Pith. sign in

REVIEW 4 major objections 5 minor 39 references

A Drift Stable Quantum Federated Learning for Intelligent Services

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A quantum federated learning framework that replaces fixed local optimizers with controller-tuned, per-step SPSA updates and a proximal drift penalty improves stability, generalization, and client fairness on fraud and genome benchmarks.

desk verdict A genuinely new recipe for drift-stable quantum federated learning, but single-seed point estimates and a partly circular meta-objective mean the central claim is not yet load-bearing. read the letter →

arxiv 2607.21647 v1 pith:BPDWHI6L submitted 2026-07-22 cs.LG cs.AI

Shanika Iroshi Nanayakkara , Shiva Raj Pokhrel This is my paper
classification cs.LGcs.AI
keywords quantumfederatedlearningclientdriftdeepunfoldingSPSAproximalregularizationnon-IIDdatafairnessvariationalcircuits
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that the usual way of training quantum neural networks in a federated setting — every client running the same fixed SPSA optimizer schedule — is the real source of client drift, instability, and unfairness. Its proposed framework, DUQFL-Prox, turns local training into a deep-unfolded, controller-guided sequence of SPSA steps: a shared controller picks the learning rate and perturbation scale at each step from the client's optimization state, a proximal term pulls local parameters toward the global model, and each client uploads the unfold step that best validates locally. On financial-fraud and genomic-classification benchmarks, this combination reports stronger mean client test accuracy, a near-zero train-test gap, and a much smaller client fairness gap than fixed-optimizer quantum federated baselines. The paper establishes that local QNN optimization is a learnable trajectory, not a fixed routine, and that learning it improves aggregation-compatible updates.

What carries the argument

The load-bearing object is the 'unfolded proximal SPSA block': a single SPSA update applied to the proximal objective Li(θ) + μ/2||θ − θ^(t)||², with learning rate η and perturbation scale δ supplied by a shared controller. The controller is a clipped log-linear map from an optimization-state feature vector (unfold progress, round progress, recent loss, displacement, data fraction, heterogeneity) to (η, δ). After K blocks, each client uploads the checkpoint with the lowest local validation loss rather than the last state, and the controller is periodically updated by an outer SPSA finite-difference step on a validation-based meta-loss that includes fairness, communication, and stability pena

What would settle it

On the BAF benchmark, rerun DUQFL-Prox with local validation splits randomly reduced to, say, ten samples per client (or with the controller updated using training loss instead of validation loss). If the near-zero train-test gap and the fairness gap (0.0193) revert to the FedProx-QFL/DEFAULT-QFL ranges, the reported advantage is driven by validation-proxy quality, not by the unfolded optimizer. A second check: compute the correlation between the selected unfold index k* and the index of the truly best local test checkpoint; chance-level correlation would falsify the checkpoint-selection mecha

Watch

Extended reading notes

Core claim

The central discovery is that deep-unfolded local optimization with a proximal drift penalty and validation-based checkpoint selection makes quantum federated learning drift-stable under heterogeneous, non-IID client data. Concretely, the paper shows that replacing a fixed local SPSA schedule with controller-generated per-unfold hyperparameters — a bilevel design where the controller is itself refined by an outer SPSA step against post-aggregation validation loss — yields the best client-level generalization and the lowest fairness gap on both evaluated tasks, even though on the genome task a baseline still wins on final global accuracy. In the paper's own framing, DUQFL-Prox is not an accur

Load-bearing premise

Each client's local validation split is a reliable mirror of that client's test distribution, so the validation-selected checkpoint and the validation-driven meta-loss are not being steered by noise; the paper itself notes the BAF validation curves are noisier than test curves because the validation split has very few positive fraud samples.

Editorial extensions

If this is right

  • Quantum federated learning can move away from fixed optimizer schedules: controller-guided per-step hyperparameters are learnable and shared across clients, adding no communication overhead beyond the upload of validation-selected checkpoints.
  • If the near-zero train-test gap holds, the aggregated model transfers to individual client distributions far better than fixed-optimizer QFL, directly addressing the over-specialization that plagues non-IID federated learning.
  • The fairness-gap reductions imply that minority or hard clients are not sacrificed for global accuracy, which matters for fraud detection and genomics where rare patterns are the point.
  • On the genome task, trading a small drop in final global accuracy for a large improvement in client-level generalization suggests that QFL evaluation should be reported as a profile of client-level metrics, not just one global number.

