Handling privacy-sensitive medical data with federated learning: Challenges and future directions

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

No existing work extensively reviews the use of FL in IoMT networks and applications.

“Although FL has been well studied in different domains, to the best of our knowledge, no existing work extensively reviews the use of FL in IoMT networks and applications.”

May combine two claims.

A general FL-trained model can achieve performance levels comparable to models trained on centrally hosted datasets and superior to models that only see isolated single-institutional data.

“Independently of the FL training strategy, a recent analysis for the medical field has shown that a general FL-trained model can achieve performance levels comparable to the ones trained on centrally hosted datasets and superior to models that only see isolated single-institutional data [27], [28].”

• cites [27] sourceNot checked yet

  Not checked in the free run.

• cites [28] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Federated learning enables collaborative training of a global model over a decentralized network via local updates, without exchanging private data, thereby keeping data local and preserving privacy.

“With FL, a global model is trained collaboratively by each agent of the system (e.g., hospitals, IoMTs, or health care centers) over the decentralized network.This is done via local updates, without exchanging private data.”

• cites [7] sourceNot checked yet

  Not checked in the free run.

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

FL requires computationally robust devices.

“FL also requires computationally robust devices.”

May combine two claims.

FL improves patient privacy and reduces low latency during the collaboration of multiple hospitals or entities.

“It also improves patient privacy and reduces low latency during the collaboration of multiple hospitals/entities [7].”

• cites [7] sourceNot checked yet

  Not checked in the free run.

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

The TFL framework increases the classification time.

“It also shows that the TFL framework increases the classification time.”

• cites [44] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

FedHealth can improve classification performance compared to the non-federated model and the traditional machine learning model.

“The experiment results demonstrate that FedHealth can improve the classification performance compared to the non-federated model and the traditional machine learning model.”

• cites [43] sourceNot checked yet

  Not checked in the free run.

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

Although FL provides privacy-preservation by allowing clients to keep data local, there are still model security and data leakage risks that would compromise the security of the FL system and the data privacy of clients.

“Although FL is the enabling technology in this scenario and provides a privacy-preservation capability by allowing the clients to keep the data on local devices, there are still model security and data leakage risks that would compromise the security of the FL system and the data privacy of clients.”

• AssumptionRests on an assumption: Security problems in FL are primarily caused by curious or malicious attackers targeting vulnerabilities of the FL system, which can lead to significant performance drop and sometimes model invalidation.
• AssumptionRests on an assumption: In healthcare and IoMT scenarios, an attacker can directly manipulate the model and data of a local affiliation, resulting in a malicious update of the global model or mislabeled data.

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

FL has made a huge contribution to the fight against COVID-19 without sharing patient data.

“Moreover, recently FL has made a huge contribution to the fight against COVID-19 without sharing patient data.”

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

When the training data is IID, the difference between FL and non-FL is within 3%.

“Also, the authors showed that when the training data is IID, the difference between FL and non-FL is within 3%.”

• cites [37] sourceNot checked yet

  Not checked in the free run.

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

Human activity recognition with semi-supervised FL is not affected by non-IID data and can achieve comparable accuracy to supervised FL.

“Their experimental results show that human activity recognition with semi-supervised FL is not affected by the non-IID data and can achieve comparable accuracy to that of the supervised FL.”

• cites [40] Zhao, 2020, Semi-supervised federated learning for activity recognitionNot checked yet

  Not checked in the free run.

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

FL requires less communication and computational resources while using less complex models, but at the cost of lower accuracy.

“Then, they study the trade-off between the communication cost and the model accuracy, which indicates that FL requires less communication and computational resources while using less complex models but at the cost of lower accuracy.”

• cites [37] sourceNot checked yet

  Not checked in the free run.

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

FTL outperforms the non-TFL approach in terms of the Area Under The Curve (AUC) metric and maintains the same accuracy.

“The results demonstrate that the FTL outperforms the non-TFL approach in terms of the Area Under The Curve (AUC) metric and maintains the same accuracy.”

• cites [44] sourceNot checked yet

  Not checked in the free run.

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

FL plays an important role in facilitating healthcare services.

“From the presented state-of-the-art on FL-IoMT, we can notice that FL plays an important role in facilitating healthcare services.”

May combine two claims.

