Pith. sign in

REVIEW 3 major objections 6 minor 47 references

Enhancing Convergence, Privacy and Fairness for Wireless Personalized Federated Learning: Quantization-Assisted Min-Max Fair Scheduling

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Wireless personalized federated learning can treat the rounding errors of low-bit quantization as part of its differential-privacy budget, and a min-max scheduler built on the resulting convergence bounds improves accuracy, worst-client…

desk verdict Quantization-assisted DP for wireless PFL is a real idea, but the optimal scheduling guarantee is not proved as written. read the letter →

arxiv 2506.02422 v1 pith:2MJZX5BO submitted 2025-06-03 cs.DC cs.LG

Xiyu Zhao , Qimei Cui , Ziqiang Du , Weicai Li , Xi Yu , Wei Ni , Ji Zhang , Xiaofeng Tao
show 1 more author
Ping Zhang
This is my paper
classification cs.DCcs.LG
keywords personalizedfederatedlearningwirelessdifferentialprivacyquantizationmin-maxfairnessschedulingconvergenceanalysisOFDMA
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 in wireless personalized federated learning (WPFL) over OFDMA subchannels, the rounding errors of low-bit quantization can be counted as a privacy resource rather than a nuisance: by adding Gaussian noise and then quantizing, a given differential-privacy budget can be met with less injected noise than noise-only schemes. The paper also derives per-client convergence upper bounds that fold in quantization error, DP noise, and imperfect wireless channels, and it uses the worst of those bounds as a scheduling objective. Minimizing that worst bound—through client selection, subchannel allocation, power control, and adaptive choice of learning rates and the personalization weight—decouples into a polynomial-time matching problem plus a convex per-client optimization, which is what makes a min-max fair scheduler practical. Reported experiments on MNIST, FMNIST, and CIFAR-10 show the scheduler beating round-robin, random, and fixed-configuration alternatives by 87.08% in accuracy, 16.21% in worst-client test loss, and 38.37% in a standard fairness index under a CNN model. A reader should care because privacy, communication efficiency, and fairness usually pull against each other in federated learning; the claim is that quantization error, which is already present for communication efficiency, can be made to help with privacy and fairness at the same time.

What carries the argument

The load-bearing object is the quantization-assisted Gaussian mechanism $M_Q(u_n,D_n)=Q(u_n(D_n)+z_n)$, whose output-distribution analysis (Theorem 1) turns the quantization step size—equivalently the maximum quantization error $E^L_{\max}$—into a term of the differential-privacy budget, so coarser quantization buys a smaller required $\sigma_{DP}$. The second load-bearing object is the per-round personalized-model convergence bound $\Phi_n^{t+1}$ from Theorem 3, which separates the shared aggregation error $\Theta_L^t$ from client-specific terms in the federated-learning rate $\eta_{F,n}^t$, the personalized-learning rate $\eta_{P,n}^{t+1}$, and the FL-PL weighting coefficient $\lambda_n^{t+1}$. That separation creates the nested structure the scheduler exploits: minimizing $\Theta_L^t$ over client selection, subchannel allocation, and power control is a maximum-weight bipartite matching, while minimizing each $\Phi_n^{t+1}$ becomes a convex program in the personalized-learning rate over the feasible regions $\Omega_0^{t+1}\cup\Omega_1^{t+1}$ when $\mu<2$ and the common convergence rate $\epsilon_P^{t+1}$ is chosen in $[1-\mu^2/4,1)$. Together these pieces convert the min-max fairness objective into a sequentially decoupled policy with polynomial overall complexity.

What would settle it

Audit the quantization-assisted Gaussian mechanism by computing the empirical maximum divergence between outputs produced from two datasets that differ in one sample, using the paper's stated clipping threshold $C$, quantization bits $R$, and noise scale $\sigma_{DP}$; if the measured privacy loss exceeds the claimed $(\epsilon_Q,\delta_Q)$ budget, the central claim that quantization errors can be counted toward the differential-privacy budget is wrong.

