Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Enhancing Gradient Variance and Differential Privacy in Quantum Federated Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that adaptive Gaussian noise in quantum federated learning can counteract vanishing gradient variance, with convergence bounds whose quantum variance term shrinks as 3σ²/(2^{2n}−1).

desk verdict The central theorem is algebraically invalid and the convergence bounds are vacuous, so the theory collapses; the adaptive-noise idea is plausible but the paper does not meet the bar. read the letter →

arxiv 2509.05377 v1 pith:IN6GZ5B3 submitted 2025-09-04 quant-ph cs.CR

Duc-Thien Phan , Minh-Duong Nguyen , Quoc-Viet Pham , Huilong Pi This is my paper · ORCID
classification quant-phcs.CR
keywords quantumfederatedlearningdifferentialprivacyadaptivenoisebarrenplateaugradientvarianceneuralnetworkclient-levelconvergenceanalysis
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 tries to establish that the exponential shrinking of gradient variance as the number of qubits grows—a serious obstacle in quantum machine learning—can be managed by adding differential-privacy noise adaptively. It proposes ADP-QFL, in which each client injects Gaussian noise with variance that decays over training rounds, estimates and filters local models before transmission, and sends only updates within a threshold of the global model. The paper's central quantitative claim is Theorem 1: the variance of an n-qubit mini-batch gradient is bounded by 3σ²/(2^{2n}−1), so the quantum-induced variance term in the convergence bounds falls exponentially with qubit count. If this is right, QFL can keep strong privacy and communication efficiency while still converging, and noise injection doubles as a way to escape sharp minima caused by vanishing gradients.

What carries the argument

The central object is the n-qubit gradient variance identity Var[∇L] = f(ρ,H,U)/(2^{2n}−1) imported from the barren-plateau literature, recast by Theorem 1 as a mini-batch bound E[‖ḡ^t − g^t_{n-Qubit}‖²] ≤ 3σ²/(2^{2n}−1). It carries the argument because every subsequent convergence bound uses this term as the quantum-induced noise ϑ, and the adaptive schedule σ_t² = σ_0²/(1+αt) is what the algorithm controls to regulate it. The model-estimation condition ∥φ̂_u−φ_u∥ ≤ b/λ is the second mechanism: it compresses transmissions and filters out client updates that deviate too far from the global objective.

What would settle it

Measure the empirical variance of the per-round quantum gradient in ADP-QFL's 8-qubit QCNN at fixed batch size and compare it with the bound 3σ²/(2^{16}−1) predicted by Theorem 1; if the variance does not follow that exponential scaling, or if it scales as 1/B with batch size, the bound and the convergence terms built on it are contradicted.

Watch

Extended reading notes

Core claim

The paper claims that adding client-level adaptive Gaussian noise to local quantum neural network updates in a federated setting simultaneously addresses three problems: privacy leakage from predictable gradient descent, entrapment in sharp minima due to barren plateaus, and performance loss from fixed noise. The mechanism is a noise schedule σ_t² = σ_0²/(1+αt) that gives strong perturbation early and fine-tuning later, plus a model-estimation filter ∥φ̂_u−φ_u∥ ≤ b/λ. The formal anchor is Theorem 1, which bounds the deviation between the expected gradient and the n-qubit batch gradient by 3σ²/(2^{2n}−1), and both convergence theorems substitute this term for the gradient variance ϑ. The pape

Load-bearing premise

The results stand on the assumption that the barren-plateau variance formula for an n-qubit circuit applies directly to the mini-batch stochastic gradients of the algorithm, and that the Gaussian noise added for differential privacy behaves exactly like stochastic gradient noise in the convergence analysis.

Editorial extensions

If this is right

  • If the bound in Theorem 1 is correct, the variance a QFL client contributes to the global update shrinks exponentially with qubit count, so at scale gradient noise becomes both a privacy feature and a training hazard.
  • The convex convergence bound contains an explicit quantum term (2τ²+3τ+1)η_l²·3σ²/(2^{2n}−1), meaning increasing qubits reduces this particular error source.
  • The non-convex analysis says that adding noise raises the factor κ, lowering the probability of being trapped at sharp minimizers, so adaptive noise is presented not only as a privacy cost but as an optimization mechanism.
  • The model-estimation threshold filters out client updates that deviate too far from the global model; experiments show this improves convergence speed and final accuracy relative to QFL without estimation and to classical differentially private federated learning.

