Federated learning for smart healthcare: A survey

May combine two claims.

A centralized AI solution for smart healthcare is not efficient in communication latency due to raw data transmission and cannot achieve high network scalability.

“Given the increasing volumes of health data and the growth of IoMT devices in modern healthcare networks, this centralized solution is not efficient in terms of communication latency due to raw data transmission and cannot achieve high network scalability.”

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

Traditional centralized AI for smart healthcare requires centralized data collection and processing that may be infeasible in realistic healthcare scenarios due to the high scalability of modern healthcare networks and growing data privacy concerns.

“Traditionally, AI techniques require centralized data collection and processing that may be infeasible in realistic healthcare scenarios due to the high scalability of modern healthcare networks and growing data privacy concerns.”

May combine two claims.

The dataset of a single medical site may be insufficient to run the AI model, which can prevent proper health data training.

“In realistic healthcare systems, the dataset of a single medical site (e.g., a clinical lab) may be insufficient to run the AI model which can prevent a proper health data training [26].”

• cites [26] [26] M. Frid-Adar, I. Diamant, E. Klang, M. Amitai, J. Goldberger, and H. Greenspan, 'Gan-based synthetic medical image augmentation for increased CNN performance in liver lesion classification,' Neurocomputing , vol. 321, pp. 321-331, 2018. -Not checked yet

  Not checked in the free run.

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

Because FL attracts large computation and dataset resources from multiple health data clients, health data training quality, such as accuracy, would be significantly improved, which might not be achieved by centralized AI with less data and limited computational capabilities.

“Moreover, since FL attracts large computation and dataset resources from a number of health data clients to train AI models, the health data training quality, e.g., accuracy, would be significantly improved which might not be achieved by using centralized AI approaches with less data and limited computational capabilities.”

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

The offloading process requires transmit power of medical devices, which poses new challenges on battery and hardware designs on devices.

“The offloading process also requires transmit power of medical devices which in turn poses new challenges on battery and hardware designs on devices.”

May combine two claims.

The use of traditional AI-based approaches for smart healthcare requires open data sharing with cloud or data centres, which makes health information vulnerable to privacy attacks.

“the use of traditional AI-based approaches for realizing smart healthcare require open data sharing with cloud or data centres which makes health information vulnerable to privacy attacks.”

• cites [24] [24] J.-J. Yang, J.-Q. Li, and Y. Niu, 'A hybrid solution for privacy preserving medical data sharing in the cloud environment,' Future Generation Computer Systems , vol. 43, pp. 74-86, 2015. -Not checked yet

  Not checked in the free run.

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

Health data transfer consumes much network bandwidth, which is likely to cause network congestion when the number of devices increases.

“Moreover, the health data transfer also consumes much network bandwidth, which is likely to cause network congestion when the number of devices increases.”

May combine two claims.

Adversaries can gain unauthorized access to AI training centres for data retrieval, or third parties such as cloud providers can gain control over data and modify data patterns without user consent.

“Indeed, adversaries can gain unauthorized access to the AI training centres for data retrieval, or third parties such as cloud providers can gain the control over the data and modify data patterns without the consent of users [24].”

• cites [24] [24] J.-J. Yang, J.-Q. Li, and Y. Niu, 'A hybrid solution for privacy preserving medical data sharing in the cloud environment,' Future Generation Computer Systems , vol. 43, pp. 74-86, 2015. -Not checked yet

  Not checked in the free run.

May combine two claims.

In FL-based smart healthcare, only local updates such as model gradients are required by the central server for AI training, while local health data are kept at local medical sites and devices.

“In the FL-based smart healthcare system, only the local updates such as model gradients are required by the central server for the AI training, while the local health data are kept at local medical sites and devices.”

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

The latest advances in FL such as resource-aware FL, secure and privacy-enhanced FL, incentive-aware FL, and personalized FL have not been fully explored.

“However, the latest advances in FL such as resource-aware FL, secure and privacy-enhanced FL, incentive-aware FL, and personalized FL have not been fully explored.”

May combine two claims.

FL reduces the risks of leakage of sensitive user information to external third-party, thus providing a higher degree of user privacy.

“This would reduce the risks of the leakage of sensitive user information to the external third-party, and thus providing a higher degree of user privacy [30].”

• cites [30] [30] 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 , vol. 8, no. 3, pp. 1817-1829, 2021. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

FL potentially prevents disclosing sensitive user information and user preference, and thus mitigates privacy leakage risks.

“This potentially prevents disclosing sensitive user information and user preference, and thus mitigates privacy leakage risks.”

May combine two claims.

FL saves much network bandwidth and mitigates the possibility of network congestion in massive healthcare networks.

“As a result, FL also save much network bandwidth and mitigate possibility of network congestion in massive healthcare networks.”

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

An attack can attempt to modify or change local updates, leading to aggregation bias at the server.

“Further, an attack can make attempt to modify or change the local updates which leads to aggregation bias at the server.”

May combine two claims.

Without involvement of multiple devices in training, FL becomes inefficient in smart healthcare, where contributed computation from different devices is highly important to improve health data training.

