CoChat
Sign Up
Cited by 119Privacy-Preserving Technologies in DataBlockchain Technology Applications and SecurityInternet of Things and AI

A Blockchain‐Based Federated Learning Method for Smart Healthcare

Open access
Chang, Yuxia, Fang, Chen, Sun, Wenzhuo·2021·Computational Intelligence and Neuroscience
DOI 10.1155/2021/4376418PMC PMC8635913
Open this paper in CoChat — free accountOpen PDF

Abstract

The development of artificial intelligence and worldwide epidemic events has promoted the implementation of smart healthcare while bringing issues of data privacy, malicious attack, and service quality. The Medical Internet of Things (MIoT), along with the technologies of federated learning and blockchain, has become a feasible solution for these issues. In this paper, we present a blockchain‐based federated learning method for smart healthcare in which the edge nodes maintain the blockchain to resist a single point of failure and MIoT devices implement the federated learning to make full of the distributed clinical data. In particular, we design an adaptive differential privacy algorithm to protect data privacy and gradient verification‐based consensus protocol to detect poisoning attacks. We compare our method with two similar methods on a real‐world diabetes dataset. Promising experimental results show that our method can achieve high model accuracy in acceptable running time while also showing good performance in reducing the privacy budget consumption and resisting poisoning attacks.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes a blockchain-based federated learning method for smart healthcare that combines adaptive differential privacy with a gradient-verification consensus protocol. On a real-world diabetes dataset, it achieves 82.7% accuracy at privacy budget ε=3 while reducing privacy-budget consumption and limiting poisoning-attack success to under 20% in later training.

Key findings

  • With adaptive DP, the method exhausts ε=2 and ε=3 in the 36th and 53rd rounds and reaches 78.5% and 82.7% accuracy, respectively; at ε=3 this is only slightly below original FL (84.5%) and BlockFL (84%).
  • At 50 training rounds, running time is 1047 s for original FL, 1702 s for BlockFL, and 1624 s for the proposed method.
  • To reach 80%, 82%, and 84% model accuracy, the method reduces privacy-budget consumption by 57%, 96%, and 81%, respectively, compared with a conventional fixed-clipping DP method.
  • Under a label-flipping poisoning attack with 30% poisoned MIoT devices, the method limits attack success rate to less than 20% in later training, while original FL and BlockFL are almost always above 50%; four MIoT devices are blacklisted by the end of training.

Methods

  • Design: a blockchain-based FL framework where edge nodes maintain the blockchain and MIoT devices train locally; it uses an adaptive differential privacy algorithm that adjusts the gradient clipping threshold based on RMSProp-style prior knowledge, and a gradient-verification consensus protocol based on multi-KRUM and Algorand-like committee selection.
  • Data and sample: experiments use the American National Institute of Diabetes and Digestive and Kidney Diseases diabetes dataset with eight medical predictor variables and a target variable, split 70:30 into training and testing; the dataset is shuffled and divided evenly among 20 simulated MIoT devices.
  • Model and analysis: a CNN with two 5×5 convolution layers, a fully connected layer, and a softmax output layer (1,663,370 parameters) is trained; results compare accuracy, running time, privacy-budget consumption, and poisoning-attack resistance against BlockFL and original FL.

Limitations

  • The paper does not include a dedicated limitations section; it only states future work to continue exploring and advancing the method with public partners.
  • Evident from the text: evaluation is limited to one real-world diabetes dataset and a simulation of 20 MIoT devices, with no real-world clinical deployment reported.
  • Evident from the text: the consensus security assumes no more than 1/3 of miners are malicious, and the method incurs blockchain-related computation overhead and some accuracy loss from DP noise relative to non-private baselines.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1155/2021/4376418

Cites

28 of 28 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 16 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    The proposed method shows good performance in resisting poisoning attacks.

    “Promising experimental results show that our method can achieve high model accuracy in acceptable running time while also showing good performance in reducing the privacy budget consumption and resisting poisoning attacks.”
    Quote in the paperSubstantial claim
  • Findingown data

    The proposed method shows good performance in reducing privacy budget consumption.

    “Promising experimental results show that our method can achieve high model accuracy in acceptable running time while also showing good performance in reducing the privacy budget consumption and resisting poisoning attacks.”
    Quote in the paperSubstantial claim
  • Findingown data

    When training reaches 50 rounds, the running time of original FL, BlockFL, and the proposed method is 1047 s, 1702 s, and 1624 s, respectively.