Reading between the lines

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

  • Beyond the paper's claims, if the exponential variance law holds, one could tune qubit count against the privacy budget: more qubits mean less intrinsic gradient leakage, so a smaller injected noise variance may satisfy the same (ε,δ) guarantee—but the paper does not compute this trade-off.
  • Beyond the paper's claims, the fixed decay schedule σ_t² = σ_0²/(1+αt) could be replaced by a closed-loop rule that reads each client's current gradient variance and sets noise proportional to it, which would make the adaptive mechanism in Algorithm 1 match the variance term that appears in the theorems.
  • Beyond the paper's claims, the model-estimation threshold b/λ can be viewed as a sparsification rule; a directly testable extension is to measure how much it shrinks the effective client population per round and how that changes the privacy-budget accumulation defined in the paper.
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 ADP-QFL, a quantum federated learning framework that adds client-level adaptive Gaussian noise to local model updates, filters clients by a model-estimation threshold, and claims that the exponentially decaying gradient variance of n-qubit QNNs can be regulated by this noise to improve both privacy and convergence. The authors state three main theoretical contributions: an unbiased n-qubit gradient estimator with variance bound 3σ²/(2^{2n}-1) (Theorem 1), convex convergence (Theorem 2), and non-convex convergence (Theorem 3). They also report experiments on MNIST and CIFAR-10 comparing against FedAvg, FedBN, pFedMe, FedRep, and FedQNN, with ADP-QFL achieving 98.47% and 83.85% accuracy, respectively. The central theoretical claim is that QNN mini-batch gradients have exponentially decaying variance with qubit count, and that adaptive DP noise can counteract the resulting barren-plateau and sharp-minima problems.

Significance. If the theoretical claims were correct, the adaptive noise mechanism would be a meaningful step toward privacy-utility trade-offs in QFL, and the claimed exponential variance reduction would connect barren-plateau analysis to DP noise calibration in a novel way. The empirical section also reports a broad baseline comparison and statistical significance testing, which is a strength. However, the load-bearing theoretical result, Theorem 1, is not established: its proof in Appendix B is algebraically invalid. The convergence theorems inherit this unsupported gradient-variance term and do not analyze the actual DP noise injected by Algorithm 1. The experimental results, while suggestive, cannot compensate for the collapse of the theoretical contribution, and the paper does not provide reproducible code or machine-checked proofs. The significance of the paper is therefore currently not established.

