Pith. sign in

REVIEW 3 major objections 6 minor 41 references

Byzantine-robust aggregation rules such as multi-Krum can be replicated without a coordinator or consensus, yielding byte-identical aggregates once honest replicas share the same updates.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 12:45 UTC pith:RYBZDF6S

load-bearing objection Consistency claim is elementary and solid; the real contribution is the composition that makes discontinuous multi-Krum byte-identical without consensus, with robustness correctly left conditional. the 3 major comments →

arxiv 2607.10305 v1 pith:RYBZDF6S submitted 2026-07-11 cs.DC cs.CRcs.LG

Byzantine Accountability Without Consensus: Strong Eventual Consistency for Non-Associative, Stochastic, Robust Aggregation

Ryan Gillespie This is my paper
classification cs.DC cs.CRcs.LG
keywords Byzantine aggregationCRDTsStrong Eventual Consistencymulti-Krumconsensus-freeequivocation proofsfixed-point arithmeticaccountable eviction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that the usual need for a central server or a consensus protocol in front of Byzantine-robust aggregators is an architectural habit, not a requirement of the rules themselves. Multi-Krum and similar selectors need only an agreed set of contributions and an agreed exclusion predicate over identities; both can be maintained as CRDTs that converge without coordination. Everything else—scoring, selection, averaging—is then a pure function of that converged product state, and pure functions of equal states produce equal outputs. ACFA puts signed contributions in a content-addressed OR-Set, self-authenticating equivocation proofs in a grow-only set, and evaluates full-dimension multi-Krum in exact fixed-point arithmetic with hash-canonical order. The result is Strong Eventual Consistency plus permanent, vote-free eviction of equivocators. A reader cares because decentralised learning can obtain exact, auditable agreement on the filtered aggregate without the latency and coordination cost of BFT or blockchains, while still importing the statistical robustness of the original rule under stated conditions.

Core claim

Any pure total function of a product of convergent replicated data types inherits Strong Eventual Consistency, even when the function is non-monotone, non-associative, or stochastic with state-derived entropy. Applied to multi-Krum, honest replicas that have received the same contributions and the same offline-verifiable equivocation proofs therefore compute byte-identical aggregates without agreeing on arrival order or running a consensus protocol.

What carries the argument

The product-lifting theorem (any pure function of a converged product of CvRDTs inherits SEC) applied to the ACFA product state: a contribution OR-Set composed with a grow-only set of self-authenticating equivocation proofs, reduced by a fixed-point, hash-ordered multi-Krum kernel that is a pure function of the admitted set.

Load-bearing premise

The imported statistical robustness of multi-Krum only holds when the honest score gap is large enough to survive the worst-case rounding error of fixed-point arithmetic; the paper does not prove this always occurs.

What would settle it

Re-run or re-implement the 16-check falsification battery under adversarial schedules: if any two honest replicas that have received the same updates ever produce different 32-byte output roots, or if a late equivocation proof fails to re-converge them to one root, the consistency claim is false.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Coordinator-free federated or gossip learning can force exact, not merely statistical, agreement on the robust aggregate.
  • A late-arriving proof of equivocation deterministically re-converges every honest replica to the same post-eviction output with no extra protocol.
  • The same lifting applies unchanged to any other pure robust selector (for example Bulyan), so the consistency layer is rule-agnostic.
  • Accountability becomes permanent eviction of the offending key rather than per-round statistical re-filtering.
  • Tolerance-based approximate agreement is unnecessary for discontinuous selectors once exact integer arithmetic is used.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same product-of-lattices pattern could gate other discontinuous reductions that currently sit behind coordinators, such as median-of-means or clustering-based filters.
  • Multi-round training still needs an external quiescence or epoch-closure signal; the per-round safety guarantee alone does not prevent temporary forks of the round-DAG.
  • A closed-form anti-concentration bound on score gaps would turn the quantisation-margin caveat from an empirical check into a high-probability guarantee.
  • Permissionless deployments would need a Sybil-resistant identity layer in place of the assumed PKI before the construction can leave consortia.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper presents ACFA, a coordinator-free construction that replicates Byzantine-robust aggregation (full-dimension multi-Krum, optionally Bulyan) by composing a content-addressed contribution OR-Set with a grow-only set of self-authenticating equivocation proofs, then evaluating the selector as a pure total function of the converged product state under Q16.16 fixed-point arithmetic and hash-canonical order. It proves that any pure function of a product of CvRDTs inherits Strong Eventual Consistency (Theorem 7), gives a converse characterisation of realisable implementations (Theorem 8), and shows accountable permanent eviction by offline-verifiable proofs without votes (Theorem 10). Consistency (Theorem 14 Part A) is separated from imported, conditional robustness (Part B) and from multi-round quiescence. A prototype (n=10, f≤3) passes a 16-check falsification battery including three ablations that each break byte-identity, plus late-proof re-convergence and partition healing.