“Without the involvement of multiple devices in the training, the FL concept becomes inefficient in smart healthcare, where the contributed computation from different devices are highly important to improve health data training.”

May combine two claims.

An adversary can deploy data attacks to communication channels established by clients and server to steal updated information, which can interrupt model computation at the global server due to insufficient local updates to build a global model.

“Indeed, an adversary can deploy data attacks to the communication channels established by the clients and the server to steal the updated information which can interrupt the model computation at the global server due to the insufficient local updates to build a global model.”

• cites [35] [35] M. Nasr, R. Shokri, and A. Houmansadr, 'Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning,' in 2019 IEEE symposium on security and privacy (SP) , 2019, pp. 739-753. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

FL training retains model generalizability at the cost of nominal accuracy loss.

“Moreover, FL training retains the model generalizability at the cost of nominal accuracy loss.”

May combine two claims.

Compared with conventional centralized learning, FL offers a reasonable trade-off between accuracy and utility along with privacy enhancement.

“Compared with conventional centralized learning, FL is able to offer a reasonable trade-off between accuracy and utility along with privacy enhancement.”

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

Reputation is introduced in the context of FL-enabled healthcare systems to prevent unreliable updates from untrusted devices.

“To prevent unreliable updates from untrusted devices, a new concept called reputation is introduced in the context of FL-enabled healthcare systems [51].”

• cites [51] [51] J. Kang, Z. Xiong, D. Niyato, Y. Zou, Y. Zhang, and M. Guizani, 'Reliable federated learning for mobile networks,' IEEE Wireless Communications , vol. 27, no. 2, pp. 72-80, 2020. - sourceNot checked yet

  Not checked in the free run.

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

Blockchain can coordinate the calculation of a global model via block consensus among participants in a peer-to-peer manner.

“The blockchain can coordinate the calculation of global model via the block consensus among participants in a peer-to-peer manner.”

May combine two claims.

Integrating blockchain into federated settings decentralizes FL and allows elimination of a single central server in model aggregation.

“Particularly, the integration of blockchain into federated settings decentralizes the FL that allows for the elimination of a single central server in the model aggregation [54].”

• cites [54] [54] J. Passerat-Palmbach, T. Farnan, M. McCoy, J. D. Harris, S. T. Manion, H. L. Flannery, and B. Gleim, 'Blockchain-orchestrated machine learning for privacy preserving federated learning in electronic health data,' in 2020 IEEE International Conference on Blockchain (Blockchain) . IEEE, 2020, pp. 550-555. - sourceNot checked yet

  Not checked in the free run.

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

Reliable device selection is important in scenarios where local FL models are trained using low-quality and noiseless data.

“Reliable device selection is also important in scenarios, where local FL models are trained using low-quality and noiseless data.”

May combine two claims.

There are potential security attacks in FL systems such as poisoning attacks, inference, backdoor attacks, malicious server, communication bottlenecks, and free-riding attacks.

“It is important to study secure solutions for FL-enabled healthcare applications since there are potential security attacks in FL systems such as poisoning attacks, inference, backdoor attacks, malicious server, communication bottlenecks, and free-riding attacks [47-49].”

• cites [47] [47] V. Tolpegin, S. Truex, M. E. Gursoy, and L. Liu, 'Data poisoning attacks against federated learning systems,' in European Symposium on Research in Computer Security . Springer, 2020, pp. 480-501. -Not checked yet

  Not checked in the free run.

• cites [48] [48] M. Fang, X. Cao, J. Jia, and N. Gong, 'Local model poisoning attacks to byzantine-robust federated learning,' in 29th USENIX Security Symposium USENIX Security 20) , 2020, pp. 1605-1622. -Not checked yet

  Not checked in the free run.

• cites [49] [49] H. Wang, K. Sreenivasan, S. Rajput, H. Vishwakarma, S. Agarwal, J.-y. Sohn, K. Lee, and D. Papailiopoulos, 'Attack of the tails: Yes, you really can backdoor federated learning,' in Advances in Neural Information Processing Systems , vol. 33, 2020, pp. 16 070-16 084. -Not checked yet

  Not checked in the free run.

May combine two claims.

A malicious IoMT device may inject poisonous data into its local data, and updating the local model may degrade the accuracy of the global model.

“For example, a malicious IoMT device may inject poisonous data into its local data, and thus updating the fall local model may degrade the accuracy of the global model.”

May combine two claims.

A decentralized FL design over a graph for smart healthcare allows local clients to perform local updates for several iterations and then peer-to-peer communications among health clients, reducing delay of parameter exchange and communications because models need not be transferred to a remote central server.

“In such a way, the delay of parameter exchange and communications among clients can be reduced since there is no need to transfer the models to the remote central server.”

• cites [53] [53] S. Lu, Y. Zhang, and Y. Wang, 'Decentralized federated learning for electronic health records,' in 2020 54th Annual Conference on Information Sciences and Systems (CISS) . IEEE, 2020, pp. 1-5. -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 MAB approach results in significantly low training loss compared with several benchmarks and outperformance in terms of training latency.

