Poster: FedBlockParadox -- A Framework for Simulating and Securing Decentralized Federated Learning

Federico Caroli; Francesco Bleggi; Gabriele Digregorio; Michele Carminati; Stefano Longari; Stefano Zanero

REVIEW 3 major objections 4 minor 17 references

Poster: FedBlockParadox -- A Framework for Simulating and Securing Decentralized Federated Learning

T0 review · 3 major / 4 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 FedBlockParadox is a modular, open-source simulator that measures how blockchain-based federated learning performs under adversarial nodes.

desk verdict A useful modular blockchain-FL simulator whose paper undersells itself with aggregated, confounded experimental plots; the open-source code is the real artifact. read the letter →

arxiv 2506.02679 v1 pith:Z7B4QZ4K submitted 2025-06-03 cs.CR

Gabriele Digregorio , Francesco Bleggi , Federico Caroli , Michele Carminati , Stefano Zanero , Stefano Longari This is my paper

classification cs.CR

keywords federatedlearningblockchaindecentralizedsimulationframeworkadversarialattacksconsensusmechanismsmodelpoisoningrobustaggregation

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

FedBlockParadox is a simulation framework for decentralized federated learning systems built on blockchain technology. The paper argues that existing work combining federated learning with blockchain lacks flexible, configurable tools for testing robustness against adversaries, and that FedBlockParadox fills that gap. It supports pluggable consensus protocols (Proof of Work, Proof of Stake, committee), validation rules (from accept-all to accuracy thresholds and Multi-Krum), aggregation methods (FedAvg, Mean, Median), and attack models (label flipping, data poisoning, additive noise), with the number of malicious nodes as a live parameter. The authors' preliminary experiments on MNIST and CIFAR-10 report baseline accuracy curves under these settings, positioning the framework as a common testbed for fair comparisons. A sympathetic reader would care because the framework gives the community a standardized way to evaluate whether proposed defenses against poisoning and consensus attacks actually hold up.

What carries the argument

The central object is the simulation harness itself, whose pluggable components mirror the decision points of real blockchain-based federated learning systems: consensus algorithms decide which nodes certify updates, validation algorithms decide whether an update is accepted (accept-all, accuracy threshold on a global or local dataset, or Multi-Krum distance filtering), aggregation algorithms combine accepted updates (FedAvg, Mean, Median), and attack models corrupt updates from a chosen fraction of nodes. The work of this machinery is to isolate each design choice so that robustness can be attributed to a specific mechanism under a fixed experimental setting.

What would settle it

Give a specific configuration, for example 45 percent label-flipping nodes with committee consensus and FedAvg, and run it both in the simulator and on a small real deployment with the same model, data partitions, and validation rules; if the real system shows a different accuracy ordering across consensus mechanisms, or if configurations the simulator ranks as resistant collapse in practice, the claim that the simulation captures the adversarial dynamics would be falsified.

Watch

Extended reading notes

Core claim

The central claim is that a single open-source, extensible simulator can reproduce the main design choices found across the state of the art in blockchain-based decentralized federated learning and subject them to controlled adversarial conditions. On the paper's own terms, FedBlockParadox is the contribution: it lets a researcher configure the consensus mechanism, the update-validation rule, the aggregation rule, the data partition strategy, and the attack type, then run repeated rounds in which some fraction of nodes behaves maliciously and measure the resulting global model accuracy. The preliminary experiments are meant to show the tool runs end to end and produces baseline numbers for PoW, PoS, and committee consensus under label flipping, data poisoning, and additive noise, with attacker fractions up to 45 percent. The authors intend this as a research-enabling step, not as a new defense or a formal guarantee.

Load-bearing premise

The load-bearing premise is that a simulator can faithfully mimic a real blockchain-based federated learning system, including how consensus validation behaves under failure and how adversarial updates spread, so that the measured robustness transfers to actual deployments; the paper does not yet validate that equivalence.

Editorial extensions

If this is right

Researchers can run the same attack scenario across PoW, PoS, and committee consensus and directly compare which consensus mechanism preserves accuracy.
Because validation rules and aggregation rules are independent configuration axes, the framework can isolate whether robustness comes from validation (Multi-Krum, accuracy thresholds) or aggregation, rather than from the system as a whole.
New defense algorithms can be added to the framework and evaluated under the same experimental pipeline, which the authors argue enables fair comparisons across architectures.
The public MNIST and CIFAR-10 baselines under label flipping, data poisoning, and additive noise give future proposals a reference curve to beat.
The framework's attack models allow systematic study of how the proportion of malicious nodes shifts the accuracy-robustness trade-off under each consensus mechanism.