Significance. If the results hold, the paper cleanly shows that global coupling, non-associativity, and discontinuity of robust selectors do not force a coordinator or consensus for safety: only an agreed set and an agreed exclusion predicate are required, both of which are join-semilattices. The product-lifting theorem is elementary (as the authors state), but the composition of a data lattice with an evidence lattice gating a discontinuous selector under exact integer arithmetic, together with the converse (Theorem 8) and the late-proof re-convergence property (Theorem 10(iii)), is a useful and carefully scoped contribution for Byzantine-CRDT and decentralised-learning systems. Strengths include honest separation of consistency from accuracy and of safety from liveness, discharge of environmental determinism by integer arithmetic forced by Lemma 2, a falsification battery with mechanism-removing ablations, and explicit open items (quantisation-margin anti-concentration, multi-round epoch sync). The robustness leg is imported and conditional, which the paper correctly labels rather than overclaiming.

major comments (3)
  1. Theorem 14 Part B and Lemma 12 / Remark 13: Part B states that the imported multi-Krum envelope holds under |A|≥2f+3 and the quantisation-margin condition g>2β, while Remark 13 leaves a closed-form anti-concentration bound on the score-gap distribution as open and reports only empirical zero flips on 1600 configurations (including near-tie). Because multi-Krum is discontinuous (Lemma 2), a residual flip moves the aggregate by Δ/m. Please rephrase Part B so it is unambiguously a conditional corollary of [5] under a checkable margin predicate, not a free-standing robustness theorem of this paper; either supply a high-probability bound under a stated honest-input model or demote the envelope language further so that the only unconditional claim remains Part A.
  2. Section 1 (scope) and Section 7 (multi-round execution): the consensus-free claim is correctly restricted to per-round resolve over a converged admitted set, with round-closure / epoch forking delegated to an orthogonal mechanism. This is load-bearing for any multi-round training pipeline that would actually use ACFA. Elevate this limitation into the abstract and the contributions list with the same prominence as “consistency, not accuracy,” so that the title claim cannot be over-read as end-to-end multi-round coordinator-free training.
  3. Reproducibility (final paragraph and Appendix A): the prototype and harnesses are stated to be released upon publication and available on request. For a systems paper whose central empirical support is a 16/16 binary falsification battery plus three ablations, the artifact should be available to reviewers (or a permanent anonymous archive linked in the submission). Without it, the byte-identity claims and the Lemma 1 / A3 transcripts cannot be independently checked at review time.
minor comments (6)
  1. Section 5, Theorem 8: the measure-theoretic entropy argument is sketched rather than fully formalised (“measurability details are in the proof”). For journal archival, expand the Borel-set argument or move a complete proof to an appendix so the converse is checkable without reconstruction.
  2. Section 6.1 / Table 1: state explicitly that all checks are single-architecture; cross-architecture byte-identity is argued from integer semantics (Section 7.1) but not measured. A one-sentence caveat in the table caption would prevent over-reading.
  3. Remark 5 and Section 4.3: the retirement of the subsampled / Merkle-seeded kernel is well motivated (grinding, bound mismatch). Consider moving the grinding negative result (X9, 0/900) into the main text near Remark 5 so the design change is evidenced where it is introduced.
  4. Related Work (Section 8): the distinctions from BRACE, IOS/WFAgg/BALANCE, blocklace, and Brocco are useful. A short comparison table (guarantee class: statistical convergence vs byte-identity; object protected: state vs discontinuous reduction; evidence actuation: graded trust vs binary conviction) would help readers place the contribution faster.
  5. Notation: Admit / Visible / Convicted / R are defined cleanly, but the product merge is written both as ⊔1×⊔2 and as componentwise union. Pick one notation and use it consistently in Theorems 6–7.
  6. Typos / polish: “ulpscale” → “ulp-scale”; abstract “2f + 3” vs body “≥2f+3” — align; a few long sentences in Section 3.2 and Remark 13 would benefit from splitting for readability.

