Cellular Traffic Prediction via Byzantine-robust Asynchronous Federated Learning

Hui Ma; Kai Yang; Yang Jiao

REVIEW 5 major objections 5 minor 75 references

Cellular Traffic Prediction via Byzantine-robust Asynchronous Federated Learning

T0 review · 5 major / 5 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 The paper argues that BAFDP, an asynchronous federated learning algorithm with local differential privacy and L1-penalized robust aggregation, can predict cellular traffic better than eight existing methods on three real-world datasets…

desk verdict Plausible combination of known techniques with a thin Byzantine-robustness analysis and under-powered experiments; the Eq. (20) oracle concern is overstated, but the paper still needs major revisions. read the letter →

arxiv 2505.19263 v1 pith:6UIMQSMV submitted 2025-05-25 cs.LG cs.AI

Hui Ma , Kai Yang , Yang Jiao This is my paper · ORCID

classification cs.LGcs.AI

keywords cellulartrafficpredictionfederatedlearningdifferentialprivacyByzantinerobustnessdistributionallyrobustoptimizationasynchronousWassersteindistanceforecasting

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that cellular traffic prediction does not have to choose between privacy, robustness, and accuracy. It proposes BAFDP, a federated learning algorithm in which each client adds Gaussian noise to its own data, the server treats the resulting noisy distribution as the center of a Wasserstein uncertainty ball, and an L1 penalty in the aggregation step keeps a minority of malicious (Byzantine) clients from dragging the global model off course. Because the server updates as soon as any S clients report rather than waiting for all clients, the algorithm also avoids the straggler problem of synchronous federated learning. The paper reports that BAFDP achieves the best average rank among nine methods across the Milano, Trento, and LTE traffic datasets, and proves that the number of iterations needed to reach an Υ-stationary point is O(1/Υ²). A reader should care because traffic data is sensitive and base stations are heterogeneous, so a predictor that keeps raw data local, tolerates delayed or malicious updates, and still forecasts next-hour and next-day traffic would be directly usable in network operations.

What carries the argument

The engine is the Wasserstein-ball robust reformulation combined with an L1 consensus penalty. The paper constructs the uncertainty set as a 1-Wasserstein ball of radius ρ_t^i = η_i + c_3/ε_t^i around the empirical distribution of each client's Gaussian-perturbed data, so the privacy budget ε_t^i directly sets the size of the robustness radius. Under the G(ω)-Lipschitz assumption on the loss, the inner supremum over the Wasserstein ball collapses to the empirical loss plus ρ_t^i G(ω_i), which converts the distributionally robust objective into a regularized machine-learning problem. The Byzantine-robust part is the L1-norm penalty ψ||z−ω_i||_1 in the augmented Lagrangian (Eq. (16)): minimizing it pulls every local model—including malicious ones—toward the global consensus variable z, and the server's update in Eq. (20) sums sign terms over both normal clients R and Byzantine clients B. The asynchronous part is the trigger rule: the server updates z and the dual variables whenever S clients have reported, and the convergence proof in Theorem 1 covers exactly this single-loop asynchronous scheme.

What would settle it

A concrete falsifier: measure the realized Lipschitz constant G(ω) of the MLP loss during BAFDP training on one of the three datasets; if the measured value is unbounded or substantially exceeds the value silently assumed in the bound ρ_t^i G(ω_i), then the regularized objective is not an upper bound on the true worst-case loss, and the claimed O(1/Υ²) convergence and robustness guarantee would fail on that run.

Watch

Extended reading notes

Core claim

The central claim is that the three pressures that usually fight each other in federated traffic forecasting—differential privacy noise, Byzantine clients, and asynchronous delays—can be absorbed into a single distributionally robust optimization problem and solved by one single-loop algorithm. The paper takes the per-client empirical loss on Gaussian-perturbed data, builds a 1-Wasserstein ball of radius ρ_t^i = η_i + σ_{i,t} around that empirical distribution, and uses Assumption 2 (an G(ω)-Lipschitz loss) to replace the worst-case expectation over the ball by the empirical loss plus ρ_t^i G(ω_i), turning the robust problem into a regularized one. An L1 penalty ψ||z−ω_i||_1 added to the augmented Lagrangian is the Byzantine-robust mechanism: it forces all local models, normal and Byzantine, to stay close to the server's consensus variable z, and the server update in Eq. (20) applies this penalty over both the normal set R and the malicious set B. The paper then proves in Theorem 1 that reaching an Υ-stationary point of this objective takes at most O(1/Υ²) iterations, and reports experiments in which BAFDP ranks first (average rank 1.08) in RMSE and MAE against FedGRU, Fed-NTP, FedAtt, FedDA, AFL, ASPIRE-EASE, UDP, and NbAFL on three real-world datasets.

Load-bearing premise

The load-bearing premise is that every client's loss function is Lipschitz continuous with a known constant G(ω); if that constant cannot be computed or is not bounded, the Wasserstein-ball reformulation and the convergence guarantee in Theorem 1 do not follow.

Editorial extensions

If this is right

If BAFDP's reported ranking holds, operators can run federated traffic predictors that never transmit raw subscriber data and still beat the leading centralized and federated baselines on RMSE and MAE for both one-hour and one-day horizons.
The O(1/Υ²) iteration bound means the asynchronous scheme converges at the standard rate for non-convex optimization, and the paper's analysis says the bound improves as the number of active clients S per round grows.
Because the privacy budget ε_t^i is a learned variable rather than a fixed hyperparameter, deployment does not require manually tuning the privacy level; the paper's experiments show the budget converging to a small stable value as training progresses.
The Byzantine-robustness experiments show graceful degradation as the fraction of malicious clients rises, implying the L1 penalty gives a tunable trade-off between robustness and prediction accuracy.
Switching from synchronous (BSFDP) to asynchronous updates reduces the wall-clock time to reach the same training loss in the paper's experiments, which is a direct argument for asynchronous federated learning in heterogeneous base-station settings.