“The experimental results show that the MAB approach results in significantly low training loss compared with several benchmarks as well as outperformance in terms of the training latency.”

• cites [39] [39] B. Xu, W. Xia, J. Zhang, T. Q. Quek, and H. Zhu, 'Online client scheduling for fast federated learning,' IEEE Wireless Communications Letters , vol. 10, no. 7, pp. 1434-1438, 2021. -Not checked yet

  Not checked in the free run.

May combine two claims.

More lazy IoMT devices decrease learning performance; for example, no lazy client achieves 85.53% accuracy while 30% lazy clients achieve 78.80% accuracy.

“More lazy IoMT devices decrease the learning performance, e.g., no lazy client can achieve an accuracy of 85.53% while the accuracy is only 78.80% when 30% clients are lazy.”

• cites [55] [55] J. Li, Y. Shao, K. Wei, M. Ding, C. Ma, L. Shi, Z. Han, and H. V. Poor, 'Blockchain assisted decentralized federated learning (BLADE-FL): Performance analysis and resource allocation,' arXiv preprint arXiv:2101.06905 , 2021. -Not checked yet

  Not checked in the free run.

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

A secure aggregation method needs to provide security guarantees under unfavorable environments such as unauthenticated transmission channels and resource-constrained edge nodes.

“(4) It can provide security guarantees under unfavorable environments such as unauthenticated transmission channels and resource-constrained edge nodes.”

• cites [57] [57] K. A. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan, S. Patel, D. Ramage, A. Segal, and K. Seth, 'Practical secure aggregation for federated learning on user-held data,' in NIPS Workshop on Private Multi-Party Machine Learning , 2016. -Not checked yet

  Not checked in the free run.

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

A secure aggregation method needs to be robust to user participant unavailability, also known as user dropping-out.

“(3) It is robust to the user participant and unavailability (also known as user dropping-out),”

• cites [57] [57] K. A. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan, S. Patel, D. Ramage, A. Segal, and K. Seth, 'Practical secure aggregation for federated learning on user-held data,' in NIPS Workshop on Private Multi-Party Machine Learning , 2016. -Not checked yet

  Not checked in the free run.

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

Turbo-Aggregate's advantages are only evaluated via numerical results, so experiments on real health datasets can be examined using the secure aggregation methods proposed so far.

“However, such advantages are only evaluated via numerical results and thus experiments on the real health datasets can be examined by using the secure aggregation methods proposed so far.”

• cites [58] [58] J. So, B. Güler, and A. S. Avestimehr, 'Turbo-Aggregate: Breaking the quadratic aggregation barrier in secure federated learning,' IEEE Journal on Selected Areas in Information Theory , vol. 2, no. 1, pp. 479-489, 2021. -Not checked yet

  Not checked in the free run.

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

A secure aggregation method needs to work with high-dimensional updates from users.

“(1) It can work with high-dimensional updates from users,”

• cites [57] [57] K. A. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan, S. Patel, D. Ramage, A. Segal, and K. Seth, 'Practical secure aggregation for federated learning on user-held data,' in NIPS Workshop on Private Multi-Party Machine Learning , 2016. -Not checked yet

  Not checked in the free run.

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

Secure aggregation approaches can be classified into secure communication protocols, dining cryptographers based secure aggregation, homomorphic threshold encryption, and pairwise masking.

“According to [57], secure aggregation approaches can be classified into secure communication protocols, dining cryptographers (DC) based secure aggregation, homomorphic threshold encryption, and pairwise masking.”

• cites [57] [57] K. A. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan, S. Patel, D. Ramage, A. Segal, and K. Seth, 'Practical secure aggregation for federated learning on user-held data,' in NIPS Workshop on Private Multi-Party Machine Learning , 2016. -Not checked yet

  Not checked in the free run.

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

A secure aggregation method needs to be communication-efficient even with massive users.

“(2) It is communication-efficient even with massive users,”

• cites [57] [57] K. A. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan, S. Patel, D. Ramage, A. Segal, and K. Seth, 'Practical secure aggregation for federated learning on user-held data,' in NIPS Workshop on Private Multi-Party Machine Learning , 2016. -Not checked yet

  Not checked in the free run.

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

A double-masking scheme protects user privacy such that the server cannot know the user's masks.

“A double-masking scheme is also leveraged to protect the user privacy, i.e., the server cannot know the user's masks.”

• cites [57] [57] K. A. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan, S. Patel, D. Ramage, A. Segal, and K. Seth, 'Practical secure aggregation for federated learning on user-held data,' in NIPS Workshop on Private Multi-Party Machine Learning , 2016. -Not checked yet

  Not checked in the free run.

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

A three-dimensional contract-based incentive approach with differential privacy obtains lower training loss and higher training accuracy compared with seminal FL.

“Experimental results showed that the proposed three-dimensional contract-based incentive approach with differential privacy can obtain lower training loss and higher training accuracy compared with the seminal FL.”