Watch

Extended reading notes

Core claim

The paper's central claim is that quantization error is not merely tolerated but actively useful: the quantization-assisted Gaussian mechanism $M_Q(u,D)=Q(u(D)+z)$ satisfies $(\epsilon_Q,\delta_Q)$-differential privacy with a budget that explicitly depends on the maximum quantization error $E^L_{\max}$, so for a fixed privacy target the required Gaussian noise scale $\sigma_{DP}$ can be smaller than in a noise-only Gaussian mechanism. The paper further claims that the per-round convergence of each personalized model is bounded by a quantity $\Phi_n^{t+1}$ depending on transmit powers, subchannel assignments, transmission error probabilities, quantization resolutions, the federated-learning and personalized-learning rates, and the FL-PL weighting coefficient, and that the worst of these bounds across clients can be minimized. The minimization is solved by revealing a nested structure: once the shared aggregation-error term $\Theta_L^t$ is minimized by a maximum-weight matching over clients and subchannels, each client's remaining bound depends only on its own learning rate and weighting coefficient, so the min-max problem separates into a matching step and a per-client convex optimization. Under the assumption $\mu<2$ on the strong-convexity constant, the resulting policy is claimed optimal, and experiments on DNN, MLR, and CNN models over MNIST, FMNIST, and CIFAR-10 report large improvements over round-robin, random-selection, and fixed-configuration scheduling in accuracy, worst-client test loss, and fairness.

Load-bearing premise

The whole argument assumes every client's loss function is a well-behaved, bowl-shaped objective (strongly convex and smooth) with bounded gradients and with each client's ideal model close to the global ideal; the tested neural networks with ReLU activations and cross-entropy loss do not actually satisfy these conditions, so the convergence bounds and the scheduling decisions derived from them are only strictly valid under that idealized assumption.

Editorial extensions

If this is right

  • For a fixed privacy budget $(\epsilon_Q,\delta_Q)$, the required Gaussian noise scale $\sigma_{DP}$ decreases as quantization error grows, so compressing model updates and protecting privacy become complementary goals rather than competing ones.
  • The convergence bound $\Phi_n^{t+1}$ gives a computable per-client objective for fairness: scheduling decisions that minimize the worst bound directly target the worst-client expected loss.
  • Because the min-max problem decouples, the joint scheduling-and-configuration problem is solvable in polynomial time instead of as a mixed-integer program, with overall complexity $O(N^3 + N\log(1/\alpha))$.
  • Within the stated strongly-convex regime, the policy is claimed optimal, meaning no other joint choice of client selection, channel allocation, power control, learning rates, and weighting coefficients can strictly improve the worst-client convergence bound while keeping convergence rates equal across clients.

Reading between the lines

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

  • The paper does not claim it, but the same accounting principle should extend to any compression scheme with a bounded error: if sparsification or low-rank compression admits a maximum-error bound, that bound could be substituted into the Theorem 1 reasoning to reduce $\sigma_{DP}$ further.
  • A testable extension the paper leaves open is the boundary $\mu\geq 2$: the convex-region argument in Theorem 5 is restricted to $\mu<2$, so running the scheduler on a model whose empirical strong-convexity constant is at least 2 would delimit where the optimality claim holds.
  • The fairness objective is defined through the worst convergence bound; the same machinery could be re-pointed at worst-case accuracy, tail performance, or group-fairness constraints, but the paper only evaluates the loss-based fairness metric.
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 / 6 minor

Summary. This paper studies wireless personalized federated learning (WPFL) over OFDMA links and makes three intertwined claims: (i) a quantization-assisted Gaussian differential privacy mechanism that exploits quantization error as part of the privacy budget; (ii) convergence upper bounds for both the FL global model and each client's PL model under DP noise, quantization, and imperfect channels; and (iii) a min-max fair scheduling and configuration strategy that, by decoupling the min-max problem into client selection, channel allocation, power control, and learning-rate/weight adjustment, is claimed to be optimal. The experimental section reports large improvements over round-robin, random selection, and non-adjustment baselines on MNIST, FMNIST, and CIFAR10 with DNN, MLR, and CNN models, as well as comparisons against pFedMe, APPLE, FedAMP, and FedALA. The paper is well structured and contains substantial derivations in the appendices.

