Pith. sign in

REVIEW 3 major objections 4 minor 37 references

Trust-Aware Topology Learning for Dynamic Decentralized Federated Learning under Adversaries

T0 review · 3 major / 4 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 This paper claims that a decentralized federated learning protocol called DMTT, which screens peer models and topology claims through evidential and Beta-distributed trust, confines Byzantine influence in the mixing matrix to a residual δ_m

desk verdict Plausible engineering contribution with a real gap between the proved aggregation rule and the implemented one; the empirical story is stronger than the theory. read the letter →

arxiv 2608.03156 v1 pith:4ZQRUKYJ submitted 2026-08-04 cs.DC

Shubham Vaishnav , Murtaza Rangwala , Ali Beikmohammadi , Sindri Magnússon , Rajkumar Buyya This is my paper
classification cs.DC
keywords decentralizedfederatedlearningByzantinerobustnesstopologytrustmanagementevidentialdeepdynamicgraphspersonalizedmobilecomputing
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 seeks to show that decentralized federated learning can remain both safe and useful when adversaries poison not only model updates but also the topology information peers rely on. It presents DMTT, a protocol that extends the MURMURA evidential-compatibility approach to time-varying graphs by maintaining a confidence-weighted topology view built from link-reliability estimates, signed topology claims, witness corroboration, and a Beta-distributed source trust. Each device aggregates only over a trust-screened collaborator set, and the paper proves that the induced screened mixing matrices confine Byzantine influence to a bounded residual δ_max that vanishes under perfect screening. Experimentally, DMTT sustains honest-node accuracy above 0.862 on UCI HAR and 0.829 on PAMAP2 across all tested Byzantine fractions from 10% to 80%, while FedAvg collapses to chance and three robust aggregators fail to consistently beat a local-only baseline. If correct, DMTT is a decentralized protocol that profits from honest collaborators and degrades gracefully toward local-only performance in adversarial extremes rather than collapsing.

What carries the argument

The screened mixing matrix W^t, whose entries are zero outside the trust-screened collaborator set C^t_i, is the object that carries both the learning and the Byzantine-confinement argument. It is built from a composite collaboration score q^t_ij = λ1·s_model + λ2·T_topo + λ3·p_hat − λ4·c_comm, where s_model is the MURMURA evidential compatibility score, T_topo is the Beta-evidence source trust penalized by epistemic uncertainty, and p_hat is the direct link reliability; top-B selection followed by MURMURA-style weight normalization produces W^t. The theory requires W^t to be doubly stochastic (Assumption 3) so that the average iterate tracks gradient descent on the component objective, and

What would settle it

Run DMTT at 50% Byzantine fraction and directly compute the column sums of the actual W^t matrices produced by Algorithm 1; if max_j |sum_i w^t_ij − 1| stays bounded away from zero, then the doubly-stochastic Assumption 3 is violated and Theorems 1 and Proposition 1 do not govern the implemented system. Alternatively, deploy an adversary that sends bounded-norm poisoned models while maintaining a small nonzero trust score, and check whether honest-node accuracy degrades as the ρ² term of (23) predicts.

Watch

Extended reading notes

Core claim

The central claim is that trust-screened aggregation in decentralized personalized federated learning induces a mixing matrix in which Byzantine influence is provably bounded by δ_max, the largest total weight any honest node assigns to Byzantine peers, and that δ_max vanishes when screening perfectly excludes all adversaries. Under perfect screening, DMTT reduces to time-varying decentralized SGD over the honest screened graph and converges at the standard rate with the heterogeneity term τ²_scr shrunk by removing incompatible peers. The perturbation analysis shows imperfect screening adds an error term scaling with ρ² = (Δ_scr(R+ηG)+δ_max Θ_B)², so bounded Byzantine message norms and a fix

Load-bearing premise

The convergence and Byzantine-confinement guarantees require the screened mixing matrix to be doubly stochastic at every round, yet Algorithm 1's trust-normalized weights are not explicitly symmetrized or Metropolis-Hastings reweighted, so the implemented matrix may not satisfy the assumption the theorems rely on.