Reading between the lines

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

A testable extension the paper leaves implicit is a provable Byzantine tolerance bound: deriving, as a function of the penalty strength ψ, the number of clients M, and the attack magnitude, exactly how many Byzantine clients the L1 aggregation can absorb before the consensus variable is pulled away; the paper reports only empirical ratios.
Because G(ω), the Lipschitz constant of the MLP loss, is never computed or bounded in the paper, an immediate engineering extension is to estimate it online from the model weights and input sensitivity and use that estimate to set ρ_t^i, which would make the robust penalty an actual worst-case bound rather than a tuned regularizer.
The same Wasserstein-ball-plus-L1 recipe could be applied to other sensitive spatiotemporal forecasting tasks, such as electricity load or intelligent-transportation flow; a concrete check would be whether the ranking advantage over FedAvg-style baselines persists when the data are more strongly non-IID than the three traffic datasets.
If the privacy guarantee is meant to be local differential privacy of raw inputs, the robustness radius depends on the input sensitivity; a natural variant is to add the noise to gradients with clipping, which would make the Lipschitz constant bounded and could tighten the connection between the privacy budget ε and the Wasserstein radius.

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.

Desk Editor's Note

read the letter

Dear colleague,

Quick take on arXiv:2505.19263: the paper combines asynchronous federated learning, local differential privacy, Wasserstein DRO, and an L1 penalty, and applies this to cellular traffic prediction. The specific combination is new for this application, and the empirical average rank on three real datasets looks decent. However, the Byzantine-robustness argument is thin, the experiments lack any error bars, and the theory is not fully checkable from the paper.

The stress-test about Eq. (20) is, I think, a red herring. The update separates the sum over R and B, but the server can compute the sum over all received clients without knowing which are Byzantine. The notation is genuinely confusing and should be rewritten, but it is not an oracle requirement. The deeper problem is that the paper never states this explicitly, and the L1 penalty is treated as a robustness mechanism without any quantitative tolerance bound or attack model. The only robustness experiment varies the fraction of malicious clients without specifying what those clients do.

A second soft spot is Assumption 2: the objective includes G(omega), the Lipschitz constant, but the paper never computes or bounds it for the MLP used in the experiments. It looks like a placeholder, and the method as implemented may just ignore the regularizer or set it arbitrarily. The convergence theorem is relegated to an external appendix with undefined constants (mu3, mu4, c5, c6), so a referee cannot verify it from the submission alone. And all results are single runs, no variance or significance tests.

On the positive side, the literature review is thorough, the DRO reformulation is standard but executed correctly, and the asynchronous single-loop design is sensible. The application to traffic prediction is timely, and the reported improvements over eight baselines are consistent across datasets, even if we cannot trust the magnitudes.

Who is this for? Someone working on practical federated traffic prediction might find the combination useful, but the current form is not a reliable reference. I would not cite it in the next twelve months. It does deserve a serious referee, though: the idea is coherent, the gaps are fixable in principle, and the paper is not a waste of referee time. If I were the editor, I would send it to review with a request for major revisions: specify the attack model and a tolerance bound, report error bars, move the appendix into the paper, and either compute G(omega) or state how it is estimated.

Best, [Your name]

Referee Report

5 major / 5 minor

Summary. The manuscript proposes BAFDP (Byzantine-robust asynchronous federated learning with differential privacy) for cellular traffic forecasting. The method adds Gaussian noise to local input data, models the resulting uncertainty through a Wasserstein ball around the empirical distribution, converts the distributionally robust problem into a regularized empirical risk problem, and solves it with an asynchronous augmented-Lagrangian algorithm in which the privacy level epsilon is also an optimized variable. The paper claims (i) superior RMSE/MAE over eight federated-learning baselines on Milano, Trento, and LTE datasets; (ii) Byzantine robustness via an L1 consensus penalty; (iii) differential privacy guarantees; and (iv) an iteration complexity of O(1/Upsilon^2) to reach a Upsilon-stationary point. The convergence proof is deferred to a supplementary appendix.

Significance. If the stated claims were established, the paper would be a useful contribution to privacy-preserving and Byzantine-robust federated traffic prediction. The empirical comparison is fairly broad, the datasets are public, and the simultaneous optimization of the model and the privacy level is an interesting idea. The paper also provides a supplementary proof file for the convergence analysis. However, the Byzantine-robustness mechanism, as specified in Eq. (20), assumes knowledge of the Byzantine client set, which contradicts the stated threat model; the privacy mechanism is not consistently implemented in the objective used by the algorithm; and the experimental section reports single runs without attack definitions or variance estimates. These are load-bearing problems, so the advertised contributions are not established in the current form.

major comments (5)

