Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Coding-Enforced Robust Secure Aggregation for Federated Learning Under Unreliable Communication

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 A coding scheme, SecCoGC, lets federated learning reconstruct the exact global model under unreliable links while canceling arbitrarily strong zero-sum privacy noise.

desk verdict The exact-recovery mechanism is sound, but the convergence proof analyzes a different update rule than Algorithm 1 implements, so the paper's central rate guarantee is unsupported. read the letter →

arxiv 2507.07565 v5 pith:BKJD7PSP submitted 2025-07-10 cs.IT math.IT

Shudi Weng , Chao Ren , Yizhou Zhao , Ming Xiao , Mikael Skoglund This is my paper · ORCID
classification cs.ITmath.IT
keywords federatedlearningsecureaggregationzero-sumprivacynoisegradientcodingunreliablecommunicationlocaldifferentialstragglermitigationfairness
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

This paper proposes SecCoGC, a way to do privacy-preserving federated learning over links that randomly fail. The central claim is that by combining zero-sum privacy noise with a fixed cyclic gradient-coding structure, the server can reconstruct exactly the same global model it would get with full, reliable participation, and the privacy noise cancels completely no matter how strong it is. If this is right, the usual privacy-utility trade-off disappears: each client can mask its update with arbitrarily large noise, and the coding structure still yields the exact aggregate whenever enough partial sums arrive. The paper also introduces fairness as a privacy metric, gives real-field constructions of zero-sum noise, analyzes differential privacy across client-to-client, relay, and server layers, and proves a non-convex convergence rate that is independent of the privacy level.

What carries the argument

The central object is the generator matrix $A$ together with the cyclic gradient-code pair $(G,C)$. $A \in \mathbb{R}^{K \times L}$ maps $L$ independent Gaussian components to $K$ secret keys so that each column sums to zero (correctness) and $\mathrm{rank}(A) = K-1$ (security); Construction IV.3 gives equal row norms (fairness), e.g., $A$ with diagonal entries $-2$ and off-diagonal entries $1$ for $K=5$, $\lambda^2=6$. The pair $(G,C)$ satisfies $CG = \mathbf{1}_{f \times K}$, so combining received complete partial sums with any row of $C$ whose support is contained in the received set reproduces the unweighted average of all $K$ local updates. The coding structure does the work: it fixes which coefficients each contribution receives, so the zero-sum noise is forced to cancel and partial participation cannot bias the update.

What would settle it

Run SecCoGC on a small known objective (e.g., a quadratic with known gradients) with two consecutive failed rounds followed by a success. Compare the model the server computes under Algorithm 1 with the accumulated-update value Lemma VI.2 assumes, $\sum_k \sum_{r=1}^{2} \Delta\Theta_k^{t_r,I}$. If they differ, the implemented protocol is not the one whose convergence is proved.

Watch

Extended reading notes

Core claim

The discovery is that the aggregation structure, not the noise cancellation, can be made the load-bearing element. Each client $k$ masks its local update $\Delta\Theta^{t,I}_k$ by adding a secret key $N^t_k$, with $\sum_k N^t_k = 0$. Using cyclic gradient codes whose allocation and combination matrices satisfy $CG = \mathbf{1}_{f \times K}$, relays form partial sums $S^t_k = \sum_{m \in U^t_k} g_{k,m} Y^t_m$. Whenever at least $K-s$ complete partial sums reach the server, the server picks a combinator row $c_f$ whose nonzero entries lie in the received set, and the aggregation in (12) equals $\frac{1}{K}\sum_k \Delta\Theta^{t,I}_k$ exactly; the noise terms cancel by the zero-sum property. The paper calls the decoding outcome binary: either this exact global model is obtained, or the result is not a meaningful aggregate. The paper further constructs fair real-field zero-sum noise with equal per-client power, derives per-layer local differential privacy guarantees, and proves a $99.74\%$-probability $O(1/\sqrt{T})$ convergence bound under arbitrary local solvers.