Editorial extensions

If this is right

  • DMTT is the only method evaluated that consistently clears the local-only bar across both UCI HAR and PAMAP2 at every Byzantine fraction from 10% to 80%, indicating that trust-aware screening can both resist poisoning and extract collaborative benefit from honest peers.
  • Static and dynamic FedAvg collapse to chance under the combined topology-liar and Gaussian model attack, showing that unprotected decentralized averaging is extremely fragile when adversaries can also lie about topology.
  • Krum, BALANCE, and UBAR, dedicated robust aggregators, fail to consistently beat a local-only baseline, suggesting that filtering only model updates is insufficient when topology claims are manipulated.
  • The Byzantine-influence bound in Proposition 1 shows that imperfect screening adds an O(ρ²/η²) penalty, so a fixed stepsize floor is essential; the error cannot be driven to zero by shrinking the stepsize.
  • The coordinator-free ZeroMQ implementation reproduces the simulation-backend accuracy trajectories, indicating the protocol runs end-to-end on real distributed nodes without a central coordinator.

Reading between the lines

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

  • A natural stress test is to run DMTT against an adaptive adversary that deliberately maintains a small, nonzero trust score by occasionally sending truthful topology claims, directly measuring whether δ_max leaves zero and whether accuracy degrades at the rate predicted by the ρ² term in (23).
  • The doubly-stochastic assumption may be replaceable by a push-sum or gradient-tracking variant over asymmetric trust weights; if that holds, the protocol's guarantees would extend to fully asymmetric collaboration without the symmetrization step the paper only sketches.
  • The topology-trust machinery (signed claims, witness corroboration, Beta source trust) is generic and could be reused outside federated learning, for example in peer-to-peer gossip or blockchain peer selection where topology claims are as security-critical as model updates.
  • The empirical claim that surviving Byzantine weight is exactly zero deserves a direct measurement under higher mobility or sparser graphs, where the honest-neighbor pool shrinks and the screening gates might no longer converge to δ_max=0 within the round budget.
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 / 4 minor

Summary. The paper proposes DMTT, a decentralized personalized federated learning protocol that extends MURMURA to dynamic topologies under Byzantine clients who can poison both model updates and topology information. Each node maintains link-reliability estimates, signed topology claims, witness corroboration, and Beta-distributed source trust, then aggregates only over a trust-screened collaborator set using a composite score. The paper claims convergence guarantees for the screened mixing matrices, with Byzantine influence confined to a bounded residual δ_max that vanishes under perfect screening. It also reports a coordinator-free ZeroMQ implementation and experiments on UCI HAR and PAMAP2 across 100 clients with Byzantine fractions from 10% to 80%, showing that DMTT maintains honest-node accuracy above 0.862/0.829 while FedAvg and robust aggregators fail. The central claims are that trust-aware screening both provably limits Byzantine influence and outperforms local-only and robust baselines.

Significance. If the advertised guarantees hold, DMTT would be a useful contribution to decentralized FL under combined model-poisoning and topology-manipulation attacks. The empirical study is carefully structured: it includes a distributed backend validation at 10 nodes, a hyperparameter sensitivity table, three seeds, two datasets, and an honest discussion of non-monotone behavior at high Byzantine fractions. The paper also reports the useful negative result that standard robust aggregators fail under topology manipulation. However, the theoretical contribution as stated is not yet supported: the proof relies on a doubly stochastic mixing matrix that Algorithm 1 does not construct, and the δ_max bound is largely a restatement of the definition of δ_max rather than a demonstrated property of the screening mechanism. The empirical findings are promising but are obtained under a single, easily detected topology-liar attack. These issues are load-bearing for the paper's headline claims, though they appear addressable within the manuscript's scope.