Reading between the lines

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

We infer that the most informative use of FedBlockParadox will be comparative rather than absolute: its accuracy curves are best read as rankings between configurations, not as predictions for a real deployment.
A natural next step the authors do not report would be to validate the simulator against a small physical blockchain-federated-learning testbed, checking whether the configurations that survive in simulation also survive when consensus messages, forks, and validation delays are real.
Because data partitioning (iid versus non-iid) is configurable, the framework could be used to test whether attack impact depends on how skewed each node's local data is, an interaction the paper does not examine.
The framework could also be extended to model adaptive adversaries who observe validation outcomes and change their attack, which would stress-test whether Multi-Krum and committee validation remain resilient to strategic poisoning.

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

FedBlockParadox is a modular open-source simulator for blockchain-based federated learning. The genuinely new thing is the integration, not the components: PoW/PoS/committee consensus, accuracy-based and Multi-Krum validation, FedAvg/mean/median aggregation, and label-flipping/poisoning/noise attacks have all appeared separately in prior work. What the paper offers is a configurable testbed, released on GitHub, that lets researchers mix and match these pieces under adversarial conditions. That is a real contribution for a niche community, and the state-of-the-art summary is fair. The open-source release is the strongest part; it makes the work independently checkable.

The soft spot is exactly where the stress-test note lands. Figure 1 is labeled 'aggregated,' and Section 4 says the experiments varied aggregation, validation, dataset, architecture, consensus, and attacker fraction. Pooling all of that means the differences between PoW, PoS, and committee at a given attack percentage are confounded with everything else. A lower PoW bar could just reflect that those runs used CIFAR-10 or a weaker aggregator. Without per-configuration numbers, error bars, or a matched no-attack baseline, the trends in Figure 1 do not support the claim that the tool yields meaningful baseline robustness numbers. The text says the full experiments are in the repository, so the defect might be presentation rather than the artifact itself — but the manuscript as submitted does not demonstrate it. The deeper question of whether simulations faithfully capture real consensus mechanics is open, but that is a general simulator concern, not a unique flaw.

The conclusion's real-world impact sentence is speculative but typical for a poster. Citation pattern is fine; the self-citations are to the same lab's related work and are not load-bearing. No fitted parameters or circular derivations here.

Who this is for: researchers working on blockchain-based federated learning who need a common evaluation harness. They will get value from it. For the broader ML or security community, it is a niche tool poster.

Recommendation: it deserves a serious referee. I would accept it for review with a request to disaggregate the experimental reporting — a table of configurations, error bars, and matched baselines, plus a clear pointer to repository artifacts that reproduce each subplot. If the repo is complete and the code runs, this is a solid workshop/poster contribution. If the repo is empty, the central claim collapses. Check the link before accepting.

Referee Report

3 major / 4 minor

Summary. The paper presents FedBlockParadox, a modular open-source simulator for blockchain-based decentralized federated learning. The framework supports several consensus algorithms (Proof-of-Work, Proof-of-Stake, committee-based), validation methods (pass-all, accuracy-based, Multi-Krum), aggregation strategies (FedAvg, mean, median), and attack models (label flipping, data poisoning, additive noise), with configurable datasets and data partitioning. Section 4 reports preliminary experiments whose results are summarized in aggregated form in Figure 1, showing accuracy under different consensus mechanisms and malicious-node fractions for each attack category. Section 5 argues that the framework enables reproducible testing and community-driven extensibility, and states an ultimate goal of assessing real-world impact.

Significance. The paper addresses a real gap in the decentralized federated learning literature: the lack of flexible, common testbeds for evaluating robustness under adversarial conditions. The open-source release and explicitly stated configuration interface are concrete, useful contributions that could support future comparisons of blockchain-based federated learning defenses. If the tool works as advertised, its ability to vary consensus and validation mechanisms within one codebase would be valuable. However, the experimental evidence presented in the manuscript does not, in its current form, substantiate the claimed baseline robustness numbers because the reported results are confounded across datasets, architectures, aggregators, and validators. The credibility of the framework therefore rests largely on the existence of the public repository, which the paper does not document in sufficient detail to allow independent verification.

major comments (3)

