Federated learning meets blockchain in edge computing: Opportunities and challenges

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

There is no existing work that surveys the integration of FL and blockchain in edge computing.

“Although FL and blockchain in edge computing have been extensively studied in the literature, there is no any existing work to survey the integration of FL and blockchain in edge computing, to the best of our knowledge.”

May combine two claims.

Integration of FL and blockchain creates FLchain, which potentially transforms intelligent edge networks into decentralized and secure systems.

“Particularly, the integration of FL and blockchain creates a new paradigm called FLchain , which potentially transforms intelligent edge networks with decentralized and secure natures [7], [8].”

• cites [7] [7] U. Majeed and C. S. Hong, 'FLchain: Federated Learning via MECenabled Blockchain Network,' in 2019 20th Asia-Pacific Network Operations and Management Symposium (APNOMS) , Sep. 2019, pp. 1-4. -Not checked yet

  Not checked in the free run.

• cites [8] [8] Y. J. Kim and C. S. Hong, 'Blockchain-based Node-aware Dynamic Weighting Methods for Improving Federated Learning Performance,' in 2019 20th Asia-Pacific Network Operations and Management Symposium (APNOMS) , Sep. 2019, pp. 1-4. -Not checked yet

  Not checked in the free run.

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

Blockchain provides high security for FL training through immutable block ledgers.

“Moreover, blockchain provides high security for FL training through immutable block ledgers.”

May combine two claims.

Blockchain in FLchain eliminates the need for a central server, potentially alleviating communication costs and improving scalability.

“In this way, blockchain helps eliminate the need for a central server which thus potentially alleviates communication costs and achieves better intelligent network scalability.”

May combine two claims.

FL enables collaborative training of a shared AI model across multiple mobile devices without exposing raw data, enhancing privacy.

“FL allows devices and the BS to train a common model while raw datasets are kept locally at users where the data reside.”

• cites [17] [17] J. Koneˇ cn´ y, H. B. McMahan, F. X. Yu, P. Richt´ arik, A. T. Suresh, and D. Bacon, 'Federated Learning: Strategies for Improving Communication Efficiency,' arXiv:1610.05492 , Oct. 2017. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Using blockchain, FL can be implemented via decentralized data ledgers without a central server, mitigating single-point failures and enabling transparent tracing of updates.

“By using blockchain, FL can be implemented via decentralized data ledgers without requiring any central server which mitigates risks of single-point failures, and any update events and user behaviours are traced by all network entities in a transparent manner.”

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

FL has transformed many applications by providing distributed AI solutions at the network edge with cooperation of multiple mobile devices.

“Since the invention in 2016 [17], FL has transformed many applications by providing distributed AI solutions at the network edge with the cooperation of multiple mobile devices.”

• cites [17] [17] J. Koneˇ cn´ y, H. B. McMahan, F. X. Yu, P. Richt´ arik, A. T. Suresh, and D. Bacon, 'Federated Learning: Strategies for Improving Communication Efficiency,' arXiv:1610.05492 , Oct. 2017. - sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

In FLchain, each device acts as a client with equal rights to update and aggregate the learning model in a decentralized manner.

“In FLchain, each device can act as a client with equal rights to update and aggregate the learning model in a decentralized manner.”

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

Learning updates appended to immutable blocks ensure high security for training against external attacks.

“The learning updates are appended to immutable blocks for information exchange among clients during the training which ensures high security for training against external attacks.”

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

Elimination of central server in blockchain avoids single-point failures, improving reliability and stability.

“Moreover, the elimination of central server avoids the risk of single-point failures, thus improving the reliability and stability of blockchain systems.”

May combine two claims.

Decentralization of model aggregation mitigates single-point failures and reduces burden on central server, especially with numerous clients.

“The decentralization of model aggregation not only mitigates the risk of single-point failures for better training reliability but also reduces the burden posed on the central server in global model aggregation, especially when edge networks have numerous clients.”

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

Another integrated FL-blockchain architecture in [28] allows FL clients (vehicles) to collaboratively train AI models and exchange aggregated updates through blockchain miners.

“Another integrated FL-blockchain architecture is introduced in [28] which allows FL clients (i.e., vehicles) to collaboratively train AI models and exchange aggregated updates through blockchain miners.”

• cites [28] [28] S. R. Pokhrel and J. Choi, 'Federated Learning With Blockchain for Autonomous Vehicles: Analysis and Design Challenges,' IEEE Transactions on Communications , vol. 68, no. 8, pp. 4734-4746, Aug. 2020. -Not checked yet

  Not checked in the free run.

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

FL is integrated with blockchain for an MEC-based IoT network [29] to coordinate IoT users to train neural networks and share training parameters with MEC servers at BSs for solving data relaying optimization.

“FL is also integrated with blockchain for an MEC-based IoT network [29], aiming to coordinate IoT users to train neural networks and share training parameters with MEC servers at BSs for cooperatively solving a data relaying optimization problem.”

• cites [29] [29] Y. Lu, X. Huang, K. Zhang, S. Maharjan, and Y. Zhang, 'Low-latency Federated Learning and Blockchain for Edge Association in Digital Twin empowered 6G Networks,' IEEE Transactions on Industrial Informatics , 2020. -Not checked yet

  Not checked in the free run.

May combine two claims.

Information on learning updates is secure in immutable blocks, which improves the security of federated learning training in edge networks.

“Particularly, the information on learning updates is secure in a form of immutable blocks which improves the security of FL training in edge networks [29].”