Significance. If the theoretical chain is fully established, the paper would be a valuable integrated contribution: it combines privacy accounting, convergence analysis, and scheduling in a single WPFL framework, and the idea of treating quantization error as a privacy resource is timely and potentially useful. The paper also ships a fairly extensive evaluation against both DP baselines and PFL baselines, including communication-overhead measurements. The main value of the paper, however, rests on the claim that Algorithm 2 provides an optimal min-max fair scheduling strategy; that claim is currently not fully supported by the proofs as written. The convergence and privacy analyses are long and not machine-checked, and the experimental validation does not by itself fill the gap because the scheduling objective is derived from the theoretical bounds.

major comments (3)
  1. [Section VI-A, Eq. (32) and Problem P1] The conversion of the min-max problem P into the max-min problem P1 is asserted without proof. In general, max over variables of min over clients is at most min over clients of max over variables, and equality requires a saddle-point or minimax-theorem condition that is not established. The clients are coupled not only through the common term Theta_L^t and the shared channel variables c^t, p^t, but also through constraint C1, which forces epsilon_{P,n}^{t+1} = epsilon_P^{t+1} for all n; this coupling is precisely the kind of structure that can prevent a trivial interchange of min and max. The subsequent decoupling into P2 and P4, and the statement in Section VI-D that 'the solution to Problem P1 is optimal', inherit this gap. The authors should either prove the equality under their assumptions or downgrade the optimality claim to a heuristic policy whose performance is validated empirically.
  2. [Appendix F, Theorem 5] The convexity proof of Phi_n^{t+1} on Omega_0^{t+1} union Omega_1^{t+1} is incomplete. The displayed second derivative has a denominator a0 * eta^4 * (b0 - mu*eta + eta^2)^3, and the sign of this denominator over the feasible set is not discussed. The four terms H1-H4 are asserted to be positive 'by analyzing the monotonicity', but that analysis is not shown. Since Theorem 5 is the basis for the claim that Problem P7 is solved to global optimality by convex optimization methods (Section VI-D), this is a load-bearing gap. A complete analytical proof, or an explicit statement that parts of the convexity verification are numerical rather than analytical, is required.
  3. [Section V, Assumption 1, and Section VII] The convergence theorems and the scheduling objective depend on Assumption 1, which requires every local loss function F_n to be mu-strongly convex and L-smooth with bounded stochastic gradients. However, the experimental evaluation in Section VII uses ReLU-based CNNs with cross-entropy loss, which are neither strongly convex nor L-smooth. The footnote that estimates L and mu empirically provides numerical constants, but it does not establish that Assumption 1 holds for these models. Consequently, Theorems 2-4 and the bound Phi_n^{t+1} in Eq. (30c) may not describe the actual training dynamics in the CNN experiments. The authors should either restrict the formal claims to models satisfying Assumption 1, justify the assumption through a smoothed or regularized surrogate objective, or provide direct evidence, such as tracking the derived bound against the measured E[||tilde{varpi}_n^t - varpi_n^*||^2], that the analysis remains predictive for CNNs.
minor comments (6)
  1. [Eq. (7)] The definition of E_max^G uses Delta_L/2; it should be Delta_G/2 = beta_G * C.
  2. [Section VI-A, Problem P1] The quantifier order in P1 is written as max_{n in N} min_{...}, which is not the intended max-min reformulation; the maximization should be over the optimization variables and the minimization over clients. Please correct the notation and define the strategy sets explicitly.
  3. [Section VII, Fig. 3] The legend entry 'Error-free (t_n, U = 0)' is undefined; it presumably means rho_{n,L}^t = 0 and rho_{n,G}^{t+1} = 0.
  4. [References and Related Work] Reference [18] contains the typo 'inn Proc.', and several sentences in Related Work are ungrammatical, e.g., 'None has taken fairness among the PL models of different participants.'
  5. [Theorem 1] The subsequent '-DP' expression with log(psi/psi_1) and log(psi'/psi'_1) is undefined if psi_1 or psi'_1 vanishes; the authors should specify conditions on sigma_DP, C, and R that exclude this case.
  6. [Algorithm 2, line 8] Algorithm 2 sets epsilon_P^{t+1} = epsilon_P, but Theorem 5's convexity guarantee holds only for epsilon_P in specific intervals; the algorithm should either check or enforce that epsilon_P lies in the required range.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: convergence and scheduling results are derived from stated assumptions and validated against external baselines; self-citations are non-load-bearing.