    “when the training reaches 50 rounds, the running time of original FL, BlockFL, and our method is 1047 s, 1702 s, and 1624 s, respectively.”
    Quote in the paperSubstantial claim
  • Findingreasoning

    The running time of BlockFL and the proposed method is greater than that of original FL.

    “the running time of BlockFL and our method is greater than that of original FL.”
    Quote in the paperSubstantial claim
  • Findingreasoning

    Due to the designed adaptive DP algorithm, the local training phase of the proposed method consumes slightly more time than that of BlockFL.

    “due to the designed adaptive DP algorithm, the local training phase of our method consumes slightly more time than that of BlockFL”
    Quote in the paperNot a substantial claim
  • May combine two claims. 26
  • Classified as probably not a substantial claim. 8

  • May combine two claims.

    Findingown data

    The proposed method achieves high model accuracy in acceptable running time.

    “Promising experimental results show that our method can achieve high model accuracy in acceptable running time while also showing good performance in reducing the privacy budget consumption and resisting poisoning attacks.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    Original FL and BlockFL achieve higher model accuracy than the proposed method.

    “Original FL and BlockFL achieve higher model accuracy than our method”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    When the privacy budget ε=3, the proposed method achieves 82.7% accuracy, which is only slightly lower than 84.5% of original FL and 84% of BlockFL.

    “when the privacy budget ε=3, our method achieves 82.7% accuracy, which is only slightly lower than 84.5% of original FL and 84% of BlockFL.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The larger the privacy budget, the higher the model accuracy but the lower the level of privacy preservation simultaneously.

    “It can be seen that the larger the privacy budget, the higher the model accuracy but the lower the level of privacy preservation simultaneously.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The method exhausts the privacy budget ε=2 and ε=3 in the 36th and 53rd rounds and achieves model accuracy of 78.5% and 82.7%, respectively.

    “Our method exhausts the privacy budget ε=2 and ε=3 in the 36th and 53rd rounds and achieves model accuracy of 78.5% and 82.7%, respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The local training time of BlockFL and the proposed method does not increase with the number of devices.

    “but their local training time does not increase with the number of devices.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Due to the limit of privacy budget ε=3, the proposed method converges in 53 rounds, but original FL and BlockFL cannot converge even within 70 rounds.

    “due to the limit of privacy budget ε=3, our method converges in 53 rounds, but the other two methods cannot converge even within 70 rounds.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingreasoning

    In the first 10 rounds, Ct keeps unchanged because it is fixed as 3 in the initial training stage.

    “in the first 10 rounds, Ct keeps unchanged, which is because we fix it as 3 in the initial training stage according to equation (6).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    When the model accuracy is 80%, 82%, and 84%, the proposed method reduces the privacy budget by 57%, 96%, and 81%, respectively, compared with the conventional method.

    “when the model accuracy is 80%, 82%, and 84%, our method reduces the privacy budget by 57%, 96%, and 81%, respectively, compared with the conventional method.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    The curves of the two methods almost overlap at the beginning, but the increase of the privacy budget of the proposed method gradually decreases in the later stage of training, while the privacy budget of the conventional method still increases linearly.

    “the curves of the two methods almost overlap at the beginning, but the increase of the privacy budget of our method gradually decreases in the later stage of training, while the privacy budget of the conventional method still increases linearly.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Differential privacy technology needs less computation overhead than Shamir secret sharing and threshold Paillier encryption, making it more suitable for MIoT devices with limited resources.

    “In contrast, differential privacy technology needs less computation overhead, which is more suitable for MIoT devices with limited resources.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Shamir secret sharing and threshold Paillier encryption both consume large computation overhead.

    “References [18, 21] used Shamir secret sharing and threshold Paillier encryption to protect local gradients, respectively, but both consume large computation overhead.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [18] Shayan M. Fung C. Yoon C. J. M. Beschastnikh I. Biscotti: a ledger for private and secure peer-to-peer machine learning 2018 https://arxiv.org/abs/1811.09904

      Not checked in the free run.

    • cites [21] Weng J. Weng J. Zhang J. Li M. Zhang Y. Luo W. Deepchain: auditable and privacy-preserving deep learning with blockchain-based incentive IEEE Transactions on Dependable and Secure Computing 2019 10.1109/tdsc.2019.2952332 source

      Not checked in the free run.