• cites [29] [29] Y. Lu, X. Huang, K. Zhang, S. Maharjan, and Y. Zhang, 'Low-latency Federated Learning and Blockchain for Edge Association in Digital Twin empowered 6G Networks,' IEEE Transactions on Industrial Informatics , 2020. -Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Achieving synchronization between training data organization and blockchain storage during FLchain training is an important issue.

“Given that fact, how to achieve synchronization among training data organization and blockchain storage during the FLchain training is an important issue.”

• AssumptionRests on an assumption: Blockchain transaction storage requires exact content match, where the hash value of a transaction must be unique to its content.

May combine two claims.

Repeated information exchange among miners requires large bandwidth and results in high communication latencies, potentially degrading FLchain training performance.

“Moreover, the repeated information exchange among miners in the mining group also requires large bandwidth resources and results in high communication latencies which potentially degrades the overall performance of FLchain training, e.g., delayed model aggregation.”

• AssumptionRests on an assumption: Appending a new transaction to blockchain requires running a mining puzzle (e.g., PoW) that is complicated and requires vast computing and storage resources.

May combine two claims.

In FLchain training, training examples may not be fully deterministic; for example, clients may have different image datasets with varying pixel resolution that can be updated continuously through sensing environments.

“In the FLchain training, the training examples may not be fully deterministic. For example, in federated online image classification tasks, clients may have different image datasets with varying pixel resolution which can be also updated continuously through sensing environments.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

FL is of paramount importance in realizing full intelligence in next-generation mobile edge networks because a BS cannot collect all data from distributed mobile devices.

“As a BS cannot collect all data from distributed mobile devices for AI/ML training, FL is of paramount importance in realizing full intelligence in the next generation of mobile edge networks.”

Probably not a substantial claim: may not pass S2 (checkable).

With its decentralized and secure nature, FLchain is promising for enabling intelligent and scalable edge networks with security guarantees.

“With its decentralized and secure nature, FLchain is promising to provide attractive solutions for enabling intelligent and scalable edge networks with security guarantees.”

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

The key idea behind blockchain is decentralization: data on blockchain is not controlled by any single entity.

“The key idea behind the blockchain concept is its decentralization, that is, data on blockchain is not controlled by any single entity [22].”

• cites [22] [22] D. C. Nguyen, P. N. Pathirana, M. Ding, and A. Seneviratne, 'Blockchain for 5G and Beyond Networks: A State of The Art Survey,' Journal of Network and Computer Applications , vol. 166, p. 102693, Sep. 2020. -Not checked yet

  Not checked in the free run.

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

Classical federated learning relies on a centralized server, such as an MEC server, for model aggregation, and this centralized architecture suffers from malfunction and scalability issues in ubiquitous wireless networks.

“From the learning procedure, we can see that the classical FL relies on a centralized server, i.e., an MEC server, for model aggregation, which suffers from malfunction and scalability issues in ubiquitous wireless networks.”

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

FL is particularly attractive for intelligent edge services such as smart transportation and smart healthcare.

“Following the recent advances in mobile hardware and increasing concerns of user privacy, FL is particularly attractive for various intelligent edge services such as smart transportation and smart healthcare [19].”

• cites [19] [19] Z. Du, C. Wu, T. Yoshinaga, K.-L. A. Yau, Y. Ji, and J. Li, 'Federated Learning for Vehicular Internet of Things: Recent Advances and Open Issues,' IEEE Open Journal of the Computer Society , vol. 1, pp. 45-61, 2020. -Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

FL enables distributed learning by allocating AI functions to local devices to build a shared global model at an aggregator.

“Instead of offloading all raw data to a cloud or a data center to perform AI training, FL enables distributed learning by allocating AI functions, e.g., AI model training, directly to the local devices to build a shared global model at an aggregator, e.g., an MEC server at a base station (BS) or an access point (AP) [18].”

• cites [18] [18] T. Li, A. K. Sahu, A. Talwalkar, and V. Smith, 'Federated Learning: Challenges, Methods, and Future Directions,' IEEE Signal Processing Magazine , vol. 37, no. 3, pp. 50-60, May 2020. -Not checked yet

  Not checked in the free run.

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

The server constructs a new global model by aggregating all updates from clients in a fashion that minimizes a loss function.

“By aggregating all updates from clients, the server constructs a new global model in a fashion that minimizes a loss function [31].”

• cites [31] [31] F. Sattler, S. Wiedemann, K.-R. M¨ uller, and W. Samek, 'Robust and Communication-Efficient Federated Learning From Non-i.i.d. Data,' IEEE Transactions on Neural Networks and Learning Systems , pp. 1-14, 2019. -Not checked yet

  Not checked in the free run.

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

The MEC server can select different subsets of devices over different update rounds using client scheduling to improve training quality.

“In practice, the MEC server can select different subsets of devices over different update rounds by using a suitable client scheduling mechanism to improve its training quality [28].”

• cites [28] [28] S. R. Pokhrel and J. Choi, 'Federated Learning With Blockchain for Autonomous Vehicles: Analysis and Design Challenges,' IEEE Transactions on Communications , vol. 68, no. 8, pp. 4734-4746, Aug. 2020. -Not checked yet

  Not checked in the free run.

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

An integrated FL-blockchain scheme proposed in [27] performs both local training and global model updating at mobile devices via blockchain in MEC network.

“For example, an integrated FL-blockchain scheme is proposed in [27], where both local training and global model updating are performed at mobile devices via the blockchain in the MEC network.”

• cites [27] [27] H. Kim, J. Park, M. Bennis, and S.-L. Kim, 'Blockchained On-Device Federated Learning,' IEEE Communications Letters , vol. 24, no. 6, pp. 1279-1283, Jun. 2020. - sourceNot checked yet

  Not checked in the free run.