major comments (3)
  1. [Section VI-B, Assumption 3; Lemma 2 and Eq. (24)] The convergence analysis assumes that the screened mixing matrix W_t is doubly stochastic (W_t 1 = 1, 1^T W_t = 1^T), and the proof of Lemma 2 relies on W_t J = J W_t = J so that the average iterate evolves as noiseless SGD. However, Algorithm 1 constructs W_t only through Eq. (17), normalizing weights over C_i^t ∪ {i}; this guarantees row stochasticity but not column stochasticity. The paper states that double stochasticity can be enforced 'by symmetrizing the trust-weighted scores, or by adopting a Metropolis-Hastings reweighting,' but no line of Algorithm 1 or the companion algorithms performs this enforcement. The conclusion's future-work sentence admits that relaxing the doubly stochastic assumption is needed. Therefore Theorems 1, 2, and Proposition 1 govern a different aggregation rule than the one implemented and evaluated. A concrete column-sum check of W_t in the implementation
  2. [Section VI-E, Assumption 7 and Proposition 1] The bounded-residual result is largely definitional. Assumption 7 defines δ_max as the supremum of the total aggregation weight that honest nodes assign to Byzantine peers, and then assumes this quantity is bounded. Proposition 1 then returns δ_max Θ_B as the Byzantine drift term in the convergence bound. The statement that Byzantine influence 'vanishes under perfect screening' is a restatement of the definition of perfect screening (δ_max = 0, Δ_scr = 0), not a derived property of the Beta trust mechanism or the composite-score gates. The analysis does not prove that the trust updates drive δ_max to zero under the stated adversary model; the empirical δ_max = 0 observation is for one specific attack. To support the headline claim, the paper needs either a theorem showing that the trust and screening rules make δ_max small under the adversary model of Section III-C, or a clear downgrade
  3. [Section V-E and Section VII-C] The only topology-manipulation attack evaluated is the omission liar: Byzantine nodes claim only other Byzantine nodes, so the self-inclusion check fires immediately for any honest neighbor that communicated with them. This is a very weak topology attack. The adversary model in Section III-C explicitly allows false edge additions/removals, replay, flooding, and other arbitrary topology claims. A Byzantine node that includes honest nodes in its claims but lies about other edges, or colludes to corroborate false edge additions, would not trigger the self-inclusion contradiction and is not tested. Thus the empirical claim that DMTT resists 'topology-manipulation attacks' is narrower than the stated threat model. The central empirical contribution would be stronger if at least one non-omission topology attack were evaluated.
minor comments (4)
  1. [Table III] The text says the hyperparameter sensitivity is evaluated on both datasets, but the table shows only a single set of accuracy values with no dataset label. Please clarify which dataset is reported or split the table into two panels.
  2. [Eq. (18)] The 'coordinator-free' synchronization uses a shared start time t_start announced once at experiment start. This one-time bootstrap is not a coordinator in the learning path, but it should be acknowledged explicitly as an initial setup assumption.
  3. [Section VI-C and Theorem 1] Theorem 1 bounds convergence of the component-average objective f, not per-client personalized objectives. Since the paper emphasizes personalized FL, it would be helpful to state explicitly that the guarantee is for the average over the compatible component, and to discuss whether individual F_i are also controlled.
  4. [Algorithm 1, line 11] The 'stale neighbors' timeout is not defined in Table II or in the algorithm pseudocode. Please specify the timeout parameter or state that it is a fixed system constant.

Circularity Check

1 steps flagged · score 6.0 of 10

The advertised Byzantine-confinement proof reduces to the definition/assumption of δ_max; the empirical study and Theorem 1 are otherwise self-contained, but Assumption 3 is an unenforced gap rather than a circular step.

  1. self definitional [Section VI-E, Assumption 7 and Proposition 1 (echoed in Abstract)]
    "Also define δ_max = sup_t max_{i∈H} Σ_{j∈B} w^t_{ij}, and assume δ_max is uniformly bounded and sufficiently small relative to the connectivity of the screened honest graph. ... Define the worst-case per-step drift magnitude: ρ := Δ_scr(R+ηG) + δ_maxΘ_B."

    The headline claim that screened mixing matrices 'confine Byzantine influence to a bounded residual δ_max that vanishes under perfect screening' is a restatement of Assumption 7, which defines δ_max as the surviving Byzantine aggregation weight and assumes it bounded. Proposition 1 then feeds that same δ_max into its drift bound ρ; it never derives δ_max's boundedness or its vanishing. 'Perfect screening' is literally the condition δ_max=0, so the 'vanishing' is assumed, not proved. The theorem's output is therefore an arithmetic bound parameterized by an input quantity that is defined to be exactly the advertised residual.

