Pith. sign in

REVIEW 3 major objections 5 minor 53 references

Privacy-Preserving Quantized Federated Learning with Diverse Precision

T0 review · 3 major / 5 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 The paper claims that a stochastic quantizer can give differential privacy and minimum quantization distortion simultaneously, and that a federated learning algorithm built on it keeps its utility when devices use different bit widths.

desk verdict The DP guarantee at the core of this paper doesn't follow from Lemma 3, but the distortion analysis and system design are solid enough to warrant review with major revision. read the letter →

arxiv 2507.00920 v2 pith:OMNULYTA submitted 2025-07-01 cs.LG eess.SP

Dang Qua Nguyen , Morteza Hashemi , Erik Perrins , Sergiy A. Vorobyov , David J. Love , Taejoon Kim This is my paper · ORCID
classification cs.LGeess.SP
keywords federatedlearningdifferentialprivacystochasticquantizationheterogeneityprivacy-utilitytradeoffconvergenceanalysisdeepleakagefromgradientsclustersizeoptimization
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 lets many devices train a shared model without uploading raw data, but the model updates themselves can reveal private training examples, and real devices vary in how many bits they can send. The paper's central claim is that the randomness of a stochastic quantizer can be tuned so that a quantized update satisfies differential privacy and minimizes expected distortion at the same time, with distortion that stays bounded even when the privacy requirement becomes very strict. On top of that quantizer, the paper builds a federated learning loop in which groups of devices using different bit resolutions participate each round, with fusion weights and cluster sizes chosen to reduce a convergence-error bound. A sympathetic reader would care because previous approaches either add unbounded Laplace noise for privacy or ignore the mixture of quantization resolutions, and the paper shows numerically that its combined design keeps test accuracy around 80 percent where the benchmark LaplaceSQ-FL drops to about 41 percent.

What carries the argument

The central object is the stochastic quantizer $Q_{b,\epsilon_1}$ with levels uniformly spaced across the input range, equipped with the closed-form rounding probabilities of Lemma 3. It is the solution to a bounded linear program that minimizes $\mathbb{E}[|Q(a)-a|^2]$ while respecting $e^{-\epsilon_1}\le p/(1-p)\le e^{\epsilon_1}$; the same ratio constraint is what bounds the privacy-loss likelihood ratio. The second mechanism is the SNR-proportional fusion rule $\omega_{m,u}\propto \theta_{m,u}$, where $\theta_{m,u}$ inverts the sum of quantization distortion and link-noise variance, and the third is the cluster-size optimization (27) that minimizes the linear-in-cluster-size deviation term $\Delta$ in the convergence bound of Theorem 1.

What would settle it

Compute, for a realistic clipped local update, the maximum over neighboring data sets of $\Pr[\text{quantizer output}\mid A]/\Pr[\text{quantizer output}\mid A']$; if this ratio exceeds $e^{d\,\epsilon_{1,m,u}}$ on any coordinate, or if changing a single training sample moves any coordinate across more than one quantization cell, the claimed per-round privacy loss does not follow from Lemma 3 alone.

Watch

Extended reading notes

Core claim

The core claim is that one mechanism—the stochastic quantizer $Q_{b,\epsilon_1}$—can serve as both compressor and privacy protector. For an input $a$ between adjacent levels $q_i$ and $q_{i+1}$, Lemma 3 proves that the probability $p^\star$ of rounding to $q_i$ should be $e^{\epsilon_1}/(e^{\epsilon_1}+1)$ when $q_i$ is closer to $a$, and $1/(e^{\epsilon_1}+1)$ otherwise, and that this choice is the exact minimizer of expected quantization distortion subject to the output-likelihood-ratio constraint (5b). The distortion formula (7) then depends monotonically on $\epsilon_1$ and remains bounded as $\epsilon_1\to 0$, in contrast to the LaplaceSQ baseline whose variance and distortion grow without bound. Algorithm 1 applies this quantizer coordinate-wise to clipped local model updates; applying the sequential-composition lemma gives per-round privacy loss $\epsilon_{m,u}=d\,\epsilon_{1,m,u}$ in equation (10). The paper's experiments support the conclusion that this protected update resists deep-leakage-from-gradients inversion and yields better training loss and test accuracy than the benchmark.