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

FLchain provides unique security features for federated learning training through immutable blockchain, enabling reliable intelligent decentralized edge networks.

“Particularly, FLchain provides unique security features for the FL training offered by immutable blockchain, which thus enables reliable intelligent decentralized edge networks.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

The proposed FLchain architecture is more generic and covers the most important features of current FL-blockchain system designs.

“Inspired by the existing integrated FL-blockchain schemes, this paper introduces a more generic FL-blockchain architecture or FLchain, which covers the most important features of current FL-blockchain system designs.”

May combine two claims.

Most current FLchain schemes analyze communication costs by considering on-device training latency, latency of communication with miners, and mining latency, while the work in [29] focuses on edge computing latency and parameter transfer communication latency.

“Most of current FLchain schemes analyze communication costs by considering on-device training latency, latency of communication with miners, and mining latency [27], [28], while the work in [29] focuses on edge computing latency and parameter transfer communication latency.”

• cites [27] [27] H. Kim, J. Park, M. Bennis, and S.-L. Kim, 'Blockchained On-Device Federated Learning,' IEEE Communications Letters , vol. 24, no. 6, pp. 1279-1283, Jun. 2020. - sourceNot checked yet

  Not checked in the free run.

• cites [28] [28] S. R. Pokhrel and J. Choi, 'Federated Learning With Blockchain for Autonomous Vehicles: Analysis and Design Challenges,' IEEE Transactions on Communications , vol. 68, no. 8, pp. 4734-4746, Aug. 2020. - sourceNot checked yet

  Not checked in the free run.

• cites [29] [29] Y. Lu, X. Huang, K. Zhang, S. Maharjan, and Y. Zhang, 'Low-latency Federated Learning and Blockchain for Edge Association in Digital Twin empowered 6G Networks,' IEEE Transactions on Industrial Informatics , 2020. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Most current FLchain solutions rely on the PoW consensus mechanism, which often requires large bandwidth and energy resources to implement the mining process.

“Most of current solutions [27], [28] rely on the PoW consensus mechanism, which often requires large bandwidth and energy resources to implement the mining process.”

• cites [27] [27] H. Kim, J. Park, M. Bennis, and S.-L. Kim, 'Blockchained On-Device Federated Learning,' IEEE Communications Letters , vol. 24, no. 6, pp. 1279-1283, Jun. 2020. - sourceNot checked yet

  Not checked in the free run.

• cites [28] [28] S. R. Pokhrel and J. Choi, 'Federated Learning With Blockchain for Autonomous Vehicles: Analysis and Design Challenges,' IEEE Transactions on Communications , vol. 68, no. 8, pp. 4734-4746, Aug. 2020. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Elimination of central server for FLchain model aggregation potentially alleviates communication costs and attracts more mobile users, enhancing scalability.

“Moreover, the elimination of a central server for FLchain model aggregation potentially alleviates communication costs and attract more mobile users to participate in data training based on its decentralized network topology, which in return enhances the scalability of mobile edge networks.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

FLchain still has drawbacks such as possible latency from blockchain mining, resource management, and security from curious miners, which need to be considered in system designs.

“In addition to benefits, FLchain still remains some drawbacks such as possible latency from blockchain mining, resource management, and security from curious miners, which need to be considered in system designs.”

May combine two claims.

MEC servers are often selected to perform mining to maintain the blockchain network, while mobile devices can participate in local training or mining, or both.

“Due to the high computational capability, MEC servers are often selected to perform mining to maintain the blockchain network, while mobile devices can participate in local training or mining, or implement both training and mining tasks [27].”

• cites [27] [27] H. Kim, J. Park, M. Bennis, and S.-L. Kim, 'Blockchained On-Device Federated Learning,' IEEE Communications Letters , vol. 24, no. 6, pp. 1279-1283, Jun. 2020. - sourceNot checked yet

  Not checked in the free run.

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

In classical FL, the global model is computed and updated by the MEC server, whereas in FLchain, global model computation is performed directly at devices in a decentralized manner via blockchain.

“It is noticed that in the classical FL, the global model w G is computed and updated by the MEC server. However, in FLchain, the global model computation is performed directly at devices in a decentralized manner via blockchain.”

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

A centralized federated learning architecture cannot attract devices located far from the server for training, which limits the learning performance of the overall system.

“Moreover, such a centralized FL architecture cannot attract devices located far from the server for training which thus limits the learning performance of the overall system.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Resource allocation is an important task to ensure optimal resource usage for FLchain data training.

“Resource allocation is an important task to ensure optimal resource usage for FLchain data training.”

Classified as probably not a substantial claim.

Blockchain's decentralized feature makes it resistant to data modifications or attacks.

“This decentralized feature makes blockchain resistant to data modifications or attacks.”

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

By using blockchain, FLchain can attract more devices in the learning process for better scalability due to its decentralized network topology.

“By using blockchain, FLchain can attract more devices in the learning process for better scalability thanks to its decentralized network topology.”

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

A similar integrated FL-blockchain architecture in [30] collaborates IoT devices and MEC servers at BSs, with a shared global model learned using distributed local datasets under blockchain control providing model verification.

“A similar integrated FL-blockchain architecture is considered in [30] to collaborate IoT devices and MEC servers at BSs. In particular, a shared global model is learned using distributed local datasets under the control of a blockchain which also provides model verification in a decentralized manner.”