Circularity Check

0 steps flagged

No significant circularity: product-lifting SEC is elementary from purity + CvRDT convergence; self-citations are scaffolding, not load-bearing for the central claim.

full rationale

The paper’s strongest consistency claim (Theorem 7 / Theorem 14 Part A) is that any pure total function of a converged product of CvRDTs evaluates identically at replicas that have received the same updates. The proof is self-contained and non-circular: product of CvRDTs is a CvRDT (Theorem 6, componentwise join), CvRDT convergence yields equal states [37], and a pure function of equal arguments yields equal values. Stochastic selectors are admitted only when entropy is state-derived; the converse (Theorem 8) characterises rather than assumes that restriction. The paper itself labels the lifting step elementary and locates its contribution in the composition (contribution OR-Set × equivocation G-Set + discontinuous multi-Krum under fixed-point arithmetic), not in a novel algebra. Self-citations to the author’s prior two-layer architecture [19] and E4 [20] supply scaffolding and an explicit reduction-class contrast (smooth trust-weighted mean vs discontinuous selector); they are not invoked as uniqueness theorems that force the present result, and the proofs of Theorems 6–8 and 10 do not depend on those works being true. Robustness bounds are imported from Blanchard et al. [5] and Mhamdi et al. [32] under stated population and quantisation-margin conditions that the paper leaves partly open (Remark 13)—conditional import, not circular self-prediction. Experimental checks are binary falsifications of byte-identity, not fitted parameters re-labelled as predictions. Score 1 reflects only routine self-citation for architectural lineage, not a reduction of the central claim to its inputs.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 2 invented entities

The consistency claim rests on standard CvRDT theory, a PKI, eventual pairwise delivery among honest replicas, and the purity/integer-arithmetic discipline of the resolve function. Robustness is imported wholesale from multi-Krum / Bulyan under an admitted-population bound and a quantisation-margin inequality that is only partially closed. No free parameters are fitted to obtain the consistency theorems; Q16.16 and m = |A|-f-2 are design choices. The invented entities are the product state and the ACFA admission/resolve functions, both of which have direct operational definitions and falsifiable byte-identity checks.

free parameters (2)
  • Q16.16 quantisation step δ = 2^{-16}
    Fixed design choice for the integer kernel; not fitted to data, but the dynamic range and the margin condition depend on it. Any other fixed-point width would require a re-stated margin lemma.
  • multi-Krum selection count m = |A| - f - 2
    Standard multi-Krum parameter fixed by the paper (Definition 4); not fitted, but the robustness envelope and the non-foldability lemma are stated for this choice.
axioms (5)
  • standard math CvRDT convergence theorem: replicas that have received the same set of updates converge to identical states regardless of order, duplication or delay.
    Invoked for the product lattice (Theorem 6) and the lifting theorem (Theorem 7); standard from Shapiro et al.
  • domain assumption PKI: every contribution and proof is signed under a public-key infrastructure that authenticates identities.
    Stated in Section 1 and Theorem 10; restricts the construction to permissioned settings and makes validity of equivocation proofs objective.
  • domain assumption Eventual pairwise delivery among honest replicas.
    Required for Theorem 10 (accountable eviction) and for SEC; standard asynchronous-network assumption, no timing bounds.
  • domain assumption Imported multi-Krum / Bulyan statistical robustness under |A| ≥ 2f+3 (resp. 4f+3) and at most f Byzantine admitted contributions, plus the quantisation-margin condition g > 2β of Lemma 12.
    Theorem 14 Part B; the paper does not re-derive the envelope and leaves anti-concentration of the score gap open.
  • domain assumption Integer (fixed-point) arithmetic is exact and identical on any architecture implementing the stated width contract.
    Forced by discontinuity (Lemma 2) and used to discharge the environmental-determinism assumption of the prior architecture; cross-architecture identity is argued, not measured.
invented entities (2)
  • ACFA product state (contribution OR-Set × equivocation-proof G-Set) independent evidence
    purpose: Replicates the two objects a robust selector needs (agreed set + agreed exclusion predicate) as a product CvRDT so that the selector can be a pure function of converged state.
    Composition of two standard CRDTs; independent evidence is the byte-identity falsification battery.
  • Admit / Resolve pure functions (Definition 3–4) independent evidence
    purpose: Deterministic admission (uniqueness + conviction) and fixed-point multi-Krum with hash-canonical order and content-hash tie-break.
    Operational definitions with explicit ablations showing each determinism mechanism is load-bearing; falsifiable by root equality.