Load-bearing premise

The proof of the privacy guarantee assumes that two neighboring data sets produce quantizer inputs that fall within the same quantization interval; in the algorithm, the input to the quantizer is a clipped multi-step SGD update whose per-coordinate change from one training sample is not shown to stay within that interval.

Editorial extensions

If this is right

  • At any fixed bit budget $b$ and per-coordinate privacy loss $\epsilon_1$, using $Q_{b,\epsilon_1}$ instead of QSGD-plus-Laplace lowers expected distortion, and in the $b=6$, $\epsilon_1=0.1$ simulation the gap is roughly 2.5 orders of magnitude.
  • With a total bit budget $B$, optimizing the number of devices selected from each precision group reduces the deviation term in the convergence bound, and in the MNIST runs raises final accuracy from 70 percent to 75 percent with fusion-weight optimization and to about 80 percent with cluster-size optimization added.
  • The per-round privacy budget scales as $d\,\epsilon_{1,m,u}$, so for high-dimensional models the mapping from desired overall privacy to the per-coordinate budget $\epsilon_{1,m,u}$ must be set with the dimension in mind.
  • The convergence bound (26) includes a term $(t-4)\Delta$ that grows with rounds, so the algorithm's guarantee is for a finite number of rounds; the paper identifies this as the reason continued training can eventually degrade utility.
  • Against the deep-leakage-from-gradients attack with $\epsilon_{1,m,u}=10^{-6}$, the proposed algorithm keeps SSIM values near 0.005 to 0.02, while the SQ-FL baseline reaches SSIM values of 0.20 to 0.33 after 20 to 40 attack iterations.

Reading between the lines

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

  • The strongest practical payoff of a bounded-distortion DP quantizer is that it can replace additive Laplace noise inside other compression pipelines: any stochastic quantizer whose ratio constraint (5b) is satisfied cell-wise gives the same privacy guarantee with distortion that does not grow as $\epsilon_1\to 0$.
  • The paper's privacy composition in (10) counts $d$ independent coordinates; treating the $d$-dimensional update as a single mechanism with one global likelihood-ratio bound could produce a smaller privacy loss, but doing so would require a vector-level proof that the paper does not provide.
  • Because the scalar DP lemma assumes neighboring inputs lie in the same quantization interval, a natural next step is to prove a per-coordinate sensitivity bound for clipped, multi-step SGD updates, which would turn the per-round privacy budget into an end-to-end guarantee.
  • The cluster-size optimization is formulated as a static one-shot allocation; an online variant that estimates the effective SNR $\theta_{m,u}$ from received updates and reallocates clusters across rounds would extend the method to non-stationary link noise and is not tested 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

3 major / 5 minor

Summary. The paper proposes a differentially private stochastic quantizer for federated learning with heterogeneous quantization resolutions. The quantizer maps each input coordinate to one of two neighboring quantization levels with probabilities chosen by solving a distortion-minimization problem subject to a likelihood-ratio constraint (5b), yielding the closed form p* in Lemma 3. This quantizer is embedded in Algorithm 1, where each device clips its multi-step SGD model difference, applies the stochastic quantizer, and transmits the result through a noisy link; the fusion center aggregates with optimized weights. The paper also derives a convergence bound and a cluster-size optimization problem, and reports simulations comparing the method with a LaplaceSQ-FL baseline. The central claimed contribution is that Algorithm 1 provides per-round user-level DP with privacy loss epsilon_{m,u} = d epsilon_{1,m,u} in Eq. (10).

Significance. If the DP claim were valid, the proposed bounded-distortion quantizer would be a genuine improvement over Laplace-based quantization for privacy-preserving FL, and the fusion-weight and cluster-size optimization are reasonable engineering contributions that are supported by the derivations. The closed-form solution to (5), the distortion analysis in Section II-C, and the fusion weight derivation in Section III-C are internally consistent. The paper also ships a supplement (GitHub link in [46]) for standard derivations, which is a strength. However, the DP guarantee is the load-bearing claim of the paper, and it is not established: Lemma 3 only bounds likelihood ratios for inputs lying in the same quantization interval, whereas the actual quantizer input in Algorithm 1 is a clipped multi-step local update whose coordinates can span many intervals after a single sample change. The empirical DLG attack results in Section V-B do not compensate for the missing formal guarantee. The central privacy-utility advantage over LaplaceSQ-FL therefore rests on an unproven premise.