• cites [30] [30] Y. Lu, X. Huang, K. Zhang, and Maharjan, 'Blockchain Empowered Asynchronous Federated Learning for Secure Data Sharing in Internet of Vehicles,' IEEE Transactions on Vehicular Technology , vol. 69, no. 4, pp. 4298-4311, Apr. 2020. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Replication of blocks to entire network allows all clients to verify and trace training progress, ensuring high trust and transparency.

“The replication of blocks to the entire network also allows all clients to verify and trace the training progress so as to ensure high trust and transparency for the FLchain system.”

May combine two claims.

Model updates are verified and audited by blockchain, while gradient collecting and parameter update are monitored by all participants over the ledger network, which ensures fairness in the federated learning process.

“Also, model updates are verified and audited by blockchain, while gradient collecting and parameter update are monitored by all participants over the ledger network, which thus ensures fairness in the FL process.”

May combine two claims.

Blockchain has provided unique features such as decentralization and trustworthiness for enabling secure incentive solutions in federated learning training where model updates are verified and audited by blockchain, while gradient collecting and parameter update are monitored by all participants over the ledger network, which ensures fairness in the federated learning process.

“Blockchain has provided unique features (e.g., decentralization, trustworthiness) for enabling secure incentive solutions in FL training where model updates are verified and audited by blockchain, while gradient collecting and parameter update are monitored by all participants over the ledger network, which thus ensures fairness in the FL process [50], [51].”

• cites [50] [50] J. Weng, J. Weng, J. Zhang, M. Li, Y. Zhang, and W. Luo, 'DeepChain: Auditable and Privacy-Preserving Deep Learning with Blockchain-based Incentive,' IEEE Transactions on Dependable and Secure Computing , 2019. -Not checked yet

  Not checked in the free run.

• cites [51] [51] I. Martinez, S. Francis, and A. S. Hafid, 'Record and Reward Federated Learning Contributions with Blockchain,' in Proceedings of the International Conference on Cyber-Enabled Distributed Computing and Knowledge Discovery (CyberC) , Oct. 2019, pp. 50-57. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

To improve the performance of global model training, incentivization is highly necessary to recruit more mobile users with high-quality data sources.

“To improve the performance of global model training, incentivization is highly necessary to recruit more mobile users with high-quality data sources.”

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

In the FLchain scheme of [67], data owners and data requestors can achieve reliable and fast data sharing while sharing records are stored on the blockchain for tracing.

“As a result, data owners and data requestors can achieve reliable and fast data sharing while the sharing records are stored on the blockchain for tracing.”

• cites [67] [67] Y. Lu, X. Huang, Y. Dai, S. Maharjan, and Y. Zhang, 'Blockchain and Federated Learning for Privacy-Preserved Data Sharing in Industrial IoT,' IEEE Transactions on Industrial Informatics , vol. 16, no. 6, pp. 4177-4186, Jun. 2020. -Not checked yet

  Not checked in the free run.

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

In the IoV FLchain framework of [30], an immutable blockchain ledger is deployed on top of the edge vehicular network to verify model parameter updates and store them in blocks in a decentralized manner.

“Particularly, to provide the security and reliability of the vehicular data sharing, an immutable blockchain ledger is deployed on the top of the edge vehicular network to implement the verification of model parameter updates and store them in blocks in a decentralized manner.”

• cites [30] [30] Y. Lu, X. Huang, K. Zhang, and Maharjan, 'Blockchain Empowered Asynchronous Federated Learning for Secure Data Sharing in Internet of Vehicles,' IEEE Transactions on Vehicular Technology , vol. 69, no. 4, pp. 4298-4311, Apr. 2020. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable).

To ensure the robustness and safety of FLchain systems, building security and privacy protection mechanisms is of paramount importance.

“To ensure the robustness and safety of FLchain systems, building security and privacy protection mechanisms is of paramount importance.”

May combine two claims.

FLchain enables highly scalable intelligent edge data crowdsensing solutions and provides high degrees of privacy and security based on FL and blockchain integration.

“FLchain not only enables highly scalable intelligent edge data crowdsensing solutions but also provides high degrees of privacy and security based on FL and blockchain integration.”

May combine two claims.

Without a proper incentive mechanism, clients may not be willing to join data training, which would reduce the scalability of the designed federated learning system.

“Without a proper incentive mechanism, clients may not be willing to join the data training which would reduce the scalability of the designed FL system [45].”

• cites [45] [45] Y. Zhan, P. Li, Z. Qu, D. Zeng, and S. Guo, 'A Learning-Based Incentive Mechanism for Federated Learning,' IEEE Internet of Things Journal , vol. 7, no. 7, pp. 6360-6368, Jul. 2020. - sourceNot checked yet

  Not checked in the free run.

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

A differential privacy approach is applied to locally generated models with artificial noise to reduce the possibility of individual record identification.

“A differential privacy approach is applied to locally generated models with artificial noise to reduce the possibility of individual record identification.”

• cites [57] [57] P. C. M. Arachchige, P. Bertok, I. Khalil, D. Liu, S. Camtepe, and M. Atiquzzaman, 'A Trustworthy Privacy Preserving Framework for Machine Learning in Industrial IoT Systems,' IEEE Transactions on Industrial Informatics , vol. 16, no. 9, pp. 6092-6102, Sep. 2020. -Not checked yet

  Not checked in the free run.

May combine two claims.

Although federated learning promotes collaborative learning while preserving user privacy, it still faces a challenge of incentivizing clients to participate in the federated learning process and contribute their data and computation resources.

“Although FL has promoted collaborative learning while preserving user privacy, it still faces a challenge of incentivizing clients to participate in the FL process and contribute their data as well as computation resources.”