pith-pipeline@v1.1.0-grok45 · 23724 in / 3870 out tokens · 32400 ms · 2026-07-14T12:45:28.535258+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Byzantine Accountability Without Consensus: Strong Eventual Consistency for Non-Associative, Stochastic, Robust Aggregation." pith.science (2026). https://pith.science/paper/RYBZDF6S

@misc{pith2026260710305,
  author       = {Pith},
  title        = {Pith review of: Byzantine Accountability Without Consensus: Strong Eventual Consistency for Non-Associative, Stochastic, Robust Aggregation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RYBZDF6S}},
  note         = {Machine review of arXiv:2607.10305}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Byzantine-robust aggregation rules such as multi-Krum assume a central coordinator, and decentralising them is obstructed by the rules themselves: they are globally coupled, non-associative, and discontinuous, so an ulpscale perturbation can flip the selected subset, moving the output by a non-vanishing amount. None of this prevents coordinator-free replication, because a robust rule needs no agreed order of contributions, only an agreed set and an agreed exclusion predicate, both of which converge without consensus. ACFA (Accountable Consensus-Free Aggregation) replicates a content-addressed OR-Set of signed contributions and a grow-only set of self-authenticating equivocation proofs, offline-verifiable by anyone. Aggregation is a deterministic pure function of the converged product state: fixed-point integer arithmetic over a hash-canonical order, ties broken by content hash. We prove that any pure function of a converged product of CRDTs (non-monotone, non-associative, or stochastic) inherits Strong Eventual Consistency, together with its converse; the contribution is the composition of a data lattice with an evidence lattice applied to a robust selector, not the elementary lifting step. A prototype (10 nodes, 3 Byzantine) passes 16/16 falsification checks: byte-identical roots under adversarial gossip, deterministic re-convergence after late equivocation proofs, partition recovery, and three byte-identity-breaking ablations. The guarantee is consistency, not accuracy; robustness is imported, conditional on 2f + 3 admitted contributions (at most f Byzantine) and a stated quantisation-margin condition.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 9 linked inside Pith

  1. [1]

    The blocklace: A Byzantine-repelling and universal conflict-free replicated data type.arXiv preprint arXiv:2402.08068, 2024

    Paulo Sérgio Almeida and Ehud Shapiro. The blocklace: A Byzantine-repelling and universal conflict-free replicated data type.arXiv preprint arXiv:2402.08068, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    Delta state replicated data types.Journal of Parallel and Distributed Computing, 111:162–173, 2018

    Paulo Sérgio Almeida, Ali Shoker, and Carlos Baquero. Delta state replicated data types.Journal of Parallel and Distributed Computing, 111:162–173, 2018

    2018

  3. [3]

    Composition in state-based replicated data types.Bulletin of the European Association for Theoretical Computer Science (EATCS), 123, 2017

    Carlos Baquero, Paulo Sérgio Almeida, Alcino Cunha, and Carla Ferreira. Composition in state-based replicated data types.Bulletin of the European Association for Theoretical Computer Science (EATCS), 123, 2017

    2017

  4. [4]

    A little is enough: Circumventing defenses for distributed learning

    Gilad Baruch, Moran Baruch, and Yoav Goldberg. A little is enough: Circumventing defenses for distributed learning. InAdvances in Neural Information Processing Systems 32 (NeurIPS), 2019