Load-bearing premise

The convergence proof assumes the server, after a series of failed rounds, rebuilds the model from all the local training done since the last successful round; the pseudocode instead transmits only the training done in the final round of that series, so the guarantee may describe a different protocol.

Editorial extensions

If this is right

  • Under a failure pattern where at least $K-s$ complete partial sums arrive, the server's update is exactly the full-participation global model, so random client dropout no longer biases the optimization trajectory.
  • Privacy noise strength no longer enters the convergence or accuracy picture, so clients can use arbitrarily strong masking without paying a utility price.
  • Failed rounds leave no partial information: the server either reconstructs the exact global model or obtains a meaningless combination, so an adversary cannot peel off partial sums from failed aggregations.
  • The fair construction (equal noise power per client, zero-sum columns) provides a privacy metric that can hold uniformly across clients under heterogeneous networks.
  • The analysis supports an $O(1/\sqrt{T})$ convergence rate with probability $99.74\%$ under arbitrary local solvers and privacy levels for algorithms with binary global-model recovery.

Reading between the lines

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

  • A failed round's output is deliberately meaningless, which suggests using failure detection as a privacy amplifier: the set of failed combiner rows could be chosen according to an adaptive privacy budget, something the paper does not explore.
  • The equal-power fair construction could be rescaled per client to match heterogeneous learning weights, giving a weighted fairness notion that the paper leaves for future work.
  • If the convergence-analysis mismatch is real, a protocol variant that actually transmits the accumulated update after failures would match the theorem; alternatively, the theorem's random $R_{t_s}$ bound could be rederived for incremental updates, a testable extension.
  • The coding structure also gives a natural integrity check for aggregation: any server-side deviation from the fixed combinator structure is detectable, which could be combined with verifiable aggregation.
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 SecCoGC, a secure aggregation scheme for federated learning over unreliable communication. Clients mask local updates with zero-sum privacy noise, relay the masked updates through a cyclic gradient code, and the server combines complete partial sums to reconstruct the global model exactly while canceling all noise. The paper also constructs fair real-field zero-sum noises, provides local differential privacy guarantees at several protocol layers under channel outages, and gives a convergence analysis claiming O(1/sqrt(T)) with probability 99.74%. Experiments on MNIST and CINIC-10 compare SecCoGC with private and non-private baselines under varying privacy levels, heterogeneity, and network asymmetry.

Significance. If the mechanism works as intended, the paper's contribution is meaningful: it combines cyclic gradient coding with zero-sum noise so that, on a successful round, the server obtains the exact global model while the privacy noise cancels completely, and it permits arbitrarily strong privacy without sacrificing utility. The algebraic core of the scheme, Eqs. (12)-(13), is correct for a round in which clients start from the global model, and the fair noise construction in Construction IV.3 is simple and elegant. The multi-layer privacy analysis is detailed. The main weakness is that the convergence theorem analyzes an update rule different from the one implemented in Algorithm 1, so the central theoretical claim is not established for the protocol as written. The issue is fixable by aligning the algorithm with the analysis, which is why I recommend major revision rather than rejection.