full rationale

The paper's derivation chain is self-contained rather than circular. The convergence upper bounds (Theorems 1–4) are obtained from explicitly stated Assumption 1 (strong convexity, L-smoothness, bounded gradients, bounded optimum distance), the quantization-assisted Gaussian mechanism, and standard DP composition results; the bounds are not fitted to the experimental curves. The scheduling policy in Section VI is a design-then-validate flow: it minimizes the derived per-round convergence bound via client selection, channel allocation, power control, and learning-rate adjustment, and the reported accuracy/fairness gains are measured against external scheduling and PFL baselines, not against the bound itself. The empirically estimated L and µ (Section VII footnote) are hyperparameters used inside the proposed optimizer, not fitted predictions of the reported outcomes. The paper contains self-citations (e.g., [16], [29], [45]) and cites a co-authored work [45] for the interior-point complexity bound, but none of these citations carries the load of the main derivation; the central results are derived from first-principles inequalities and standard references such as [1], [21], [22], [37], [43], and [46]. Two concerns raised by the skeptic are mathematical-correctness risks rather than circularity: the equivalence between the original min-max problem P and the max-min problem P1 is asserted without proof in Section VI-A, and Theorem 5's convexity proof in Appendix F asserts positivity of H1–H4 by monotonicity without displaying the analysis. These are unproven or incomplete steps, but they do not reduce the paper's conclusions to its own inputs by construction, and they do not involve fitted parameters being renamed as predictions. The overall circularity score is therefore 0.

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

The central claim rests on a standard set of FL convergence assumptions, a clipping-based sensitivity bound, and a QAM BER model. No new physical entities are introduced. The most fragile elements are the empirical assignment of L and µ to non-convex networks and the arbitrary constants in the bounds.

free parameters (5)
  • σ_DP = 0.001 to 0.024 depending on model and T0 (Table II)
    Standard deviation of Gaussian noise added to clipped local models; chosen via one-dimensional search to satisfy the target privacy budget (ϵ_Q, δ_Q) using Theorem 1, as described in Section IV. It enters all convergence bounds and the scheduling objective.
  • Clipping threshold C = 3 (MLR), 7 (DNN), 20 (CNN)
    Hand-set per model in Section VII; controls the sensitivity bound and the quantization range, and appears in every theorem.
  • L and µ (smoothness and convexity constants) = e.g., L=1.32, µ=0.27 for DNN/MNIST; L=0.29, µ=0.05 for CNN/FMNIST
    Empirically estimated from data via gradient differences (Section VII footnote); used in learning rate formulas and in the convergence analysis. For CNN/ReLU models these assumptions are not theoretically valid.
  • Arbitrary constants φ1, φ2, ϕ1, ϕ2
    Positive constants introduced in Theorems 2 and 3 to make inequalities hold. Their values affect the tightness of the bounds and the FL learning rate η_F = µ/(2(1+ϕ1)L^2), but no specific values are given.
  • ε_P (target PL contraction factor) = Input to Algorithm 2
    The target per-round PL convergence factor set as a constant in Algorithm 2; used to compute λ via (37). The feasible range is restricted by µ according to Theorem 5.
assumptions (5)
  • domain assumption Assumption 1: each local loss F_n is µ-strongly convex and L-smooth; stochastic gradients have bounded second moment E||∇F_n||^2 ≤ G0^2; and ||u*_n - ω*|| ≤ M.
    Assumption 1 in Section V is the foundation of Lemmas 1-3 and Theorems 2-4. It is standard in FL convergence analysis but is not satisfied by ReLU-based CNNs used in the experiments.
  • domain assumption Clipping (2) ensures ||u_n|| ≤ C, so for adjacent datasets the per-coordinate sensitivity is at most 2C.
    Used in the proof of Theorem 1 (Appendix A) to bound the probability ratio and derive δ_Q. The bound is deterministic because quantization saturates at the range [−C−3σ_DP, C+3σ_DP].
  • domain assumption M-QAM BER expression (13) from [38] models transmission errors as independent bit errors, and the error indicator vectors s are used to mix erroneous and error-free components.
    The channel model in Section III-B and Lemma 1 proof assumes this BER formula and independence of bit errors across elements and rounds.
  • ad hoc to paper Empirically estimated L and µ are valid for the actual models, including CNN/ReLU.
    The paper estimates L and µ from empirical gradient ratios (Section VII) and then applies the strong-convexity analysis to non-convex models. This is an unverified assumption about the loss landscape.
  • standard math Privacy composition theorem [37, Thm 3.16] composes per-round privacy losses over T0 rounds.
    Theorem 1 uses composition to convert the single-round max-divergence bound into the cumulative (ϵ_Q, δ_Q) guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Convergence, Privacy and Fairness for Wireless Personalized Federated Learning: Quantization-Assisted Min-Max Fair Scheduling." pith.science (2026). https://pith.science/paper/2MJZX5BO