full rationale

The paper's core convergence theorem (Theorem 1, Eq. 22) is a standard time-varying decentralized SGD analysis and is not circular: under its explicit assumptions (smoothness, bounded variance, contraction, bounded heterogeneity) it yields a genuine bound, and the empirical comparisons against FedAvg, Krum, BALANCE, UBAR, and local-only are externally meaningful and self-contained. The circularity is concentrated in the Byzantine-confinement guarantee: Proposition 1's drift term ρ is defined as Δ_scr(R+ηG)+δ_maxΘ_B, with δ_max already defined in Assumption 7 as the total Byzantine weight and assumed bounded/small. Thus the assertion 'we prove that the induced screened mixing matrices confine Byzantine influence to a bounded residual δ_max' reduces by construction to the definition of δ_max; perfect screening δ_max=0 is an assumed regime, and the empirical δ_max=0 is a measurement of that same defined quantity. Separately, Assumption 3 (doubly stochastic W_t) is not enforced anywhere in Algorithm 1, whose Eq. (17) only normalizes over C_i^t ∪ {i}, guaranteeing row stochasticity but not column stochasticity; the appendix proofs explicitly use W_t J = J W_t = J. This is a serious correctness/support gap, and the paper's Future Work sentence admits it, but it is not itself a circular reduction. Self-citations to MURMURA [4] are used as a modular personalization component and are not the load-bearing support for the Byzantine claim, so they do not raise the circularity score. Taking the definitional δ_max step as a partial circularity of the central theoretical claim, while recognizing the substantial independent empirical content, yields score 6.

Assumptions & free parameters 9 free parameters · 9 assumptions · 0 invented entities

The guarantees rest on nine explicit assumptions and a large set of hand-set hyperparameters. The most consequential assumption is that W_t is doubly stochastic, which the implementation does not enforce; the most consequential defined quantity is δ_max, which is assumed rather than derived. The empirical results depend on multiple thresholds, several of which are unreported.

free parameters (9)
  • Composite score weights (λ1=0.4, λ2=0.3, λ3=0.2, λ4=0.1) = 0.4, 0.3, 0.2, 0.1
    Table II; Eq. (15). Chosen by hand, determines the balance between model compatibility, topology trust, link reliability, and communication cost in collaborator selection.
  • Model-compatibility gate γ=0.6 = 0.6
    Section IV-E; a peer must satisfy s_model ≥ γ·s_self. Directly controls screening and contributes to the empirical δ_max=0 claim.
  • Topology-trust gate τ_trust=0.25 = 0.25
    Section IV-E; peers below this trust threshold are excluded. Sensitivity table shows robustness in 0.10-0.50, but the default is a hand-set choice.
  • Norm-sanity ratio κ_norm=5 = 5
    Section IV-E; rejects peer models with norm greater than 5 times the local norm. Combined with Assumption 9 this sets Θ_B=5R and strongly influences zero Byzantine weight.
  • Beta evidence weights and forgetting factor = λ=0.9; w_d=w_x=1; w_c, w_f, w_q, w_s unreported
    Algorithm 4, Eqs. (6)-(7); governs how quickly source trust collapses. The reported honest/liar trust separation depends on these choices.
  • Edge-acceptance threshold τ_add and witness minimum m
    Algorithm 2 uses τ_add and m to commit topology edges; values are not in Table II, so the topology view depends on unpublished choices.
  • Forwarding threshold τ_fwd, hop limit H, and distance penalty ξ
    Algorithms 2 and 3; control claim propagation and witness weighting. Values are unreported.
  • Collaborator budget B=5 and self-weight ω=0.5 = B=5, ω=0.5
    Table II; bounds the collaborator set and sets the own-model weight in aggregation.
  • Uncertainty penalty coefficients τ_U=0.3 and η=5 = 0.3, 5
    Eq. (10); controls how strongly epistemic uncertainty in the Beta reputation penalizes topology trust.
assumptions (9)
  • domain assumption Assumption 1: each local objective F_i is L-smooth
    Section VI-B(a); standard in SGD analysis, but unverified for the evidential loss used in experiments.
  • domain assumption Assumption 2: unbiased stochastic gradients with bounded variance σ²
    Section VI-B(b); standard for SGD, but no clipping or projection details in Algorithm 1 guarantee it.
  • ad hoc to paper Assumption 3: screened mixing matrix W_t is doubly stochastic
    Section VI-B(c); Eq. (17) uses client-specific MURMURA trust weights, and the paper only says double stochasticity can be enforced by symmetrization or Metropolis-Hastings. No algorithmic step does so, so the proof may not apply to the implementation.
  • domain assumption Assumption 4: per-step consensus contraction of the screened honest graph with constant p>0
    Section VI-B(d); requires every screened graph to contract uniformly every round, stronger than the window connectivity used in prior work.
  • domain assumption Assumption 5: bounded screened heterogeneity τ²_scr
    Section VI-C(a); bounds heterogeneity among screened collaborators and defines the quantity screening is supposed to reduce.
  • domain assumption Assumption 6: bounded stochastic gradient second moment G²
    Invoked in Proposition 1 and Theorem 2. The paper admits it is restrictive and does not hold for unregularized least-squares or logistic regression on unbounded data.
  • ad hoc to paper Assumption 7: bounded screening error Δ²_scr and bounded surviving Byzantine influence δ_max
    Section VI-E(b); δ_max is the supremum of Byzantine aggregation weight, assumed small. This is the quantity the abstract presents as a theorem output, making the core confinement result partly an assumption.
  • domain assumption Assumption 8: Byzantine messages have norm at most Θ_B almost surely
    Introduced in Proposition 1; in the paper it is enforced by the norm-sanity gate plus Assumption 9, giving Θ_B=κ_norm R.
  • domain assumption Assumption 9: iterates stay in a bounded domain ∥θ∥≤R
    Used to derive Θ_B. The paper says it is enforced by projected SGD or weight decay, but no projection step appears in Algorithm 1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Trust-Aware Topology Learning for Dynamic Decentralized Federated Learning under Adversaries." pith.science (2026). https://pith.science/paper/4ZQRUKYJ