major comments (3)
  1. [Section VI, Algorithm 1, Eq. (115a)] The convergence analysis does not match the implemented protocol. Theorem VI.1 and Lemma VI.2 analyze the accumulated local update ΔΘ^{ts,R_ts I}_k over R_ts I local iterations since the last successful global model (Eq. (115a)). In Algorithm 1, however, line 14 sets ΔΘ^{t,I}_k = Θ^{t,I}_k − Θ^{t,0}_k for the current round only, and after a failed round the next round continues from the previous local model (Section III and line 9). The server aggregation (12) therefore sums only the per-round increments of the successful round; local progress made in failed rounds is neither transmitted nor retained by the server. Consequently, Eq. (115a) describes an update rule that Algorithm 1 does not execute, and the stated O(1/sqrt(T)) convergence with probability 99.74% is not established for SecCoGC as written. The protocol should be revised so that clients transmit cumulative deltas since the last successful global model, or the analysis must be re-done for the actual update rule.
  2. [Section VI, footnote 13, Theorem VI.1] The failure probability P_O is a load-bearing parameter: it appears in the geometric distribution of R_ts in Lemma VI.1 and in the final bound (44). The paper does not derive P_O for SecCoGC; footnote 13 refers the reader to the authors' companion manuscript [35], an arXiv preprint. For a self-contained journal paper, the expression for P_O under the SecCoGC channel model and code parameters should be included, or the theorem should state the result conditionally on P_O with its derivation properly referenced.
  3. [Section VI, Lemma VI.2, Appendix H] Lemma VI.2's bound (43) is finite only when the denominator 1−2η²G²R_ts²∥a∥₁² is positive. With the learning rate η = G^{-1}√(K/T) used in Appendix H, this condition is 2(K/T)R_ts²∥a∥₁² < 1. Since R_ts is a geometric random variable with unbounded support, this condition is not guaranteed for every realized R_ts, and the event where it fails is not included in the 99.74% probability event C1∩C2 from Lemma VI.1. The theorem should either state a sufficient condition on T and ∥a∥₁ or incorporate the violation event into the probability bound.
minor comments (5)
  1. [Algorithm 1] Algorithm 1 contains several pseudocode problems: line 2's while condition is malformed, line 9 sets Θ^{t+1,0}_k = Θ^{t,I}_k before Θ^{t,I}_k is computed on lines 11–13, and line 32 increments an unused variable r. The pseudocode should be aligned with the prose in Section III.
  2. [Lemma VI.1] Lemma VI.1 states Pr(C1,C2)≈0.99872≈0.9974, which contains two inconsistent numerical values; Theorem VI.1 uses 99.74%. Please correct this to a single value with a consistent derivation.
  3. [Section II-A and Algorithm 1, line 14] Eq. (4) defines ΔΘ^{t,I}_k = Θ^{t,I}_k − Θ^t, while Algorithm 1 line 14 uses Θ^{t,0}_k. After a failed round the relationship between Θ^t and Θ^{t,0}_k is ambiguous and should be clarified.
  4. [References [21], [32], [35]] Several load-bearing derivations are deferred to the authors' own manuscripts: Lemma VI.2 is obtained by substitution into [21, Lemma 3], P_O is deferred to [35], and the complete proofs are placed in [32]. For a journal submission, the key derivations should be self-contained or properly referenced to published versions.
  5. [Throughout] There are numerous typographical and formatting errors, e.g., 'interation', 'Guassian', 'efficient', and 'axillary', as well as inconsistent notation such as the unexplained η = 1/L√(K/T) in Appendix H versus the theorem statement. A careful proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the core coding identity and privacy analysis are self-contained, though the convergence proof has a non-circular protocol-mismatch gap.

full rationale

I find no step in the derivation chain that reduces to its own inputs. The central identity (12)-(13) rests on the cyclic-gradient-code property CG = 1_{f x K} from [34] together with the zero-sum condition (6); both are stated assumptions, not fitted or predicted quantities. The privacy theorems (V.1-V.5) are derived from the Gaussian mechanism and Construction IV.3, with details deferred to the paper's own supplementary file [32] rather than to a prior result that already contains the target claim. Lemmas IV.2-IV.4 are self-contained algebraic arguments. The convergence section is the only area with substantial self-citation: the failure probability P_O is said to be 'identical to the derivations in [35]', and Lemma VI.2 is obtained by substituting (42) into the authors' [21, Lemma 3]. These are reusable results from the authors' prior work, but Theorem VI.1 is stated for any algorithm with binary recovery and failure probability P_O, so P_O functions as an input parameter rather than as a fitted prediction; the cited lemma is parameter-free and does not presuppose the target convergence rate. Thus the self-citations are load-bearing but not circular under the definitions used here. Separately, there is a substantive correctness gap that is not circularity: Algorithm 1 line 14 transmits Delta_Theta^{t,I}_k = Theta^{t,I}_k - Theta^{t,0}_k, i.e. the current round's increment, whereas the proof of Theorem VI.1 in Eq. (115a) uses an accumulated update over all R_ts I local iterations since the last successful global model. The stated O(1/sqrt(T)) convergence with probability 99.74% is therefore not established for the protocol as written. That is a proof-protocol mismatch, not an input-equivalent derivation, and it does not make the paper's coding or privacy claims circular.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central reconstruction claim depends on standard GC properties and the zero-sum noise construction; the convergence claim adds strong FL assumptions and an unverified learning-rate condition; the privacy analysis at the global-model layer adds a Gaussian-updates assumption that is not generally true. The paper introduces no new physical entities, only design parameters and analytical assumptions.