major comments (3)
  1. [Section II-B, constraint (5b) and following sentence] The DP guarantee of the proposed quantizer is explicitly restricted: the sentence after (5b) states the ratio bound holds for a, a' in [q_i, q_{i+1}). In Algorithm 1, Step 12 applies Q_{b_m, epsilon_{1,m,u}} to bv_{m,u,t}, a clipped multi-step SGD update whose per-coordinate entries can lie anywhere in [-C, C], i.e., in any of the 2^{b_m}-1 quantization cells. Replacing one training sample changes the stochastic gradients, and a coordinate of bv_{m,u,t} can cross cell boundaries. For two neighboring inputs lying in different cells, some output quantization level is reachable from one input but not the other, so the likelihood ratio required by Definition 2 is infinite or far exceeds e^{epsilon_{1,m,u}}. Lemma 3 therefore does not imply that Step 12 is epsilon_{1,m,u}-DP on the actual inputs, and the privacy claim of Algorithm 1 does not follow.
  2. [Section III-B, Eq. (10)] The step epsilon_{m,u} = d epsilon_{1,m,u} applies Lemma 1 (sequential composition) to the d coordinates of the quantized update. This composition step is only valid if each coordinate mechanism is epsilon_{1,m,u}-DP with respect to the same pair of neighboring data sets. Here a single data-point change perturbs every coordinate of bv_{m,u,t} through L local SGD updates, and the per-coordinate sensitivity of this clipped multi-step update is not bounded in units of the cell width (2C)/(2^{b_m}-1). Without a per-coordinate sensitivity analysis in cell-width units, the d-fold composition in Eq. (10) is unjustified. This is a second load-bearing gap in the central privacy claim.
  3. [Section V-B, Figs. 2b-2c and Table I] The DLG attack experiments are presented as validation of privacy protection at epsilon_{1,m,u}=10^{-6}, but they only show resistance to one specific gradient-inversion attack in a favorable setting (single data point, L=1, noiseless update). Attack simulations cannot establish the worst-case likelihood-ratio bound required by Definition 2; they are not a substitute for a correct DP proof. The formal guarantee must come from a sensitivity analysis of the actual quantizer input, which is absent.
minor comments (5)
  1. [Section II-B, Eq. (4)] The interval [a, a] uses the same symbol a for the interval endpoints and for the scalar input a; this makes the definition harder to read and should be changed to, e.g., [A, \bar{A}] or [a_min, a_max].
  2. [Section III-C, Eq. (11)] The scalars denoted zeta_1, ..., zeta_n in inequality (11) are later replaced by omega_{m,u}; the notation is not harmful but conflicts with the variance parameter zeta in Assumption 3. Consider using different symbols for the generic inequality and for the application.
  3. [Section IV-A, Theorem 1 and Eq. (26)] The bound in (26) contains the term (t-4)Delta, which grows linearly with t, so the bound diverges as t increases. The text states that the first two terms diminish and that 'learning utility improves over time'; this is inconsistent with the unbounded third term. The authors should clarify that Theorem 1 is not a convergence guarantee in the usual sense and should reconcile the statement with the divergence caused by Delta.
  4. [Section V-C.1] The benchmark description contains a typo: 'sate-of-the-art' should be 'state-of-the-art'.
  5. [Appendix B, Eqs. (21), (23), and (33a)] The symbol e appears inside bounds such as L^3 gamma^2 zeta eta_t^2 e and e^{...} without being defined; it is presumably Euler's number, but it should be stated explicitly or replaced with exp(1) in a notation table.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the quantizer is a constrained-optimization construction, and the main privacy concern is a correctness gap rather than a derivation that assumes its own conclusion.