Some IoMT devices offload their model training task to edge gateways, helping the FL training process take advantage of cloud and edge computing, and using IoMT devices as simple data collectors.

“For this reason, some IoMT devices offload their model training task to the edge gateways; therefore, helping the FL training process to take advantage of the cloud and edge computing, and using the IoMT devices as simple data collectors.”

May combine two claims.

FL requires less communication overhead, and hence it can replace traditional learning approaches in IoMT applications.

“It can be learned from [37] that FL requires less communication overhead, and hence it can replace traditional learning approaches in IoMT applications.”

• cites [37] sourceNot checked yet

  Not checked in the free run.

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

The shared global model trained on the FL server fails in personalization due to different characteristics or daily activity patterns of users; hence, it is important to have a fine-grained or personalized model.

“Although the efficiency of FL application for healthcare systems, the shared global model trained on the FL server fails in personalization due to the different characteristics or daily activity patterns of users; hence, it is important to have a fine-grained or personalized model.”

Classified as probably not a substantial claim.

FL can play a significant role in various industry domains that involve sensitive personal data, such as IoMT for healthcare.

“Since federated learning comes with a privacy-preserving attribute, it can play a significant role in various industry domains that involve sensitive personal data, such as IoMT for healthcare.”

• AssumptionRests on an assumption: Medical data is highly sensitive and must be protected by appropriate confidentiality procedures.

May combine two claims.

Federated learning still suffers from information leakage issues and, having multiple parties, extends the attack surface.

“FL, despite a different training process compared to traditional ML, still suffers from information leakage issues and, having multiple parties, extends the attack surface.”

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

FL solves the problem of having inadequate data by providing a trust factor between heterogeneous domains.

“For example, along with the data privacy and delocalization problems, FL also solves the problem of having inadequate data by providing a trust factor between heterogeneous domains.”

May combine two claims.

Edge computing helps to train the model closer to end-users and may conduct inference much faster than in the cloud.

“As shown in Fig. 2, edge computing helps to train the model closer to the endusers and hence may conduct inference much faster than in the cloud [46].”

• cites [46] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

ResNet-18 has the best performance in training both with FL and without FL approaches, while ResNeXt shows the highest efficiency on images with COVID-19 labels.

“The experimental results show that ResNet-18 has the best performance in training both with FL and without FL approaches, while ResNeXt shows the highest efficiency on images with COVID-19 labels.”

• cites [49] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

In IoMT applications, intercepted data could be a patient's personal or healthcare information, which presents a severe ethical problem.

“In IoMT applications, the intercepted data could be a patient's personal or healthcare information, which presents a severe ethical problem.”

May combine two claims.

FL provides a highly trained ML model without the risk of exposing training data since the data are kept where they are generated.

“It provides a highly trained ML model without the risk of exposing training data since they are kept where they are generated.”

• cites [7] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Edge computing is an efficient solution to address the limited hardware of IoMT devices and reduce network congestion and latency that occur with the cloud.

“It is an efficient solution to address these issues and reduce the network congestion and latency that occur with the cloud.”

May combine two claims.

The FedHome framework is effective in terms of accuracy, computation, and communication overhead.

“The experimental results on human activity recognition demonstrate the effectiveness of the FedHome framework in terms of accuracy, computation, and communication overhead.”

• cites [46] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Transfer learning avoids learning from scratch and solves the problem of insufficient training data, and can be effective in healthcare scenarios.

“As a result, the recent technique of transfer learning, which avoids learning from scratch and solves the problem of insufficient training data, can be effective in this scenario [42].”

• cites [42] sourceNot checked yet

  Not checked in the free run.

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

The proposed model in [35] converges faster and with less communication overhead compared to an alternative distributed algorithm.

“The theoretical and experimental comparisons show that the proposed model converges faster and with less communication overhead compared to an alternative distributed algorithm.”

• cites [35] sourceNot checked yet

  Not checked in the free run.

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

Some clients may deliberately try to extract information from other parties, degrade performance, or bring the system down.

“Some clients, for example, may deliberately try to extract information from other parties, degrade performance, or bring the system down.”

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

While standards are necessary for security and privacy purposes, they can make sharing and using health data challenging.

“While standards are necessary for security and privacy purposes, they can make sharing and using health data challenging.”

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

Operating FL systems at a larger scale might make collaborative agreements impractical, so it is more reasonable not to have trustworthiness assumptions.