free parameters (3)
  • lambda (privacy noise std) = 0.05/0.1 (MNIST), 0.03/0.05 (CINIC-10)
    Controls zero-sum noise power in Construction IV.3; user-set privacy level, not fitted to data, but central to the privacy claims.
  • gamma (off-diagonal nonzeros in A) = 2 (Example IV.2)
    Design parameter in Construction IV.3 determining the sparsity of the generator matrix; chosen by hand.
  • s (coding parameter, nonzeros per row of G) = 7 (experiments)
    Controls the number of neighbors each client sends to and the straggler tolerance K-s; set without further specification in Section VII.
assumptions (6)
  • domain assumption Local loss functions are G-smooth, lower bounded, unbiased gradients with bounded variance, and bounded dissimilarity (Assumptions VI.1-VI.3).
    Standard non-convex FL assumptions used for Theorem VI.1; cited from [21].
  • domain assumption Inter-client and uplink channels are independent Bernoulli random variables.
    Communication model in Section II-C; all subsequent probability calculations depend on link independence.
  • domain assumption Secret keys are generated by a key management server and each client knows only its own key; keys are Gaussian in Construction IV.3.
    Threat model and zero-sum noise setup in Sections III-IV; the server or clients cannot use key correlations adaptively.
  • ad hoc to paper Local model updates are i.i.d. Gaussian with covariance zeta^2 I_D for Theorem V.5.
    Assumed 'to facilitate theoretical analysis' in Section V-C; not true for SGD updates, and the privacy guarantees epsilon(6), epsilon(7) rely on this.
  • ad hoc to paper The learning rate and local steps satisfy 1 - 2*eta^2*G^2*R_ts^2*||a||_1^2 > 0 for every realized R_ts, so Lemma VI.2's bound is finite.
    Unstated; the inequality is required for Lemma VI.2 but the paper does not verify it for the random R_ts values.
  • standard math The cyclic gradient code from [34] has the property that any K-s complete rows allow a combination row c with CG = 1.
    Known gradient coding result; invoked in Section II-D and used in (12)-(13).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Coding-Enforced Robust Secure Aggregation for Federated Learning Under Unreliable Communication." pith.science (2026). https://pith.science/paper/BKJD7PSP

@misc{pith2026250707565,
  author       = {Pith},
  title        = {Pith review of: Coding-Enforced Robust Secure Aggregation for Federated Learning Under Unreliable Communication},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BKJD7PSP}},
  note         = {Machine review of arXiv:2507.07565}
}
read the original abstract

This work studies privacy-preserving federated learning (ppFL) under unreliable communication. In ppFL, zero-sum privacy noises enables privacy protection without sacrificing model accuracy, effectively overcoming the privacy-utility trade-off. However, in practice, unreliable communication can randomly disrupt the coordination of zero-sum noises, leading to aggregation errors and unpredictable partial participation, which severely harm the model accuracy and learning performance. To overcome these challenges, we propose a robust coding-enforced structured secure aggregation method, termed secure cooperative gradient coding (SecCoGC), which enables exact reconstruction of the global model under unreliable communication while allowing for arbitrarily strong privacy preservation. In this paper, a complete problem formulation and constructions of real-field zero-sum privacy noise are presented, and fairness is introduced as a privacy metric. Privacy across all protocol layers in SecCoGC is evaluated, accounting for the correlation among privacy noises and their linear combination under unreliable communication. Moreover, a distinct convergence analysis for the FL algorithm with a binary outcome for global model recovery is provided. Experimental results demonstrate that SecCoGC achieves strong resilience to unreliable communication while maintaining varying levels of privacy preservation, yielding test accuracy improvements of up to 20%-70% over existing benchmark methods.