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

An FLchain-based vehicular data sharing framework for the Internet of Vehicles (IoV) is proposed in [30].

“Another FLchain-based vehicular data sharing framework in edge computing is proposed for the Internet of Vehicles (IoV) in [30], as illustrated in Fig. 5 with three key phases: node selection, local training, and global model computation.”

• cites [30] [30] Y. Lu, X. Huang, K. Zhang, and Maharjan, 'Blockchain Empowered Asynchronous Federated Learning for Secure Data Sharing in Internet of Vehicles,' IEEE Transactions on Vehicular Technology , vol. 69, no. 4, pp. 4298-4311, Apr. 2020. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Blockchain has emerged as an attractive tool to facilitate transparent economic mechanism designs to boost federated learning training in FLchain systems.

“Recently, blockchain has emerged as an attractive tool to facilitate transparent economic mechanism designs, so as to boost the FL training in FLchain systems.”

May combine two claims.

Traditionally, centralized AI/ML architectures are often used to realize intelligent mobile crowdsensing [74], but they require direct access to sensing data which raises privacy concerns.

“Traditionally, centralized AI/ML architectures are often used to realize intelligent mobile crowdsensing [74], but they require direct access to sensing data which raises privacy concerns.”

• cites [74] [74] Z. Zhou, H. Liao, B. Gu, K. M. S. Huq, S. Mumtaz, and J. Rodriguez, 'Robust Mobile Crowd Sensing: When Deep Learning Meets Edge Computing,' IEEE Network , vol. 32, no. 4, pp. 54-60, Jul. 2018. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Blockchain would be useful to build trust between users and MEC servers for reliable content caching without the need for extra network management infrastructure.

“Meanwhile, blockchain would be useful to build trust between users and MEC servers for reliable content caching without the need for extra network management infrastructure.”

May combine two claims.

The benefits of learning incentivization are twofold: attracting more clients to join the training and ensuring that clients are honest in local model training and update exchange.

“The benefits of learning incentivization are twofold, including attracting more clients to join the training and ensuring that clients are honest in local model training and update exchange.”

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

FL is incorporated with a hierarchical blockchain to build knowledge sharing in edge-based vehicular systems [68].

“FL is also incorporated with a hierarchical blockchain to build a knowledge sharing in edge-based vehicular systems [68].”

• cites [68] [68] H. Chai, S. Leng, Y. Chen, and K. Zhang, 'A Hierarchical BlockchainEnabled Federated Learning Algorithm for Knowledge Sharing in Internet of Vehicles,' IEEE Transactions on Intelligent Transportation Systems , 2020. -Not checked yet

  Not checked in the free run.

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

The utilization of a single server to handle all sensed data is not efficient especially when data are often distributed across a large-scale network.

“Moreover, the utilization of a single server to handle all sensed data is not efficient especially when data are often distributed across a large-scale network.”

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

For privacy protection, differential privacy techniques can be employed in FLchain so that model computation is considered differentially private if the output is independent of particular data points from input data.

“For privacy protection, differential privacy techniques can be employed in FLchain so that model computation is considered to be differentially private if the output is independent to particular data points from input data [57], [60].”

• cites [57] [57] P. C. M. Arachchige, P. Bertok, I. Khalil, D. Liu, S. Camtepe, and M. Atiquzzaman, 'A Trustworthy Privacy Preserving Framework for Machine Learning in Industrial IoT Systems,' IEEE Transactions on Industrial Informatics , vol. 16, no. 9, pp. 6092-6102, Sep. 2020. -Not checked yet

  Not checked in the free run.

• cites [60] [60] S. Lugan, P. Desbordes, E. Brion, L. X. Ramos Tormo, A. Legay, and B. Macq, 'Secure Architectures Implementing Trusted Coalitions for Blockchained Distributed Learning (TCLearn),' IEEE Access , vol. 7, pp. 181 789-181 799, 2019. -Not checked yet

  Not checked in the free run.

May combine two claims.

Lightweight blockchain platforms based on DPoS consensus have been adopted to support model updating and block mining in FLchain, which has the potential to mitigate energy consumption for FL training.

“Lightweight blockchain platforms have been adopted based on DPoS consensus to support model updating and block mining in FLchain [42], which has the potential to mitigate energy consumption for FL training.”

• cites [42] [42] Y. Lu, X. Huang, K. Zhang, S. Maharjan, and Y. Zhang, 'Communication-Efficient Federated Learning for Digital Twin Edge Networks in Industrial IoT,' IEEE Transactions on Industrial Informatics , 2020. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The costs of verification for gradient updates during the incentive process have not been considered.

“However, the costs of verification for gradient updates during the incentive process have not been considered.”

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

An FLchain scheme is studied in [65] to construct a mobile crowdsensing system based on Unmanned Aerial Vehicles (UAVs).

“For example, an FLchain scheme is studied in [65] to construct a mobile crowdsensing system based on Unmanned Aerial Vehicles (UAVs), as indicated in Fig. 6.”

• cites [65] [65] Y. Wang, Z. Su, N. Zhang, and A. Benslimane, 'Learning in the Air: Secure Federated Learning for UAV-Assisted Crowdsensing,' IEEE Transactions on Network Science and Engineering , 2020. - sourceNot checked yet

  Not checked in the free run.

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

An FLchain framework is studied in [69] to achieve secure IoT data collaboration among IoT users.

“To achieve a secure IoT data collaboration among IoT users, an FLchain framework is also studied in [69] among private data center, a public data center, and a blockchain network.”