@misc{pith2026260803156,
  author       = {Pith},
  title        = {Pith review of: Trust-Aware Topology Learning for Dynamic Decentralized Federated Learning under Adversaries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4ZQRUKYJ}},
  note         = {Machine review of arXiv:2608.03156}
}
abstract

In dynamic mobile decentralized federated learning (DFL), adversaries can poison both model updates and the topology information devices use to choose collaborators. We present DMTT (Dynamic MURMURA with Trusted Topology), a decentralized personalized FL protocol built on MURMURA, which uses evidential deep learning to down-weight distribution-mismatched peers, extended here to time-varying graphs under topology-manipulation attacks. Each device maintains a confidence-weighted local topology view from link-reliability estimates, signed topology claims, witness corroboration, and a Beta-distributed source-trust model, then aggregates only over a trust-screened collaborator set using a composite score fusing model compatibility, topology trust, and link reliability. We prove the screened mixing matrices confine Byzantine influence to a bounded residual $\delta_{max}$ that vanishes under perfect screening, and implement DMTT as a coordinator-free distributed system with each client running as an independent ZeroMQ process synchronized by a shared wall-clock epoch. On UCI HAR and PAMAP2, each partitioned across 100 mobile clients with Dirichlet heterogeneity, DMTT sustains honest-node accuracy above 0.862 (UCI HAR) and 0.829 (PAMAP2) across all tested adversary fractions (10 to 80%), nearly matching no-attack accuracy at low fractions and degrading gracefully toward local-only performance at extremes; static and dynamic FedAvg collapse to chance at every fraction, and robust aggregators (Krum, BALANCE, UBAR) fail to consistently beat a local-only baseline, while DMTT is the only method that clears this bar across both datasets at all fractions, with surviving Byzantine aggregation weight empirically zero throughout, consistent with $\delta_{max}$=0. The protocol runs end-to-end on real nodes via a coordinator-free ZeroMQ backend on the Melbourne Research Cloud.

Figures

Figures reproduced from arXiv: 2608.03156 by the authors.

Figure 1
Figure 1. • PULL (bind): Receives MODEL_STATE and TOPO_CLAIM messages from neighbours. A single socket handles both message types, distinguished by a 1-byte type field in the header. • PUSH ×|C t i | (connect): Sends MODEL_STATE and TOPO_CLAIM to each current collaborator. Sockets are created on first use so that the collaborator set can change without re-binding [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 1
Figure 1. Updated MURMURA architecture. The unchanged simulation backend (single OS process, shared memory) and the new distributed backend, in which each client and the passive monitor run as independent OS processes. Model weights travel via MODEL_STATE PUSH/PULL pairs; DMTT nodes additionally exchange signed neighbourhood observations as TOPO_CLAIM messages. Round boundaries are set by a shared wall-clock epoch; no coordin… view at source ↗
Figure 2
Figure 2. Honest-node accuracy versus round under the combined topology-liar + Gaussian ( [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figures from the paper (1 more)
Figure 3
Figure 3. Figure 3: Distributed-backend verification at 10 nodes. Solid lines: simulation backend; dashed lines: ZeroMQ distributed backend (real OS processes, wall￾clock synchronisation). Thin lines show individual seeds; bold lines show the mean over 2 seeds. Rows: DMTT (top) and FedAvg…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 27 canonical work pages