• cites [62] [62] M. Wu, D. Ye, J. Ding, Y. Guo, R. Yu, and M. Pan, 'Incentivizing differentially private federated learning: A multi-dimensional contract approach,' IEEE Internet of Things Journal , 2021. -Not checked yet

  Not checked in the free run.

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

A non-cooperative game-based distributed algorithm for cross-silo FL is proved to be Nash-equilibrium and has properties such as social welfare, individual rationality, and no need for a third-party payment entity.

“The proposed game is proved to be not only Nash-equilibrium but also has nice properties such as social welfare, individual rationality, and no need for a third-party payment entity.”

• cites [70] [70] M. Tang and V. W. Wong, 'An incentive mechanism for cross-silo federated learning: A public goods perspective,' in INFOCOM 2021 IEEE Conference on Computer Communications , 2021. -Not checked yet

  Not checked in the free run.

May combine two claims.

The conventional FL global model only captures statistical characteristics of different IoMT devices but is hard to provide distinct personal styles.

“A very first challenge is that the global model only captures the statistical characteristics of different IoMT devices but is hard to provide distinct personal styles.”

May combine two claims.

FL-SIGN-DP sends only 1.76 Mb while standard FL sends up to 56.48 Mb, while maintaining a reasonable level of privacy protection.

“The FL-SIGN-DP is shown to consume a very small amount of bandwidth, i.e., FL-SIGN-DP sends only 1.76 Mb while the standard FL sends up to 56.48 Mb, while FL-SIGN-DP can maintain a reasonable level of privacy protection.”

• cites [66] [66] R. Kerkouche, G. Acs, C. Castelluccia, and P. Genevès, 'Privacy-preserving and bandwidth-efficient federated learning: an application to in-hospital mortality prediction,' in ACM Conference on Health, Inference, and Learning , 2021. -Not checked yet

  Not checked in the free run.

May combine two claims.

Differentially private FL frameworks can avoid transmitting large amounts of data from distributed health organizations to a central entity and protect patient data against potential privacy attacks.

“However, such differently private FL works in [63] and [64] are promising as their FL frameworks can avoid the need for transmitting a large amount of data from distributed health organizations to a central entity and also project the patient data against potential privacy attacks.”

• cites [63] [63] M. Malekzadeh, B. Hasircioglu, N. Mital, K. Katarya, M. E. Ozfatura, and D. Gündüz, 'Dopamine: Differentially private federated learning on medical data,' arXiv preprint arXiv:2102.13314 , 2021. -Not checked yet

  Not checked in the free run.

• cites [64] [64] O. Choudhury, A. Gkoulalas-Divanis, T. Salonidis, I. Sylla, Y. Park, G. Hsu, and A. Das, 'Differential privacy-enabled federated learning for sensitive health data,' arXiv preprint arXiv:1910.02578 , 2020. -Not checked yet

  Not checked in the free run.

May combine two claims.

Based on local model updates from an IoMT device, an attacker can infer and recover local data using construction attacks and/or predict the existence of a data sample in the local training dataset via membership inference attacks.

“For example, based on the local model updates from an IoMT device, the attacker can infer and recover the local data using the construction attacks [60] and/or predict the existence of a data sample in the local training dataset such as blood type, disease name, gender information, and any other private information, which can be done via membership inference attacks [61].”

• cites [60] [60] L. Zhu and S. Han, 'Deep leakage from gradients,' in Federated Learning . Springer, 2020, pp. 17-31. - sourceNot checked yet

  Not checked in the free run.

• cites [61] [61] M. Nasr, R. Shokri, and A. Houmansadr, 'Machine learning with membership privacy using adversarial regularization,' in Proceedings of the 2018 ACM SIGSAC Conference on Computer and Communications Security , 2018, pp. 634-646. - sourceNot checked yet

  Not checked in the free run.

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

A differentially private FL method using DP-SGD and homomorphic encryption secure aggregation achieves performance close to the centralized approach while significantly outperforming FL with parallel differential privacy and FL with centralized differential privacy on the DR dataset for diabetic retinopathy.

“Experimental results show that the performance of the proposed method is close the that of the centralized approach, while significantly outperforming several benchmarks such as FL with parallel differential privacy, where different hospitals apply differential privacy on their datasets individually and FL with centralized differential privacy, where a trusted entity is required to apply different”

• cites [63] [63] M. Malekzadeh, B. Hasircioglu, N. Mital, K. Katarya, M. E. Ozfatura, and D. Gündüz, 'Dopamine: Differentially private federated learning on medical data,' arXiv preprint arXiv:2102.13314 , 2021. -Not checked yet

  Not checked in the free run.

May combine two claims.

FL has privacy concerns such as membership attacks, unintentional data leakage, and generative adversarial network based inference attacks.

“Despite the great potential in improving user data privacy, FL also has its own privacy concerns such as membership attacks, unintentional data leakage, generative adversarial network based inference attacks [50].”

• cites [50] [50] V. Mothukuri, R. M. Parizi, S. Pouriyeh, Y. Huang, A. Dehghantanha, and G. Srivastava, 'A survey on security and privacy of federated learning,' Future Generation Computer Systems , vol. 115, pp. 619-640, 2021. - sourceNot checked yet

  Not checked in the free run.