full rationale

The paper's central derivation chain is not circular. Lemma 3 solves the optimization in (5), where the constraint (5b) is exactly the likelihood-ratio condition needed for the scalar quantizer to be epsilon1-DP when two inputs lie in the same quantization cell; the closed-form p* is then derived, not assumed. The distortion bound in (7) follows algebraically from p* and the objective in (5a), and the convergence bound in Theorem 1 is an upper bound obtained from the stated Assumptions 1-5 plus standard FedPAQ-style arguments, with no target result inserted as an assumption. The self-cited supplement [46] contains the proof of Lemma 7, but that lemma is a supporting standard bound with explicitly stated assumptions, and the paper's main privacy-quantization contribution does not reduce to that citation. The reviewer-identified weakness, namely that Algorithm 1 applies the scalar quantizer to clipped multi-step local updates whose per-coordinate values can cross multiple quantization cells, is a genuine correctness gap in the derivation of epsilon_{m,u}=d epsilon_{1,m,u} in (10); however, it is not circularity: no equation or definition in the paper makes (10) true by construction, and sequential composition (Lemma 1) is an external, independent result. The DLG simulations are empirical evidence for the specific attack settings and do not substitute for a formal proof, but again this is a validity concern, not an input-equivalent derivation.

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

The central claim rests on five standard FL assumptions (smoothness, strong convexity, unbiased gradients, clipping-error bound, and bounded bias) plus an unstated assumption that neighboring data sets change each quantizer input by at most one quantization cell. The latter is not derived and is likely false, which breaks the privacy guarantee. There are no fitted parameters: the DP budget, clipping constant, and noise variances are inputs; cluster sizes are optimized.

assumptions (6)
  • domain assumption Assumption 1: Each local loss function is convex, differentiable, and γ-smooth.
    Invoked for the convergence analysis in Lemmas 5 and 7; standard in FL theory.
  • domain assumption Assumption 2: Each local loss function is µ-strongly convex.
    Used to obtain the contraction factor in Theorem 1; excludes non-convex models like neural networks.
  • domain assumption Assumption 3: The stochastic gradient is unbiased with bounded variance ζ.
    Used throughout the proof of Lemma 5; standard SGD assumption.
  • domain assumption Assumption 4: The clipping error is bounded by Λ times the squared norm.
    Used in Lemma 5 to bound the effect of ℓ1 clipping.
  • domain assumption Assumption 5: The bias and sub-optimality gap cross-correlation is bounded by a constant ϑ.
    This assumption leads to the linearly growing (t-4)Δ term in Theorem 1; it is a strong assumption that is not verified.
  • ad hoc to paper Implicit: The per-coordinate input to the quantizer changes by at most one quantization interval for neighboring data sets.
    Needed for the DP claim, but never stated or derived; the mechanism's DP is only shown for inputs within the same cell.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Privacy-Preserving Quantized Federated Learning with Diverse Precision." pith.science (2026). https://pith.science/paper/OMNULYTA

@misc{pith2026250700920,
  author       = {Pith},
  title        = {Pith review of: Privacy-Preserving Quantized Federated Learning with Diverse Precision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OMNULYTA}},
  note         = {Machine review of arXiv:2507.00920}
}
read the original abstract

Federated learning (FL) has emerged as a promising paradigm for distributed machine learning, enabling collaborative training of a global model across multiple local devices without requiring them to share raw data. Despite its advancements, FL is limited by factors such as: (i) privacy risks arising from the unprotected transmission of local model updates to the fusion center (FC) and (ii) decreased learning utility caused by heterogeneity in model quantization resolution across participating devices. Prior work typically addresses only one of these challenges because maintaining learning utility under both privacy risks and quantization heterogeneity is a non-trivial task. In this paper, our aim is therefore to improve the learning utility of a privacy-preserving FL that allows clusters of devices with different quantization resolutions to participate in each FL round. Specifically, we introduce a novel stochastic quantizer (SQ) that is designed to simultaneously achieve differential privacy (DP) and minimum quantization error. Notably, the proposed SQ guarantees bounded distortion, unlike other DP approaches. To address quantization heterogeneity, we introduce a cluster size optimization technique combined with a linear fusion approach to enhance model aggregation accuracy. Numerical simulations validate the benefits of our approach in terms of privacy protection and learning utility compared to the conventional LaplaceSQ-FL algorithm.

Figures

Figures reproduced from arXiv: 2507.00920 by the authors.