[Section IV.B (Eq. (20)), Section III, Algorithm 1] The Byzantine-robustness claim hinges on the server being able to separate active clients into the sets R and B in the z-update (Eq. (20)). Section III states that the identity of malicious clients is 'a priori unknown to the server,' and Algorithm 1 takes only the dataset D = {D_1, ..., D_M} as input. If the server possessed such an oracle, it could simply exclude B and train on R alone, making the robust aggregation unnecessary. Since no Byzantine-detection or filtering mechanism is described, the stated update cannot be executed under the paper's own threat model. Section VI.E reports only the ratio of malicious clients (Table IV) and never specifies what the Byzantine clients actually do, so the experiments do not demonstrate robustness against any concrete attack.
[Section III.B (Eq. (4)), Section IV.A (Eqs. (14)-(15)), Algorithm 1] The local differential privacy mechanism is not consistently defined. Section III.B says clients perturb their local data and train on the noisy data D_tilde_i, and Eq. (4) defines the empirical loss on that noisy data. However, Eq. (14) drops the noise term entirely by first-order Taylor expansion, and Eq. (15) defines g_omega(omega_i) as the clean-data empirical loss (1/N) * sum_j L(x_{i,j}, y_{i,j}, omega_i). Algorithm 1 then updates omega and epsilon from the resulting clean-data objective and never injects noise. The paper also claims 'theoretical privacy guarantees' but provides no theorem accounting for the fact that epsilon_i^t is chosen adaptively during training; the (epsilon, delta)-DP definition in Section III.A and Eq. (2) only state the definition. As written, the implemented algorithm may not be differentially private at all.
[Section VI.A (Table I), Section VI.E (Table IV)] The empirical comparison in Table I reports a single RMSE/MAE value per method, dataset, and horizon, with no standard deviations, confidence intervals, or significance tests. The 'average rank' column is therefore not robust evidence of superiority. Table IV, which is the only Byzantine-robustness experiment, varies the malicious-client ratio but does not describe the Byzantine attack model (e.g., what messages the malicious clients send), and it does not report the performance of RSA and DP-RSA at ratio 0.3 as it does for BAFDP. The claim that BAFDP is superior under Byzantine attacks is not supported by the presented evidence.
[Assumption 2, Section IV.A (Eqs. (10), (13), (15))] The DRO reformulation and the implemented regularizer depend on the Lipschitz constant G(omega_i), which is treated as a known quantity in Assumption 2 and appears explicitly in the objective (Eq. (15)). The paper never states how G(.) is computed or upper-bounded for the MLP used in the experiments, nor does it give a practical procedure for obtaining it from the model weights and data. Without such a value, the objective minimized in Algorithm 1 is not well defined, and the equivalence between the DRO problem (9) and the regularized problem (15) cannot be verified in practice.
[Section IV.C (Theorem 1, Setting 1)] The convergence-rate claim T(Upsilon) ~ O(1/Upsilon^2) is central to the theoretical contribution, but the proof is entirely deferred to a supplementary file [67], and the theorem statement references quantities that are not defined in the main text (for example, c5 in the second term of the bound, and beta, c_bar, c6, c7 described only as 'constants'). The step-size choices in Setting 1 and the statement of Theorem 1 also appear garbled (e.g., ambiguous denominators in alpha_omega). Moreover, the bound contains no dependence on the number of Byzantine clients B, so even a valid proof would not quantify Byzantine robustness. As presented, the convergence theorem cannot be checked from the manuscript alone.

minor comments (5)

[Section VI.E.3] The heading 'Coverage rate' does not match the content of the paragraph, which discusses training loss and convergence time; the heading should be corrected or the content revised.
[Section IV.A (Eq. (9))] Equation (9) uses E_{D_i} inside a supremum over D_tilde_i with W1(D_i, D_tilde_i) <= rho, while Eq. (10) and the surrounding text switch between E_{D_i} and E_{D_tilde_i}; this notation should be made consistent to avoid ambiguity about which distribution is being optimized.
[Section VI.E.2] The text states that the MLP model has a size of 440 MB; for a multilayer perceptron used in traffic prediction, this seems implausibly large and should be verified or clarified.
[Section VI.D] The sentence 'immediately sends the updated global model back to the server' should presumably read 'back to the client,' since the server broadcasts to clients.
[Section VI.B and Figure captions] The text in Section VI.B says the privacy-level trend is observed in Fig. 4, but Fig. 4 is the training-loss comparison and Fig. 3 is the privacy-level plot; the cross-reference should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's DRO reformulation and convergence analysis are not constructed from their own outputs, and the performance claims rest on independent empirical comparisons.

full rationale

The derivation chain from Eq. (5) to Eq. (15) is a standard Wasserstein distributionally robust optimization bound: Proposition 1 uses Lipschitz continuity (Assumption 2) to replace the worst-case expectation by the empirical loss plus a radius times the Lipschitz constant, and the subsequent Taylor expansion is stated as an approximation, not as the target result. The convergence claim in Theorem 1 is presented as an O(1/Upsilon^2) iteration complexity bound derived from stated Assumptions 3 and 4; it is not fitted to, or inferred from, the experimental tables. The empirical superiority claim in Table I is a benchmark comparison on three datasets, so the reported RMSE/MAE values are not by construction equal to any fitted parameter or assumption. There are self-citations ([17], [18], [56], [59], and the supplementary [67]), but they are used for motivation, for explicitly stated boundedness assumptions, or for the location of the proof, and none of them imports an unverified result that makes a claimed output identical to an input. Two non-circularity concerns do exist and should be weighed as correctness risks rather than circularity: Eq. (20) appears to require the server to split active clients into R and B even though Section III says Byzantine identities are a priori unknown, and the proof of Theorem 1 is outsourced to a self-published supplementary file. These affect rigor and reproducibility, but they do not exhibit the specific reduction, equation-equals-equation, or fitted-parameter-renamed-as-prediction structure that would constitute circularity.