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

Incentive mechanisms designed for FL can be categorized by data contribution, device reputation, and resource allocation.

“According to a recent survey [67], incentive mechanisms designed for FL can be categorized by different aspects, including the device's data contribution, device reputation, and resource allocation.”

• cites [67] [67] Y. Zhan, J. Zhang, Z. Hong, L. Wu, P. Li, and S. Guo, 'A survey of incentive mechanism design for federated learning,' IEEE Transactions on Emerging Topics in Computing , 2021. -Not checked yet

  Not checked in the free run.

May combine two claims.

An alpha-weighted update solves label heterogeneity by aggregating overlapping label information at the central server, increasing classification accuracy by about 16.7% on Animals-10 and by average increases of 9.153% and 11.01% on human activity recognition datasets.

“The proposed approach is examined on the Animals-10 dataset [75], showing that the classification accuracy is increased by about 16.7%. This approach is also tested on the human activity recognition dataset with an average increase of 9.153% 11.01% in [76] and [77], respectively.”

• cites [74] [74] G. K. Gudur, B. S. Balaji, and S. K. Perepu, 'Resource-constrained federated learning with heterogeneous labels and models,' arXiv preprint arXiv:2011.03206 , 2020. -Not checked yet

  Not checked in the free run.

• cites [76] [76] G. K. Gudur and S. K. Perepu, 'Resource-constrained federated learning with heterogeneous labels and models for human activity recognition,' in Deep Learning for Human Activity Recognition: Second International Workshop, DL-HAR 2020, Held in Conjunction with IJCAI-PRICAI 2020, Kyoto, Japan, January 8, 2021, Proceedings . Springer Nature, 2020, p. 57. -Not checked yet

  Not checked in the free run.

• cites [77] [77] G. Krishna Gudur and S. K. Perepu, 'Federated learning with heterogeneous labels and models for mobile activity monitoring,' arXiv e-prints , pp. arXiv-2012, 2020. -Not checked yet

  Not checked in the free run.

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

Personalized FL faces the challenge of heterogeneous computing resources and network conditions of different FL devices.

“Another challenge is heterogeneous computing resources and network conditions of different FL devices.”

May combine two claims.

A lightweight data perturbation method in FL for EHRs analytic management perturbs training-related data and can defend against model memorization attacks.

“To ensure privacy for model parameters in the FL process, a lightweight data perturbation method is considered to perturb the training-related data, which thus can defend model memorization attack in the learning.”

• cites [65] [65] M. Hao, H. Li, G. Xu, Z. Liu, and Z. Chen, 'Privacy-aware and resource-saving collaborative learning for healthcare in cloud computing,' in ICC 2020-2020 IEEE International Conference on Communications (ICC) , 2020, pp. 1-6. - sourceNot checked yet

  Not checked in the free run.

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

Label heterogeneity is a new challenge in FL design where each FL device has its own definition of data labels and learning model, independent of definitions in other devices and the central server.

“The work in [74] identifies a new challenge in FL design, which is called label heterogeneity. Concretely, each FL device has its own definition of data labels and learning model, which is independent of the definitions in other devices and the central server.”

• cites [74] [74] G. K. Gudur, B. S. Balaji, and S. K. Perepu, 'Resource-constrained federated learning with heterogeneous labels and models,' arXiv preprint arXiv:2011.03206 , 2020. -Not checked yet

  Not checked in the free run.

May combine two claims.

FedHome achieves 95.87% accuracy while personalized FL with MLP and CNN achieve only 92.31% and 91.77%, respectively.

“For example, FedHome can achieve the accuracy of 95.87% while the personalized FL with multi-layer perceptron (MLP) and convolutional neural network (CNN) can only achieve 92.31% and 91.77%, respectively.”

• cites [72] [72] Q. Wu, X. Chen, Z. Zhou, and J. Zhang, 'FedHome: Cloud-edge based personalized federated learning for in-home health monitoring,' IEEE Transactions on Mobile Computing , 2020. -Not checked yet

  Not checked in the free run.

May combine two claims.

SplitNN implements vanilla FL with split learning by offloading the weights of the split layer without requiring raw data sharing.

“By splitting the network into multiple sections that are then transmitted to distributed clients, the network training can be implemented by offloading the weights of the split layer, while raw data sharing is not required.”

• cites [84] [84] P. Vepakomma, O. Gupta, T. Swedish, and R. Raskar, 'Split learning for health: Distributed deep learning without sharing raw patient data,' arXiv preprint arXiv:1812.00564 , 2018. -Not checked yet

  Not checked in the free run.

The quote only partly states this claim.

A personalized FL framework for pain prediction using face images with a lightweight CNN and local last-layer parameters outperforms conventional FL on the UNBC-McMaster Shoulder Pain dataset.

“The proposed model is evaluated by using the UNBC-McMaster Shoulder Pain dataset [79], showing that the proposed personalized FL with privacy preserving outperforms the conventional FL approach, which shares all the learning updates with the aggregation server, which in turn increases the probability of data breach.”