“However, operating FL systems on a larger scale might be impractical to establish collaborative agreements and it is more reasonable not to have trustworthiness assumptions.”

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

For FL consortia in which all parties are bounded by an enforceable collaboration agreement, some nefarious motivations can be disregarded.

“For FL consortia in which all parties are bounded by an enforceable collaboration agreement, we can disregard some of the more nefarious motivations, such as deliberated attempts to extract sensitive information or to intentionally corrupt the model.”

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

Precious data often remain confined to individual institutions and are rarely leveraged for analysis, hindering the application of deep learning algorithms in healthcare.

“In view of the above arguments, precious data often remain confined to individual institutions and are rarely leveraged for analysis, hindering the application of deep learning algorithms in healthcare.”

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

The model can be manipulated by attackers inducing additional memorization through gradient-ascent-style attacks.

“Besides, the model can be manipulated by attackers inducing additional memorization through gradient-ascent-style attacks.”

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

Personal health information is highly sensitive, containing diagnostic and healthcare related information and identifiable details about individuals.

“In fact, personal health information is considered to be highly sensitive, containing not only diagnostic and healthcare related information but also identifiable details about individuals.”

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

Bonawitz et al. introduce a secure aggregation protocol for protecting individual model updates.

“Bonawitz et al. [62] introduce a secure aggregation protocol for protecting individual model updates.”

• cites [62] sourceNot checked yet

  Not checked in the free run.

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

The applied secure aggregation is a lossless method that can retain the original accuracy with a very high privacy guarantee.

“The applied secure aggregation is a lossless method that can retain the original accuracy with a very high privacy guarantee.”

• cites [62] sourceNot checked yet

  Not checked in the free run.

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

Some sophisticated countermeasures can be avoided if all parties are deemed trustworthy.

“At the same time, it can be noted how some sophisticated countermeasures can be avoided if all parties are deemed trustworthy.”

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

Privacy problems are exacerbated in medical companies by strict laws and regulations designed to prevent re-identification and data breaches.

“Lastly, privacy problems are exacerbated in medical companies by strict laws and regulations designed to prevent the risk of re-identification and data breaches.”

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

If standards are respected, FL can foster the application of deep learning in digital health and healthcare informatics, making it possible to use existing health datasets while protecting user privacy.

“If these standards are respected, FL can foster the application of deep learning in digital health and healthcare informatics, making it possible the use of existing health datasets while protecting user privacy.”

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

Data privacy is one of the public's most important concerns because data breaches can result in reduced public trust.

“Even from the consumer perspective, data privacy is one of the public's most important concerns because data breaches can result in reduced public trust.”

• AssumptionRests on an assumption: Data breaches can result in reduced public trust.

May combine two claims.

In Bonawitz et al.'s secure aggregation solution, the central server cannot see local updates but can observe the exact aggregated results at each round.

“In this solution, the central server is not able to see any local updates, but it can still observe the exact aggregated results at each round.”

• cites [62] sourceNot checked yet

  Not checked in the free run.

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

Li et al. introduced locally differentially private algorithms in the context of meta-learning while providing provable learning guarantees in convex settings.

“Li et al. [65] introduced locally differentially private algorithms in the context of meta-learning while providing provable learning guarantees in convex settings.”

• cites [65] Li, 2019, Differentially private meta-learningNot checked yet

  Not checked in the free run.

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

Deployment of FL on blockchain technology can establish data security and eliminate trust issues.

“Another approach to establish data security and eliminate the trust issues is the deployment of FL on blockchain technology [71]- [73].”

• cites [71] sourceNot checked yet

  Not checked in the free run.

• cites [72] sourceNot checked yet

  Not checked in the free run.

• cites [73] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The differential privacy approach comes with many hyperparameters that need careful selection and affect communication and accuracy.

“This type of approach comes with a considerable number of hyperparameters that need to be carefully selected and that affect communication and accuracy.”

• cites [63] McMahan, 2017, Learning differentially private recurrent language modelsNot checked yet

  Not checked in the free run.

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

PRCL reduces local training overhead by outsourcing the middle part to the cloud server while providing similar accuracy to other state-of-the-art approaches.

“The simulation results show that PRCL reduces the local training overhead because outsourcing the middle part to the cloud server while providing similar accuracy to other state-of-the-art approaches.”