major comments (4)
  1. [Appendix B, Eq. (22)] The proof of Theorem 1 is algebraically invalid. Eq. (22) starts from the barren-plateau formula Var[∇L] = f(ρ,H,U)/(2^{2n}-1) and rewrites it as Var[∇L] = f/(2^1-1) · (2^1-1)/(2^{2n}-1) = 3/(2^{2n}-1) Var[∇L]. This equates a quantity to a scaled version of itself. The only consistent reading is V_n = 3 V_1/(2^{2n}-1), a relation between variances for different qubit counts, not a bound on the per-sample gradient variance σ² appearing in Eq. (12). The theorem's bound 3σ²/(2^{2n}-1) is never derived from the cited barren-plateau result.
  2. [Appendix B, Eq. (23)] The batch-gradient construction is incorrect. Eq. (23) defines g_t^{n-Qubit} = 3/(2^{2n}-1) Σ_{i=1}^B g_i, then claims it equals \bar g_t + 3/(2^{2n}-1) Σ Δg_i. But expanding the left-hand side gives 3B/(2^{2n}-1)\bar g_t + 3/(2^{2n}-1)ΣΔg_i, so the \bar g_t term is missing a factor B. Thus the quantity is not a batch average, and the subsequent application of the L2 Weak Law in Eq. (24) does not follow. Even if the barren-plateau formula were applicable, this algebra would not yield Eq. (12).
  3. [Section V.C, Theorems 2 and 3] The convergence theorems substitute the unproven Theorem 1 bound for ∥ϑ∥², but ϑ in the appendices is described as the difference between batch-wise and stochastic gradients, not the adaptive DP parameter noise G_t added in Algorithm 1, Eq. (4). No step connects the Gaussian mechanism variance σ_t² = σ_0²/(1+αt) to the term 3σ²/(2^{2n}-1). Consequently, the claimed ability of adaptive DP noise to regulate gradient variance and improve convergence is not modeled in the proofs. The convergence analysis is for a different algorithm than the one implemented.
  4. [Section V.C, Theorem 3, Eq. (14)] The non-convex convergence bound is vacuous because the second and third terms grow with T: the right-hand side contains 24T η_l τ L σ_g²/κ and 40T η_l² τ(τ-1)L²σ_g²/κ (in Eq. (89)), so as the number of communication rounds increases, the bound worsens rather than converging. Additionally, κ = 1 - Σ_t P_sharp 1(∇L(θ^{(m)}) + ϑ < L) can be zero or negative, and no condition guarantees κ > 0. The theorem is also stated under µ-strong convexity even though it is presented as a non-convex result, which is internally inconsistent.
minor comments (5)
  1. [Table I and Appendix C] The notation ϑ is described as 'Gradient variance' in Table I, but in Appendix C it is used as a random vector with ϑ ∼ N(0, Var[∇L]) and then ∥ϑ∥² appears. The dimension and distribution should be specified consistently.
  2. [Section IV.A, Eq. (5)] The initialization for σ² in Algorithm 1, line 2, is given as 8T(2L+b)² log(1/δ)/(K²ε²), but the adaptive schedule σ_t² = σ_0²/(1+αt) is never composed with Lemma 1, and no end-to-end privacy accounting is provided. The relation between σ_0 and the privacy guarantee should be clarified.
  3. [Section VI, Figure 4] Figure 4's caption says 'convergence time (log scale)' but the y-axis is labeled 'Accuracy (%)'; this is confusing and should be corrected.
  4. [Algorithm 1, line 15] The threshold condition ∥\hat ϕ_u - ϕ_u∥ ≤ b/λ uses \hat ϕ_u, but the construction of \hat ϕ_u is not fully specified; it is only said to be obtained via 'sparsification and estimate'. A precise definition is needed for the filter to be reproducible.
  5. [Appendix G, proof of Theorem 3] Equation (86) carries a factor τ∥ϑ∥² in the drift bound, and then Eq. (87) writes T τ² ∥ϑ∥². The algebra is not transparent, and the role of B in the batch size is absent throughout the appendix.

Circularity Check

3 steps flagged · score 8.0 of 10

Theorem 1's exponential variance bound is derived by equating the target variance to a constant multiple of itself (Appendix B, Eq. 22), and the convergence theorems inherit this self-referential term in place of the actual adaptive DP noise.

  1. self definitional [Appendix B, Eq. (22)]
    "When n = 1, the system represents similar to conventional 2-bit systems. Thus, we can represent the gradient variance of the n-qubit system as follows: Var[∇L]= E[(g − ¯g)2]] = f (ρ, H, U) 2^1 − 1 · 2^1 − 1 2^{2n} − 1 = 3 2^{2n} − 1Var[∇L], (22)"

    The displayed equation has the very quantity to be bounded, Var[∇L], on both sides. Starting from the cited barren-plateau formula Var[∇L] = f/(2^{2n}−1), the proof substitutes the n=1 version into itself and concludes Var[∇L] = 3/(2^{2n}−1) Var[∇L]. For any nonzero variance this is algebraically inconsistent except at n=1; it is not a bound. Even reading it charitably as a relation V_n = 3V_1/(2^{2n}−1) between variances at different n, Theorem 1 requires a bound on the per-sample gradient variance σ² appearing in Eq. (12). No step connects f(ρ,H,U) or V_1 to that σ², so the exponential decay is assumed by construction rather than derived.