@misc{pith2026250602422,
  author       = {Pith},
  title        = {Pith review of: Enhancing Convergence, Privacy and Fairness for Wireless Personalized Federated Learning: Quantization-Assisted Min-Max Fair Scheduling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2MJZX5BO}},
  note         = {Machine review of arXiv:2506.02422}
}
read the original abstract

Personalized federated learning (PFL) offers a solution to balancing personalization and generalization by conducting federated learning (FL) to guide personalized learning (PL). Little attention has been given to wireless PFL (WPFL), where privacy concerns arise. Performance fairness of PL models is another challenge resulting from communication bottlenecks in WPFL. This paper exploits quantization errors to enhance the privacy of WPFL and proposes a novel quantization-assisted Gaussian differential privacy (DP) mechanism. We analyze the convergence upper bounds of individual PL models by considering the impact of the mechanism (i.e., quantization errors and Gaussian DP noises) and imperfect communication channels on the FL of WPFL. By minimizing the maximum of the bounds, we design an optimal transmission scheduling strategy that yields min-max fairness for WPFL with OFDMA interfaces. This is achieved by revealing the nested structure of this problem to decouple it into subproblems solved sequentially for the client selection, channel allocation, and power control, and for the learning rates and PL-FL weighting coefficients. Experiments validate our analysis and demonstrate that our approach substantially outperforms alternative scheduling strategies by 87.08%, 16.21%, and 38.37% in accuracy, the maximum test loss of participating clients, and fairness (Jain's index), respectively.

Figures

Figures reproduced from arXiv: 2506.02422 by the authors.

Figure 1
Figure 1. The timeline of WPFL in the t-th round. ψ ′ = (1 − q)ψ ′ 1 + qQ  3σDP − E max L σDP  ; (24c) ψ ′ 1 = Q  2C + 3σDP − E max L σDP  , (24d) where q is the mini-batch sampling rate, and T0 is the maximum number of rounds in which each client can upload its FL local model due to privacy concerns. Proof. See Appendix A. When δQ = 0, MQ satisfies T0 max{ln ψ ψ1 , ln ψ ′ ψ′ 1 }-DP, which can be readily proved by substit… view at source ↗
Figure 2
Figure 2. PL model accuracy vs. the maximum number of rounds [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Accuracy of the PL model concerning the maximum number of rounds [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of fairness and the maximum test loss of all participating [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Comparison of accuracy vs. aggregation time [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 7
Figure 7. Figure 7: Comparison of fairness and maximum test loss of all participating [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 44 canonical work pages

  1. [28]

    Joint privacy enhancement and quantization in federated learning,

    N. Lang, E. Sofer, T. Shakedet al., “Joint privacy enhancement and quantization in federated learning,”IEEE Trans. Signal Process., vol. 71, pp. 295–310, 2023

    work page 2023

  2. [30]

    P2cefl: Privacy-preserving and commu- nication efficient federated learning with sparse gradient and dithering quantization,

    G. Wang, Q. Qi, R. Hanet al., “P2cefl: Privacy-preserving and commu- nication efficient federated learning with sparse gradient and dithering quantization,”IEEE Trans. Mob. Comput., 2024

    work page 2024

  3. [1]

    Ditto: Fair and robust federated learning through personalization,

    T. Li, S. Hu, A. Beiramiet al., “Ditto: Fair and robust federated learning through personalization,” inProc. 38th Int. Conf. Mach. Learn., vol. 139, 2021, pp. 6357–6368

    work page 2021

  4. [2]

    Over-the-air clustered federated learning,

    H. U. Sami and B. G ¨uler, “Over-the-air clustered federated learning,” IEEE Trans. Wirel. Commun., vol. 23, no. 7, pp. 7877–7893, 2023

    work page 2023

  5. [3]

    User-centric federated learning: Trading off wireless resources for personalization,

    M. Mestoukirdi, M. Zecchin, D. Gesbert, and Q. Li, “User-centric federated learning: Trading off wireless resources for personalization,” IEEE trans. mach. learn. commun. netw., vol. 1, pp. 346–359, 2023

    work page 2023

  6. [4]