• cites [68] sourceNot checked yet

  Not checked in the free run.

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

When a client's trust value is lower than 60%, the request from this client will be prohibited.

“If lower the respective target values, the lower the trust value of the client, and when such a trust value is lower than 60%, the request from this client will be prohibited.”

• cites [70] sourceNot checked yet

  Not checked in the free run.

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

Bhowmick et al. proposed a relaxed version of local privacy by limiting the power of potential adversaries.

“For stronger privacy guarantees, Bhowmick et al. [64] have proposed a relaxed version of local privacy by limiting the power of potential adversaries.”

• cites [64] Bhowmick, 2018, Protection against reconstruction and its applications in private federated learningNot checked yet

  Not checked in the free run.

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

A blockchain-based verification scheme can improve the reliability of the federated training process.

“In addition, the blockchain-based verification scheme can also improve the reliability of the federated training process.”

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

EPPDA preserves user privacy, requires less computing and communication resources than alternatives such as [62], and provides acceptable fault tolerance in the event of user disconnection.

“Adopting the homomorphisms of secret sharing, it preserves user privacy and requires less computing and communication resources than alternatives as [62], and is able to provide acceptable fault tolerance in the event of user disconnection.”

• cites [69] sourceNot checked yet

  Not checked in the free run.

• cites [62] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

In PRCL, confidentiality is preserved because the cloud cannot access the plaintext gradient and an attacker can only get the perturbed data due to the added noise.

“At the same time, confidentiality is preserved since the cloud cannot access the plaintext gradient and an attacker can only get the perturbed data due to the added noise.”

• cites [68] sourceNot checked yet

  Not checked in the free run.

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

Bhowmick et al.'s relaxed local privacy provides stronger guarantees than global privacy and better model performance than strict local privacy.

“Providing stronger guarantees than global privacy, it has better model performance than strict local privacy.”

• cites [64] Bhowmick, 2018, Protection against reconstruction and its applications in private federated learningNot checked yet

  Not checked in the free run.

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

EPPDA, based on secret sharing to resist the reverse attack, can covertly aggregate user-trained models without disclosing the user model.

“Similar results have been achieved by EPPDA [69], an efficient privacy-preserving data aggregation method for FL that, based on secret sharing to resist the reverse attack, can covertly aggregate user-trained models without disclosing the user model.”

• cites [69] sourceNot checked yet

  Not checked in the free run.

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

Wang et al. proposed a defensive strategy based on comparing the uploaded model's accuracy and size.

“Wang et al.proposed a defensive strategy based on comparing the uploaded model's accuracy and size [70].”

• cites [70] sourceNot checked yet

  Not checked in the free run.

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

TrustFed uses Industrial IoT devices as federated learning candidates and smart contract technology over the Ethereum blockchain to maintain participants' reputations.

“For example, [76] presents TrustFed, which uses Industrial IoT devices as federated learning candidates and smart contract technology over the Ethereum blockchain to maintain participants' reputations.”

• cites [76] sourceNot checked yet

  Not checked in the free run.

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

DITEN integrates blockchain and FL in edge networks and uses deep neural networks as a strategy scheduler to ensure privacy of user data and enhance learning security.

“Lu et al. [77] designed a new system, called DITEN, that integrates blockchain and FL in edge networks and uses Deep Neural Networks (DNN) as a strategy scheduler to ensure the privacy of user data and enhance learning security.”

• cites [77] sourceNot checked yet

  Not checked in the free run.

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

Qu et al.'s approach guarantees block generation efficiency, enabling decentralized privacy protection while preventing single point failure.

“In such a way, block generation efficiency is guaranteed, enabling decentralized privacy protection while preventing single point failure.”

• cites [72] sourceNot checked yet

  Not checked in the free run.

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

Blockchain FL offers advantages such as decentralization, non-tempering, open autonomy, and anonymous traceability.

“Such a solution offers several advantages such as decentralization, non-tempering, open autonomy, and anonymous traceability [75].”

• cites [75] sourceNot checked yet

  Not checked in the free run.

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

Qu et al. propose storing only the pointer of the global updates on-chain while a distributed hash table is used to save the data.

“To empower FL with this approach, Qu et al. [72] propose to store only the pointer of the global updates on-chain, while a distributed hash table is used to save the data.”