Figure 1
Figure 1. Privacy-preserving quantized FL network with quantization resolution [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. a) Expected distortions of the proposed quantizer [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. a) Training loss of Algorithm 1+FWO+CSO, Algorithm 1+FWO, Algorithm 1, and LaplaceSQ-FL when [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: a) Testing accuracy of Algorithm 1+FWO+CSO, Algorithm 1+FWO, Algorithm 1, and LaplaceSQ-FL when [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 46 canonical work pages

  1. [25]

    Randomized quantization is all you need for differential privacy in federated learning,

    Y . Youn, Z. Hu, J. Ziani, and J. Abernethy, “Randomized quantization is all you need for differential privacy in federated learning,” 2023. [Online]. Available: https://arxiv.org/abs/2306.11913

    arXiv 2023

  2. [26]

    QMGeo: Differentially private federated learning via stochastic quantization with mixed truncated geometric dis- tribution,

    Z. Wang and M. C. Gursoy, “QMGeo: Differentially private federated learning via stochastic quantization with mixed truncated geometric dis- tribution,” in Proc. IEEE International Conference on Communications , Denver, CO, USA, June 2024, pp. 2300–2305

    work page 2024

  3. [46]

    Supplement to Privacy-Preserving Quantized Federated Learning with Diverse Precision,

    D. Q. Nguyen, E. Perrins, M. Hashemi, S. A. V orobyov, D. J. Love, and T. Kim, “Supplement to Privacy-Preserving Quantized Federated Learning with Diverse Precision,” 2025. [Online]. Available: https: //github.com/quandku/Supplement-to-DPSQ-

    work page 2025

  4. [1]

    Federated learning: Challenges, methods, and future directions,

    T. Li, A. K. Sahu, A. Talwalkar, and V . Smith, “Federated learning: Challenges, methods, and future directions,” IEEE Signal Processing Magazine, vol. 37, no. 3, pp. 50–60, 2020

    2020

  5. [2]

    Federated learning and wireless commu- nications,

    Z. Qin, G. Y . Li, and H. Ye, “Federated learning and wireless commu- nications,” IEEE Wireless Communications, vol. 28, no. 5, pp. 134–140, 2021

    work page 2021

  6. [3]

    Federated learning: A signal processing perspective,

    T. Gafni, N. Shlezinger, K. Cohen, Y . C. Eldar, and H. V . Poor, “Federated learning: A signal processing perspective,” IEEE Signal Processing Magazine, vol. 39, no. 3, pp. 14–41, 2022

    work page 2022

  7. [4]

    Pretraining federated text models for next word prediction,

    J. Stremmel and A. Singh, “Pretraining federated text models for next word prediction,” in Advances in Information and Communication: Proceedings of the 2021 Future of Information and Communication Conference (FICC), Volume 2 . Springer, 2021, pp. 477–488

    work page 2021

  8. [5]

    FedNLP: Benchmarking federated learning methods for natural language processing tasks,

    B. Y . Lin, C. He, Z. Zeng, H. Wang, Y . Huang, C. Dupuy, R. Gupta, M. Soltanolkotabi, X. Ren, and S. Avestimehr, “FedNLP: Benchmarking federated learning methods for natural language processing tasks,”

    work page

Show all 53 references
  1. [6]

    Federated learning for connected and automated vehicles: A survey of existing approaches and challenges,

    V . P. Chellapandi, L. Yuan, C. G. Brinton, S. H. ˙Zak, and Z. Wang, “Federated learning for connected and automated vehicles: A survey of existing approaches and challenges,” IEEE Transactions on Intelligent Vehicles, vol. 9, no. 1, pp. 119–137, 2024

    2024

  2. [7]

    Federated learning for internet of things: A comprehensive survey,

    D. C. Nguyen, M. Ding, P. N. Pathirana, A. Seneviratne, J. Li, and H. V . Poor, “Federated learning for internet of things: A comprehensive survey,” IEEE Communications Surveys & Tutorials , vol. 23, no. 3, pp. 1622–1658, 2021

    2021

  3. [8]

    Scoring aided federated learning on long-tailed data for wireless iomt based healthcare system,

    L. Zhang, Y . Wu, L. Chen, L. Fan, and A. Nallanathan, “Scoring aided federated learning on long-tailed data for wireless iomt based healthcare system,” IEEE Journal of Biomedical and Health Informatics , vol. 28, no. 6, pp. 3341–3348, 2023

    2023

  4. [9]

    Federated learning for medical applications: A taxonomy, current trends, challenges, and future research directions,

    A. Rauniyar, D. H. Hagos, D. Jha, J. E. H ˚akeg˚ard, U. Bagci, D. B. Rawat, and V . Vlassov, “Federated learning for medical applications: A taxonomy, current trends, challenges, and future research directions,” IEEE Internet of Things Journal , vol. 11, no. 5, pp. 7374–7398, 2024