  2. renaming known result [Theorem 1 (Eq. 12) and Appendix B, Eqs. (21)–(23)]
    "We follow the Barren Plateaus theorem [14], and have Var[∇L] = E[(g − ¯g)^2] = f(ρ, H, U)/(2^{2n}−1) ... Applying the gradient descent with batch size B, we have: g_t^{n-Qubit}= 3/(2^{2n}−1) Σ_{i=1}^B g_t^i = ... = ¯g_t + 3/(2^{2n}−1) Σ Δg_t^i. (23)"

    The n-dependent denominator 2^{2n}−1 in Theorem 1 is imported verbatim from the cited barren-plateau variance formula. The proof then defines the n-qubit mini-batch estimator by pre-multiplying the sum with exactly 3/(2^{2n}−1) instead of the standard batch average 1/B. This drops the factor B in the ¯g term (Σ_{i=1}^B ¯g = B¯g), and the subsequent L2 weak-law step compares a first-moment inequality with the squared expectation in Eq. (12). The claimed mini-batch variance is therefore not obtained from the stochastic-gradient decomposition; it is the cited formula renormalized and relabeled as a new prediction.

1 more flagged steps
  1. other [Theorem 2 (Eq. 13), Theorem 3 (Eq. 14), Appendix E (Eqs. 61–62), Appendix C]
    "ϑ represents the noise between the batch-wise gradient ∇ℓ(θ) and the stochastic gradient ∇L(θ), and can be characterized by ϑ ∼ N(0, Var[∇L]). ... Apply Theorem 1, we have: L(¯θ(R)) − L(θ∗) ≤ O(∥θ(0)−θ∗∥²/...) + O(η²_l τ(τ−1)Lσ²_∗) + O(η_l τ σ²_∗) + O((2τ²+3τ+1)η²_l · 3σ²/(2^{2n}−1)). (62)"

    In the convergence proofs the only variance term retained is the stochastic-gradient noise ϑ, and it is explicitly identified with Var[∇L]. Appendix E then replaces ∥ϑ∥² by the Theorem 1 expression 3σ²/(2^{2n}−1) ('Apply Theorem 1'), so the convergence bounds inherit the self-referential result. The actual adaptive DP noise of Algorithm 1, G_t with variance σ_t² = σ_0²/(1+αt), never appears anywhere in these proofs; no step connects that Gaussian parameter noise to Var[∇L] or to 3σ²/(2^{2n}−1). Thus the advertised convergence–privacy trade-off is the circular quantum-variance term substituted for the algorithm's own noise mechanism, not an analysis of the mechanism itself.

full rationale

The central theoretical claim of the paper is Theorem 1, which asserts that an n-qubit mini-batch gradient has variance bounded by 3σ²/(2^{2n}−1). The proof in Appendix B is self-referential by the paper's own equations: Eq. (22) writes Var[∇L] = 3/(2^{2n}−1) Var[∇L], i.e., the target quantity is equated to a constant multiple of itself. This is not a derivation of a bound on the per-sample variance σ² stated in Eq. (12); it is the cited barren-plateau formula (from [14]/[42]) with an invalid normalization, and even the charitable reading V_n = 3V_1/(2^{2n}−1) does not yield Theorem 1. Eq. (23) then defines the n-qubit mini-batch estimator by multiplying the sum by 3/(2^{2n}−1), dropping the batch-size factor, so the claimed estimator is constructed to reproduce the input formula rather than derived from the decomposition in Definition 3. Theorems 2 and 3 are load-bearing on this result: their 'quantum-induced gradient variance' term is exactly the Theorem 1 expression substituted for ∥ϑ∥², while the adaptive DP noise σ_t² = σ_0²/(1+αt) from Algorithm 1 is absent from the convergence bounds. The self-citation at [15] is incidental and not load-bearing, so the circularity is not of the self-citation type; it is a definitional/fitted-input circularity in the mathematical core. Because the paper's central contribution—exponential quantum gradient-variance reduction and its use to justify adaptive noise—reduces by construction to the input barren-plateau variance formula with an erroneous normalization, the circularity score is high, 8.