    2019

  5. [5]

    Machine learning with adversaries: Byzantine tolerant gradient descent

    Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. InAdvances in Neural Information Processing Systems 30 (NeurIPS), pages 119–129, 2017

    2017

  6. [6]

    A composable CRDT layer for Byzantine-resilient deterministic reconstruction.arXiv preprint arXiv:2606.18966, 2026

    Amos Brocco. A composable CRDT layer for Byzantine-resilient deterministic reconstruction.arXiv preprint arXiv:2606.18966, 2026

    Pith/arXiv arXiv 2026

  7. [7]

    Decoupling trust in Byzantine CRDTs: Fine-grained post-compromise handling without breaking causality.arXiv preprint arXiv:2606.31759, 2026

    Amos Brocco. Decoupling trust in Byzantine CRDTs: Fine-grained post-compromise handling without breaking causality.arXiv preprint arXiv:2606.31759, 2026

    Pith/arXiv arXiv 2026

  8. [8]

    Casper the friendly finality gadget.arXiv preprint arXiv:1710.09437, 2017

    Vitalik Buterin and Virgil Griffith. Casper the friendly finality gadget.arXiv preprint arXiv:1710.09437, 2017

    Pith/arXiv arXiv 2017

  9. [9]

    Díaz-Redondo, and Manuel Fernández-Veiga

    Diego Cajaraville-Aboy, Ana Fernández-Vilas, Rebeca P. Díaz-Redondo, and Manuel Fernández-Veiga. Byzantine- robust aggregation for securing decentralized federated learning.IEEE Access, 13:190947–190963, 2025

    2025

  10. [10]

    FLTrust: Byzantine-robust federated learning via trust bootstrapping

    Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. FLTrust: Byzantine-robust federated learning via trust bootstrapping. InProceedings of the Network and Distributed System Security Symposium (NDSS), 2021

    2021

  11. [11]

    Polygraph: Accountable Byzantine agreement

    Pierre Civit, Seth Gilbert, and Vincent Gramoli. Polygraph: Accountable Byzantine agreement. InProceedings of the 41st IEEE International Conference on Distributed Computing Systems (ICDCS), pages 403–413, 2021

    2021

  12. [12]

    Marczak, Peter Alvaro, Joseph M

    Neil Conway, William R. Marczak, Peter Alvaro, Joseph M. Hellerstein, and David Maier. Logic and lattices for distributed programming. InProceedings of the 3rd ACM Symposium on Cloud Computing (SoCC), 2012

    2012

  13. [13]

    Accountability and reconfiguration: Self-healing lattice agreement

    Luciano Freitas de Souza, Petr Kuznetsov, Thibault Rieutord, and Sara Tucci-Piergiovanni. Accountability and reconfiguration: Self-healing lattice agreement. In25th International Conference on Principles of Distributed Systems (OPODIS), volume 217 ofLIPIcs, pages 25:1–25:23, 2021

    2021

  14. [14]

    Lynch, Shlomit S

    Danny Dolev, Nancy A. Lynch, Shlomit S. Pinter, Eugene W. Stark, and William E. Weihl. Reaching approximate agreement in the presence of faults.Journal of the ACM, 33(3):499–516, 1986

    1986

  15. [15]

    Rusu, Rachita Chhaparia, Yani Donchev, Adhiguna Kuncoro, Marc’Aurelio Ranzato, Arthur Szlam, and Jiajun Shen

    Arthur Douillard, Qixuan Feng, Andrei A. Rusu, Rachita Chhaparia, Yani Donchev, Adhiguna Kuncoro, Marc’Aurelio Ranzato, Arthur Szlam, and Jiajun Shen. DiLoCo: Distributed low-communication training of language models, 2023. arXiv preprint arXiv:2311.08105

    Pith/arXiv arXiv 2023

  16. [16]

    Byzantine-robust federated learning over ring-all-reduce distributed computing

    Minghong Fang, Zhuqing Liu, Xuecen Zhao, and Jia Liu. Byzantine-robust federated learning over ring-all-reduce distributed computing. InCompanion Proceedings of the ACM Web Conference 2025 (WWW Companion), pages 961–965, 2025