  1. [1]

    Decentralized federated learning with unreliable communications,

    H. Ye, L. Liang, and G. Y . Li, “Decentralized federated learning with unreliable communications,”IEEE journal of selected topics in signal processing, vol. 16, no. 3, pp. 487–500, 2022

    work page 2022

  2. [2]

    Decentralized personalized federated learning,

    S. Kharrat, M. Canini, and S. Horvath, “Decentralized personalized federated learning,”arXiv preprint arXiv:2406.06520, 2024

    arXiv 2024

  3. [3]

    Decentralized di- rected collaboration for personalized federated learning,

    Y . Liu, Y . Shi, Q. Li, B. Wu, X. Wang, and L. Shen, “Decentralized di- rected collaboration for personalized federated learning,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 23 168–23 178

    work page 2024

  4. [4]

    Evidential trust-aware model personalization in decentralized federated learning for wearable iot,

    M. Rangwala, R. O. Sinnott, and R. Buyya, “Evidential trust-aware model personalization in decentralized federated learning for wearable iot,”arXiv preprint arXiv:2512.19131, 2025

    arXiv 2025

  5. [5]

    Topology learning for hetero- geneous decentralized federated learning over unreliable d2d networks,

    Z. Wu, Z. Xu, D. Zeng, J. Li, and J. Liu, “Topology learning for hetero- geneous decentralized federated learning over unreliable d2d networks,” IEEE Transactions on Vehicular Technology, vol. 73, no. 8, pp. 12 201– 12 206, 2024

    work page 2024

  6. [6]

    Discovering network topology in the presence of byzantine faults,

    M. Nesterenko and S. Tixeuil, “Discovering network topology in the presence of byzantine faults,”IEEE Transactions on Parallel and Dis- tributed Systems, vol. 20, no. 12, pp. 1777–1789, 2009

    work page 2009

  7. [7]

    Unified breakdown analysis for byzantine robust gossip,

    R. Gaucher, A. Dieuleveut, and H. Hendrikx, “Unified breakdown analysis for byzantine robust gossip,”arXiv preprint arXiv:2410.10418, 2024

    arXiv 2024

  8. [8]

    Decentralized federated averaging,

    T. Sun, D. Li, and B. Wang, “Decentralized federated averaging,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 4, pp. 4289–4301, 2023

    work page 2023

Show all 37 references
  1. [9]

    Decentralized federated learning: Balancing communication and computing costs,

    W. Liu, L. Chen, and W. Zhang, “Decentralized federated learning: Balancing communication and computing costs,”IEEE Transactions on Signal and Information Processing over Networks, vol. 8, pp. 131–143, 2022

    2022

  2. [10]

    Fully decentralized joint learning of personalized models and collaboration graphs,

    V . Zantedeschi, A. Bellet, and M. Tommasi, “Fully decentralized joint learning of personalized models and collaboration graphs,” inProceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, ser. Proceedings of Machine Learning Research...

    2020

  3. [11]

    Learning to collaborate in decentralized learning of personalized models,

    S. Li, T. Zhou, X. Tian, and D. Tao, “Learning to collaborate in decentralized learning of personalized models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 9766–9775

    2022

  4. [12]

    Personalized federated learning with inferred collaboration graphs,

    R. Ye, Z. Ni, F. Wu, S. Chen, and Y . Wang, “Personalized federated learning with inferred collaboration graphs,” inInternational conference on machine learning. PMLR, 2023, pp. 39 801–39 817

    2023

  5. [13]

    Evidential deep learning to quantify classification uncertainty,

    M. Sensoy, L. Kaplan, and M. Kandemir, “Evidential deep learning to quantify classification uncertainty,”Advances in Neural Information Processing Systems, vol. 31, 2018