• cites [69] [69] B. Yin, H. Yin, Y. Wu, and Z. Jiang, 'FDC: A Secure Federated Deep Learning Mechanism for Data Collaborations in the Internet of Things,' IEEE Internet of Things Journal , vol. 7, no. 7, pp. 6348-6359, Jul. 2020. - sourceNot checked yet

  Not checked in the free run.

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

Several literature works have been presented to design incentive mechanisms for FLchain systems, with the common objective of attracting more clients and datasets in AI training and improving the robustness of overall FLchain learning.

“Several literature works have been presented to design incentive mechanisms for FLchain systems, with the common objective of attracting more clients and datasets in AI training and improving the robustness of the overall FLchain learning [46], [47], [49].”

• cites [46] [46] K. Toyoda and A. N. Zhang, 'Mechanism Design for An Incentive-aware Blockchain-enabled Federated Learning Platform,' in Proceedings of the IEEE International Conference on Big Data (Big Data) , Dec. 2019, pp. 395-403. -Not checked yet

  Not checked in the free run.

• cites [47] [47] M. H. Ur Rehman, K. Salah, E. Damiani, and D. Svetinovic, 'Towards Blockchain-Based Reputation-Aware Federated Learning,' in Proceedings of the IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS) , Jul. 2020, pp. 183-188. -Not checked yet

  Not checked in the free run.

• cites [49] [49] X. Bao, C. Su, Y. Xiong, W. Huang, and Y. Hu, 'FLChain: A Blockchain for Auditable Federated Learning with Trust and Incentive,' in Proceedings of the 5th International Conference on Big Data Computing and Communications (BIGCOM) , Aug. 2019, pp. 151-159. -Not checked yet

  Not checked in the free run.

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

Simulations with human movement datasets confirm high performance in training accuracy and high-security degrees for a wearable sensor network in the FLchain framework of [69].

“Simulations with human movement datasets confirm high performance in training accuracy and high-security degrees for a wearable sensor network.”

• cites [69] [69] B. Yin, H. Yin, Y. Wu, and Z. Jiang, 'FDC: A Secure Federated Deep Learning Mechanism for Data Collaborations in the Internet of Things,' IEEE Internet of Things Journal , vol. 7, no. 7, pp. 6348-6359, Jul. 2020. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

In FLchain, attacks can attempt to train a local model using designed falsified data to replace the global model and modify parameter values during model transmission, aiming to manipulate the training output.

“In FLchain, attacks can make attempt to train a local model using designed falsified data to replace the global model and modify parameter values during model transmission, aiming to manipulate the training output.”

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

A new optimization algorithm proposed in [86] for FL-based IoT networks, called FetchSGD, can train high-quality models for communication efficiency.

“Moreover, a new optimization algorithm is proposed in [86] for FL-based IoT networks, called FetchSGD, that can train high-quality models for communication efficiency.”

• cites [86] [86] D. Rothchild, A. Panda, E. Ullah, N. Ivkin, I. Stoica, V. Braverman, J. Gonzalez, and R. Arora, 'FetchSGD: Communication-Efficient Federated Learning with Sketching,' in International Conference on Machine Learning , Nov. 2020, pp. 8253-8265. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Mining pool centralization may result in challenging selfish behaviors which poses threats to the system throughput.

“However, the mining pool centralization may result in challenging selfish behaviors which poses threats to the system throughput.”

May combine two claims.

The fake parameter update from an adversary is a security bottleneck [81] which can impersonate a certain client and broadcast fake parameters.

“The fake parameter update from an adversary is also a security bottleneck [81] which can impersonate a certain client and broadcast fake parameters.”

• cites [81] [81] J. Zhang, J. Zhang, J. Chen, and S. Yu, 'GAN Enhanced Membership Inference: A Passive Local Attack in Federated Learning,' in Proceedings of the 2020 IEEE International Conference on Communications (ICC) , Dublin, Ireland, Jun. 2020, pp. 1-6. - sourceNot checked yet

  Not checked in the free run.

May combine two claims. The quote only partly states this claim.

Evaluation results confirm a high utility of UAVs and low aggregation error while reducing convergence latency for the differential privacy and incentive mechanism approach in [65].

“Evaluation results confirm a high utility of UAVs and low aggregation error while reducing convergence latency.”

• cites [65] [65] Y. Wang, Z. Su, N. Zhang, and A. Benslimane, 'Learning in the Air: Secure Federated Learning for UAV-Assisted Crowdsensing,' IEEE Transactions on Network Science and Engineering , 2020. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Increased traffic congestion potentially incurs high delays and higher probabilities of data loss during communication rounds, which results in slow and inaccurate FL training convergence rates.

“This potentially incurs high delays and higher probabilities of data loss during the communication rounds, which results in slow and inaccurate FL training convergence rates.”

May combine two claims.

In the UAV crowdsensing FLchain scheme of [65], blockchain is employed to build a decentralized learning network by interconnecting UAVs with data task publishers through an immutable ledger, which ensures that data training and model exchange among UAVs are monitored and traced for attack detection and data modification prevention.

“Blockchain is employed to build a decentralized learning network by interconnecting UAVs with data task publishers through an immutable ledger, which ensures that data training and model exchange among UAVs are monitored and traced for attack detection and data modification prevention.”

• cites [65] [65] Y. Wang, Z. Su, N. Zhang, and A. Benslimane, 'Learning in the Air: Secure Federated Learning for UAV-Assisted Crowdsensing,' IEEE Transactions on Network Science and Engineering , 2020. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

In a 51% attack, miners can modify block data or prevent blocks from appending to the blockchain, which interrupts communication among users (e.g., FL clients) and FL owners (e.g., MEC servers).