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

The central claim rests on several unverified assumptions about the loss function, the Lipschitz constant, and the boundedness of variables, plus a poorly specified Byzantine aggregation rule. The paper does not report the values of most free parameters, making the method difficult to implement or verify.

free parameters (5)

Lipschitz constant G(omega) = not reported
The DRO regularizer rho*G(omega) in Eq (15) requires the Lipschitz constant of the MLP, but the paper does not specify how it is computed or bounded.
L1 penalty weight psi = not reported
psi controls the Byzantine robustness regularizer in Eq (16); no value or tuning procedure is given.
Privacy budget upper bound a = varied 10-70 (Tables II, III)
The budget a is a user-set parameter that strongly affects the regularizer and prediction accuracy; the paper tunes it per dataset and horizon.
Step sizes alpha_omega, alpha_epsilon, alpha_z, alpha_lambda, alpha_phi = not reported
Step sizes are claimed to follow Theorem 1, but the actual values used in the experiments are not disclosed.
Regularization parameters a_t1, a_t2 = not reported
Eq (17) defines a regularized Lagrangian with a_t1 and a_t2, which are not defined in the main text.

assumptions (5)

domain assumption Assumption 1: Light-tailed data distribution
Used to derive the Wasserstein radius eta_i in Eq (8); unverified for cellular traffic data.
ad hoc to paper Assumption 2: Loss L(.) is G(omega)-Lipschitz continuous
The DRO reformulation in Proposition 1 relies on this; for an MLP with unbounded weights, this is not guaranteed and G(omega) is not computed.
standard math Assumption 3: Gradient Lipschitz of augmented Lagrangian
Standard for convergence analysis, but unverified for the non-convex MLP objective.
domain assumption Assumption 4: Bounded variables (||omega_i|| <= mu_1, etc.)
Assumed to hold during the algorithm; not enforced in experiments, so the convergence proof may not apply to the actual training process.
ad hoc to paper Server can aggregate updates from unknown Byzantine clients via the L1 penalty in Eq (20)
The algorithm's implementability and robustness depend on this, but no robustness guarantee or attack model is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Cellular Traffic Prediction via Byzantine-robust Asynchronous Federated Learning." pith.science (2026). https://pith.science/paper/6UIMQSMV

@misc{pith2026250519263,
  author       = {Pith},
  title        = {Pith review of: Cellular Traffic Prediction via Byzantine-robust Asynchronous Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6UIMQSMV}},
  note         = {Machine review of arXiv:2505.19263}
}

read the original abstract

Network traffic prediction plays a crucial role in intelligent network operation. Traditional prediction methods often rely on centralized training, necessitating the transfer of vast amounts of traffic data to a central server. This approach can lead to latency and privacy concerns. To address these issues, federated learning integrated with differential privacy has emerged as a solution to improve data privacy and model robustness in distributed settings. Nonetheless, existing federated learning protocols are vulnerable to Byzantine attacks, which may significantly compromise model robustness. Developing a robust and privacy-preserving prediction model in the presence of Byzantine clients remains a significant challenge. To this end, we propose an asynchronous differential federated learning framework based on distributionally robust optimization. The proposed framework utilizes multiple clients to train the prediction model collaboratively with local differential privacy. In addition, regularization techniques have been employed to further improve the Byzantine robustness of the models. We have conducted extensive experiments on three real-world datasets, and the results elucidate that our proposed distributed algorithm can achieve superior performance over existing methods.

Figures

Figures reproduced from arXiv: 2505.19263 by the authors.

**Figure 1.** The diagram of BAFDP. added or removed for the i-th client. The randomized mechanism M : X → R with domain X and range R satisfies (ϵ, δ)-differential privacy, if for any two adjacent datasets Di , D′ i ∈ X and for any measurable sets S ⊆ R, P[M(Di) ∈ S] ≤ e ϵP[M(D ′ i ) ∈ S] + δ, (2) where ϵ denotes the bound of all outputs on Di and D ′ i . δ is the ratio of the probabilities for Di and D ′ i cannot be bounded by… view at source ↗

**Figure 2.** Visualization of prediction and ground truth on two open-sourced [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Visualization of privacy level changes during the training process on [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 5.** Figure 5: Comparison of RMSE for synchronous and asynchronous distributed [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: The comparison of MAE for synchronous and asynchronous dis [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 7.** Figure 7: Size to be transmitted in federated learning framework in terms of [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

**Figure 8.** Figure 8: the training loss within different robustness levels [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

75 extracted references · 64 canonical work pages

[67]

The suplementary content,

H. Ma, K. Yang, and Y . Jiao, “The suplementary content,” 2025. [Online]. Available: https://github.com/maggiemh/BAFDP/blob/main/ auxiliary.pdf

work page 2025
[1]

A communication-efficient federated learning scheme for IoT-based traffic forecasting,

C. Zhang, L. Cui, S. Yu, and J. J. Q. Yu, “A communication-efficient federated learning scheme for IoT-based traffic forecasting,” IEEE Internet Things J. , vol. 9, pp. 11 918–11 931, 2022

work page 2022
[2]

HCP: Heterogeneous computing platform for federated learning based collaborative content caching towards 6G networks,

Z. M. Fadlullah and N. Kato, “HCP: Heterogeneous computing platform for federated learning based collaborative content caching towards 6G networks,” IEEE Trans. Emerg. Top. Comput. , vol. 10, pp. 112–123, 2022

work page 2022
[3]

Unsupervised deep learning for iot time series,

Y . Liu, Y . Zhou, K. Yang, and X. Wang, “Unsupervised deep learning for iot time series,” IEEE Internet of Things Journal , vol. 10, no. 16, pp. 14 285–14 306, 2023

work page 2023
[4]

AI-empowered network root cause analysis for 6G,

C. Qiu, K. Yang, J. Wang, and S. Zhao, “AI-empowered network root cause analysis for 6G,” IEEE Network, pp. 1–9, 2023

work page 2023
[5]

TimeAutoML: Autonomous representation learning for multivariate irregularly sampled time series,

Y . Jiao, K. Yang, S. Dou, P. Luo, S. Liu, and D. Song, “TimeAutoML: Autonomous representation learning for multivariate irregularly sampled time series,” arXiv e-prints, pp. arXiv–2010, 2020

work page 2010
[6]

Anomaly detection in event-triggered traffic time series via similarity learning,