[Section 4, Figure 1] The results in Figure 1 are aggregated across datasets (MNIST and CIFAR-10), model architectures, aggregation methods, validation processes, and malicious-node fractions. Accuracy differences between PoW, PoS, and Committee therefore cannot be attributed to the consensus mechanism: for example, a lower PoW bar at a given attack fraction could simply reflect that those runs used a harder dataset or a weaker aggregator. The claim that Figure 1 provides baseline accuracy numbers under varying malicious-node fractions is not supported unless per-configuration results are shown. Please report disaggregated results (a table or available online appendix with one row per configuration) with error bars and a matched no-attack baseline for each configuration, or explicitly relabel Figure 1 as illustrative and remove the baseline claim.
[Section 4] The experimental setup is under-specified: the paper does not give the number of simulation rounds, the number of participating nodes, the iid/non-iid partition parameters, the attack magnitudes (e.g., noise variance or poisoning fraction), the validation accuracy thresholds, or the exact model architectures used for each dataset. Without these details, the experiments cannot be reproduced from the manuscript alone, which contradicts the reproducibility claim in Section 5. The repository link is a step in the right direction, but the paper should either include a configuration table in an appendix or reference specific files in the repository and explain what each configuration does.
[Section 5] The paper claims that the framework can assess 'the real-world impact of blockchain-enabled decentralization,' but it does not discuss the fidelity of the simulation to real blockchain mechanics. The simulation abstracts away or simplifies consensus failure modes, network delays, fork resolution, mining difficulty, and incentive structures, all of which can materially affect robustness in actual blockchain-based federated learning. Without a statement of what is abstracted and without validation against at least one reference implementation, the real-world-impact claim is unsupported. Please add a limitations discussion that explicitly lists the abstractions and any validation performed.

minor comments (4)

[Title and header] The paper contains formatting errors in the title and venue name ('FedBlockP aradox', 'DIMV A '25'); these should be corrected in the camera-ready version.
[Figure 1] The three panels use different y-axis ranges (0–1 for additive noise, 0–0.8 for the others), and the captions do not indicate which parameters (dataset, architecture, aggregation) are held fixed or varied. Adding a table of the configurations actually pooled would help the reader understand the figure.
[Section 3] The validation algorithms are listed as Pass-Weights and Pass-Gradients, but the paper does not explain whether these are simply unconditional acceptance or whether 'pass' refers to a syntactic check; one sentence of clarification would avoid ambiguity.
[Section 3] The dataset settings mention 'the percentage of independent and identically distributed (iid) versus non-iid partitions' but do not describe how the non-iid partitions are generated (e.g., Dirichlet distribution, label imbalance). Adding a brief description would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: FedBlockParadox is a modular simulator whose claims are supported by the released implementation and demonstration experiments; self-citations are motivational only.

full rationale

The paper's central claim is the release of FedBlockParadox, a configurable simulation framework for decentralized federated learning. There is no derivation chain, no fitted parameter, and no uniqueness theorem invoked. Section 3 enumerates configuration options (consensus, validation, aggregation, attacks, datasets) as engineering features, not as results derived from other results. Section 4 reports preliminary aggregate accuracy plots and explicitly defers the full breakdown to the repository; even if the aggregation confounds variables, that is an evidentiary limitation, not a circular argument. The only same-author references are [2], [4], and [14], cited in the introduction as examples of federated-learning application domains; they do not justify any load-bearing step in the framework's design or evaluation. No equation in the paper defines one quantity in terms of the target it is supposed to predict. Accordingly, there are no circular steps.

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

The central claim (a functional simulator) depends on three unvalidated domain assumptions: simulation fidelity, coverage of the design space, and the representativeness of the benchmark datasets. These are assumptions inherited from the literature rather than derived, and they are not tested in the poster.

assumptions (3)

domain assumption Blockchain-based decentralized federated learning can be adequately simulated with the modeled components (consensus, validation, aggregation, attacks).
The framework's usefulness relies on fidelity of the simulation to real systems; no validation against real deployment is provided.
domain assumption The selected state-of-the-art designs in the State of the Art section (refs [6], [7], [8], [12], [15], [16]) define the relevant configuration space.
Section 2 reviews these works and Section 3 derives the configurable options from them; the review is not systematic and may miss other valid designs.
domain assumption Standard datasets MNIST and CIFAR-10 are adequate to validate attack resilience.
Section 4 uses these datasets without justification for their representativeness of real federated learning workloads.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Poster: FedBlockParadox -- A Framework for Simulating and Securing Decentralized Federated Learning." pith.science (2026). https://pith.science/paper/Z7B4QZ4K