Reading between the lines

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

  • I would predict that the controller's advantage is strongest when clients have similar validation-set sizes; with highly skewed or tiny validation splits, the best-unfold selection should degrade toward the baseline, which the paper's own BAF validation-noise observation hints at.
  • A cheap, testable extension would be to replace the hand-built feature vector with a learned summary of the loss trajectory; if the framework's robustness comes from the controller rather than the features, the two should behave similarly after enough rounds.
  • Because the outer loop estimates gradients by running two perturbed federated rounds, per-round cost roughly triples; a surrogate meta-loss that avoids full federated rollouts would be needed before the method scales to hundreds of clients.
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 / 5 minor

Summary. The paper proposes DUQFL-Prox, a quantum federated learning (QFL) framework in which each client runs a deep-unfolded SPSA trajectory whose per-step learning rate and perturbation scale are generated by a shared controller. A proximal term penalizes deviation from the broadcast global model, and each client uploads the validation-preferred unfolded checkpoint rather than the final iterate. The controller is periodically refined by an outer SPSA step using a post-aggregation meta-loss that includes validation loss, fairness, communication, and stability terms. The framework is evaluated on a bank-fraud (BAF) dataset and a genomic-classification (Genome) task, with claims of improved global accuracy, mean client test accuracy, generalization, and client fairness over Default-QFL, FedProx-QFL, and ablation variants.

Significance. If the empirical claims hold, the paper makes a useful contribution to QFL by moving away from fixed local optimizers and providing a mechanism to reduce client drift in quantum settings. The ablation structure (DUQFL-last, DUQFL-best, DUQFL-drift, DUQFL-Prox) is well designed and isolates the contributions of checkpoint selection and drift control. The trace-logging protocol and the supplementary hardware validation on IBM quantum hardware are also valuable strengths. However, the central empirical evidence is currently based on single fixed-seed runs of a triply stochastic pipeline, with no variance estimates, and the validation-based meta-objective partially determines the fairness and stability numbers that are later reported as findings. The significance of the contribution is therefore conditional on strengthening the experimental evidence and clarifying the role of the validation signal.