Figures

Figures reproduced from arXiv: 2507.07565 by the authors.

Figure 1
Figure 1. An overview of the proposed SecCoGC method with simplified [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the number of training rounds [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Performance comparison under varying privacy levels, network conditions, and data heterogeneity on the MNIST dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Performance comparison under varying privacy levels, network conditions, and data heterogeneity on the CINIC dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 34 canonical work pages

  1. [35]

    Cooperative Gradient Coding

    S. Weng, M. Xiao, C. Ren, and M. Skoglund, “Cooperative gradient coding,” arXiv preprint arXiv:2507.05230, 2025

    work page Pith review arXiv 2025

  2. [1]

    Federated learning- empowered mobile network management for 5g and beyond networks: From access to core,

    J. Lee, F. Solat, T. Y. Kim, and H. V. Poor, “Federated learning- empowered mobile network management for 5g and beyond networks: From access to core,” IEEE Communications Surveys & Tutorials, vol. 26, no. 3, pp. 2176–2212, 2024

    work page 2024

  3. [2]

    Advances and open challenges in federated foundation models,

    C. Ren, H. Yu, H. Peng, X. Tang, B. Zhao, L. Yi, A. Z. Tan, Y. Gao, A. Li, X. Li et al., “Advances and open challenges in federated foundation models,” IEEE Communications Surveys & Tutorials, 2025

    work page 2025

  4. [3]

    The algorithmic foundations of differential privacy,

    C. Dwork, A. Roth et al., “The algorithmic foundations of differential privacy,” Foundations and Trends® in Theoretical Computer Science, vol. 9, no. 3–4, pp. 211–407, 2014

    work page 2014

  5. [4]

    Privacy preserving semi-decentralized mean estimation over intermittently-connected networks,

    R. Saha, M. Seif, M. Yemini, A. J. Goldsmith, and H. V. Poor, “Privacy preserving semi-decentralized mean estimation over intermittently-connected networks,” IEEE Transactions on Signal Processing, 2024

    work page 2024

  6. [5]

    Attack-aware noise calibration for differential privacy,

    B. Kulynych, J. F. Gomez, G. Kaissis, F. du Pin Calmon, and C. Troncoso, “Attack-aware noise calibration for differential privacy,” Advances in Neural Information Processing Systems, vol. 37, pp. 134 868–134 901, 2024

    work page 2024

  7. [6]

    Locally differen- tially private online federated learning with correlated noise,

    J. Zhang, L. Zhu, D. Fay, and M. Johansson, “Locally differen- tially private online federated learning with correlated noise,” IEEE Transactions on Signal Processing, 2025

    work page 2025

  8. [7]

    Optimizing Privacy-Utility Trade-off in Decentralized Learning with Generalized Correlated Noise

    A. Rodio, Z. Chen, and E. G. Larsson, “Optimizing privacy- utility trade-off in decentralized learning with generalized correlated noise,” 2025. [Online]. A vailable: https://arxiv.org/ abs/2501.14644

    work page Pith review arXiv 2025

Show all 37 references
  1. [8]

    Dap-fl: Federated learning flourishes by adaptive tuning and secure aggregation,

    Q. Chen, Z. Wang, J. Chen, H. Yan, and X. Lin, “Dap-fl: Federated learning flourishes by adaptive tuning and secure aggregation,” IEEE Transactions on Parallel and Distributed Systems, vol. 34, no. 6, pp. 1923–1941, 2023

    1923

  2. [9]

    Practical secure aggregation for privacy-preserving machine learning,

    K. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan, S. Patel, D. Ramage, A. Segal, and K. Seth, “Practical secure aggregation for privacy-preserving machine learning,” in proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, 2017, ...