    2018

  6. [14]

    A unified theory of decentralized sgd with changing topology and local updates,

    A. Koloskova, N. Loizou, S. Boreiri, M. Jaggi, and S. Stich, “A unified theory of decentralized sgd with changing topology and local updates,” inInternational conference on machine learning. PMLR, 2020, pp. 5381–5393

    2020

  7. [15]

    Distributed optimization over time-varying directed graphs,

    A. Nedi ´c and A. Olshevsky, “Distributed optimization over time-varying directed graphs,”IEEE Transactions on Automatic Control, vol. 60, no. 3, pp. 601–615, 2014. 15

    2014

  8. [16]

    Stochastic gradient-push for strongly convex functions on time-varying directed graphs,

    A. Nedi ´c and A. Olshevsky, “Stochastic gradient-push for strongly convex functions on time-varying directed graphs,”IEEE Transactions on Automatic Control, vol. 61, no. 12, pp. 3936–3947, 2016

    2016

  9. [17]

    Refined convergence and topology learning for decentralized sgd with heterogeneous data,

    B. Le Bars, A. Bellet, M. Tommasi, E. Lavoie, and A.-M. Kermarrec, “Refined convergence and topology learning for decentralized sgd with heterogeneous data,” inInternational Conference on Artificial Intelli- gence and Statistics. PMLR, 2023, pp. 1672–1702

    2023

  10. [18]

    Privacy-preserving decentralized federated learning over time-varying communication graph,

    Y . Lu, Z. Yu, and N. Suri, “Privacy-preserving decentralized federated learning over time-varying communication graph,”ACM Transactions on Privacy and Security, vol. 26, no. 3, pp. 1–39, 2023

    2023

  11. [19]

    Dystop: Dynamic staleness control and topology construction for asynchronous decentralized federated learning,

    Y . Shi, Q. Ma, Y . Xu, J. Zhou, M. Hu, Y . Liao, and H. Xu, “Dystop: Dynamic staleness control and topology construction for asynchronous decentralized federated learning,”IEEE Transactions on Mobile Com- puting, 2026

    2026

  12. [20]

    Decentralized federated learning over time-varying and heterogeneous mobile computing networks,

    B. Li, W. Gao, X. Deng, J. Xie, Z. Xiong, M. Siew, B. Guo, S. Mao, and Z. Han, “Decentralized federated learning over time-varying and heterogeneous mobile computing networks,”IEEE Transactions on Mobile Computing, vol. 25, no. 5, pp. 6688–6704, 2026

    2026

  13. [21]

    Bridge: Byzantine-resilient decen- tralized gradient descent,

    C. Fang, Z. Yang, and W. U. Bajwa, “Bridge: Byzantine-resilient decen- tralized gradient descent,”IEEE Transactions on Signal and Information Processing over Networks, vol. 8, pp. 610–626, 2022

    2022

  14. [22]

    Byzantine-resilient decentralized stochastic gradient descent,

    S. Guo, T. Zhang, H. Yu, X. Xie, L. Ma, T. Xiang, and Y . Liu, “Byzantine-resilient decentralized stochastic gradient descent,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 6, pp. 4096–4106, 2021

    2021

  15. [23]

    Byzantine-resilient decentralized stochas- tic optimization with robust aggregation rules,

    Z. Wu, T. Chen, and Q. Ling, “Byzantine-resilient decentralized stochas- tic optimization with robust aggregation rules,”IEEE transactions on signal processing, vol. 71, pp. 3179–3195, 2023

    2023

  16. [24]

    Byzantine-robust decentralized learning via clippedgossip,

    L. He, S. P. Karimireddy, and M. Jaggi, “Byzantine-robust decentralized learning via clippedgossip,”arXiv preprint arXiv:2202.01545, 2022

    2022 arXiv

  17. [25]

    Byzantine-robust decentralized federated learning,

    M. Fang, Z. Zhang, Hairi, P. Khanduri, J. Liu, S. Lu, Y . Liu, and N. Gong, “Byzantine-robust decentralized federated learning,” inPro- ceedings of the 2024 ACM SIGSAC Conference on Computer and Communications Security, 2024, pp. 2874–2888

    2024

  18. [26]

    A decentralized federated learning framework via committee mechanism with convergence guar- antee,