S. Dou, K. Yang, Y . Jiao, C. Qiu, and K. Ren, “Anomaly detection in event-triggered traffic time series via similarity learning,” IEEE Trans. on Dependable Secur. Comput. , vol. 01, pp. 1–15, 2024

work page 2024
[7]

Active learning for wireless iot intrusion detection,

K. Yang, J. Ren, Y . Zhu, and W. Zhang, “Active learning for wireless iot intrusion detection,” IEEE Wireless Communications, vol. 25, no. 6, pp. 19–25, 2018

work page 2018

Show all 75 references

[8]

Pc2a: Predicting collective contextual anomalies via lstm with deep generative model,

S. Dou, K. Yang, and H. V . Poor, “Pc2a: Predicting collective contextual anomalies via lstm with deep generative model,”IEEE Internet of Things Journal, vol. 6, no. 6, pp. 9645–9655, 2019

2019
[9]

LNTP: An end-to-end online prediction model for network traffic,

L. Zhang, H. Zhang, Q. Tang, P. Dong, Z. Zhao, Y . Wei, J. Mei, and K. Xue, “LNTP: An end-to-end online prediction model for network traffic,” IEEE Network, vol. 35, no. 1, pp. 226–233, 2021

2021
[10]

Toward QoS prediction based on temporal transformers for IoT applications,

A. Hameed, J. Violos, A. Leivadeas, N. Santi, R. Gr ¨unblatt, and N. Mitton, “Toward QoS prediction based on temporal transformers for IoT applications,” IEEE Trans. Netw. Serv. Manage., vol. 19, no. 4, pp. 4010–4027, 2022

2022
[11]

Cellular traffic prediction via deep state space models with attention mechanism,

H. Ma, K. Yang, and M.-O. Pun, “Cellular traffic prediction via deep state space models with attention mechanism,” Comput. Commun., vol. 197, pp. 276–283, 2023

2023
[12]

Optimiza- tion design for federated learning in heterogeneous 6G networks,

B. Luo, P. Han, P. Sun, X. Ouyang, J. Huang, and N. Ding, “Optimiza- tion design for federated learning in heterogeneous 6G networks,” IEEE Network, vol. 37, no. 2, pp. 38–43, 2023

2023
[13]

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 Proc. Adv. Neural Inf. Proces. Syst. , 2020

2020
[14]

Differential privacy in deep learning: An overview,

T. Ha, T. K. Dang, T. T. Dang, T. A. Truong, and M. T. Nguyen, “Differential privacy in deep learning: An overview,” in Int. Conf. Adv. Comput. Appl., 2019, pp. 97–102

2019
[15]

A robust game- theoretical federated learning framework with joint differential privacy,

L. Zhang, T. Zhu, P. Xiong, W. Zhou, and P. S. Yu, “A robust game- theoretical federated learning framework with joint differential privacy,” IEEE Trans. Knowl. Data Eng. , vol. 35, no. 4, pp. 3333–3346, 2023

2023
[16]

Distributionally robust federated learning for differentially private data,

S. Shi, C. Hu, D. Wang, Y . Zhu, and Z. Han, “Distributionally robust federated learning for differentially private data,” in Proc. Int. Conf. Distrib. Comput. Syst. , 2022, pp. 842–852

2022
[17]

Distributed robust optimization (DRO) part I: Framework and example,

K. Yang, J. Huang, Y . Wu, X. Wang, and M. Chiang, “Distributed robust optimization (DRO) part I: Framework and example,” Optim. Eng., vol. 15, pp. 35–67, 2012

2012
[18]

Distributed robust optimization for communication networks,

K. Yang, Y . Wu, J. Huang, X. Wang, and S. Verd ´u, “Distributed robust optimization for communication networks,” in Proc. IEEE INFOCOM , 2008, pp. 1157–1165

2008
[19]

Machine learning with adversaries: Byzantine tolerant gradient descent,

P. Blanchard, E. M. E. Mhamdi, R. Guerraoui, and J. Stainer, “Machine learning with adversaries: Byzantine tolerant gradient descent,” in Proc. Adv. Neural Inf. Proces. Syst. , 2017, pp. 119–

2017
[20]

Byzantine-robust distributed learning: Towards optimal statistical rates,

D. Yin, Y . Chen, K. Ramchandran, and P. L. Bartlett, “Byzantine-robust distributed learning: Towards optimal statistical rates,” in Proc. Int. Conf. Mach. Learn., vol. 80, 2018, pp. 5636–5645. [Online]. Available: http://proceedings.mlr.press/v80/yin18a.html JOURNAL OF LATEX ...

2018
[21]

Certifiably Byzantine-robust federated conformal prediction,

M. Kang, Z. Lin, J. Sun, C. Xiao, and B. Li, “Certifiably Byzantine-robust federated conformal prediction,” in Int. Conf. Learn. Represent., 2024. [Online]. Available: https://openreview.net/forum?id= RuW7jnyYyP

2024
[22]

RSA: byzantine- robust stochastic aggregation methods for distributed learning from heterogeneous datasets,

L. Li, W. Xu, T. Chen, G. B. Giannakis, and Q. Ling, “RSA: byzantine- robust stochastic aggregation methods for distributed learning from heterogeneous datasets,” in AAAI Conf. Artif. Intell. , 2019, pp. 1544– 1551

2019
[23]

Bridging differential privacy and Byzantine- robustness via model aggregation,

H. Zhu and Q. Ling, “Bridging differential privacy and Byzantine- robustness via model aggregation,” in Int. Joint Conf. Artif. Intell., 2022, pp. 2427–2433

2022
[24]