    2024

  5. [10]

    Learning differentially private recurrent language models,

    H. B. McMahan, D. Ramage, K. Talwar, and L. Zhang, “Learning differentially private recurrent language models,” 2018. [Online]. Available: https://arxiv.org/abs/1710.06963

    2018 arXiv

  6. [11]

    OpenFedLLM: Training large language models on decentral- ized private data via federated learning,

    R. Ye, W. Wang, J. Chai, D. Li, Z. Li, Y . Xu, Y . Du, Y . Wang, and S. Chen, “OpenFedLLM: Training large language models on decentral- ized private data via federated learning,” inProc. the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , New York, NY , USA...

    2024

  7. [12]

    The distributed discrete gaussian mechanism for federated learning with secure aggregation,

    P. Kairouz, Z. Liu, and T. Steinke, “The distributed discrete gaussian mechanism for federated learning with secure aggregation,” in Proc. the 38th International Conference on Machine Learning , vol. 139, 18–24 Jul 2021, pp. 5201–5212

    2021

  8. [13]

    Deep leakage from gradients,

    L. Zhu, Z. Liu, and S. Han, “Deep leakage from gradients,” in Proc. the 33rd International Conference on Neural Information Processing Systems, Red Hook, NY , USA, December 2019

    2019

  9. [14]

    Evaluating gradient inversion attacks and defenses in federated learning,

    Y . Huang, S. Gupta, Z. Song, K. Li, and S. Arora, “Evaluating gradient inversion attacks and defenses in federated learning,”Advances in Neural Information Processing Systems , vol. 34, pp. 7232–7241, 2021

    2021

  10. [15]

    Inverting gradients - how easy is it to break privacy in federated learning?

    J. Geiping, H. Bauermeister, H. Dr ¨oge, and M. Moeller, “Inverting gradients - how easy is it to break privacy in federated learning?” in Advances in Neural Information Processing Systems , vol. 33, 2020, pp. 16 937–16 947

    2020

  11. [16]

    The algorithmic foundations of differential privacy,

    C. Dwork and A. Roth, “The algorithmic foundations of differential privacy,”Found. Trends Theor. Comput. Sci., vol. 9, no. 3–4, p. 211–407, Aug 2014

    2014

  12. [17]

    Federated learning with differential privacy: Algorithms and performance analysis,

    K. Wei, J. Li, M. Ding, C. Ma, H. H. Yang, F. Farokhi, S. Jin, T. Q. S. Quek, and H. Vincent Poor, “Federated learning with differential privacy: Algorithms and performance analysis,” IEEE Transactions on Information Forensics and Security , vol. 15, pp. 3454–3469, 2020

    2020

  13. [18]

    Differentially private Aircomp federated learning with power adaptation harnessing receiver noise,

    Y . Koda, K. Yamamoto, T. Nishio, and M. Morikura, “Differentially private Aircomp federated learning with power adaptation harnessing receiver noise,” in IEEE Global Communications Conference , Taipei, Taiwan, December 2020, pp. 1–6

    2020

  14. [19]

    Privacy for free: Wireless federated learning via uncoded transmission with adaptive power control,

    D. Liu and O. Simeone, “Privacy for free: Wireless federated learning via uncoded transmission with adaptive power control,” IEEE Journal on Selected Areas in Communications, vol. 39, no. 1, pp. 170–185, 2021

    2021

  15. [20]

    Privacy amplification for federated learning via user sampling and wireless aggregation,

    M. S. E. Mohamed, W.-T. Chang, and R. Tandon, “Privacy amplification for federated learning via user sampling and wireless aggregation,” IEEE Journal on Selected Areas in Communications, vol. 39, no. 12, pp. 3821– 3835, 2021

    2021

  16. [21]

    Federated learning with local differential privacy: Trade-offs between privacy, utility, and communica- tion,

    M. Kim, O. G ¨unl¨u, and R. F. Schaefer, “Federated learning with local differential privacy: Trade-offs between privacy, utility, and communica- tion,” in Proc. International Conference on Acoustics, Speech and Signal Processing (ICASSP), Ontario, Canada, June 2021, pp. 2650–2654

    2021

  17. [22]

    User- level privacy-preserving federated learning: Analysis and performance optimization,