major comments (4)
  1. [Section V-A, Figs. 3 and 6] All headline results are single-run point estimates. The pipeline is stochastic at three nested levels: finite-shot QNN measurements (Table II, shots=1024), client-side SPSA perturbations (Eq. (10)), and the outer Rademacher SPSA controller update (Eq. (31)). With one trajectory, the reported margins (BAF global accuracy 0.6504 vs 0.5344, fairness gap 0.0193 vs 0.0958; Genome mean client accuracy ~0.80 vs ~0.79) cannot be distinguished from random variation. Please report mean and standard deviation (or confidence intervals) over multiple seeds for each method and dataset, and state whether differences in global accuracy, mean client accuracy, train–test gap, and fairness gap are statistically significant. This is load-bearing for the paper's central claim.
  2. [Section III-I, Eq. (29); Section V-B] The outer meta-loss explicitly includes λ_fair Ω_fair and λ_stab Ω_stab, and the paper then reports a low fairness gap and stability as empirical findings. Since the controller is trained to minimize these quantities on validation data, the comparison is partly by construction. To separate the algorithmic benefit from the objective design, report results for a variant with λ_fair = λ_stab = 0, or restrict the fairness/stability claims to test-set metrics and state the caveat explicitly. Relatedly, the same validation data are used both to select the uploaded checkpoint via Eq. (11) and to update the controller via Eq. (29); the paper itself concedes in Section V-B that the BAF validation split contains very few positive fraud samples and yields noisier validation curves than test curves. The near-zero train–test gap and low fairness gap could therefore be artifacts of validation noise ra
  3. [Section VI and Appendix C] The conclusion states that DUQFL-Prox is supported by 'three theoretical observations' and refers to 'Detailed statements and proofs provided in Appendix A.' However, Appendix A contains only elementary algebra and no theorem, and Appendix C explicitly disclaims a convergence guarantee. In particular, the third observation—that, under smoothness, the outer SPSA update provides a descent direction in expectation—is asserted without a formal statement or proof. Please either provide a precise theorem with assumptions and a proof, or remove the theoretical-support claim and present Section VI as a qualitative stability discussion. The current cross-reference to proofs that do not exist is misleading.
  4. [Abstract and Section V-C] The abstract's broad statement that DUQFL-Prox 'improves stability, generalization, and client fairness compared with standard QFL baselines' is not fully supported by the Genome results, where FedProx-QFL achieves the highest final global accuracy (~0.85 vs ~0.82–0.83). The body of the paper does qualify this in Section V-C, but the abstract and the cross-dataset summary in Section V-D should make explicit that the Genome advantage is at the client-level generalization/fairness level rather than global accuracy. Without this caveat, the central claim overstates the empirical picture.
minor comments (5)
  1. [Section I] The organization paragraph contains a placeholder 'Section??' for the discussion section; this should be fixed before final submission.
  2. [Table III, Section IV-C] The Genome dataset description states 10,000 training and 2,000 test records but does not specify a validation split. Since Eq. (11) and Eq. (29) rely on local and global validation data, the experimental protocol should state how validation sets were constructed for Genome.
  3. [Figures 4 and 5] The captions are inconsistent with the text: Figure 4 is captioned as 'Epoch-wise ROC-AUC, PR-AUC, MCC, and specificity' while Figure 5 is captioned simply 'BAF Data.' Please align captions with the metrics actually plotted and the referencing in Section V-B.
  4. [Table I] The table contains formatting artifacts, e.g., '✓ indicates' with a missing space and the use of '△' and '×' symbols. Clean up the table and ensure all symbols render consistently.
  5. [Section IV-A] The reproducibility protocol is described in detail, but no code repository or data access link is provided. If the trace logs and experiment configuration are intended to support reproducibility, a public artifact link should be included.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported test-set results are not fit targets, and no equation-level identity between the optimized meta-loss and the reported fairness/stability metrics is shown.

full rationale

The central empirical claim is that DUQFL-Prox improves test accuracy, train–test gap, and client fairness gap relative to baselines. Those numbers are computed on test splits (BAF: 1,250 test samples; Genome: 2,000 test records) that are not used by the local objectives (Eq. 8), the validation-based checkpoint selection (Eq. 11, which uses local validation sets V_i), or the outer meta-loss (Eq. 29, which uses D_val). The controller is indeed fitted to validation behavior, so a low validation fairness or stability value would be partly an optimized objective rather than an emergent finding. However, the paper's reported fairness gap is explicitly a test-side P90−P10 spread, and Ω_fair and Ω_stab are not defined in the text as identical to that test metric or to the reported train–test gap, so no by-construction reduction can be exhibited. The near-zero train–test gap is mechanistically explained by validation-based best-unfold checkpoint selection, which is a legitimate model-selection strategy rather than a fit to the test set. Section V-B concedes that the BAF validation split contains very few positive fraud samples and produces noisy validation curves; this is a reliability limitation for checkpoint selection, but it does not make the test comparison circular. Similarly, Section VI and Appendix C explicitly disclaim a general convergence theorem, which is a support weakness but not circularity. The self-citations in the paper ([25], [31]) appear in related-work context and are not load-bearing: the derivation of DUQFL-Prox follows from the paper's own equations in Sections III-D through III-I. Under the rule that circularity must be demonstrated by a quotable reduction (e.g., Eq. X = Eq. Y by construction, or a fitted parameter renamed as a prediction), no significant circularity is found.