• cites [72] sourceNot checked yet

  Not checked in the free run.

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

A blockchain-based FL should mitigate the overhead of frequent communications in blockchain consensus, which can cause excessive latency.

“A blockchain-based FL, however, should mitigate the overhead of frequent communications in the blockchain consensus, which can cause excessive latency.”

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

TrustFed can identify and eliminate outliers in the training distributions before combining the model updates.

“The solution can also identify and eliminate outliers in the training distributions before combining the model updates.”

• cites [76] sourceNot checked yet

  Not checked in the free run.

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

Two key aspects must be considered for an effective reduction of communication overhead: the total number of communication rounds and the size of messages transmitted in each round.

“Two key aspects must be considered for an effective reduction of the communication overhead: the total number of communication rounds, and the size of messages transmitted in each round.”

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

FedAvg is based on averaging local stochastic gradient descent updates for the primal problem.

“The most commonly used method for FL, FedAvg [29], is based on averaging local stochastic gradient descent (SGD) updates for the primal problem.”

• cites [29] McMahan, 2017, Communication-efficient learning of deep networks from decentralized dataNot checked yet

  Not checked in the free run.

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

In practice, FL solutions for IoMT suffer from several issues and challenges that should be considered when designing FL-based approaches for healthcare services.

“However, in practice, these solutions suffer from several issues and challenges that should be considered when designing FL-based approaches for healthcare services.”

May combine two claims.

Some recent techniques can partially solve privacy issues, but the more secure the techniques, the lower the accuracy of the final model may be.

“Some recent techniques can partially solve these privacy issues [61], however, the more secure the techniques, the lower the accuracy of the final model may be [80].”

• cites [61] sourceNot checked yet

  Not checked in the free run.

• cites [80] sourceNot checked yet

  Not checked in the free run.

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

A high volume of aggregated data from different parties must be processed by the server.

“We can mention two main obstacles: (i) a high volume of aggregated data from different parties to be processed by the server; (ii) none of these parties fully trust each others (including the server), thus fearing data leakage.”

May combine two claims.

Asynchronous decentralized SGD methods are inherently robust to stragglers but may suffer from slow convergence due to the use of stale models.

“Although these asynchronous methods are inherently robust to stragglers, they may suffer from slow convergence due to the use of stale models.”

• cites [82] Dutta, 2018, Slow and stale gradients can win the race: Error-runtime trade-offs in distributed SGDNot checked yet

  Not checked in the free run.

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

A blockchain-based approach can jointly solve the problem of single-point-of-failure and central trust.

“Moreover, decentralizing can be combined with blockchain technology, where a blockchain-based approach can jointly solve the problem of single-point-of-failure and central trust.”

May combine two claims.

Federated networks of IoMT can comprise a massive number of devices, and communication can be many orders of magnitude slower than local computation due to limited bandwidth, energy, and power.

“Federated networks of IoMT can indeed comprise a massive number of devices (e.g., sensors), and communication in the network can be many orders of magnitude slower than local computation due to limited resources, such as bandwidth, energy, and power [86].”

• cites [86] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

FedAvg works well for non-convex problems but does not exhibit convergence guarantees and can diverge in practical settings when data are heterogeneous.

“While FedAvg has been shown to work well for non-convex problems, it does not exhibit convergence guarantees and can diverge in practical settings when data are heterogeneous [32] (see Section V-D for how to address this problem).”

• cites [32] Li, 2020, Federated optimization in heterogeneous networksNot checked yet

  Not checked in the free run.

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

With either peer-to-peer or blockchain solutions, the problem is converted to the pursuit of fault tolerance for the clients.

“With either solution, i.e., peer-to-peer or blockchain, the problem is converted to the pursuit of fault tolerance for the clients.”

May combine two claims.

Other recent methods use secure multiparty computation or differential privacy, often providing privacy at the cost of reduced model performance or system efficiency.

“Other recent methods aiming to enhance the privacy of federated learning use tools such as secure multiparty computation or differential privacy.These approaches often provide privacy at the cost of reduced model performance or system efficiency [62].”

• cites [62] sourceNot checked yet

  Not checked in the free run.

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

The features of FL make it uniquely suited for sensitive data as in IoMT systems, but the architecture and model sharing must address further challenges compared to other fields.

