot,” IEEE Transactions on Industrial Informatics , vol. 17, no. 12, pp. 8475–8484, 2021

    2021

  13. [21]

    Deep-reinforcement-learning-assisted client selection in nonorthogonal- multiple-access-based federated learning,

    R. Albelaihi, A. Alasandagutti, L. Yu, J. Yao, and X. Sun, “Deep-reinforcement-learning-assisted client selection in nonorthogonal- multiple-access-based federated learning,” IEEE Internet of Things Jour- nal, vol. 10, no. 17, pp. 15 515–15 525, 2023

    2023

  14. [22]

    Deep reinforcement learning: A brief survey,

    K. Arulkumaran, M. P. Deisenroth, M. Brundage, and A. A. Bharath, “Deep reinforcement learning: A brief survey,” IEEE Signal Processing Magazine, vol. 34, no. 6, pp. 26–38, 2017

    2017

  15. [23]

    A buffered semi-asynchronous mechanism with mab for efficient federated learning,

    C. Wang, Q. Wu, Q. Ma, and X. Chen, “A buffered semi-asynchronous mechanism with mab for efficient federated learning,” in 2022 Inter- national Conference on High Performance Big Data and Intelligent Systems (HDIS), 2022, pp. 180–184

    2022

  16. [24]

    Online client selection for asynchronous federated learning with fairness considera- tion,

    H. Zhu, Y . Zhou, H. Qian, Y . Shi, X. Chen, and Y . Yang, “Online client selection for asynchronous federated learning with fairness considera- tion,” IEEE Transactions on Wireless Communications , vol. 22, no. 4, pp. 2493–2506, 2022

    2022

  17. [25]

    A novel reputation-aware client selection scheme for federated learning within mobile environments,

    Y . Wang and B. Kantarci, “A novel reputation-aware client selection scheme for federated learning within mobile environments,” in 2020 IEEE 25th International Workshop on Computer Aided Modeling and Design of Communication Links and Networks (CAMAD) . IEEE, 2020, pp. 1–6

    2020

  18. [26]

    Block-racs: Towards reputation- aware client selection and monetization mechanism for federated learn- ing,

    Z. Batool, K. Zhang, and M. Toews, “Block-racs: Towards reputation- aware client selection and monetization mechanism for federated learn- ing,” ACM SIGAPP Applied Computing Review , vol. 23, no. 3, pp. 49– 65, 2023

    2023

  19. [27]

    The collective wisdom in the covid-19 research: Comparison and synthesis of epidemi- ological parameter estimates in preprints and peer-reviewed articles,

    Y . Wang, Z. Cao, D. D. Zeng, Q. Zhang, and T. Luo, “The collective wisdom in the covid-19 research: Comparison and synthesis of epidemi- ological parameter estimates in preprints and peer-reviewed articles,” International Journal of Infectious Diseases , vol. 104, pp. 1–6, 2021

    2021

  20. [28]

    Reputation-aware federated learning client selection based on stochastic integer programming,

    X. Tan, W. C. Ng, W. Y . B. Lim, Z. Xiong, D. Niyato, and H. Yu, “Reputation-aware federated learning client selection based on stochastic integer programming,” IEEE Transactions on Big Data , 2022

    2022

  21. [29]

    Active federated learning,

    J. Goetz, K. Malik, D. Bui, S. Moon, H. Liu, and A. Kumar, “Active federated learning,” 2019. [Online]. Available: https://arxiv.org/abs/1909.12641

    2019 arXiv

  22. [30]

    Towards understanding biased client selection in federated learning,

    Y . Jee Cho, J. Wang, and G. Joshi, “Towards understanding biased client selection in federated learning,” in Proceedings of The 25th International Conference on Artificial Intelligence and Statistics , ser. Proceedings of Machine Learning Research, G. Camps-Valls, F. J. R. Ru...

    2022

  23. [31]

    Client selection in federated learning based on gradients importance,

    O. Marnissi, H. E. Hammouti, and E. H. Bergou, “Client selection in federated learning based on gradients importance,” in AIP Conference Proceedings, vol. 3034, no. 1. AIP Publishing, 2024

    2024

  24. [32]

    Gradient coreset for federated learning,

    D. Sivasubramanian, L. Nagalapatti, R. Iyer, and G. Ramakrish- nan, “Gradient coreset for federated learning,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2024, pp. 2648–2657

    2024

  25. [33]

    Federated learning with client selection and gradient compression in heterogeneous edge systems,

    Y . Xu, Z. Jiang, H. Xu, Z. Wang, C. Qian, and C. Qiao, “Federated learning with client selection and gradient compression in heterogeneous edge systems,” IEEE Transactions on Mobile Computing , 2023

    2023

  26. [34]

    Fedcor: Correlation-based active client selection strategy for heteroge- neous federated learning,

    M. Tang, X. Ning, Y . Wang, J. Sun, Y . Wang, H. Li, and Y . Chen, “Fedcor: Correlation-based active client selection strategy for heteroge- neous federated learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 10 102–10 111

    2022

  27. [35]

    Cifar-10 (canadian institute for advanced research),

    A. Krizhevsky, V . Nair, and G. Hinton, “Cifar-10 (canadian institute for advanced research),” 2009. [Online]. Available: http://www.cs.toronto.edu/ kriz/cifar.html

    2009

  28. [36]

    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

    2017

  29. [37]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” 2015. [Online]. Available: https://arxiv.org/abs/1409.1556

    2015 arXiv

Pith tools

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