Assumptions & free parameters 7 free parameters · 8 assumptions · 1 invented entities

The central claims rest on unproved or misapplied results: the barren plateau variance formula is treated as an axiom and then 'derived' again; the DP noise bound is cited from a different meta-learning context without verification for this algorithm; the non-convex proof invokes an undefined probability Psharp; and the convergence analysis ignores the actual adaptive DP noise schedule.

free parameters (7)
  • sigma_0 (initial noise variance) = not reported
    In the adaptive noise schedule sigma_t^2 = sigma_0^2/(1+alpha t); values used in experiments are not stated.
  • alpha (noise decay rate) = not reported
    Controls how fast DP noise decreases; no value or sensitivity analysis given.
  • b (estimation error) = not reported
    Threshold parameter in model selection ||phi_hat_u - phi_u|| <= b/lambda.
  • lambda (regularization) = not reported
    Together with b defines the client selection threshold; no values reported.
  • Psharp = undefined
    Probability of being trapped in sharp minimizers introduced in Theorem 3; no definition, no estimator, no experiment.
  • sigma_g^2 / sigma_*^2 (heterogeneity bound) = not estimated
    Assumed bounded but never computed from data; appears in convergence bounds.
  • QCNN architecture depth (3 conv-pooling pairs, 64 parameters) = 64 trainable parameters
    Empirically determined for best trade-off; no ablation.
assumptions (8)
  • domain assumption L-smoothness of local objectives (Assumption 1)
    Standard in FL analysis but not verified for QCNN loss surfaces.
  • domain assumption mu-strong convexity (Assumption 2)
    QCNN losses are generally non-convex; used in Theorem 2.
  • domain assumption Bounded gradient dissimilarity at optimum (Assumption 3)
    Heterogeneity bound needed for client drift terms.
  • ad hoc to paper Barren plateau variance formula Var[grad L] = f(rho,H,U)/(2^{2n}-1) from [14,42]
    Used as starting point of Theorem 1; the 'derivation' of Eq (22) rewrites this formula rather than proving it.
  • domain assumption Unbiased gradient assumption from [34, Assumption 2]
    Imported to justify Theorem 1.
  • ad hoc to paper Lemma 1 DP noise variance formula from [32]
    Cited without proof; its applicability to client-level DP with model perturbation and adaptive schedule is not established.
  • domain assumption Noise added by different clients is IID and independent of model parameters
    Stated in Section IV-C; required for DP but not justified for quantum clients with correlated device noise.
  • ad hoc to paper The L2 Weak Law application in Appendix B
    Misapplied: the weak law bounds averages, not the scaled gradient estimator in Eq (23).
invented entities (1)
  • Psharp (probability of convergence to sharp minimizers)
    purpose: Introduced in Theorem 3 to degrade the convergence bound when gradients are trapped in sharp minima; never defined or estimated.
    No falsifiable handle; it is an ad hoc parameter that adjusts the bound to the desired conclusion. The condition inside the indicator is also dimensionally inconsistent (comparing ||grad L|| + variance to L, which has units of loss).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Gradient Variance and Differential Privacy in Quantum Federated Learning." pith.science (2026). https://pith.science/paper/IN6GZ5B3

@misc{pith2026250905377,
  author       = {Pith},
  title        = {Pith review of: Enhancing Gradient Variance and Differential Privacy in Quantum Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IN6GZ5B3}},
  note         = {Machine review of arXiv:2509.05377}
}
read the original abstract

Upon integrating Quantum Neural Network (QNN) as the local model, Quantum Federated Learning (QFL) has recently confronted notable challenges. Firstly, exploration is hindered over sharp minima, decreasing learning performance. Secondly, the steady gradient descent results in more stable and predictable model transmissions over wireless channels, making the model more susceptible to attacks from adversarial entities. Additionally, the local QFL model is vulnerable to noise produced by the quantum device's intermediate noise states, since it requires the use of quantum gates and circuits for training. This local noise becomes intertwined with learning parameters during training, impairing model precision and convergence rate. To address these issues, we propose a new QFL technique that incorporates differential privacy and introduces a dedicated noise estimation strategy to quantify and mitigate the impact of intermediate quantum noise. Furthermore, we design an adaptive noise generation scheme to alleviate privacy threats associated with the vanishing gradient variance phenomenon of QNN and enhance robustness against device noise. Experimental results demonstrate that our algorithm effectively balances convergence, reduces communication costs, and mitigates the adverse effects of intermediate quantum noise while maintaining strong privacy protection. Using real-world datasets, we achieved test accuracy of up to 98.47\% for the MNIST dataset and 83.85\% for the CIFAR-10 dataset while maintaining fast execution times.

Figures

Figures reproduced from arXiv: 2509.05377 by the authors.