    2017

  3. [10]

    Securing secure aggregation: Mitigating multi-round privacy leakage in federated learning,

    J. So, R. E. Ali, B. Güler, J. Jiao, and A. S. A vestimehr, “Securing secure aggregation: Mitigating multi-round privacy leakage in federated learning,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 8, 2023, pp. 9864–9873

    2023

  4. [11]

    Privacy-preserving deep learning via additively homomorphic encryption,

    L. T. Phong, Y. Aono, T. Hayashi, L. Wang, and S. Moriai, “Privacy-preserving deep learning via additively homomorphic encryption,” IEEE Transactions on Information Forensics and Security, vol. 13, no. 5, pp. 1333–1345, 2018

    2018

  5. [12]

    Turbo-aggregate: Breaking the quadratic aggregation barrier in secure federated learning,

    J. So, B. Güler, and A. S. A vestimehr, “Turbo-aggregate: Breaking the quadratic aggregation barrier in secure federated learning,” IEEE Journal on Selected Areas in Information Theory, vol. 2, no. 1, pp. 479–489, 2021

    2021

  6. [13]

    Information theoretic secure aggregation with user dropouts,

    Y. Zhao and H. Sun, “Information theoretic secure aggregation with user dropouts,” IEEE Transactions on Information Theory, vol. 68, no. 11, pp. 7471–7484, 2022

    2022

  7. [14]

    Fundamental limits of hierarchical secure aggregation with cyclic user association,

    X. Zhang, Z. Li, K. Wan, H. Sun, M. Ji, and G. Caire, “Fundamental limits of hierarchical secure aggregation with cyclic user association,” 2025. [Online]. A vailable: https: //arxiv.org/abs/2503.04564

    2025

  8. [15]

    Mds variable generation and secure sum- mation with user selection,

    Y. Zhao and H. Sun, “Mds variable generation and secure sum- mation with user selection,” IEEE Transactions on Information Theory, vol. 71, no. 4, pp. 3129–3141, 2025

    2025

  9. [16]

    Over-the-air feder- ated learning with privacy protection via correlated additive perturbations,

    J. Liao, Z. Chen, and E. G. Larsson, “Over-the-air feder- ated learning with privacy protection via correlated additive perturbations,” in 2022 58th Annual Allerton Conference on Communication, Control, and Computing (Allerton), 2022, pp. 1–8

    2022

  10. [17]

    Cod- edpaddedfl and codedsecagg: Straggler mitigation and secure aggregation in federated learning,

    R. Schlegel, S. Kumar, E. Rosnes, and A. G. i Amat, “Cod- edpaddedfl and codedsecagg: Straggler mitigation and secure aggregation in federated learning,” IEEE Transactions on Com- munications, vol. 71, no. 4, pp. 2013–2027, 2023

    2013

  11. [18]

    SCAFFOLD: Stochastic controlled averaging for federated learning,

    S. P. Karimireddy, S. Kale, M. Mohri, S. Reddi, S. Stich, and A. T. Suresh, “SCAFFOLD: Stochastic controlled averaging for federated learning,” in International Conference on Machine Learning, vol. 119, 2020, pp. 5132–5143

    2020

  12. [19]

    Tackling the objective inconsistency problem in heterogeneous federated optimization,

    J. Wang, Q. Liu, H. Liang, G. Joshi, and H. V. Poor, “Tackling the objective inconsistency problem in heterogeneous federated optimization,” in Advances in Neural Information Processing Systems, vol. 33, 2020, pp. 7611–7623

    2020

  13. [20]

    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, 2021

    2021

  14. [21]

    Heterogeneity- aware client sampling: A unified solution for consistent federated learning,

    S. Weng, C. Ren, M. Xiao, and M. Skoglund, “Heterogeneity- aware client sampling: A unified solution for consistent federated learning,” arXiv preprint:2505.11304, 2025

    2025

  15. [22]

    Feder- ated learning in heterogeneous networks with unreliable com- munication,

    P. Zheng, Y. Zhu, Y. Hu, Z. Zhang, and A. Schmeink, “Feder- ated learning in heterogeneous networks with unreliable com- munication,” IEEE Transactions on Wireless Communications, vol. 23, no. 4, pp. 3823–3838, 2023

    2023

  16. [23]

    Communication- efficient device scheduling for federated learning using stochastic optimization,