    K. Wei, J. Li, M. Ding, C. Ma, H. Su, B. Zhang, and H. V . Poor, “User- level privacy-preserving federated learning: Analysis and performance optimization,” IEEE Transactions on Mobile Computing , vol. 21, no. 9, pp. 3388–3401, 2022

    2022

  18. [23]

    Time-varying noise perturbation and power control for differential-privacy-preserving wireless federated learning,

    D. Q. Nguyen and T. Kim, “Time-varying noise perturbation and power control for differential-privacy-preserving wireless federated learning,” in Proc. the 57th Asilomar Conference on Signals, Systems, and Com- puters, Pacific Grove, CA, USA, October 2023, pp. 656–660

    2023

  19. [24]

    Breaking the communication- privacy-accuracy trilemma,

    W.-N. Chen, P. Kairouz, and A. Ozgur, “Breaking the communication- privacy-accuracy trilemma,” inAdvances in Neural Information Process- ing Systems, vol. 33, 2020, pp. 3312–3324

    2020

  20. [27]

    Joint privacy en- hancement and quantization in federated learning,

    N. Lang, E. Sofer, T. Shaked, and N. Shlezinger, “Joint privacy en- hancement and quantization in federated learning,” IEEE Transactions on Signal Processing , vol. 71, pp. 295–310, 2023

    2023

  21. [28]

    Dynamic aggregation for heterogeneous quantization in federated learning,

    S. Chen, C. Shen, L. Zhang, and Y . Tang, “Dynamic aggregation for heterogeneous quantization in federated learning,” IEEE Transactions on Wireless Communications, vol. 20, no. 10, pp. 6804–6819, 2021

    2021

  22. [29]

    Federated learning with heterogeneous quantization bit allocation and aggregation for internet of things,

    S. Chen, L. Li, G. Wang, M. Pang, and C. Shen, “Federated learning with heterogeneous quantization bit allocation and aggregation for internet of things,” IEEE Internet of Things Journal , vol. 11, no. 2, pp. 3132–3143, 2024

    2024

  23. [30]

    Quantized federated learning under transmission delay and outage constraints,

    Y . Wang, Y . Xu, Q. Shi, and T.-H. Chang, “Quantized federated learning under transmission delay and outage constraints,” IEEE Journal on Selected Areas in Communications , vol. 40, no. 1, pp. 323–341, 2022

    2022

  24. [31]

    QSGD: communication-efficient SGD via gradient quantization and encoding,

    D. Alistarh, D. Grubic, J. Z. Li, R. Tomioka, and M. V ojnovic, “QSGD: communication-efficient SGD via gradient quantization and encoding,” in Proc. the 31st International Conference on Neural Information Pro- cessing Systems, Red Hook, NY , USA, December 2017, p. 1707–1718

    2017

  25. [32]

    FedPAQ: A communication-efficient federated learning method with periodic averaging and quantization,

    A. Reisizadeh, A. Mokhtari, H. Hassani, A. Jadbabaie, and R. Pedarsani, “FedPAQ: A communication-efficient federated learning method with periodic averaging and quantization,” in International Conference on Artificial Intelligence and Statistics , June 2020, pp. 2021–2031

    2020

  26. [33]

    Quantized federated learning under transmission delay and outage constraints,

    Y . Wang, Y . Xu, Q. Shi, and T.-H. Chang, “Quantized federated learning under transmission delay and outage constraints,” IEEE Journal on Selected Areas in Communications , vol. 40, pp. 323–341, 2021

    2021

  27. [34]

    The dual simplex algorithm for bounded variables,

    H. M. Wagner, “The dual simplex algorithm for bounded variables,” Naval Research Logistics Quarterly , vol. 5, no. 3, pp. 257–261, 1958

    1958

  28. [35]

    D. G. Luenberger and Y . Ye, Linear and Nonlinear Programming . Springer, 1984, vol. 2

    1984

  29. [36]

    A new polynomial-time algorithm for linear program- ming,

    N. Karmarkar, “A new polynomial-time algorithm for linear program- ming,” in Proc. the 16th Annual ACM Symposium on Theory of Com- puting, New York, NY , USA, December 1984, p. 302–311

    1984

  30. [37]

    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 Proc. the 20th International Conference on Artificial Intelligence and Statistics , vol. 54, Florida, USA, April 2017, pp. 1273–1282