@misc{pith2026250602679,
  author       = {Pith},
  title        = {Pith review of: Poster: FedBlockParadox -- A Framework for Simulating and Securing Decentralized Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z7B4QZ4K}},
  note         = {Machine review of arXiv:2506.02679}
}

read the original abstract

A significant body of research in decentralized federated learning focuses on combining the privacy-preserving properties of federated learning with the resilience and transparency offered by blockchain-based systems. While these approaches are promising, they often lack flexible tools to evaluate system robustness under adversarial conditions. To fill this gap, we present FedBlockParadox, a modular framework for modeling and evaluating decentralized federated learning systems built on blockchain technologies, with a focus on resilience against a broad spectrum of adversarial attack scenarios. It supports multiple consensus protocols, validation methods, aggregation strategies, and configurable attack models. By enabling controlled experiments, FedBlockParadox provides a valuable resource for researchers developing secure, decentralized learning solutions. The framework is open-source and built to be extensible by the community.

Figures

Figures reproduced from arXiv: 2506.02679 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 12 canonical work pages

[1]

http://yann

The mnist database of handwritten digits. http://yann. lecun. com/exdb/mnist/
[2]

In: 2024 20th International Conference on Wireless and Mo- bile Computing, Networking and Communications (WiMob)

Boiano, A., Di Gennaro, M., Barbieri, L., Carminati, M., Nicoli, M., Redondi, A., Milasheuski, U., Kianoush, S., Savazzi, S., Aillet, A.S., Santos, D.R., Serio, L.: A secure and trustworthy network architecture for federated learning health- care applications. In: 2024 20th International Conference on Wireless and Mo- bile Computing, Networking and Commun...

work page arXiv 2024
[3]

IEEe Access9, 63229–63249 (2021)

Bouacida, N., Mohapatra, P.: Vulnerabilities in federated learning. IEEe Access9, 63229–63249 (2021)

work page 2021
[4]

In: 2024 IEEE 99th Vehicular Technology Conference (VTC2024-Spring)

Digregorio, G., Cainazzo, E., Longari, S., Carminati, M., Zanero, S.: Evaluating the impact of privacy-preserving federated learning on can intrusion detection. In: 2024 IEEE 99th Vehicular Technology Conference (VTC2024-Spring). pp. 1–7. IEEE (2024)

work page 2024
[5]

Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)

2009
[6]

IEEE Network 35(1), 234–241 (2020)

Li, Y., Chen, C., Liu, N., Huang, H., Zheng, Z., Yan, Q.: A blockchain-based decen- tralized federated learning framework with committee consensus. IEEE Network 35(1), 234–241 (2020)

work page 2020
[7]

IEEE Computa- tional Intelligence Magazine17(3), 26–33 (2022)

Ma, C., Li, J., Shi, L., Ding, M., Wang, T., Han, Z., Poor, H.V.: When federated learning meets blockchain: A new distributed learning paradigm. IEEE Computa- tional Intelligence Magazine17(3), 26–33 (2022)

work page 2022
[8]

Knowledge-Based Systems284, 111338 (2024)

Mao, Q., Wang, L., Long, Y., Han, L., Wang, Z., Chen, K.: A blockchain-based framework for federated learning with privacy preservation in power load forecast- ing. Knowledge-Based Systems284, 111338 (2024)

work page 2024

Show all 17 references

[9]

In: Artificial intelligence and statistics

McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273–1282. PMLR (2017)

2017
[10]

IEEE Internet of Things Journal 10(24), 21942–21958 (2023)

Moore, E., Imteaj, A., Rezapour, S., Amini, M.H.: A survey on secure and pri- vate federated learning using blockchain: Theory and application in resource- constrained computing. IEEE Internet of Things Journal 10(24), 21942–21958 (2023)

2023
[11]

IEEE Internet of Things Journal8(16), 12806– 12825 (2021)

Nguyen, D.C., Ding, M., Pham, Q.V., Pathirana, P.N., Le, L.B., Seneviratne, A., Li, J., Niyato, D., Poor, H.V.: Federated learning meets blockchain in edge comput- ing: Opportunities and challenges. IEEE Internet of Things Journal8(16), 12806– 12825 (2021)