Differentially private byzantine-robust federated learning,

X. Ma, X. Sun, Y . Wu, Z. Liu, X. Chen, and C. Dong, “Differentially private byzantine-robust federated learning,” IEEE Trans. Parallel Dis- tributed Syst., vol. 33, no. 12, pp. 3690–3701, 2022

2022
[25]

MetaSTNet: Multimodal meta-learning for cellular traffic conformal prediction,

H. Ma and K. Yang, “MetaSTNet: Multimodal meta-learning for cellular traffic conformal prediction,” IEEE Trans. Netw. Sci. Eng., vol. 11, no. 2, pp. 1999–2011, 2024

1999
[26]

TimeMixer: Decomposable multiscale mixing for time series forecasting,

S. Wang, H. Wu, X. Shi, T. Hu, H. Luo, L. Ma, J. Y . Zhang, and J. ZHOU, “TimeMixer: Decomposable multiscale mixing for time series forecasting,” in Proc. Int. Conf. Learn. Represent. , 2024. [Online]. Available: https://openreview.net/forum?id=7oLshfEIC2

2024
[27]

FedGRU: Privacy- preserving traffic flow prediction via federated learning,

Y . Liu, S. Zhang, C. Zhang, and J. J. Q. Yu, “FedGRU: Privacy- preserving traffic flow prediction via federated learning,” in IEEE Int. Conf. Intell. Transp. Syst. , 2020, pp. 1–6

2020
[28]

Privacy-preserving traffic flow prediction: A federated learning approach,

Y . Liu, J. J. Q. Yu, J. Kang, D. Niyato, and S. Zhang, “Privacy-preserving traffic flow prediction: A federated learning approach,” IEEE Internet Things J., vol. 7, pp. 7751–7763, 2020

2020
[29]

Fed-NTP: A federated learning algo- rithm for network traffic prediction in V ANET,

S. S. Sepasgozar and S. Pierre, “Fed-NTP: A federated learning algo- rithm for network traffic prediction in V ANET,” IEEE Access, vol. 10, pp. 119 607–119 616, 2022

2022
[30]

Centralized and federated learning for predictive VNF autoscaling in multi-domain 5G networks and beyond,

T. Subramanya and R. Riggio, “Centralized and federated learning for predictive VNF autoscaling in multi-domain 5G networks and beyond,” IEEE Trans. Netw. Serv. Manage. , vol. 18, pp. 63–78, 2021

2021
[31]

FASTGNN: A topological information protected federated learning approach for traffic speed forecasting,

C. Zhang, S. Zhang, J. J. Q. Yu, and S. Yu, “FASTGNN: A topological information protected federated learning approach for traffic speed forecasting,” IEEE Trans. Ind. Inform. , vol. 17, pp. 8464–8474, 2021

2021
[32]

Short-term traffic flow prediction based on graph convolutional networks and federated learning,

M. Xia, D. Jin, and J. Chen, “Short-term traffic flow prediction based on graph convolutional networks and federated learning,” IEEE Trans. Intell. Transp. Syst. , pp. 1–13, 2022

2022
[33]

Privacy-preserving cross-area traffic forecasting in ITS: A transferable spatial-temporal graph neural network approach,

Y . Qi, J. Wu, A. K. Bashir, X. Lin, W. Yang, and M. D. Alshehri, “Privacy-preserving cross-area traffic forecasting in ITS: A transferable spatial-temporal graph neural network approach,” IEEE Trans. Intell. Transp. Syst., pp. 1–14, 2022

2022
[34]

Federated optimization in heterogeneous networks,

T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V . Smith, “Federated optimization in heterogeneous networks,” in Proc. Mach. Learn. Syst., 2020

2020
[35]

Learning private neural language modeling with attentive aggregation,

S. Ji, S. Pan, G. Long, X. Li, J. Jiang, and Z. Huang, “Learning private neural language modeling with attentive aggregation,” in Proc. Int. Jt. Conf. Neural Networks , 2019, pp. 1–8

2019
[36]

Dual attention- based federated learning for wireless traffic prediction,

C. Zhang, S. Dang, B. Shihada, and M. Alouini, “Dual attention- based federated learning for wireless traffic prediction,” in Proc. IEEE INFOCOM, 2021, pp. 1–10

2021
[37]

Differential privacy for deep and federated learning: A survey,

A. E. Ouadrhiri and A. Abdelhadi, “Differential privacy for deep and federated learning: A survey,” IEEE Access, vol. 10, pp. 22 359–22 380, 2022

2022
[38]

Privacy in deep learning: A survey,

F. Mireshghallah, M. Taram, P. Vepakomma, A. Singh, R. Raskar, and H. Esmaeilzadeh, “Privacy in deep learning: A survey,” arXiv preprint arXiv:2004.12254, 2020

2004 arXiv
[39]

Privacy- preserving approach PBCN in social network with differential privacy,

H. Huang, D. Zhang, F. Xiao, K. Wang, J. Gu, and R. Wang, “Privacy- preserving approach PBCN in social network with differential privacy,” IEEE Trans. Netw. Serv. Manage. , vol. 17, pp. 931–945, 2020

2020
[40]

Fed-GLOSS-DP: Fed- erated, global learning using synthetic sets with record level differential privacy,

H. Wang, D. Chen, R. Kerkouche, and M. Fritz, “Fed-GLOSS-DP: Fed- erated, global learning using synthetic sets with record level differential privacy,” arXiv preprint arXiv:2302.01068 , 2023

2023 arXiv
[41]

Performance-enhanced federated learning with differential privacy for Internet of things,

X. Shen, Y . Liu, and Z. Zhang, “Performance-enhanced federated learning with differential privacy for Internet of things,” IEEE Internet Things J., vol. 9, no. 23, pp. 24 079–24 094, 2022

2022
[42]