    Ensemble federated learning with non-iid data in wireless networks,

    Z. Zhao, J. Wang, W. Hong, T. Q. Quek, Z. Ding, and M. Peng, “Ensemble federated learning with non-iid data in wireless networks,” IEEE Trans. Wirel. Commun., vol. 23, no. 4, pp. 3557–3571, 2024

    work page 2024

  7. [5]

    Hierarchical personalized federated learning over massive mobile edge computing networks,

    C. You, K. Guo, H. H. Yang, and T. Q. Quek, “Hierarchical personalized federated learning over massive mobile edge computing networks,”IEEE Trans. Wirel. Commun., vol. 22, no. 11, pp. 8141–8157, 2023

    work page 2023

  8. [6]

    FedMD: Heterogenous federated learning via model distillation,

    D. Li and J. Wang, “FedMD: Heterogenous federated learning via model distillation,” inProc. NeurIPS Workshop, 2019, pp. 1–8

    work page 2019

Show all 47 references
  1. [7]

    Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach,

    A. Fallah, A. Mokhtari, and A. Ozdaglar, “Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach,”Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, pp. 3557–3568, 2020

    2020

  2. [8]

    Personalized federated learning with differ- ential privacy and convergence guarantee,

    K. Wei, J. Li, C. Maet al., “Personalized federated learning with differ- ential privacy and convergence guarantee,”IEEE Trans. Inf. Forensics Security, vol. 18, pp. 4488–4503, 2023

    2023

  3. [9]

    Semi-synchronous personalized federated learning over mobile edge networks,

    C. You, D. Feng, K. Guoet al., “Semi-synchronous personalized federated learning over mobile edge networks,”IEEE Trans. Wirel. Commun., vol. 22, no. 4, pp. 2262–2277, 2022

    2022

  4. [10]

    Personalized federated learning with moreau envelopes,

    C. T Dinh, N. Tran, and J. Nguyen, “Personalized federated learning with moreau envelopes,”Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, pp. 21 394–21 405, 2020

    2020

  5. [11]

    Federated optimization in heterogeneous networks,

    T. Li, A. K. Sahu, M. Zaheeret al., “Federated optimization in heterogeneous networks,”Proc. 3rd Conf. Mach. Learn. Syst. (MLSys), vol. 2, pp. 429–450, 2020

    2020

  6. [12]

    Personalized cross-silo federated learning on non-IID data,

    Y . Huang, L. Chu, Z. Zhouet al., “Personalized cross-silo federated learning on non-IID data,” inProc. AAAI Conf. Artif. Intell., vol. 35, no. 9, 2021, pp. 7865–7873

    2021

  7. [13]

    Adapt to adaptation: Learning personalization for cross-silo federated learning,

    J. Luo and S. Wu, “Adapt to adaptation: Learning personalization for cross-silo federated learning,” inProc. 31th Int. Joint Conf. Artif. Intell. (IJCAI), vol. 2022, 2022, pp. 2166–2173

    2022

  8. [14]

    FedALA: Adaptive local aggregation for personalized federated learning,

    J. Zhang, Y . Hua, H. Wanget al., “FedALA: Adaptive local aggregation for personalized federated learning,” inProc. AAAI Conf. Artif. Intell., vol. 37, no. 9, 2023, pp. 11 237–11 244

    2023

  9. [15]

    Federated multi-task learning with non- stationary and heterogeneous data in wireless networks,

    H. Zhang, M. Tao, Y . Shiet al., “Federated multi-task learning with non- stationary and heterogeneous data in wireless networks,”IEEE Trans. Wirel. Commun., vol. 23, no. 4, pp. 2653–2667, 2023

    2023

  10. [16]

    Overview of AI and Communication for 6G Network: Fundamentals, Challenges, and Future Research Opportu- nities,

    Q. Cui, X. You, N. Weiet al., “Overview of AI and Communication for 6G Network: Fundamentals, Challenges, and Future Research Opportu- nities,”Sci China Inf Sci, vol. 68, no. 7, p. 171301, 2025

    2025

  11. [17]

    A joint learning and communications framework for federated learning over wireless networks,

    M. Chen, Z. Yang, W. Saadet al., “A joint learning and communications framework for federated learning over wireless networks,”IEEE Trans. Wirel. Commun., vol. 20, no. 1, pp. 269–283, 2020

    2020

  12. [18]

    Integrating over-the-air federated learning and non-orthogonal multiple access: What role can ris play?