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

DUQFL-Prox introduces no new physical entities. Its free parameters are hand-set or unreported optimization hyperparameters (μ, initial lr/perturbation, meta-loss weights, outer SPSA settings, clipping bounds), plus unspecified feature-engineering choices for the controller. The key assumptions are the reliability of local validation sets, the smoothness of the meta-objective for outer SPSA, and the validity of the simulator as a stand-in for hardware. These are load-bearing because the entire stability/fairness advantage is measured through validation-guided selection and meta-optimization.

free parameters (6)
  • Proximal coefficient μ = 1e-2
    Hand-set in Table II; controls the strength of the drift penalty in Eq. (19) and directly shapes all local trajectories.
  • Initial SPSA learning rate and initial perturbation = 0.13
    Table II; hand-chosen initialization for the local optimizer/controller.
  • Meta-loss weights λ_fair, λ_comm, λ_stab = not reported
    Eq. (29) defines the outer objective with three weighted terms, but the weights are never specified; the controller update is therefore not fully determined.
  • Outer SPSA hyperparameters α_out and c_out = not reported
    Algorithm 3 requires an outer learning rate and perturbation radius; neither value is reported, so the controller adaptation step is underspecified.
  • Clipping bounds η_min, η_max, δ_min, δ_max = not reported
    Equations (17)-(18) clip the generated learning rate and perturbation scale; the bounds are not given, but they constrain the optimizer behavior.
  • Controller feature vector z and heterogeneity indicator = not specified
    Section III.E says z includes 'a client heterogeneity indicator' but no formula is given; this is a modeling choice that the controller depends on.
assumptions (5)
  • domain assumption Two-evaluation SPSA provides a usable descent direction for local QNN loss under finite-shot measurement noise
    The whole method builds on SPSA as the inner optimizer; no convergence or bias analysis for the finite-shot quantum setting is provided.
  • domain assumption Each client has a local validation set V_i that is representative of that client's test distribution
    Best-unfold selection in Eq. (11) and the meta-objective in Eq. (29) rely on validation losses; the paper itself notes the BAF validation split has very few positive samples.
  • ad hoc to paper The post-aggregation meta-objective is smooth enough that outer SPSA provides a descent direction in expectation
    Appendix B only derives the SPSA finite-difference formula; the smoothness assumption is asserted in the conclusion but never verified or proven.
  • ad hoc to paper The unfolded checkpoint with minimum local validation loss is the best model to upload for aggregation
    This is the core design choice of DUQFL-Prox; it is tested indirectly through ablations but is assumed as a general principle in Section III.G.
  • domain assumption Qiskit Aer simulator results are a faithful proxy for algorithm comparison on real quantum hardware
    All training comparisons are simulator-based; the hardware experiment is only post-training checkpoint execution on ibm_fez.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Drift Stable Quantum Federated Learning for Intelligent Services." pith.science (2026). https://pith.science/paper/BPDWHI6L

@misc{pith2026260721647,
  author       = {Pith},
  title        = {Pith review of: A Drift Stable Quantum Federated Learning for Intelligent Services},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BPDWHI6L}},
  note         = {Machine review of arXiv:2607.21647}
}
read the original abstract

Quantum federated learning enables distributed clients to train quantum neural networks without sharing local data, making it promising for privacy-aware intelligent services. Intelligent services in this context refer to privacy-sensitive distributed decision systems, such as fraud detection and genomic classification, where reliable and fair client-level learning is as important as the accuracy of the aggregate model. However, heterogeneous client data and noisy quantum optimization often cause unstable local updates, client drift, and unfair performance between clients. This paper proposes DUQFL-Prox, a drift-stable quantum federated learning framework based on deep-unfolded local optimization. Instead of using a fixed local optimizer, each client performs adaptive unfolded SPSA updates, while a proximal term keeps the local model close to the global model. A lightweight controller learns step-specific optimization parameters to improve post-aggregation performance. Experiments on financial fraud and genomic classification tasks show that DUQFL-Prox improves stability, generalization, and client fairness compared with standard QFL baselines. The results suggest that deep-unfolded quantum federated learning can support more reliable and fair intelligent services in heterogeneous distributed environments.

Figures

Figures reproduced from arXiv: 2607.21647 by the authors.