“These miners can modify block data or even prevent blocks from appending to the blockchain, which interrupts the communication among users (e.g., FL clients) and FL owners (e.g., MEC servers).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Recent studies have found that blockchain still has its own security issues.

“Although blockchain can provide decentralization and traceability for FL training in edge computing, recent studies have found that blockchain still has its own security issues.”

• AssumptionRests on an assumption: Blockchain can provide decentralization and traceability for FL training in edge computing.

May combine two claims.

Data features embedded in model updates can be leaked which thus can reveal private user information.

“Although FL can preserve privacy for users in AI training, data features embedded in model updates can be leaked which thus can reveal private user information [75].”

• cites [75] [75] V. Kulkarni, M. Kulkarni, and A. Pant, 'Survey of Personalization Techniques for Federated Learning,' arXiv:2003.08673 , Mar. 2020. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionRests on an assumption: FL can preserve privacy for users in AI training.

May combine two claims.

Implementation results prove the feasibility of the proposed FLchain in [77] with high training accuracy, security protection, and low system latency.

“Implementation results prove the feasibility of the proposed FLchain with high training accuracy, security protection, and low system latency.”

• cites [77] [77] Y. Zhao, J. Zhao, L. Jiang, R. Tan, D. Niyato, Z. Li, L. Lyu, and Y. Liu, 'Privacy-Preserving Blockchain-Based Federated Learning for IoT Devices,' IEEE Internet of Things Journal , pp. 1-1, Aug. 2020. -Not checked yet

  Not checked in the free run.

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

The FetchSGD scheme reduces the amount of communication required each round while still conforming to the training quality requirements of the federated setting.

“In this way, the proposed scheme reduces the amount of communication required each round, while still conforming to the training quality requirements of the federated setting.”

• cites [86] [86] D. Rothchild, A. Panda, E. Ullah, N. Ivkin, I. Stoica, V. Braverman, J. Gonzalez, and R. Arora, 'FetchSGD: Communication-Efficient Federated Learning with Sketching,' in International Conference on Machine Learning , Nov. 2020, pp. 8253-8265. -Not checked yet

  Not checked in the free run.

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

An incentive-based crowdsensing/crowdsourcing architecture is designed in [76] based on FL for IoT in edge computing.

“Further, an incentivebased crowdsensing/crowdsourcing architecture is designed in [76] based on FL for IoT in edge computing.”

• cites [76] [76] S. R. Pandey, N. H. Tran, M. Bennis, Y. K. Tun, A. Manzoor, and C. S. Hong, 'A Crowdsourcing Framework for On-Device Federated Learning,' IEEE Transactions on Wireless Communications , vol. 19, no. 5, pp. 3241-3256, May 2020. -Not checked yet

  Not checked in the free run.

May combine two claims.

Forking attacks [79] generate divergent blockchains which results in a change in the block update rule and thus invalidates synchronized block verification in blockchain.

“Another security issue is forking attacks [79] wherein divergent blockchains are generated which results in a change in the block update rule and thus invalidates the synchronized block verification in blockchain.”

• cites [79] [79] S. Wang, C. Wang, and Q. Hu, 'Corking by Forking: Vulnerability Analysis of Blockchain,' in Proceedings of the IEEE Conference on Computer Communications , Paris, France, Apr. 2019, pp. 829-837. -Not checked yet

  Not checked in the free run.

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

Other attacks such as double spending attacks and reentrancy attacks on smart contracts [80] are issues to be solved before integrating blockchain in FL systems.

“Other attacks such as double spending attacks and reentrancy attacks on smart contracts [80] are issues to be solved before integrating blockchain in FL systems.”

• cites [80] [80] S. Zhang and J.-H. Lee, 'Mitigations on Sybil-based Double-spend Attacks in Bitcoin,' IEEE Consumer Electronics Magazine , 2020. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Smart contracts can be integrated in the miner network to perform reliable authentication for model updates from distributed nodes.

“Furthermore, smart contracts can be also integrated in the miner network to perform reliable authentication for model updates from distributed nodes.”

May combine two claims.

Implementing lightweight block validation schemes can mitigate block mining delays and boost the overall FLchain training speed.

“mitigate block mining delays is desired to boost the overall FLchain training speed, which can be done by implementing lightweight block validation schemes [39].”

• cites [39] [39] Y. Liu, K. Wang, Y. Lin, and W. Xu, 'LightChain: A lightweight blockchain system for industrial Internet of things,' IEEE Transactions on Industrial Informatics , vol. 15, no. 6, pp. 3571-3581, Jun. 2019. - sourceNot checked yet

  Not checked in the free run.

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

Another edge crowdsourcing scheme is investigated in [77] based on an integrated FL-blockchain architecture in IoT environments.

“Another edge crowdsourcing scheme is investigated in [77] based on an integrated FL-blockchain architecture in IoT environments.”

• cites [77] [77] Y. Zhao, J. Zhao, L. Jiang, R. Tan, D. Niyato, Z. Li, L. Lyu, and Y. Liu, 'Privacy-Preserving Blockchain-Based Federated Learning for IoT Devices,' IEEE Internet of Things Journal , pp. 1-1, Aug. 2020. -Not checked yet

  Not checked in the free run.

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

A solution proposed in [82] uses a mining pool strategy to enhance the efficiency of the mining process by solving security bottlenecks such as 51% vulnerability with reduced block and transaction propagation delays.

“To overcome these challenges, a solution is proposed in [82] using a mining pool strategy to enhance the efficiency of the mining process, by solving security bottlenecks such as 51% vulnerability with reduced block and transaction propagation delays.”