• cites [78] [78] O. Rudovic, N. Tobis, S. Kaltwang, B. Schuller, D. Rueckert, J. F. Cohn, and R. W. Picard, 'Personalized federated deep learning for pain estimation from face images,' arXiv preprint arXiv:2101.04800 , 2021. -Not checked yet

  Not checked in the free run.

• cites [79] [79] P. Lucey, J. F. Cohn, K. M. Prkachin, P. E. Solomon, and I. Matthews, 'Painful data: The UNBC-McMaster shoulder pain expression archive database,' in 2011 IEEE International Conference on Automatic Face and Gesture Recognition (FG) , 2011, pp. 57-64. -Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

FL can be an ideal candidate for assisting COVID-19 detection by its federation and privacy protection features.

“FL can be an ideal candidate for assisting COVID-19 detection, by its federation and privacy protection features.”

May combine two claims.

The FL-based approach for in-home health monitoring achieves a high accuracy of 95.41%, an increase by 7.49% over the standalone CNN scheme with low communication costs, in both balanced and imbalanced data cases.

“Extensive experiments are implemented using a realistic human activity dataset, showing that the FL-based approach can achieve a high accuracy of 95.41%, an increase by 7.49% over the standalone CNN scheme with low communication costs, in both the balanced and imbalanced data cases.”

• cites [72] [72] Q. Wu, X. Chen, Z. Zhou, and J. Zhang, 'FedHome: Cloud-edge based personalized federated learning for in-home health monitoring,' IEEE Transactions on Mobile Computing , 2020. -Not checked yet

  Not checked in the free run.

May combine two claims.

Simulations from 34,006 CT scan slices of 89 subjects verify a high COVID-19 X-ray image classification and low data loss in the FL algorithm.

“Simulations from 34,006 CT scan slices (images) of 89 subjects verify a high COVID-19 X-ray image classification and low data loss in the FL algorithm running.”

• cites [112] [112] R. Kumar, A. A. Khan, S. Zhang, J. Kumar, T. Yang, N. A. Golalirz, Zakria, I. Ali, S. Shafiq, and W. Wang, 'Blockchain-federated-learning and deep learning models for COVID-19 detection using CT imaging,' arXiv:2007.06537 [cs, eess] , 2020, arXiv: 2007.06537. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

An FL model for federated brain imaging using DNNs for brain tumour segmentation achieves similar segmentation performance with the ideal centralized scheme while a degree of privacy protection is achieved.

“Simulations are conducted using a brain tumour dataset containing multiparametric pre-operative MRI scans of 285 subjects, indicating a similar segmentation performance with the ideal centralized scheme, while a degree of privacy protection is achieved.”

• cites [99] [99] W. Li, F. Milletarì, D. Xu, N. Rieke, J. Hancox, W. Zhu, M. Baust, Y. Cheng, S. Ourselin, M. J. Cardoso brain tumour segmentation,' in International Workshop on Machine Learning in Medical Imaging et al. , 'Privacy-preserving federated , 2019, pp. 133-141. - sourceNot checked yet

  Not checked in the free run.

• AssumptionThe authors state this as an assumption: Patients only trust their local hospital, and hospitals are non-malicious and non-colluding, while the server might be honest but curious.

The quote only partly states this claim.

Simulations for an FL setting with 10 hospital sites show better precision and recall performance compared to non-FL approaches for obesity phenotyping.

“Simulations for an FL setting with 10 hospital sites show a better precision and recall performance compared to non-FL approaches.”

• cites [93] [93] D. Liu, D. Dligach, and T. Miller, 'Two-stage federated phenotyping and patient representation learning,' arXiv preprint arXiv:1908.05596 , 2019. -Not checked yet

  Not checked in the free run.

May combine two claims.

An FL approach for medical imaging with GANs yields an accuracy score of 0.9722, achieving a performance enhancement by 0.13% compared to non-FL schemes on prostate cancer-related images.

“Through the simulations on prostate cancer-related images, the proposed FL scheme can yield an accuracy score of 0.9722, achieving a performance enhancement by 0.13% in comparison to non-FL schemes.”

• cites [98] [98] Z. Yan, J. Wicaksana, Z. Wang, X. Yang, and K.-T. Cheng, 'Variation-aware federated learning with multi-source decentralized medical image data,' IEEE Journal of Biomedical and Health Informatics , 2020.Not checked yet

  Not checked in the free run.

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

Patient clustering captures similar diagnoses and geographical locations to train a neural network model to forecast mortality based on drug features.

“Moreover, to support the federated mortality prediction, patient clustering is adopted in [86], where patient subgroups capture similar diagnoses and geographical locations to train a neural network model to forecast mortality based on drug features.”

• cites [86] [86] L. Huang, A. L. Shea, H. Qian, A. Masurkar, H. Deng, and D. Liu, 'Patient clustering improves efficiency of federated machine learning to predict mortality and hospital stay time using distributed electronic medical records,' Journal of biomedical informatics , vol. 99, p. 103291, 2019. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

A disease prediction method using FL on a large nationwide health insurance dataset distributed over 99 medical sites achieves competitive performance in terms of high accuracy rates and privacy protection compared to centralized learning and local training without federation.