Figure 1
Figure 1. High-level DUQFL setup with adaptive local QNN [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of DUQFL-Prox. The server broadcasts global QNN parameters to clients, which perform deep-unfolded SPSA [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Federated performance comparison on the BAF dataset. DUQFL-Prox improves final global accuracy and mean client [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Epoch-wise ROC-AUC, PR-AUC, MCC, and specificity trajectories on the BAF dataset. DUQFL-Prox shows more [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: BAF Data mean client test accuracy, train–test gap, client fairness gap, precision, recall, F1-score, ROC-AUC, PR-AUC, MCC, and specificity. A. Results on the BAF Dataset The BAF dataset represents a highly imbalanced financial fraud-detection task. This setting is cha…
Figure 6
Figure 6. Figure 6: Federated performance comparison on the Genome dataset. FedProx-QFL achieves the strongest final global accuracy, [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Simulator and real IBM quantum hardware validation [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 9
Figure 9. Figure 9: Ablation analysis of DUQFL variants on the Genome non-IID setting. DUQFL-last uploads the final unfolded local [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 11
Figure 11. Figure 11: Trace-level diagnostics of local optimization in [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: Unfold-level behaviour of DUQFL-Prox. The nor [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 6 linked inside Pith

  1. [1]

    Federated optimization in heterogeneous networks,

    T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V . Smith, “Federated optimization in heterogeneous networks,”Proceedings of Machine Learning and Systems, vol. 2, pp. 429–450, 2020

    2020

  2. [2]

    Quantum federated learning based on gradient descent,

    K. Yu, X. Zhang, Z. Ye, G.-D. Guo, and S. Lin, “Quantum federated learning based on gradient descent,”arXiv preprint arXiv:2212.12913, 2022

    arXiv 2022

  3. [3]

    Quantum federated learning with decen- tralized data,

    R. Huang, X. Tan, and Q. Xu, “Quantum federated learning with decen- tralized data,”IEEE Journal of Selected Topics in Quantum Electronics, vol. 28, no. 4, pp. 1–10, 2022

    2022

  4. [4]

    Quantum federated learning: A comprehensive survey,

    D. C. Nguyen, M. R. Uddin, S. Shaon, R. Rahman, O. Dobre, and D. Niyato, “Quantum federated learning: A comprehensive survey,” arXiv preprint arXiv:2508.15998, 2025

    arXiv 2025

  5. [5]

    Shadows of quantum machine learning,

    S. Jerbi, C. Gyurik, S. C. Marshall, R. Molteni, and V . Dunjko, “Shadows of quantum machine learning,”Nature Communications, vol. 15, no. 1, p. 5676, 2024

    2024

  6. [6]

    A holistic and hybrid service selection strategy for mec- based uav last-mile delivery systems,

    J. Xuet al., “A holistic and hybrid service selection strategy for mec- based uav last-mile delivery systems,”IEEE Transactions on Services Computing, vol. 17, no. 6, pp. 3022–3036, 2024

    2024

  7. [7]

    Drone-as-a-service: Research challenges and direc- tions,

    A. Hamdiet al., “Drone-as-a-service: Research challenges and direc- tions,”Proceedings of the IEEE, vol. 113, no. 5, pp. 416–442, 2025

    2025

  8. [8]

    A deep reinforcement learning approach for composing moving iot services,

    A. G. Neiat, A. Bouguettaya, and M. Bahutair, “A deep reinforcement learning approach for composing moving iot services,”IEEE Transac- tions on Services Computing, vol. 15, no. 5, pp. 2538–2550, 2022

    2022

Show all 39 references
  1. [9]

    Advances and open problems in federated learning,

    P. Kairouz and H. B. McMahan, “Advances and open problems in federated learning,”Foundations and trends in machine learning, vol. 14, no. 1-2, pp. 1–210, 2021

    2021

  2. [10]

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

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

    2022

  3. [11]

    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,” in15th{USENIX} Symposium on Operating Systems Design and Implementation ({OSDI} 21), 2021, pp. 19–35

    2021

  4. [12]

    Towards understanding biased client selection in federated learning,

    Y . J. Cho, J. Wang, and G. Joshi, “Towards understanding biased client selection in federated learning,” inInternational Conference on Artificial Intelligence and Statistics. PMLR, 2022, pp. 10 351–10 375

    2022

  5. [13]

    Mitigating group bias in federated learning for heterogeneous devices,

    K. Selialia, Y . Chandio, and F. M. Anwar, “Mitigating group bias in federated learning for heterogeneous devices,” inProceedings of the 2024 ACM conference on fairness, accountability, and transparency, 2024, pp. 1043–1054

    2024

  6. [14]

    Multivariate stochastic approximation using a simultaneous perturbation gradient approximation,