    2025

  17. [17]

    Byzantine-robust decentralized federated learning

    Minghong Fang, Zifan Zhang, Hairi, Prashant Khanduri, Jia Liu, Songtao Lu, Yuchen Liu, and Neil Zhenqiang Gong. Byzantine-robust decentralized federated learning. InProceedings of the 2024 ACM SIGSAC Conference on Computer and Communications Security (CCS), 2024

    2024

  18. [18]

    Victoria Luzón, and Francisco Herrera

    Mario García-Márquez, Nuria Rodríguez-Barroso, M. Victoria Luzón, and Francisco Herrera. Krum federated chain (KFC): Using blockchain to defend against adversarial attacks in federated learning.arXiv preprint arXiv:2502.06917, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Conflict-free replicated data types for neural network model merging: A two-layer architecture enabling CRDT-compliant model merging across 26 strategies, 2026

    Ryan Gillespie. Conflict-free replicated data types for neural network model merging: A two-layer architecture enabling CRDT-compliant model merging across 26 strategies, 2026. arXiv preprint arXiv:2605.19373. 13

    Pith/arXiv arXiv 2026

  20. [20]

    Recursive trust-as-data binding for byzantine fault tolerant conflict-free replicated data type synchronisation, 2026

    Ryan Gillespie. Recursive trust-as-data binding for byzantine fault tolerant conflict-free replicated data type synchronisation, 2026. UK Patent Application GB2608127.3; implemented in thecrdt-mergelibrary v0.9.5+

    2026

  21. [21]

    What every computer scientist should know about floating-point arithmetic.ACM Computing Surveys, 23(1):5–48, 1991

    David Goldberg. What every computer scientist should know about floating-point arithmetic.ACM Computing Surveys, 23(1):5–48, 1991

    1991

  22. [22]

    PeerReview: Practical accountability for distributed systems

    Andreas Haeberlen, Petr Kuznetsov, and Peter Druschel. PeerReview: Practical accountability for distributed systems. InProceedings of the 21st ACM SIGOPS Symposium on Operating Systems Principles (SOSP), pages 175–188, 2007

    2007

  23. [23]

    Hellerstein and Peter Alvaro

    Joseph M. Hellerstein and Peter Alvaro. Keeping CALM: When distributed consistency is easy.Communications of the ACM, 63(9):72–81, 2020

    2020

  24. [24]

    Nicholas J. Higham. The accuracy of floating point summation.SIAM Journal on Scientific Computing, 14(4):783–799, 1993

    1993

  25. [25]

    On conflict-free replicated data types and equivocation in Byzantine setups.arXiv preprint arXiv:2109.10554, 2021

    Florian Jacob, Saskia Bayreuther, and Hannes Hartenstein. On conflict-free replicated data types and equivocation in Byzantine setups.arXiv preprint arXiv:2109.10554, 2021

    Pith/arXiv arXiv 2021

  26. [26]

    Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, Rafael G

    Peter Kairouz, H. Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, Rafael G. L. D’Oliveira, Hubert Eichner, Salim El Rouayheb, David Evans, Josh Gardner, Zachary Garrett, Adrià Gascón, Badih Ghazi, Phillip B. Gibbons, Marco Gruteser, Zaid Harchaoui, Cha...

    2021

  27. [27]

    Making CRDTs Byzantine fault tolerant

    Martin Kleppmann. Making CRDTs Byzantine fault tolerant. InProceedings of the 9th Workshop on Principles and Practice of Consistency for Distributed Data (PaPoC), pages 8–15, 2022

    2022

  28. [28]

    Stich, and Martin Jaggi

    Anastasia Koloskova, Sebastian U. Stich, and Martin Jaggi. Decentralized stochastic optimization and gossip algorithms with compressed communication. InProceedings of the 36th International Conference on Machine Learning (ICML), volume 97, pages 3478–3487, 2019

    2019

  29. [29]

    Lindsey Kuper and Ryan R. Newton. LVars: Lattice-based data structures for deterministic parallelism. In Proceedings of the 2nd ACM SIGPLAN Workshop on Functional High-Performance Computing (FHPC), pages 71–84, 2013