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    Adding noise to a dataset can hide private information but at the same time lowers data utility.

    “From the above definitions, we can see that the private information in a dataset can be hidden by adding noise, but at the same time, the noise will lower the data utility.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Blockchain can replace the central server to decentralize the coordination process in federated learning, thereby resisting single points of failure and illegal tampering attacks.

    “blockchain can replace the central server to decentralize the coordination process in FL, thus resisting single points of failure and illegal tampering attacks.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Federated learning allows multiple devices to train machine learning models collaboratively without sharing their raw data.

    “Federated learning (FL) [1], as a distributed machine learning framework, can allow multiple devices to train machine learning models collaboratively without sharing their raw data”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1] McMahan B. Moore E. Ramage D. Hampson S. Arcas B. A. y. Communication-efficient learning of deep networks from decentralized data Proceedings of the 20 th International Conference on Artificial Intelligence and Statistics (AISTATS) 2017 April 2017 Fort Lauderdale, FL, USA 1273 1282

      Not checked in the free run.

  • May combine two claims.

    Interpretationnone given

    The value of the clipping threshold is important to the federated learning model: too large a value adds excessive noise and too small a value over clips the gradient, both causing serious accuracy loss.

    “The value of the threshold is important to the FL model: too large a value will add excessive noise and too small a value will over clip the gradient, both of which will cause serious accuracy loss.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The proposed consensus protocol based on gradient verification encourages reliable MIoT devices and edge nodes to contribute their data and computing power to federated learning.

    “We design an efficient consensus protocol based on gradient verification, which encourages reliable MIoT devices and edge nodes to contribute their data and computing power to federated learning.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Edge devices have no incentive to contribute dataset and computing power to federated learning.

  • May combine two claims.

    Interpretationreasoning

    Attackers cannot launch targeted attacks on specific verifiers because the output of consistent hashing cannot be predicted by attackers in advance.

    “The output of consistent hashing cannot be predicted by attackers in advance, so attackers cannot launch targeted attacks on specific verifiers.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The output of consistent hashing cannot be predicted by attackers in advance.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Study design and conductnone given

    In each training round, the reputation value of the MIoT device whose local gradient is identified as qualified and the verifier who returns the correct verification result both increase by 1; otherwise their reputation value decreases by 1.

    “in each round of training, the reputation value of the MIoT device whose local gradient is identified as qualified and the verifier who returns the correct verification result will both increase by 1; otherwise, their reputation value will decrease by 1.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The consensus protocol in the blockchain involves time-consuming operations such as block generation, verification, and broadcast, so blockchain-based FL methods achieve security attributes such as auditability, reliability, and resistance to a single point of failure at the cost of some computation overhead.

    “This is because the consensus protocol in the blockchain involves time-consuming operations such as block generation, verification, and broadcast. Therefore, blockchain-based FL methods (i.e., BlockFL and our method) achieve a series of security attributes such as auditability, reliability, and resistance to a single point of failure at the cost of some computation overhead”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Study design and conductnone given

    When the reputation value decreases to zero, the entity is put into the blacklist and prohibited from participating in the consensus.

    “When the reputation value decreases to zero, the entity (e.g., MIoT device or edge node) is put into the blacklist and prohibited from participating in the consensus.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductcitation

    A block is determined valid and broadcasted to arrive at consensus only when more than 2/3 of the verifiers agree on it; otherwise an empty block is created.

    “Only when more than 2/3 of the verifiers agree on the block, the block is determined to be valid and broadcasted to arrive at a consensus in the blockchain through the popular gossip protocol [29]. Otherwise, an empty block is created.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [29] Buchman E. Tendermint: byzantine fault tolerance in the age of blockchains 2016 Guelph, Canada University of Guelph Ph.D. thesis

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    The prior knowledge E[g2]0=0 in the first training round results in C1=0, which cannot be used for gradient clipping.

    “Note that the prior knowledge 𝔼g˜20=0 in the first training round will result in C1=β𝔼g˜20=0, which cannot be used for gradient clipping.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The higher model accuracy of original FL and BlockFL than the proposed method is because the proposed method adds DP noise on the gradient while the other two methods preserve the raw gradient.

    “this is because our method adds DP noise on the gradient while the other two methods preserve the raw gradient.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Because the adversary cannot obtain the state of the block until it is generated, they cannot predict the output of the consistent hashing and launch targeted attacks.