    J. C. Spall, “Multivariate stochastic approximation using a simultaneous perturbation gradient approximation,”IEEE transactions on automatic control, vol. 37, no. 3, pp. 332–341, 2002

    2002

  7. [15]

    SPSA Optimizer,

    IBM Quantum, “SPSA Optimizer,” https://quantum.cloud.ibm.com/docs/ api/qiskit/qiskit algorithms.optimizers.SPSA, accessed: 2026-05-17

    2026

  8. [16]

    Quantum federated learning through blind quantum computing,

    W. Li, S. Lu, and D.-L. Deng, “Quantum federated learning through blind quantum computing,”Science China Physics, Mechanics & As- tronomy, vol. 64, no. 10, p. 100312, 2021

    2021

  9. [17]

    Cryptoqfl: Quantum federated learning on encrypted data,

    C. Chu, L. Jiang, and F. Chen, “Cryptoqfl: Quantum federated learning on encrypted data,” inProc. IEEE International Conference on Quantum Computing and Engineering (QCE), vol. 1, 2023, pp. 1231–1237

    2023

  10. [18]

    Quantum federated learning with entanglement controlled circuits and superposition coding,

    W. J. Yun, J. P. Kim, H. Baek, S. Jung, J. Park, M. Bennis, and J. Kim, “Quantum federated learning with entanglement controlled circuits and superposition coding,”arXiv preprint arXiv:2212.01732, 2022

    2022 arXiv

  11. [19]

    Real time adaptive pid controller based on quantum neural network for nonlinear systems,

    Y . F. Hanna, A. A. Khater, M. El-Bardini, and A. M. El-Nagar, “Real time adaptive pid controller based on quantum neural network for nonlinear systems,”Engineering Applications of Artificial Intelligence, vol. 126, p. 106952, 2023

    2023

  12. [20]

    Qfnn-ffd: Quantum federated neural network for financial fraud detection,

    N. Innan, A. Marchisio, M. Bennai, and M. Shafique, “Qfnn-ffd: Quantum federated neural network for financial fraud detection,” in2025 IEEE International Conference on Quantum Software (QSW). IEEE, 2025, pp. 41–47

    2025

  13. [21]

    Quanfraud: Quantum state verification scheme for fraud detection in iot-assisted quantum-blockchain networks,

    S. Majumder, S. Ray, M. Dasgupta, P. Bhattacharya, T. R. Gadekallu, and G. Srivastava, “Quanfraud: Quantum state verification scheme for fraud detection in iot-assisted quantum-blockchain networks,”IEEE Transactions on Services Computing, vol. 19, no. 1, pp. 616–627, 2026

    2026

  14. [22]

    Dopiv: Post- quantum secure identity-based data outsourcing with public integrity verification in cloud storage,

    X. Zhang, J. Zhao, C. Xu, H. Wang, and Y . Zhang, “Dopiv: Post- quantum secure identity-based data outsourcing with public integrity verification in cloud storage,”IEEE Transactions on Services Comput- ing, vol. 15, no. 1, pp. 334–345, 2022

    2022

  15. [23]

    Oqfl: An optimized quantum-based federated learning framework for defending against adversarial attacks in intelligent transportation systems,

    W. Yamany, N. Moustafa, and B. Turnbull, “Oqfl: An optimized quantum-based federated learning framework for defending against adversarial attacks in intelligent transportation systems,”IEEE Trans- actions on Intelligent Transportation Systems, 2021

    2021

  16. [24]

    Federated quantum neural network with quantum teleportation for resource optimization in future wireless communication,

    B. Narottama and S. Y . Shin, “Federated quantum neural network with quantum teleportation for resource optimization in future wireless communication,”IEEE Transactions on Vehicular Technology, 2023

    2023

  17. [25]

    Quantum federated learning experiments in the cloud with data encoding,

    S. R. Pokhrel, N. Yash, J. Kua, G. Li, and L. Pan, “Quantum federated learning experiments in the cloud with data encoding,”arXiv preprint arXiv:2405.00909, 2024

    2024 arXiv

  18. [26]

    Non-iid quantum federated learning with one-shot communi- cation complexity,

    H. Zhao, “Non-iid quantum federated learning with one-shot communi- cation complexity,”Quantum Machine Intelligence, vol. 5, no. 1, p. 3, 2023