    2013

  30. [30]

    Can decentralized algorithms outperform centralized algorithms? A case study for decentralized parallel stochastic gradient descent

    Xiangru Lian, Ce Zhang, Huan Zhang, Cho-Jui Hsieh, Wei Zhang, and Ji Liu. Can decentralized algorithms outperform centralized algorithms? A case study for decentralized parallel stochastic gradient descent. In Advances in Neural Information Processing Systems 30 (NeurIPS), 2017

    2017

  31. [31]

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas

    H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. Communication- efficient learning of deep networks from decentralized data. InProceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), pages 1273–1282, 2017

    2017

  32. [32]

    The hidden vulnerability of distributed learning in Byzantium

    El Mahdi El Mhamdi, Rachid Guerraoui, and Sébastien Rouault. The hidden vulnerability of distributed learning in Byzantium. InProceedings of the 35th International Conference on Machine Learning (ICML), pages 3521–3530, 2018

    2018

  33. [33]

    Rundungsfehleranalyse einiger verfahren zur summation endlicher summen.Zeitschrift für Angewandte Mathematik und Mechanik (ZAMM), 54(1):39–51, 1974

    Arnold Neumaier. Rundungsfehleranalyse einiger verfahren zur summation endlicher summen.Zeitschrift für Angewandte Mathematik und Mechanik (ZAMM), 54(1):39–51, 1974

    1974

  34. [34]

    Improve accuracy of builtin sum() for float inputs (compensated summation, CPython 3.12)

    Python Software Foundation. Improve accuracy of builtin sum() for float inputs (compensated summation, CPython 3.12). CPython issue 100425, github.com/python/cpython/issues/100425, 2022

    2022

  35. [35]

    Merkle-CRDTs: Merkle-DAGs meet CRDTs.arXiv preprint arXiv:2004.00107, 2020

    Hector Sanjuan, Samuli Poyhtari, Pedro Teixeira, and Ioannis Psaras. Merkle-CRDTs: Merkle-DAGs meet CRDTs.arXiv preprint arXiv:2004.00107, 2020. 14

    Pith/arXiv arXiv 2004

  36. [36]

    A comprehensive study of convergent and commutative replicated data types

    Marc Shapiro, Nuno Preguiça, Carlos Baquero, and Marek Zawirski. A comprehensive study of convergent and commutative replicated data types. Technical Report RR-7506, INRIA, 2011

    2011

  37. [37]

    Conflict-free replicated data types

    Marc Shapiro, Nuno Preguiça, Carlos Baquero, and Marek Zawirski. Conflict-free replicated data types. In Proceedings of the 13th International Symposium on Stabilization, Safety, and Security of Distributed Systems (SSS), volume 6976 ofLecture Notes in Computer Science, pages 386–400. Springer, 2011

    2011

  38. [38]

    Manipulating the Byzantine: Optimizing model poisoning attacks and defenses for federated learning

    Virat Shejwalkar and Amir Houmansadr. Manipulating the Byzantine: Optimizing model poisoning attacks and defenses for federated learning. InProceedings of the Network and Distributed System Security Symposium (NDSS), 2021

    2021

  39. [39]

    Eventually consistent.Communications of the ACM, 52(1):40–44, 2009

    Werner Vogels. Eventually consistent.Communications of the ACM, 52(1):40–44, 2009

    2009

  40. [40]

    Byzantine-resilient decentralized stochastic optimization with robust aggregation rules.IEEE Transactions on Signal Processing, 71:3179–3195, 2023

    Zhaoxian Wu, Tianyi Chen, and Qing Ling. Byzantine-resilient decentralized stochastic optimization with robust aggregation rules.IEEE Transactions on Signal Processing, 71:3179–3195, 2023

    2023

  41. [41]

    Bartlett

    Dong Yin, Yudong Chen, Kannan Ramchandran, and Peter L. Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. InProceedings of the 35th International Conference on Machine Learning (ICML), pages 5650–5659, 2018. A Falsifier Transcript Summary The battery runs aspython3 falsifier.py (CPU only; about four seconds on commodity h...

    2018