Figure 1
Figure 1. Demonstration of rapid variance decay in QNNs [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed ADP-QFL framework. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparison of test accuracy on MNIST and CIFAR [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of convergence time (log scale) on [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Training accuracy of ADP-QFL under different [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 7
Figure 7. Figure 7: Test accuracy versus communication rounds for [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 6
Figure 6. Figure 6: Performance of ADP-QFL with varying numbers of [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 39 canonical work pages

  1. [28]

    Federated quantum machine learning with differential privacy,

    R. Rofougaran, S. Yoo, H.-H. Tseng, and S. Y .-C. Chen, “Federated quantum machine learning with differential privacy,” in 2024 IEEE International Conference on Acoustics, Speech and Signal Processing , Seoul, Korea, 2024, pp. 9811–9815

    work page 2024

  2. [34]

    Tackling the ob- jective inconsistency problem in heterogeneous federated optimization,

    J. Wang, Q. Liu, H. Liang, G. Joshi, and H. V . Poor, “Tackling the ob- jective inconsistency problem in heterogeneous federated optimization,” in Advances in Neural Information Processing Systems , vol. 34, Virtual, 2020, pp. 7611–7623

    work page 2020

  3. [1]

    tinyRadar for fitness: A contactless framework for edge computing,

    S. S. Yadav, R. Agarwal, K. Bharath, S. Rao, and C. S. Thakur, “tinyRadar for fitness: A contactless framework for edge computing,” IEEE Transactions on Biomedical Circuits and Systems , vol. 17, no. 2, pp. 192–201, 2023. 13

    work page 2023

  4. [2]

    faaShark: An end-to-end network traffic analysis system atop serverless computing platforms,

    H. Zhao, S. Pan, Z. Cai, X. Chen, L. Jin, H. Gao, S. Wan, R. Ma, and H. Guan, “faaShark: An end-to-end network traffic analysis system atop serverless computing platforms,” IEEE Transactions on Network Science and Engineering, vol. 11, no. 3, pp. 2473–2484, 2024

    work page 2024

  5. [3]

    Quantum computing for finance,

    D. Herman, C. Googin, X. Liu, Y . Sun, A. Galda, I. Safro, M. Pistoia, and Y . Alexeev, “Quantum computing for finance,” Nature Reviews Physics, vol. 5, no. 8, pp. 450–465, 2023

    work page 2023

  6. [4]

    Challenges and opportunities in quantum machine learning,

    M. Cerezo, G. Verdon, H.-Y . Huang, L. Cincio, and P. J. Coles, “Challenges and opportunities in quantum machine learning,” Nature Computational Science, vol. 2, no. 9, pp. 567–576, 2022

    work page 2022

  7. [5]

    Transitioning from federated learn- ing to quantum federated learning in internet of things: A comprehensive survey,

    C. Qiao, M. Li, Y . Liu, and Z. Tian, “Transitioning from federated learn- ing to quantum federated learning in internet of things: A comprehensive survey,”IEEE Communications Surveys & Tutorials, vol. 26, p. 1, 2024

    work page 2024

  8. [6]

    Secure delegated variational quantum algorithms,

    Q. Li, J. Quan, J. Shi, S. Zhang, and X. Li, “Secure delegated variational quantum algorithms,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , vol. 43, p. 1, 2024

    work page 2024

Show all 43 references
  1. [7]

    Quantum machine learning with differential privacy,

    W. M. Watkins, S. Y .-C. Chen, and S. Yoo, “Quantum machine learning with differential privacy,” Scientific Reports , vol. 13, no. 1, p. 2453, 2023

    2023

  2. [8]

    An efficient simulation for quantum secure multiparty computation,

    K. Sutradhar and H. Om, “An efficient simulation for quantum secure multiparty computation,” Scientific Reports, vol. 11, no. 1, p. 2206, 2021

    2021

  3. [9]

    Calibrating noise to sensitivity in private data analysis,

    C. Dwork, F. McSherry, K. Nissim, and A. Smith, “Calibrating noise to sensitivity in private data analysis,” Journal of Privacy and Confiden- tiality, vol. 7, no. 3, pp. 17–51, 2016

    2016

  4. [10]