    2023

  19. [27]

    Optimizing quantum federated learning based on federated quantum natural gradient descent,

    J. Qi, X.-L. Zhang, and J. Tejedor, “Optimizing quantum federated learning based on federated quantum natural gradient descent,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  20. [28]

    Foundations of quantum federated learning over classical and quantum networks,

    M. Chehimi, S. Y .-C. Chen, W. Saad, D. Towsley, and M. Debbah, “Foundations of quantum federated learning over classical and quantum networks,”IEEE Network, 2023

    2023

  21. [29]

    Quantumfed: A federated learning framework for collaborative quantum training,

    Q. Xia and Q. Li, “Quantumfed: A federated learning framework for collaborative quantum training,” in2021 IEEE Global Communications Conference (GLOBECOM). IEEE, 2021, pp. 1–6

    2021

  22. [30]

    Adpfl: A privacy-preserving federated learning framework through adaptive model pruning on non-iid data,

    D. Ning, Y .-F. Ge, E. Bertino, Z. Zheng, Y . Jiang, and H. Wang, “Adpfl: A privacy-preserving federated learning framework through adaptive model pruning on non-iid data,”IEEE Transactions on Services Computing, pp. 1–15, 2026

    2026

  23. [31]

    New insights on unfold- ing and fine-tuning quantum federated learning,

    S. I. Nanayakkara and S. R. Pokhrel, “New insights on unfold- ing and fine-tuning quantum federated learning,”arXiv preprint arXiv:2506.20016, 2025

    2025 arXiv

  24. [32]

    Deep unfolding-based weighted av- eraging for federated learning under device and statistical heterogeneous environments,

    A. Nakai-Kasai and T. Wadayama, “Deep unfolding-based weighted av- eraging for federated learning under device and statistical heterogeneous environments,”IEICE Transactions on Communications, vol. 108, no. 4, pp. 411–420, 2024

    2024

  25. [33]

    Spsa - qiskit algorithms documentation,

    Qiskit Community, “Spsa - qiskit algorithms documentation,” https://qiskit-community.github.io/qiskit-algorithms/stubs/qiskit algorithms.optimizers.SPSA.html, 2025, accessed: 2026-05-10

    2025

  26. [34]

    An overview of the simultaneous perturbation method for efficient optimization,

    J. C. Spall, “An overview of the simultaneous perturbation method for efficient optimization,”Johns Hopkins apl technical digest, vol. 19, no. 4, pp. 482–492, 1998

    1998

  27. [35]

    Deep unfolding: Recent developments, theory, and design guidelines,

    N. Shlezinger, S. Segarra, Y . Zhang, D. Avrahami, Z. Davidov, T. Rout- tenberg, and Y . C. Eldar, “Deep unfolding: Recent developments, theory, and design guidelines,”arXiv preprint arXiv:2512.03768, 2025

    2025

  28. [36]

    Scaffold: Stochastic controlled averaging for federated learn- ing,

    S. P. Karimireddy, S. Kale, M. Mohri, S. Reddi, S. Stich, and A. T. Suresh, “Scaffold: Stochastic controlled averaging for federated learn- ing,” pp. 5132–5143, 2020

    2020

  29. [37]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  30. [38]

    Turning the tables: Biased, imbalanced, dynamic tabular datasets for ml evaluation,

    S. Jesus, J. Pombal, D. Alves, A. Cruz, P. Saleiro, R. Ribeiro, J. Gama, and P. Bizarro, “Turning the tables: Biased, imbalanced, dynamic tabular datasets for ml evaluation,”Advances in Neural Information Processing Systems, vol. 35, pp. 33 563–33 575, 2022

    2022

  31. [39]

    Genomic benchmarks: a collection of datasets for genomic sequence classification,

    K. Gre ˇsov´a, V . Martinek, D. ˇCech´ak, P. ˇSimeˇcek, and P. Alexiou, “Genomic benchmarks: a collection of datasets for genomic sequence classification,”BMC Genomic Data, vol. 24, no. 1, p. 25, 2023. . JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, JUNE 2025 13 TABLE V: Summa...

    2023

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.