    “Since the adversary cannot obtain the state of the block until it is generated, they cannot predict the output of the consistent hashing and launch targeted attacks.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The adversary cannot obtain the state of the block until it is generated.

  • Classified as probably not a substantial claim.

    Premisecitation

    The privacy accountant proposed by Abadi et al. is used to compute the privacy loss.

    “In this paper, we use the privacy accountant proposed by Abadi et al. [24] to compute the privacy loss”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [24] Abadi M. Chu A. Goodfellow I. Deep learning with differential privacy Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security October 2016 Vienna, Austria 308 318 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    As the value of E[g2]t−1 decreases as the model converges, the local clipping threshold Ct will also decrease, making the differential privacy noise less and contributing to model convergence in the later training stage.

    “Since the value of 𝔼g˜2t−1 in equation (6) decreases as the model converges, the local clipping threshold Ct will also decrease, making the DP noise ξ ~ N(0, (Ctσ)2I) in equation (7) less, which contributes to the convergence of the model in the later training stage.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    In the execution of DP-based deep learning methods, the value range of integer λ is usually 0 ≤ λ ≤ 100.

    “Note that, in the execution of DP-based deep learning methods [24], the value range of integer λ is usually 0 ≤ λ ≤ 100.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [24] Abadi M. Chu A. Goodfellow I. Deep learning with differential privacy Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security October 2016 Vienna, Austria 308 318 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Each edge node identifies and removes malicious gradients uploaded by its associated MIoT devices so as to only aggregate qualified gradients to generate the global model and achieve reliable federated learning.

    “Each edge node identifies and removes malicious gradients uploaded by its associated MIoT devices so as to only aggregate qualified gradients to generate the global model and achieve reliable FL.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Clinical data of some edge devices may be poisoned to mislead the federated learning process.
    • Rests on an assumption: Clinical data collected by MIoT devices may be of low quality and even poisoned by adversaries, and local gradients trained on this kind of data will deviate from the global convergence trend.

  • May combine two claims.

    Interpretationown data

    The proposed method uses the same fixed clipping threshold as the conventional method due to insufficient prior knowledge at the beginning but adopts the adaptive DP algorithm in the latter to reduce the consumption of the privacy budget.

    “This proves that our method uses the same fixed clipping threshold as the conventional method due to insufficient prior knowledge at the beginning but adopts the adaptive DP algorithm in the latter to reduce the consumption of the privacy budget.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The proposed method uses a more efficient consistent hashing protocol and gradient verification method than BlockFL's PoS consensus.

    “while our method uses a more efficient consistent hashing protocol and gradient verification method.”
    Quote in the paperSubstantial claim
31 of 31

  • The paper relies on this without support.

    Model parameters stored in the blockchain may still be stolen by attackers to infer the original private clinical data.

    “(1) the model parameters stored in the blockchain may still be stolen by attackers to infer the original private clinical data”

  • The paper relies on this without support.

    Clinical data of some edge devices may be poisoned to mislead the federated learning process.

    “(2) clinical data of some edge devices may be poisoned to mislead the FL process”

  • The paper relies on this without support.

    Edge devices have no incentive to contribute dataset and computing power to federated learning.

    “(3) edge devices have no incentive to contribute dataset and computing power to FL.”

  • The authors state this as an assumption.

    At any point, no more than one third of the miners are malicious.

    “we assume that, at any point, no more than 1/3 of the miners are malicious.”

  • The paper relies on this without support.

    Clinical data collected by MIoT devices may be of low quality and even poisoned by adversaries, and local gradients trained on this kind of data will deviate from the global convergence trend.

    “the clinical data they collect may be of low quality and even be poisoned by adversaries, and then the local gradient trained on this kind of data will deviate from the global convergence trend.”

  • The paper relies on this without support.

    The adversary cannot obtain the state of the block until it is generated.

    “Since the adversary cannot obtain the state of the block until it is generated”

  • The authors state this as an assumption.

    In each training round, the transaction pool contains R gradients, of which f are Byzantine.

    “Assume that R is the total number of gradients in the transaction pool and f is the number of Byzantine gradients.”

  • The paper relies on this without support.

    The output of consistent hashing cannot be predicted by attackers in advance.

    “The output of consistent hashing cannot be predicted by attackers in advance”
8 of 8

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Chang, Fang, and Sun proposed an FL-chain system with adaptive differential privacy that adjusts noise on the gradient to balance privacy and accuracy in a MIoT environment, with a gradient verification-based consensus protocol preventing malicious attacks and single point of failure.