• cites [82] [82] P. Silva, 'Impact of Geo-Distribution and Mining Pools on Blockchains: A Study of Ethereum - Practical Experience Report and Ongoing PhD Work,' in Proceedings of the 50th Annual IEEE-IFIP International Conference on Dependable Systems and Networks-Supplemental Volume (DSN-S) , Jun. 2020, pp. 73-74. -Not checked yet

  Not checked in the free run.

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

The research in [31] proposes a new efficient communication protocol that is able to compress uplink and downlink communications while remaining high robustness to the increased number of clients, and data distribution.

“The research in [31] proposes a new efficient communication protocol that is able to compress uplink and downlink communications while remaining high robustness to the increased number of clients, and data distribution.”

• cites [31] [31] F. Sattler, S. Wiedemann, K.-R. M¨ uller, and W. Samek, 'Robust and Communication-Efficient Federated Learning From Non-i.i.d. Data,' IEEE Transactions on Neural Networks and Learning Systems , pp. 1-14, 2019. -Not checked yet

  Not checked in the free run.

May combine two claims.

The characteristics of different IoT devices are different, ranging from hardware (CPU, memory), network protocols (e.g., 4G/5G and WiFi) to battery, all of which contribute to the heterogeneity of computation and communication of IoT devices and directly affect the one-device AI model training time and learning quality [85].

“Indeed, the characteristics of different IoT devices are different, ranging from hardware (CPU, memory), network protocols (e.g., 4G/5G and WiFi) to battery, all of which contribute to the heterogeneity of computation and communication of IoT devices and directly affect the one-device AI model training time and learning quality [85].”

• cites [85] [85] Z. Chen, P. Tian, W. Liao, and W. Yu, 'Zero Knowledge Clustering Based Adversarial Mitigation in Heterogeneous Federated Learning,' IEEE Transactions on Network Science and Engineering , 2020. -Not checked yet

  Not checked in the free run.

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

51% attack is a critical security problem in blockchain platforms [78] where a group of powerful miners can take control of the mining with more than 50% of the mining power of the network.

“For example, 51% attack is a critical security problem in blockchain platforms [78] where a group of powerful miners can take control of the mining with more than 50% of the mining power of the network.”

• cites [78] [78] X. Li, P. Jiang, T. Chen, X. Luo, and Q. Wen, 'A survey on the security of blockchain systems,' Future Generation Computer Systems , vol. 107, pp. 841-853, Jun. 2020. -Not checked yet

  Not checked in the free run.

May combine two claims.

Devices with limited computing capabilities and poor connection can take much longer to complete their local training and model update to the server, which can lead to serious delay in global aggregation.

“For example, devices with limited computing capabilities and poor connection can take much longer to complete their local training and model update to the server, which can lead to serious delay in global aggregation.”

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

To mitigate forking probability in blockchain, a large deviation theory is employed in [79] to analyze the vulnerability of blockchain networks incurred by intentional forks and propose strategic planning mechanisms for preventing forking attacks.

“Moreover, to mitigate the forking probability in blockchain, a large deviation theory is employed in [79] to analyze the vulnerability of blockchain networks incurred by intentional forks and propose strategic planning mechanisms for preventing the forking attacks.”

• cites [79] [79] S. Wang, C. Wang, and Q. Hu, 'Corking by Forking: Vulnerability Analysis of Blockchain,' in Proceedings of the IEEE Conference on Computer Communications , Paris, France, Apr. 2019, pp. 829-837. -Not checked yet

  Not checked in the free run.

May combine two claims.

Plagiarism in FLchain causes unfairness among learning clients and significantly reduces overall FL training performance.

“This issue not only causes unfairness among learning clients but also significantly reduces the overall FL training performance.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Giving higher priority in the verification stage to clients with higher reputation due to frequent update training contribution helps mitigate delay and bias in model aggregation and improves overall FLchain training performance.

“For example, the client with higher reputation due to frequent update training contribution can be given a higher priority in the verification stage for fast block approval, which helps mitigate the delay and bias in the model aggregation and in turn improves the overall FLchain training performances.”

May combine two claims.

The use of blockchain introduces additional latency from block mining, which poses new challenges for FL systems because an FL client must wait for mining to complete before receiving model updates and performing its next training round.

“Moreover, the use of blockchain also introduces additional latency from block mining which poses new challenges for FL systems since an FL client needs to wait for the mining process to be completed before receiving model updates and performing its next training round.”

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

DRL-based approaches can schedule sub-channel assignment and transmission power control to improve wireless data transmission rates.

“Another DRL-based approach is also proposed in [98] to schedule the sub-channel assignment and transmission power control for improving the wireless data transmission rates.”

• cites [98] [98] H. Yang, Z. Xiong, J. Zhao, T. D. Niyato, C. Yuen, and R. Deng, 'Deep Reinforcement Learning Based Massive Access Management for UltraReliable Low-Latency Communications,' IEEE Transactions on Wireless Communications , 2021.Not checked yet

  Not checked in the free run.

May combine two claims.

Allocating more resources to devices with worse channel conditions or weaker computation capabilities can balance AI training power and improve the model aggregation rate for low-latency FL training.

“In this way, devices can achieve a balance in AI training power, thus improving the model aggregation rate for low-latency FL training.”

• cites [95] [95] W. Shi, S. Zhou, and Z. Niu, 'Device Scheduling with Fast Convergence for Wireless Federated Learning,' in Proceedings of the IEEE International Conference on Communications (ICC) , Jun. 2020, pp. 1-6. -Not checked yet

  Not checked in the free run.