“By comparing with conventional approaches such as centralized learning, local training without federation, the FL method achieves a competitive performance in terms of high accuracy rates and privacy protection.”

• cites [95] [95] D. Liu, T. A. Miller, and K. D. Mandl, 'Confederated machine learning on horizontally and vertically separated medical data for large-scale health system intelligence,' arXiv preprint arXiv:1910.02109 , 2019. -Not checked yet

  Not checked in the free run.

May combine two claims.

FL can be used for facilitating in-home health monitoring by training a global model from distributed homes under the control of a server while preventing data leakage by keeping user data locally.

“FL can be used for facilitating in-home health monitoring, by training a global model from distributed homes under the control of a server, e.g., a cloud server, while preventing data leakage by keeping user data locally [72].”

• cites [72] [72] Q. Wu, X. Chen, Z. Zhou, and J. Zhang, 'FedHome: Cloud-edge based personalized federated learning for in-home health monitoring,' IEEE Transactions on Mobile Computing , 2020. -Not checked yet

  Not checked in the free run.

The quote only partly states this claim.

A statistic channel-based FL solution that uploads a small fraction of local gradients stochastically by choosing channels with the most substantial variation improves training convergence rates compared to traditional FL schemes.

“A binary classification problem is formulated for mortality prediction using NNs in the FL simulation, showing an improvement in training convergence rates compared to traditional FL schemes.”

• cites [85] [85] R. Shao, H. He, H. Liu, and D. Liu, 'Stochastic channel-based federated learning for medical data privacy preserving,' arXiv preprint arXiv:1910.11160 , 2019. -Not checked yet

  Not checked in the free run.

May combine two claims.

Real-world FL project for COVID-19 region segmentation in chest CT with medical institutions from China, Italy, and Japan: FL-based approach captures better ground truth shapes and has less false positives compared to traditional local training.

“Compared with the traditional local training approach using 945 scans with the support of expert radiologists, the FL-based approach is able to capture better the ground truth shapes and has less false positives with the collaboration of training resources from multiple medical centres.”

• cites [118] [118] D. Yang, Z. Xu, W. Li, A. Myronenko, H. R. Roth, S. Harmon, S. Xu, B. Turkbey, E. Turkbey, X. Wang, W. Zhu, G. Carrafiello, F. Patella, M. Cariati, H. Obinata, H. Mori, K. Tamura, P. An, B. J. Wood, and D. Xu, 'Federated semi-supervised learning for COVID region segmentation in chest CT using multi-national data from China, Italy, Japan,' Medical Image Analysis , vol. 70, p. 101992, 2021. - sourceNot checked yet

  Not checked in the free run.

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

Real-world experiment at University of Pennsylvania and 19 other institutions shows FL-based approach can achieve training accuracy rate of up to 90% on image datasets, competitive with centralized scheme.

“Several preliminary real-world experiments has been conduced at the centre of Biomedical Image Computing and Analytics of the University of Pennsylvania, showing that the FL-based approach can achieve an training accuracy rate of up to 90% on image datasets, a competitive performance compared with the centralized scheme.”

• cites [119] [119] 'Federated learning for medical imaging,' 2021. [Online]. Available: https://www.intel.com/content/www/us/en/artificialintelligence/posts/federated-learning-for-medical-imaging.html -Not checked yet

  Not checked in the free run.

May combine two claims.

FL is useful for facilitating in-home health monitoring by training a global model from distributed homes under a data server while preventing data leakage by keeping user data locally.

“FL is also useful for facilitating in-home health monitoring, by training a global model from distributed homes under the control of a data server, while preventing data leakage by keeping user data locally [92], [93].”

• cites [92] [92] Y. Chen, X. Qin, J. Wang, C. Yu, and W. Gao, 'Fedhealth: A federated transfer learning framework for wearable healthcare,' IEEE Intelligent Systems , vol. 35, no. 4, pp. 83-93, 2020. - sourceNot checked yet

  Not checked in the free run.

• cites [93] [93] D. Liu, D. Dligach, and T. Miller, 'Two-stage federated phenotyping and patient representation learning,' arXiv preprint arXiv:1908.05596 , 2019. -Not checked yet

  Not checked in the free run.

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

Many new technologies are introduced to meet much stricter 6G requirements such as blockchain, compressive sensing, THz and visible light communications, 3D networking, quantum communication, and large intelligence surface.

“At the same time, many new technologies are introduced to meet much stricter 6G requirements such as blockchain, compressive sensing, THz and visible light communications, 3D networking, quantum communication, and large intelligence surface [123].”

• cites [123] [123] C. de Alwis, A. Kalla, Q.-V. Pham, P. Kumar, K. Dev, W.-J. Hwang, and M. Liyanage, 'Survey on 6G frontiers: Trends, applications, requirements, technologies and future research,' IEEE Open Journal of the Communications Society , vol. 2, pp. 836-886, 2021. -Not checked yet

  Not checked in the free run.