    J. Perazzone, S. Wang, M. Ji, and K. S. Chan, “Communication- efficient device scheduling for federated learning using stochastic optimization,” in IEEE INFOCOM 2022-IEEE Conference on Computer Communications, 2022, pp. 1449–1458

    2022

  17. [24]

    Colrel: Collaborative relaying for federated learning over intermittently connected networks,

    R. Saha, M. Yemini, E. Ozfatura, D. Gunduz, and A. Gold- smith, “Colrel: Collaborative relaying for federated learning over intermittently connected networks,” in Workshop on Federated Learning: Recent Advances and New Challenges (in Conjunction with NeurIPS 2022), 2022

    2022

  18. [25]

    Robust semi-decentralized federated learning via collaborative relaying,

    M. Yemini, R. Saha, E. Ozfatura, D. Gündüz, and A. J. Goldsmith, “Robust semi-decentralized federated learning via collaborative relaying,” IEEE Transactions on Wireless Com- munications, 2023

    2023

  19. [26]

    A unified analysis of federated learning with arbitrary client participation,

    S. Wang and M. Ji, “A unified analysis of federated learning with arbitrary client participation,” in Advances in Neural Information Processing Systems, vol. 35, 2022, pp. 19 124– 19 137

    2022

  20. [27]

    A lightweight method for tackling unknown participation statistics in federated averaging,

    S. Wang and M. Ji, “A lightweight method for tackling unknown participation statistics in federated averaging,” in International Conference on Learning Representations, vol. 104, 2024

    2024

  21. [28]

    Efficient federated learning against heterogeneous and non- stationary client unavailability,

    M. Xiang, S. Ioannidis, E. Yeh, C. Joe-Wong, and L. Su, “Efficient federated learning against heterogeneous and non- stationary client unavailability,” in Advances in Neural Infor- mation Processing Systems, vol. 37, 2024, pp. 104 281–104 328

    2024

  22. [29]

    Coded coopera- tive networks for semi-decentralized federated learning,

    S. Weng, M. Xiao, C. Ren, and M. Skoglund, “Coded coopera- tive networks for semi-decentralized federated learning,” IEEE Wireless Communications Letters, pp. 1–1, 2024

    2024

  23. [30]

    Cooperative gradient coding,

    S. Weng, C. Ren, M. Xiao, and M. Skoglund, “Cooperative gradient coding,” IEEE Transactions on Communications, pp. 1–1, 2025

    2025

  24. [31]

    On secure distributed linearly separable computation,

    K. Wan, H. Sun, M. Ji, and G. Caire, “On secure distributed linearly separable computation,” IEEE Journal on Selected Areas in Communications, vol. 40, no. 3, pp. 912–926, 2022

    2022

  25. [32]

    Secure cooperative gradient coding: Optimality, reliability, and global privacy,

    S. Weng, “Secure cooperative gradient coding: Optimality, reliability, and global privacy,” 2025. [Online]. A vailable: https://arxiv.org/abs/2507.07565

    2025 arXiv

  26. [33]

    Goldreich, Foundations of Cryptography, Volume 2

    O. Goldreich, Foundations of Cryptography, Volume 2. Cam- bridge university press Cambridge, 2004

    2004

  27. [34]

    Gradient coding: A voiding stragglers in distributed learning,

    R. Tandon, Q. Lei, A. G. Dimakis, and N. Karampatziakis, “Gradient coding: A voiding stragglers in distributed learning,” in International Conference on Machine Learning. PMLR, 2017, pp. 3368–3376

    2017

  28. [36]

    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, no. 11, pp. 2278–2324, 1998

    1998

  29. [37]

    Cinic-10 is not imagenet or cifar-10,

    L. N. Darlow, E. J. Crowley, A. Antoniou, and A. J. Storkey, “Cinic-10 is not imagenet or cifar-10,” arXiv preprint arXiv:1810.03505, 2018. Appendices Due to the page limit, all supporting proofs are detailed in the supplementary file [ 32], which can be found at https://arxiv...

    2018 arXiv

Pith tools

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