    W. Ni, Y . Liu, Z. Yanget al., “Integrating over-the-air federated learning and non-orthogonal multiple access: What role can ris play?”IEEE Trans. Wirel. Commun., vol. 21, no. 12, pp. 10 083–10 099, 2022. 15

    2022

  13. [19]

    Star-ris integrated nonorthogonal multiple access and over-the-air federated learning: Framework, analysis, and optimization,

    W. Ni, Y . Liu, Y . C. Eldaret al., “Star-ris integrated nonorthogonal multiple access and over-the-air federated learning: Framework, analysis, and optimization,”IEEE Internet Things J., vol. 9, no. 18, pp. 17 136– 17 156, 2022

    2022

  14. [20]

    Semi-federated learning for collaborative intelligence in massive iot networks,

    W. Ni, J. Zheng, and H. Tian, “Semi-federated learning for collaborative intelligence in massive iot networks,”IEEE Internet Things J., vol. 10, no. 13, pp. 11 942–11 943, 2023

    2023

  15. [21]

    Deep learning with differential privacy,

    M. Abadi, A. Chu, I. Goodfellowet al., “Deep learning with differential privacy,” inn Proc. ACM SIGSAC Conf. Comput. Commun. Secur. (CCS), 2016, pp. 308–318

    2016

  16. [22]

    Federated learning with differential pri- vacy: Algorithms and performance analysis,

    K. Wei, J. Li, M. Dinget al., “Federated learning with differential pri- vacy: Algorithms and performance analysis,”IEEE Trans. Inf. Forensics Secur., vol. 15, pp. 3454–3469, 2020

    2020

  17. [23]

    Local differential privacy-based federated learning for internet of things,

    Y . Zhao, J. Zhao, M. Yanget al., “Local differential privacy-based federated learning for internet of things,”IEEE Internet Things J., vol. 8, no. 11, pp. 8836–8853, 2020

    2020

  18. [24]

    LDP-Fed: Federated learning with local differential privacy,

    S. Truex, L. Liu, K.-H. Chowet al., “LDP-Fed: Federated learning with local differential privacy,” inProc. 3rd ACM Int. Workshop Edge Syst. Anal. Netw., 2020, pp. 61–66

    2020

  19. [25]

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

    X. Yuan, W. Ni, M. Dinget al., “Amplitude-varying perturbation for balancing privacy and utility in federated learning,”IEEE Trans. Inf. Forensics Security, vol. 18, pp. 1884–1897, 2023

    2023

  20. [26]

    Differentially private over-the-air federated learning over MIMO fading channels,

    H. Liu, J. Yan, and Y .-J. A. Zhang, “Differentially private over-the-air federated learning over MIMO fading channels,”IEEE Trans. Wirel. Commun., vol. 41, no. 11, pp. 3533–3547, 2024

    2024

  21. [27]

    FedDual: Pair-wise gossip helps federated learning in large decentralized networks,

    Q. Chen, Z. Wang, H. Wanget al., “FedDual: Pair-wise gossip helps federated learning in large decentralized networks,”IEEE Trans. Inf. Forensics Security, vol. 18, pp. 335–350, 2022

    2022

  22. [29]

    Secure and efficient federated learning with provable performance guarantees via stochastic quantization,

    X. Lyu, X. Hou, C. Ren, X. Ge, P. Yang, Q. Cui, and X. Tao, “Secure and efficient federated learning with provable performance guarantees via stochastic quantization,”IEEE Trans. Inf. Forensics Secur., vol. 19, pp. 4070–4085, 2024

    2024

  23. [31]

    Towards personalized federated learning,

    A. Z. Tan, H. Yu, L. Cuiet al., “Towards personalized federated learning,”IEEE Trans. Neural Netw. Learn. Syst., vol. 34, no. 12, pp. 9587–9603, 2022

    2022

  24. [32]

    On privacy and personalization in cross- silo federated learning,

    K. Liu, S. Hu, S. Z. Wuet al., “On privacy and personalization in cross- silo federated learning,”Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 35, pp. 5925–5940, 2022

    2022

  25. [33]

    Differentially private federated multi-task learning framework for enhancing human-to-virtual connec- tivity in human digital twin,