    “Chang, Fang, and Sun [42] proposed an FL-chain system with an adaptive differential privacy algorithm. Technically, the system adjusts the amount of noise by adding on the gradient to achieve the balance between privacy and accuracy in a MIoT environment. The key strength of this FL-chain system is the gradient verification-based consensus protocol which is implemented to prevent malicious attacks”

    From A Review of Privacy Enhancement Methods for Federated Learning in Healthcare Systems (2023) · cites this paper as [42]

  • All four FL-chain systems [42,43,44,45] from this review adopted additional privacy enhancement methods to improve privacy protection.

    “All of the four FL-chain systems [42,43,44,45] from this review adopted additional privacy enhancement methods to improve privacy protection.”

    From A Review of Privacy Enhancement Methods for Federated Learning in Healthcare Systems (2023) · cites this paper as [42]

  • The consensus process in the FL-chain systems [42,43,44,45] requires significant computational resources.

    “Besides the integration of the additional privacy enhancement methods, the consensus process in the FL-chain systems [42,43,44,45] requires significant computational resources.”

    From A Review of Privacy Enhancement Methods for Federated Learning in Healthcare Systems (2023) · cites this paper as [42]

  • Four FL-chain healthcare applications in this review demonstrate that blockchain technology is relevant to FL.

    “Four FL-chain healthcare applications [42,43,44,45] in this review prove that the blockchain technology is relevant to FL.”

    From A Review of Privacy Enhancement Methods for Federated Learning in Healthcare Systems (2023) · cites this paper as [42]

  • The existing work in this review presented approaches including adding gradient [42], extra elements [43], hashed EHR address [44], and additional rules [45].

    “The existing work in this review presented the approaches including adding gradient [42], extra elements [43], hashed EHR address [44], and additional rules [45].”

    From A Review of Privacy Enhancement Methods for Federated Learning in Healthcare Systems (2023) · cites this paper as [42]

Cites

Flagged first, then by citations

Research works printed but not linked: 4

  1. McMahanB. MooreE. RamageD. HampsonS. andArcasB. A. y. Communication-efficient learning of deep networks from decentralized data Proceedings of the 20 th International Conference on Artificial Intelligence and Statistics (AISTATS) 2017 April 2017 Fort Lauderdale FL USA 1273–1282.
  2. TramèrF. ZhangF. JuelsA. ReiterM. K. andRistenpartT. Stealing machine learning models via prediction apis Proceedings of the 25th {USENIX} Security Symposium ({USENIX} Security 16) August 2016 Austin TX USA 601–618.
  3. BlanchardP. El MhamdiE. M. GuerraouiR. andStainerJ. Machine learning with adversaries: byzantine tolerant gradient descent Proceedings of the 31st International Conference on Neural Information Processing Systems December 2017 Long Beach CA USA 118–128.
  4. BuchmanE. Tendermint: byzantine fault tolerance in the age of blockchains 2016 University of Guelph Guelph Canada Ph.D. thesis.

Cited by

Sorted by citations

Related

Authors
Chang, Y., Fang, C., & Sun, W.
Affiliations
Zhengzhou University (CN)
Corresponding author
Chang, Yuxia
ORCID
Venue
Computational Intelligence and Neuroscience
Publisher
Wiley
Published
2021-01
Volume
2021
Issue
1
Pages
4376418-4376418
ISSN
  • 1687-5265 (ISSN-L)
  • 1687-5273
DOI
10.1155/2021/4376418
PMC
PMC8635913
OpenAlex
W3216394207
Cited by
119
References
28
Citations by year
2026: 16 · 2025: 37 · 2024: 32 · 2023: 24 · 2022: 10
FWCI
5.83
Type
journal article
Language
en
Open access
Licence
Locations
  • Computational Intelligence and Neuroscience (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Science and Technology Project of Henan Province (162102310304)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
  • Physical Sciences › Computer Science › Information Systems › Blockchain Technology Applications and Security
  • Physical Sciences › Computer Science › Information Systems › Internet of Things and AI
Keywords
blockchain, federated learning, smart healthcare, Internet of Medical Things, data privacy, differential privacy, poisoning attacks, consensus algorithms, edge nodes, diabetes prediction
MeSH terms
Blockchain, Artificial Intelligence, Delivery of Health Care, Computer Security, Privacy

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review