May combine two claims.

Label distributions differ across medical institutions, making them challenging to join the federated data training.

“In this case, the label distributions differ across medical institutions, making them challenging to join the federated data training.”

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

FL privacy issues can be categorized as membership inference attacks, unintentional information leakage, and generative adversarial network.

“According to [13], FL privacy issues can be categorized as membership inference attacks, unintentional information leakage, and generative adversarial network.”

• cites [13] [13] V. Mothukuri, R. M. Parizi, S. Pouriyeh, Y. Huang, A. Dehghantanha, and G. Srivastava, 'A survey on security and privacy of federated learning,' Future Generation Computer Systems , vol. 115, pp. 619-640, 2021. - sourceNot checked yet

  Not checked in the free run.

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

Future e-health services will be advanced by AI and FL capabilities, thus enhancing the quality of life and reducing the hospitalizations for patients.

“For example, future e-health services will be advanced by AI and FL capabilities, thus enhancing the quality of life and reducing the hospitalizations for patients [124].”

• cites [124] [124] L. Mucchi, S. Jayousi, S. Caputo, E. Paoletti, P. Zoppi, S. Geli, and P. Dioniso, 'How 6G technology can change the future wireless healthcare,' in 2020 2nd 6G wireless summit (6G SUMMIT) . IEEE, 2020, pp. 1-6. -Not checked yet

  Not checked in the free run.

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

There have been many research and development activities towards future 6G wireless systems.

“Although 5G networks are not fully available and commercially deployed over the world, there have been many research and development activities towards future 6G wireless systems [123].”

• cites [123] [123] C. de Alwis, A. Kalla, Q.-V. Pham, P. Kumar, K. Dev, W.-J. Hwang, and M. Liyanage, 'Survey on 6G frontiers: Trends, applications, requirements, technologies and future research,' IEEE Open Journal of the Communications Society , vol. 2, pp. 836-886, 2021. -Not checked yet

  Not checked in the free run.

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

6G enables the availability of many applications such as Industry 5.0, intelligent healthcare, smart grid, holographic telepresence, and personalized body area networks.

“6G enables the availability of many applications such as Industry 5.0, intelligent healthcare, smart grid, holographic telepresence, and personalized body area networks.”

• cites [123] [123] C. de Alwis, A. Kalla, Q.-V. Pham, P. Kumar, K. Dev, W.-J. Hwang, and M. Liyanage, 'Survey on 6G frontiers: Trends, applications, requirements, technologies and future research,' IEEE Open Journal of the Communications Society , vol. 2, pp. 836-886, 2021. -Not checked yet

  Not checked in the free run.

May combine two claims.

Game theory and blockchain are two important tools for designing incentive mechanisms for federated learning.

“Game theory and blockchain are two important tools for designing incentive mechanisms”

• cites [32] [32] L. U. Khan, S. R. Pandey, N. H. Tran, W. Saad, Z. Han, M. N. Nguyen, and C. S. Hong, 'Federated learning for edge networks: Resource optimization and incentive mechanism,' IEEE Communications Magazine , vol. 58, no. 10, pp. 88-93, 2020. - sourceNot checked yet

  Not checked in the free run.

• cites [46] [46] J. Xu, H. Wang, and L. Chen, 'Bandwidth allocation for multiple federated learning services in wireless edge networks,' arXiv preprint arXiv:2101.03627 , 2021. -Not checked yet

  Not checked in the free run.

• cites [55] [55] J. Li, Y. Shao, K. Wei, M. Ding, C. Ma, L. Shi, Z. Han, and H. V. Poor, 'Blockchain assisted decentralized federated learning (BLADE-FL): Performance analysis and resource allocation,' arXiv preprint arXiv:2101.06905 , 2021. -Not checked yet

  Not checked in the free run.

• cites [56] [56] J. Kang, Z. Xiong, D. Niyato, S. Xie, and J. Zhang, 'Incentive mechanism for reliable federated learning: A joint optimization approach to combining reputation and contract theory,' IEEE Internet of Things Journal , vol. 6, no. 6, pp. 10 700-10 714, 2019. -Not checked yet

  Not checked in the free run.

• cites [68] [68] Y. Sarikaya and O. Ercetin, 'Motivating workers in federated learning: A stackelberg game perspective,' IEEE Networking Letters , vol. 2, no. 1, pp. 23-27, 2019. -Not checked yet

  Not checked in the free run.

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

FL is expected to play a key role in realizing large-scale and collaborative healthcare systems and allow for a shift from centralized health data analytics to distributed healthcare operations with privacy awareness.

“FL is expected to play a key role in realizing large-scale and collaborative healthcare systems and allow for a shift from centralized health data analytics to distributed healthcare operations with privacy awareness.”

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

The application of FL in smart healthcare is still at its infancy and will quickly mature in the coming years for providing intelligent and privacy-enhanced health services.

“The application of FL in smart healthcare is still at its infancy and will quickly mature in the coming years for providing intelligent and privacy-enhanced health services.”