“In conclusion, the features of FL make it uniquely suited for sensitive data as in the context of IoMT systems, but the architecture and model sharing must address further challenges compared to other fields.”

May combine two claims.

A distributed decentralized approach can make the system highly scalable but also opens up issues in gradient synchronization.

“However, while this distributed approach can help to make the system highly scalable, it also opens up some issues in gradient synchronization.”

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

None of the parties fully trust each other, including the server, and thus fear data leakage.

“We can mention two main obstacles: (i) a high volume of aggregated data from different parties to be processed by the server; (ii) none of these parties fully trust each others (including the server), thus fearing data leakage.”

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

Blockchain removes the need for a trusted curator, connects each participant through multiparty data retrieval, strengthens the data-sharing scheme, and provides some fault-tolerant guarantees.

“While blockchain removes the need for a trusted curator and connects each participant through multiparty data retrieval, it also strengthens the data-sharing scheme and provides some fault-tolerant guarantees [83].”

• cites [83] sourceNot checked yet

  Not checked in the free run.

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

Lu et al. integrated federated learning into the consensus algorithm of the blockchain to save the hash rate in industrial IoT, where blockchain enables secure data retrieval and ensures accurate model training.

“In a similar way, Lu et al. [78] integrated federated learning into the consensus algorithm of the blockchain to save the hash rate in the context of industrial IoT, in which blockchain enables secure data retrieval and ensures accurate model training.”

• cites [78] sourceNot checked yet

  Not checked in the free run.

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

Decentralized training is a potential alternative in FL to reduce high communication costs on the central server.

“Lastly, decentralized training is a potential alternative in FL to reduce high communication costs on the central server.”

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

FAug is a data augmentation scheme that can collectively learn the trade-off between privacy leakage and communication overhead using a generative adversarial network.

“To further improve the transmission of model output, the approach is combined with Federated Augmentation (FAug), a data augmentation scheme that can collectively learn the trade-off between privacy leakage and communication overhead using a Generative Adversarial Network (GAN).”

• cites [91] Jeong, 2018, Communication-efficient on-device machine learning: Federated distillation and augmentation under non-IID private dataNot checked yet

  Not checked in the free run.

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

Federated Distillation has a communication payload size much smaller than that of a benchmark FL scheme, particularly when the model size is large.

“Moreover, Jeong et al. [91] presented Federated Distillation (FD), a distributed model training algorithm whose communication payload size is much smaller than that of a benchmark FL scheme, particularly when the model size is large.”

• cites [91] Jeong, 2018, Communication-efficient on-device machine learning: Federated distillation and augmentation under non-IID private dataNot checked yet

  Not checked in the free run.

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

FedProx tolerates that any selected device performs partial work and can safely incorporate these partial updates via a proximal term.

“FedProx [32], for instance, tolerates that any selected device to perform partial work, in compliance with the underlying system's constraints, and can safely incorporate these partial updates via a proximal term.”

• cites [32] Li, 2020, Federated optimization in heterogeneous networksNot checked yet

  Not checked in the free run.

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

For medical data collected by IoMT, simple models can be likely inaccurate.

“However, for medical data collected by IoMT simple models can be likely inaccurate.”

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

Federated Distillation exchanges model output instead of model parameters, so that large local models can be adopted on the ML device.

“On top of conventional periodic communication of FL, the proposed FD exchanges, not the model parameters but the model output, so that large local models can be adopted on the ML device.”

• cites [91] Jeong, 2018, Communication-efficient on-device machine learning: Federated distillation and augmentation under non-IID private dataNot checked yet

  Not checked in the free run.

May combine two claims.

While a large number of clients can increase the diversity of the local model across them, this may degrade the performance of FL.

“In addition, while a large number of clients can increase the diversity of the local model across them, this may degrade the performance of FL [93].”

• cites [93] Zhang, 2020, Benchmarking semisupervised federated learningNot checked yet

  Not checked in the free run.

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

By selecting devices with favorable resources, FedCS provides high classification accuracy in a significantly shorter time compared to classical FL.

“By selecting the devices with favorable resources, FedCS provides high classification accuracy in a significantly shorter time compared to the classical FL.”

• cites [94] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Recent articles on decentralized training over heterogeneous data with local updating schemes are either restricted to linear models or assume full device participation.