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. V . Poor, “Federated learning with differential privacy: Algorithms and performance analysis,” IEEE Trans. Inf. Forensics Secur., vol. 15, pp. 3454–3469, 2020

2020
[43]

Personalized federated learning with differential privacy,

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

2020
[44]

Towards efficient and privacy-preserving federated deep learning,

M. Hao, H. Li, G. Xu, S. Liu, and H. Yang, “Towards efficient and privacy-preserving federated deep learning,” in IEEE Int. Conf. Commun., 2019, pp. 1–6

2019
[45]

On differential privacy for federated learning in wireless systems with multiple base stations,

N. Tavangaran, M. Chen, Z. Yang, J. M. B. da Silva Jr., and H. V . Poor, “On differential privacy for federated learning in wireless systems with multiple base stations,” IET Commun., vol. 20, no. 18, pp. 1853–1867, 2024

2024
[46]

GRNN: Generative regression neural network—a data leakage attack for federated learning,

H. Ren, J. Deng, and X. Xie, “GRNN: Generative regression neural network—a data leakage attack for federated learning,” ACM Trans. Intell. Syst. Technol., vol. 13, 2022

2022
[47]

Byzantine machine learning: A primer,

R. Guerraoui, N. Gupta, and R. Pinot, “Byzantine machine learning: A primer,” ACM Comput. Surv., 2023

2023
[48]

Byzantine fault tolerance in distributed machine learning : a survey,

D. Bouhata and H. Moumen, “Byzantine fault tolerance in distributed machine learning : a survey,” J. Exp. Theor. Artif. Intell., pp. 1–59, 2024

2024
[49]

An experimental study of Byzantine-robust aggregation schemes in federated learning,

S. Li, E. C. H. Ngai, and T. V oigt, “An experimental study of Byzantine-robust aggregation schemes in federated learning,” IEEE Trans. Big Data , pp. 1–13, 2023. [Online]. Available: http://dx.doi.org/10.1109/TBDATA.2023.3237397

2023
[50]

DRACO: byzantine-resilient distributed training via redundant gradients,

L. Chen, H. Wang, Z. Charles, and D. S. Papailiopoulos, “DRACO: byzantine-resilient distributed training via redundant gradients,” in Proc. Int. Conf. Mach. Learn. , vol. 80, 2018, pp. 902–911. [Online]. Available: http://proceedings.mlr.press/v80/chen18l.html

2018
[51]

DETOX: A redundancy-based framework for faster and more robust gradient aggregation,

S. Rajput, H. Wang, Z. Charles, and D. S. Papailiopoulos, “DETOX: A redundancy-based framework for faster and more robust gradient aggregation,” in Proc. Adv. Neural Inf. Proces. Syst. , 2019, pp. 10 320– 10 330

2019
[52]

FLTrust: Byzantine-robust federated learning via trust bootstrapping,

X. Cao, M. Fang, J. Liu, and N. Z. Gong, “FLTrust: Byzantine-robust federated learning via trust bootstrapping,” in Annu. Netw. Distrib. Syst. Secur. Symp. , 2021. [Online]. Available: https://www.ndss-symposium.org/ndss-paper/ fltrust-byzantine-robust-federated-learning-via-t...

2021
[53]

Distributed statistical machine learning in adversarial settings: Byzantine gradient descent,

Y . Chen, L. Su, and J. Xu, “Distributed statistical machine learning in adversarial settings: Byzantine gradient descent,” Proc. ACM Meas. Anal. Comput. Syst. , vol. 1, no. 2, pp. 44:1–44:25, 2017

2017
[54]

Byzantine-robust aggregation in federated learning empowered industrial IoT,

S. Li, E. C. H. Ngai, and T. V oigt, “Byzantine-robust aggregation in federated learning empowered industrial IoT,” IEEE Trans. Ind. Informatics, vol. 19, no. 2, pp. 1165–1175, 2023

2023
[55]

Learning from history for Byzantine robust optimization,

S. P. Karimireddy, L. He, and M. Jaggi, “Learning from history for Byzantine robust optimization,” in Proc. Int. Conf. Mach. Learn. , vol. 139, 2021, pp. 5311–5319

2021
[56]

Distributed distributionally robust optimization with non-convex objectives,

Y . Jiao, K. Yang, and D. Song, “Distributed distributionally robust optimization with non-convex objectives,” in Proc. Adv. Neural Inf. Proces. Syst., vol. 35, 2022, pp. 7987–7999

2022
[57]

Gradient-based bi-level optimization for deep learning: A survey,

C. Chen, X. Chen, C. Ma, Z. Liu, and X. Liu, “Gradient-based bi-level optimization for deep learning: A survey,” arXiv preprint arXiv:2207.11719, 2022

2022 arXiv
[58]

Fair resource allocation in federated learning,

T. Li, M. Sanjabi, A. Beirami, and V . Smith, “Fair resource allocation in federated learning,” in Proc. Int. Conf. Learn. Represent. , 2020. [Online]. Available: https://openreview.net/forum?id=ByexElSYDr

2020
[59]

Asynchronous distributed bilevel optimization,

Y . Jiao, K. Yang, T. Wu, D. Song, and C. Jian, “Asynchronous distributed bilevel optimization,” in Proc. Int. Conf. Learn. Represent. , 2023. [Online]. Available: https://openreview.net/forum?id=iSPPkKa2yK

2023
[60]

Agnostic federated learning,

M. Mohri, G. Sivek, and A. T. Suresh, “Agnostic federated learning,” in Proc. Int. Conf. Mach. Learn. , vol. 97, 2019, pp. 4615–4625

2019
[61]

Distributionally robust federated averaging,

Y . Deng, M. M. Kamani, and M. Mahdavi, “Distributionally robust federated averaging,” in Proc. Adv. Neural Inf. Proces. Syst. , vol. 33, 2020, pp. 15 111–15 122