2021
[12]

IEEE Internet of Things Journal7(6), 5171–5183 (2020)

Qu, Y., Gao, L., Luan, T.H., Xiang, Y., Yu, S., Li, B., Zheng, G.: Decentralized pri- vacy using blockchain-enabled federated learning in fog computing. IEEE Internet of Things Journal7(6), 5171–5183 (2020)

2020
[13]

ACM Computing Surveys55(4), 1–35 (2022)

Qu, Y., Uddin, M.P., Gan, C., Xiang, Y., Gao, L., Yearwood, J.: Blockchain- enabled federated learning: A survey. ACM Computing Surveys55(4), 1–35 (2022)

2022
[14]

In: 2024 IEEE International Conference on E-health Networking, Application & Services (HealthCom)

Santos, D.R., Aillet, A.S., Boiano, A., Milasheuski, U., Giusti, L., Di Gennaro, M., Kianoush, S., Barbieri, L., Nicoli, M., Carminati, M., Redondi, A.E.C., Savazzi, S., Serio, L.: A federated learning platform as a service for advancing stroke man- agement in european clinica...

2024
[15]

IEEE Transactions on Parallel and Distributed Systems 32(7), 1513–1525 (2020)

Shayan, M., Fung, C., Yoon, C.J., Beschastnikh, I.: Biscotti: A blockchain sys- tem for private and secure federated learning. IEEE Transactions on Parallel and Distributed Systems 32(7), 1513–1525 (2020)

2020
[16]

Mathematics11(14), 3162 (2023)

Zhou, Z., Sun, F., Chen, X., Zhang, D., Han, T., Lan, P.: A decentralized federated learning based on node selection and knowledge distillation. Mathematics11(14), 3162 (2023)

2023
[17]

ACM Computing Surveys 55(11), 1–31 (2023)

Zhu,J.,Cao,J.,Saxena,D.,Jiang,S.,Ferradi,H.:Blockchain-empoweredfederated learning: Challenges, solutions, and future directions. ACM Computing Surveys 55(11), 1–31 (2023)

2023

Pith tools

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

[1] [1]

http://yann

The mnist database of handwritten digits. http://yann. lecun. com/exdb/mnist/

[2] [2]

In: 2024 20th International Conference on Wireless and Mo- bile Computing, Networking and Communications (WiMob)

Boiano, A., Di Gennaro, M., Barbieri, L., Carminati, M., Nicoli, M., Redondi, A., Milasheuski, U., Kianoush, S., Savazzi, S., Aillet, A.S., Santos, D.R., Serio, L.: A secure and trustworthy network architecture for federated learning health- care applications. In: 2024 20th International Conference on Wireless and Mo- bile Computing, Networking and Commun...

work page arXiv 2024

[3] [3]

IEEe Access9, 63229–63249 (2021)

Bouacida, N., Mohapatra, P.: Vulnerabilities in federated learning. IEEe Access9, 63229–63249 (2021)

work page 2021

[4] [4]

In: 2024 IEEE 99th Vehicular Technology Conference (VTC2024-Spring)

Digregorio, G., Cainazzo, E., Longari, S., Carminati, M., Zanero, S.: Evaluating the impact of privacy-preserving federated learning on can intrusion detection. In: 2024 IEEE 99th Vehicular Technology Conference (VTC2024-Spring). pp. 1–7. IEEE (2024)

work page 2024

[5] [5]

Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)

2009

[6] [6]

IEEE Network 35(1), 234–241 (2020)

Li, Y., Chen, C., Liu, N., Huang, H., Zheng, Z., Yan, Q.: A blockchain-based decen- tralized federated learning framework with committee consensus. IEEE Network 35(1), 234–241 (2020)

work page 2020

[7] [7]

IEEE Computa- tional Intelligence Magazine17(3), 26–33 (2022)

Ma, C., Li, J., Shi, L., Ding, M., Wang, T., Han, Z., Poor, H.V.: When federated learning meets blockchain: A new distributed learning paradigm. IEEE Computa- tional Intelligence Magazine17(3), 26–33 (2022)

work page 2022

[8] [8]

Knowledge-Based Systems284, 111338 (2024)

Mao, Q., Wang, L., Long, Y., Han, L., Wang, Z., Chen, K.: A blockchain-based framework for federated learning with privacy preservation in power load forecast- ing. Knowledge-Based Systems284, 111338 (2024)

work page 2024