    Differential privacy in quantum computation,

    L. Zhou and M. Ying, “Differential privacy in quantum computation,” in 2017 IEEE 30th Computer Security Foundations Symposium, California, USA, 2017, pp. 249–262

    2017

  5. [11]

    Quantum differentially private sparse regression learning,

    Y . Du, M.-H. Hsieh, T. Liu, S. You, and D. Tao, “Quantum differentially private sparse regression learning,” IEEE Transactions on Information Theory, vol. 68, no. 8, pp. 5217–5233, 2022

    2022

  6. [12]

    Privacy-preserving quantum machine learning using differential privacy,

    M. Senekane, M. Mafu, and B. M. Taele, “Privacy-preserving quantum machine learning using differential privacy,” in 2017 IEEE AFRICON , Victoria and Alfred (V&A) Waterfront Cape Town, South Africa, 2017, pp. 1432–1435

    2017

  7. [13]

    Improved differential privacy noise mechanism in quantum machine learning,

    H. Yang, X. Li, Z. Liu, and W. Pedrycz, “Improved differential privacy noise mechanism in quantum machine learning,” IEEE Access, vol. 11, pp. 50 157–50 164, 2023

    2023

  8. [14]

    Barren plateaus in quantum neural network training landscapes,

    J. R. McClean, S. Boixo, V . N. Smelyanskiy, R. Babbush, and H. Neven, “Barren plateaus in quantum neural network training landscapes,”Nature Communications, vol. 9, p. 4812, 2018

    2018

  9. [15]

    Revisiting LARS for large batch training generalization of neural networks,

    K. Do, D. Nguyen, H. Nguyen, L. Tran-Thanh, and Q.-V . Pham, “Revisiting LARS for large batch training generalization of neural networks,” arXiv preprint arXiv:2309.14053 , 2023

    2023 arXiv

  10. [16]

    On large-batch training for deep learning: Generalization gap and sharp minima,

    N. S. Keskar, D. Mudigere, J. Nocedal, M. Smelyanskiy, and P. T. P. Tang, “On large-batch training for deep learning: Generalization gap and sharp minima,” in International Conference on Learning Represen- tations, Toulon, France, 2017

    2017

  11. [17]

    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 , Florida, USA, 2017, pp. 1273–1282

    2017

  12. [18]

    FedBN: Federated learning on non-IID features via local batch normalization,

    X. Li, M. Jiang, X. Zhang, M. Kamp, and Q. Dou, “FedBN: Federated learning on non-IID features via local batch normalization,” arXiv preprint arXiv:2102.07623, 2021

    2021 arXiv

  13. [19]

    Safeguarding cross-silo federated learning with local differential privacy,

    C. Wang, X. Wu, G. Liu, T. Deng, K. Peng, and S. Wan, “Safeguarding cross-silo federated learning with local differential privacy,” Digital Communications and Networks , vol. 8, no. 4, pp. 446–454, 2022

    2022

  14. [20]

    Multi-stage asynchronous federated learning with adaptive differential privacy,

    Y . Li, S. Yang, X. Ren, L. Shi, and C. Zhao, “Multi-stage asynchronous federated learning with adaptive differential privacy,”IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 46, no. 2, pp. 1243– 1256, 2024

    2024

  15. [21]

    Federated learning with sparsified model perturbation: Improving accuracy under client-level differential privacy,

    R. Hu, Y . Guo, and Y . Gong, “Federated learning with sparsified model perturbation: Improving accuracy under client-level differential privacy,” IEEE Transactions on Mobile Computing, vol. 23, no. 8, pp. 8242–8255, 2023

    2023

  16. [22]

    Towards the flatter landscape and better generalization in federated learning under client-level differential privacy,

    Y . Shi, K. Wei, L. Shen, Y . Liu, X. Wang, B. Yuan, and D. Tao, “Towards the flatter landscape and better generalization in federated learning under client-level differential privacy,”arXiv preprint arXiv:2305.00873, 2023

    2023 arXiv

  17. [23]

    Federated learning with personalized local differential privacy,

    G. Yang, S. Wang, and H. Wang, “Federated learning with personalized local differential privacy,” in 2021 IEEE 6th International Conference on Computer and Communication Systems , Chengdu, China, 2021, pp. 484–489