2020
[62]

Data-driven distributionally robust op- timization using the Wasserstein metric: Performance guarantees and tractable reformulations,

P. M. Esfahani and D. Kuhn, “Data-driven distributionally robust op- timization using the Wasserstein metric: Performance guarantees and tractable reformulations,” Math. Program., vol. 171, pp. 115–166, 2017

2017
[63]

On the rate of convergence in Wasserstein distance of the empirical measure,

N. Fournier and A. Guillin, “On the rate of convergence in Wasserstein distance of the empirical measure,” Probab. Theory Relat. Fields , vol. 162, pp. 707–738, 2015

2015
[64]

Distributionally-robust machine learning using locally differentially-private data,

F. Farokhi, “Distributionally-robust machine learning using locally differentially-private data,” Optim. Lett., vol. 16, pp. 1167–1179, 2022

2022
[65]

Bilevel optimization: Convergence analysis and enhanced design,

K. Ji, J. Yang, and Y . Liang, “Bilevel optimization: Convergence analysis and enhanced design,” in Proc. Int. Conf. Mach. Learn., 2021, pp. 4882– 4892

2021
[66]

Convergence of meta-learning with task-specific adaptation over partial parameters,

K. Ji, J. D. Lee, Y . Liang, and H. V . Poor, “Convergence of meta-learning with task-specific adaptation over partial parameters,” Proc. Adv. Neural Inf. Proces. Syst. , vol. 33, pp. 11 490–11 500, 2020

2020
[68]

Telecommunications - SMS, Call, Internet - MI,

T. Italia, “Telecommunications - SMS, Call, Internet - MI,” 2015. [Online]. Available: https://doi.org/10.7910/DVN/EGZHFV

2015 doi
[69]

Social Pulse - Milano,

SpazioDati and D. P. di Milano, “Social Pulse - Milano,” 2015. [Online]. Available: https://doi.org/10.7910/DVN/9IZALB

2015 doi
[70]

MilanoToday,

Citynews, “MilanoToday,” 2015. [Online]. Available: https://doi.org/10. 7910/DVN/QWOE1R JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13

2015
[71]

Telecommunications - SMS, Call, Internet - TN,

T. Italia, “Telecommunications - SMS, Call, Internet - TN,” 2015. [Online]. Available: https://doi.org/10.7910/DVN/QLCABU

2015 doi
[72]

Social Pulse - Trentino,

SpazioDati, “Social Pulse - Trentino,” 2015. [Online]. Available: https://doi.org/10.7910/DVN/5H0NUI

2015 doi
[73]

TrentoToday,

Citynews, “TrentoToday,” 2015. [Online]. Available: https://doi.org/10. 7910/DVN/NYQ23N

2015
[74]

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 Trans. Mob. Comput., vol. 21, no. 9, pp. 3388–3401, 2021. Hui Ma was born in Xinjiang, China. She received B.Eng. d...

2021
[129]

Available: https://proceedings.neurips.cc/paper/2017/ hash/f4b9ec30ad9f68f89b29639786cb62ef-Abstract.html

[Online]. Available: https://proceedings.neurips.cc/paper/2017/ hash/f4b9ec30ad9f68f89b29639786cb62ef-Abstract.html

2017

Pith tools

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

[1] [67]

The suplementary content,

H. Ma, K. Yang, and Y . Jiao, “The suplementary content,” 2025. [Online]. Available: https://github.com/maggiemh/BAFDP/blob/main/ auxiliary.pdf

work page 2025

[2] [1]

A communication-efficient federated learning scheme for IoT-based traffic forecasting,

C. Zhang, L. Cui, S. Yu, and J. J. Q. Yu, “A communication-efficient federated learning scheme for IoT-based traffic forecasting,” IEEE Internet Things J. , vol. 9, pp. 11 918–11 931, 2022

work page 2022

[3] [2]

HCP: Heterogeneous computing platform for federated learning based collaborative content caching towards 6G networks,

Z. M. Fadlullah and N. Kato, “HCP: Heterogeneous computing platform for federated learning based collaborative content caching towards 6G networks,” IEEE Trans. Emerg. Top. Comput. , vol. 10, pp. 112–123, 2022

work page 2022

[4] [3]

Unsupervised deep learning for iot time series,

Y . Liu, Y . Zhou, K. Yang, and X. Wang, “Unsupervised deep learning for iot time series,” IEEE Internet of Things Journal , vol. 10, no. 16, pp. 14 285–14 306, 2023

work page 2023

[5] [4]

AI-empowered network root cause analysis for 6G,

C. Qiu, K. Yang, J. Wang, and S. Zhao, “AI-empowered network root cause analysis for 6G,” IEEE Network, pp. 1–9, 2023

work page 2023

[6] [5]

TimeAutoML: Autonomous representation learning for multivariate irregularly sampled time series,

Y . Jiao, K. Yang, S. Dou, P. Luo, S. Liu, and D. Song, “TimeAutoML: Autonomous representation learning for multivariate irregularly sampled time series,” arXiv e-prints, pp. arXiv–2010, 2020

work page 2010

[7] [6]

Anomaly detection in event-triggered traffic time series via similarity learning,

S. Dou, K. Yang, Y . Jiao, C. Qiu, and K. Ren, “Anomaly detection in event-triggered traffic time series via similarity learning,” IEEE Trans. on Dependable Secur. Comput. , vol. 01, pp. 1–15, 2024

work page 2024

[8] [7]

Active learning for wireless iot intrusion detection,

K. Yang, J. Ren, Y . Zhu, and W. Zhang, “Active learning for wireless iot intrusion detection,” IEEE Wireless Communications, vol. 25, no. 6, pp. 19–25, 2018

work page 2018