    2017

  31. [38]

    Efficient mini-batch training for stochastic optimization,

    M. Li, T. Zhang, Y . Chen, and A. J. Smola, “Efficient mini-batch training for stochastic optimization,” in Proc. the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , New York, USA, August 2014, pp. 661–670

    2014

  32. [39]

    Deep learning with differential privacy,

    M. Abadi, A. Chu, I. Goodfellow, H. B. McMahan, I. Mironov, K. Tal- war, and L. Zhang, “Deep learning with differential privacy,” in Proc. ACM SIGSAC Conference on Computer and Communications Security , Vienna, Austria, October 2016, p. 308–318

    2016

  33. [40]

    Federated learning over noisy channels: Conver- gence analysis and design examples,

    X. Wei and C. Shen, “Federated learning over noisy channels: Conver- gence analysis and design examples,” IEEE Transactions on Cognitive Communications and Networking , vol. 8, no. 2, pp. 1253–1268, 2022

    2022

  34. [41]

    Optimizing federated averaging over fading channels,

    Y . Mu, C. Shen, and Y . C. Eldar, “Optimizing federated averaging over fading channels,” in IEEE International Symposium on Information Theory (ISIT), Espoo, Finland, June 2022, pp. 1277–1281

    2022

  35. [42]

    Nesterov, Introductory Lectures on Convex Optimization: A Basic Course, 1st ed

    Y . Nesterov, Introductory Lectures on Convex Optimization: A Basic Course, 1st ed. Springer Publishing Company, Incorporated, 2014

    2014

  36. [43]

    Revisiting gradient clip- ping: Stochastic bias and tight convergence guarantees,

    A. Koloskova, H. Hendrikx, and S. U. Stich, “Revisiting gradient clip- ping: Stochastic bias and tight convergence guarantees,” in International Conference on Machine Learning, Hawaii, USA, July 2023, pp. 17 343– 17 363

    2023

  37. [44]

    A guide through the zoo of biased SGD,

    Y . Demidovich, G. Malinovsky, I. Sokolov, and P. Richt ´arik, “A guide through the zoo of biased SGD,” Advances in Neural Information Processing Systems, vol. 36, pp. 23 158–23 171, December 2023

    2023

  38. [45]

    On biased compression for distributed learning,

    A. Beznosikov, S. Horv ´ath, P. Richt ´arik, and M. Safaryan, “On biased compression for distributed learning,” Journal of Machine Learning Research, vol. 24, no. 276, pp. 1–50, 2023

    2023

  39. [47]

    Amplitude- varying perturbation for balancing privacy and utility in federated learning,

    X. Yuan, W. Ni, M. Ding, K. Wei, J. Li, and H. V . Poor, “Amplitude- varying perturbation for balancing privacy and utility in federated learning,” IEEE Transactions on Information Forensics and Security , vol. 18, pp. 1884–1897, 2023

    2023

  40. [48]

    Li and X

    D. Li and X. Sun, Nonlinear Integer Programming . Springer, 2006, vol. 84

    2006

  41. [49]

    An approach to sensor network throughput enhancement by PHY-Aided MAC,

    T. Kim, D. J. Love, M. Skoglund, and Z.-Y . Jin, “An approach to sensor network throughput enhancement by PHY-Aided MAC,” IEEE Transactions on Wireless Communications, vol. 14, no. 2, pp. 670–684, 2015

    2015

  42. [50]

    Backpropagation applied to handwritten zip code recognition,

    Y . LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel, “Backpropagation applied to handwritten zip code recognition,” Neural Computation, vol. 1, no. 4, pp. 541–551, 1989

    1989

  43. [51]

    MNIST handwritten digit database,

    Y . LeCun and C. Cortes, “MNIST handwritten digit database,” 2010. [Online]. Available: http://yann.lecun.com/exdb/mnist/

    2010

  44. [52]

    Image quality assess- ment: from error visibility to structural similarity,

    Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli, “Image quality assess- ment: from error visibility to structural similarity,” IEEE Transactions on Image Processing , vol. 13, no. 4, pp. 600–612, 2004

    2004

  45. [2022]

    Available: https://arxiv.org/abs/2104.08815

    [Online]. Available: https://arxiv.org/abs/2104.08815

    arXiv

Pith tools

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