    2021

  18. [24]

    QuantumFed: A federated learning framework for collaborative quantum training,

    Q. Xia and Q. Li, “QuantumFed: A federated learning framework for collaborative quantum training,” in 2021 IEEE Global Communications Conference, Madrid, Spain, 2021, pp. 1–6

    2021

  19. [25]

    Quantum federated learning with quantum data,

    M. Chehimi and W. Saad, “Quantum federated learning with quantum data,” in 2022 IEEE International Conference on Acoustics, Speech and Signal Processing, 2022, pp. 8617–8621

    2022

  20. [26]

    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

  21. [27]

    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

  22. [29]

    AdaPDP: Adaptive personalized differential privacy,

    B. Niu, Y . Chen, B. Wang, Z. Wang, F. Li, and J. Cao, “AdaPDP: Adaptive personalized differential privacy,” in IEEE conference on computer communications, Vancouver, Canada, 2021, pp. 1–10

    2021

  23. [30]

    Experimental quantum end-to-end learning on a superconducting processor,

    X. Pan, X. Cao, W. Wang, Z. Hua, W. Cai, X. Li, H. Wang, J. Hu, Y . Song, D.-L. Denget al., “Experimental quantum end-to-end learning on a superconducting processor,”npj Quantum Information, vol. 9, p. 18, 2023

    2023

  24. [31]

    C. N. Hugh Collins. (2022) IBM Unveils 400 Qubit-Plus Quantum Processor and Next-Generation IBM Quantum System Two. [Online]. Available: https://shorturl.at/kdRXw

    2022

  25. [32]

    Task-level differentially private meta learning,

    X. Zhou and R. Bassily, “Task-level differentially private meta learning,” in Advances in Neural Information Processing Systems , vol. 35, New Orleans, USA, 2022, pp. 20 947–20 959

    2022

  26. [33]

    Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography,

    M. A. Fischler and R. C. Bolles, “Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography,” Communications of the ACM, vol. 24, no. 6, pp. 381–395, 1981

    1981

  27. [35]

    Gradient-based learning applied to document recognition,

    Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, pp. 2278 – 2324, 1998

    1998

  28. [36]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, “Learning multiple layers of features from tiny images,” University of Toronto, 2012

    2012

  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]

    Personalized federated learning with moreau envelopes,

    C. T Dinh, N. Tran, and J. Nguyen, “Personalized federated learning with moreau envelopes,” Advances in Neural Information Processing Systems, vol. 33, pp. 21 394–21 405, 2020

    2020

  31. [39]

    Exploiting shared representations for personalized federated learning,

    L. Collins, H. Hassani, A. Mokhtari, and S. Shakkottai, “Exploiting shared representations for personalized federated learning,” in Inter- national conference on machine learning , vol. 139, Virtual, 2021, pp. 2089–2099

    2021

  32. [40]

    Fedqnn: Federated learning using quantum neural networks,

    N. Innan, M. A.-Z. Khan, A. Marchisio, M. Shafique, and M. Bennai, “Fedqnn: Federated learning using quantum neural networks,” in 2024 International Joint Conference on Neural Networks (IJCNN) . IEEE, 2024, pp. 1–9

    2024

  33. [41]

    Fedexp: Speeding up federated averaging via extrapolation,

    D. Jhunjhunwala, S. Wang, and G. Joshi, “Fedexp: Speeding up federated averaging via extrapolation,” in The Eleventh International Conference on Learning Representations , Kigali, Rwanda, 2023

    2023

  34. [42]

    Connecting ansatz expressibility to gradient magnitudes and barren plateaus,

    Z. Holmes, K. Sharma, M. Cerezo, and P. J. Coles, “Connecting ansatz expressibility to gradient magnitudes and barren plateaus,” PRX Quantum, vol. 3, no. 1, p. 010313, 2022

    2022

  35. [43]

    Durrett, Probability: Theory and Examples , 4th ed., ser

    R. Durrett, Probability: Theory and Examples , 4th ed., ser. Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge Univer- sity Press, 2010. 14 APPENDIX A PRELIMINARIES We first state some preliminary lemmas that are used throughout the proof. Lemma 2 (Jense...

    2010

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.