    C. Che, X. Li, C. Chen, X. He, and Z. Zheng, “A decentralized federated learning framework via committee mechanism with convergence guar- antee,”IEEE Transactions on Parallel and Distributed Systems, vol. 33, no. 12, pp. 4783–4800, 2022

    2022

  19. [27]

    Byzantine-resilient federated learning at edge,

    Y . Tao, S. Cui, W. Xu, H. Yin, D. Yu, W. Liang, and X. Cheng, “Byzantine-resilient federated learning at edge,”IEEE Transactions on Computers, vol. 72, no. 9, pp. 2600–2614, 2023

    2023

  20. [28]

    Trail: Topology authentication in rpl,

    H. Perrey, M. Landsmann, O. Ugus, T. C. Schmidt, and M. W ¨ahlisch, “Trail: Topology authentication in rpl,”arXiv preprint arXiv:1312.0984, 2013

    2013 arXiv

  21. [29]

    Reputation propagation and updating in mobile ad hoc networks with byzantine failures,

    C. Li, M. Hurfin, and Y . Wang, “Reputation propagation and updating in mobile ad hoc networks with byzantine failures,” in2015 IEEE Trustcom/BigDataSE/ISPA, vol. 1, 2015, pp. 111–118

    2015

  22. [30]

    Privacy- preserving byzantine-robust federated learning via blockchain systems,

    Y . Miao, Z. Liu, H. Li, K.-K. R. Choo, and R. H. Deng, “Privacy- preserving byzantine-robust federated learning via blockchain systems,” IEEE Transactions on Information Forensics and Security, vol. 17, pp. 2848–2861, 2022

    2022

  23. [31]

    Sketchguard: Scaling byzantine-robust decentralized federated learning via sketch- based screening,

    M. Rangwala, F. Azzedin, R. O. Sinnott, and R. Buyya, “Sketchguard: Scaling byzantine-robust decentralized federated learning via sketch- based screening,”arXiv preprint arXiv:2510.07922, 2025

    2025 arXiv

  24. [32]

    Fully-distributed construction of byzantine-resilient dynamic peer-to-peer networks,

    A. Gupta and G. Pandurangan, “Fully-distributed construction of byzantine-resilient dynamic peer-to-peer networks,”arXiv preprint arXiv:2506.04368, 2025

    2025 arXiv

  25. [33]

    Byzantine-resilient decentralized stochas- tic optimization with robust aggregation rules,

    Z. Wu, T. Chen, and Q. Ling, “Byzantine-resilient decentralized stochas- tic optimization with robust aggregation rules,”IEEE Transactions on Signal Processing, vol. 71, pp. 3179–3195, 2023

    2023

  26. [34]

    Ma- chine learning with adversaries: Byzantine tolerant gradient descent,

    P. Blanchard, E. M. El Mhamdi, R. Guerraoui, and J. Stainer, “Ma- chine learning with adversaries: Byzantine tolerant gradient descent,” in Advances in Neural Information Processing Systems, vol. 30, 2017

    2017

  27. [35]

    A public domain dataset for human activity recognition using smartphones,

    D. Anguita, A. Ghio, L. Oneto, X. Parra, and J. L. Reyes-Ortiz, “A public domain dataset for human activity recognition using smartphones,” in European Symposium on Artificial Neural Networks (ESANN), vol. 3, 2013, p. 3

    2013

  28. [36]

    Introducing a new benchmarked dataset for activity monitoring,

    A. Reiss and D. Stricker, “Introducing a new benchmarked dataset for activity monitoring,” inProceedings of the 2012 16th Annual International Symposium on Wearable Computers (ISWC), ser. ISWC ’12. USA: IEEE Computer Society, 2012, p. 108–109. Shubham Vaishnavreceived a Ph.D. ...

    2012

  29. [37]

    Moving theU t sum to the left and multiplying by 2: T−1X t=0 E∥U t∥2 F ≤ 32η2N p2 T−1X t=0 4τ 2 scr +σ 2 + 4∥∇f( ¯θt)∥2 + 24N T p2 ρ2. We now substitute this consensus sum into the telescoped Descent Lemma (37): 3η 8 T−1X t=0 E∥∇f∥ 2 ≤(f 0 −f ⋆) + Lη2σ2T N + 2 η +L ρ2T + 64η3L...

Pith tools

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