    S. D. Okegbile, J. Cai, H. Zhenget al., “Differentially private federated multi-task learning framework for enhancing human-to-virtual connec- tivity in human digital twin,”IEEE J. Sel. Areas Commun., 2023

    2023

  26. [34]

    Personalized federated learning with differential privacy,

    R. Hu, Y . Guo, H. Liet al., “Personalized federated learning with differential privacy,”IEEE Internet Things J., vol. 7, no. 10, pp. 9530– 9539, 2020

    2020

  27. [35]

    Fair resource allocation in federated learning,

    T. Li, M. Sanjabi, A. Beiramiet al., “Fair resource allocation in federated learning,” inProc. Int. Conf. Learn. Represent., 2020, pp. 1–13

    2020

  28. [36]

    Federated learning meets multi- objective optimization,

    Z. Hu, K. Shaloudegi, G. Zhanget al., “Federated learning meets multi- objective optimization,”IEEE Trans. Netw. Sci. Eng., vol. 9, no. 4, pp. 2039–2051, 2022

    2022

  29. [37]

    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, pp. 211– 407, 2014

    2014

  30. [38]

    On the general BER expression of one-and two- dimensional amplitude modulations,

    K. Cho and D. Yoon, “On the general BER expression of one-and two- dimensional amplitude modulations,”IEEE Trans. Commun., vol. 50, no. 7, pp. 1074–1080, 2002

    2002

  31. [39]

    Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning,

    M. Nasr, R. Shokri, and A. Houmansadr, “Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning,” inProc. IEEE Symp. Secur. Privacy (SP), 2019, pp. 739–753

    2019

  32. [40]

    Model inversion attacks that exploit confidence information and basic countermeasures,

    M. Fredrikson, S. Jha, and T. Ristenpart, “Model inversion attacks that exploit confidence information and basic countermeasures,” inProc. 22nd ACM SIGSAC Conf. Comput. Commun. Secur., 2015, pp. 1322– 1333

    2015

  33. [41]

    Linear convergence of gradient and proximal-gradient methods under the polyak-łojasiewicz condition,

    H. Karimi, J. Nutini, and M. Schmidt, “Linear convergence of gradient and proximal-gradient methods under the polyak-łojasiewicz condition,” inProc. Joint Eur. Conf. Mach. Learn. Knowl. Discovery Databases. Springer, 2016, pp. 795–811

    2016

  34. [42]

    O’Searcoid,Metric spaces

    M. O’Searcoid,Metric spaces. Berlin, Germany: Springer, 2006

    2006

  35. [43]

    The Hungarian method for the assignment problem,

    H. W. Kuhn, “The Hungarian method for the assignment problem,” Naval Res. Logist. Quart., vol. 2, no. 1-2, pp. 83–97, 1955

    1955

  36. [44]

    Jungnickel and D

    D. Jungnickel and D. Jungnickel,Graphs, networks and algorithms. Berlin, Germany: Springer, 2005, vol. 3

    2005

  37. [45]

    Joint computation offloading and trajectory planning for uav-assisted edge computing,

    C. Sun, W. Ni, and X. Wang, “Joint computation offloading and trajectory planning for uav-assisted edge computing,”IEEE Trans. Wirel. Commun., vol. 20, no. 8, pp. 5343–5358, 2021

    2021

  38. [46]

    Adaptive federated learning in resource constrained edge computing systems,

    S. Wang, T. Tuor, T. Salonidiset al., “Adaptive federated learning in resource constrained edge computing systems,”IEEE J. Sel. Areas Commun., vol. 37, no. 6, pp. 1205–1221, June 2019. 16 APPENDIX A. Proof ofTheorem 1 We first focus on the DP mechanism in one communication rou...

    2019

  39. [47]

    Likewise, (64) is based on (63)

    (63) is based on the Cauchy- Schwarz inequality and (62). Likewise, (64) is based on (63). Similarly, we haveE ∥ ˆωt n,G −ω ∗ ∥2 ≤( G2 0 µ +M) 2. By plugging (62)–(64) into (61c), it readily follows that E ∥ ˜ϖt+1 n −ϖ ∗ n∥2 ≤ε t+1 P,n E ∥ ˜ϖt n −ϖ ∗ n∥2 + 1 + (λt+1 n )3 ηt+1 ...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.