“Although some recent articles have investigated decentralized training over heterogeneous data with local updating schemes, e.g., [90], they are either restricted to linear models or assume full device participation.”

• cites [90] He, 2018, Cola: Decentralized linear learningNot checked yet

  Not checked in the free run.

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

Using less complex models such as softmax regression is a viable solution for most real-world applications because user behavior phenomena are often simple enough to be captured by relatively simple models.

“Results confirm that using less complex models, such as softmax regression, is a viable solution for most real-world applications since user behavior phenomena are often simple enough to be captured by relatively simple models.”

• cites [37] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

A high number of local updating methods can reduce the total number of communication rounds, but reducing the size of messages exchanged at each round via compression schemes can be convenient.

“Although a high number of local updating methods can effectively reduce the total number of communication rounds, it can be convenient to reduce the size of messages exchanged at each round via compression schemes model, e.g., sparsification, subsampling, and quantization.”

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

Konen et al. showed that using CNNs and LSTMs, communication costs can be reduced by two orders of magnitude compared to the original federated learning algorithm.

“The authors showed how using CNNs and LSTMs, the communication costs can be reduced by two orders of magnitude compared to the original federated learning algorithm.”

• cites [88] Konen, 2016, Federated learning: Strategies for improving communication efficiencyNot checked yet

  Not checked in the free run.

May combine two claims.

IoMT devices from remote areas may be prone to drop due to poor network connectivity or battery failure, so the trained federated model will be biased towards devices with favorable network conditions.

“IoMT devices from remote areas may be proner to drop due to poor network connectivity or battery failure and, therefore, the trained federated model will be biased towards devices with favorable network conditions.”

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

A joint-announcement protocol can randomly select devices/clients from a large number of participants and is useful to alleviate the out-of-sync issue and tolerate dropout of some clients before completion of each training iteration.

“Recent studies present a jointannouncement protocol that can randomly select some devices/ clients from a large number of participants in a given training round [85].Such a joint-announcement protocol is useful to alleviate the out-of-sync issue and tolerate the dropout of some clients at some points before the completion of each training iteration.”

• cites [85] Bonawitz, 2019, Toward federated learning at scale: System designNot checked yet

  Not checked in the free run.

May combine two claims.

Little work has evaluated the impact of failures over the FL algorithm or studied directly the effect of dropped devices.

“However, while several recent studies have investigated convergence guarantees of variants of FL methods [95]- [97], little work has evaluated the impact of failures over the FL algorithm or studied directly the effect of dropped devices.”

• cites [95] Jiang, 2018, A linear speedup analysis of distributed deep learning with sparse and quantized communicationNot checked yet

  Not checked in the free run.

• cites [96] sourceNot checked yet

  Not checked in the free run.

• cites [97] Yu, 2019, On the linear speedup analysis of communication efficient momentum SGD for distributed non-convex optimizationNot checked yet

  Not checked in the free run.

May combine two claims.

Multi-task learning and meta-learning enable personalized or device-specific modeling, which is considered a valuable attempt to handle the statistical heterogeneity of the data.

“Both approaches (multi-task and meta-learning) enable personalized or device-specific modeling, which is considered a valuable attempt to handle the statistical heterogeneity of the data.”

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

Once the trained generator of the GAN is downloaded, each device can replenish the target labels until an IID training dataset is achieved.

“Once the trained generator of the GAN is downloaded, each device can replenish the target labels until an IID training dataset is achieved”

• cites [91] Jeong, 2018, Communication-efficient on-device machine learning: Federated distillation and augmentation under non-IID private dataNot checked yet

  Not checked in the free run.

May combine two claims.

Non-IID data generation violates the frequently used assumptions of independent and identically distributed data and may cause problems in modeling, theoretical analysis, and empirical evaluation of solutions.

“This paradigm of data generation violates the frequently-used assumptions of independent and identically distributed (IID) data and may cause problems in modeling, theoretical analysis, and empirical evaluation of solutions.”

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

New schemes ensuring fault tolerance would improve the deployment of FL in healthcare where IoMT devices are likely to fault, encounter challenging network conditions, and exhaust their batteries.

“Therefore, new schemes ensuring fault tolerance would improve the deployment of FL in healthcare where IoMT devices are likely to a fault, encounter challenging network conditions, and exhaust their batteries.”

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

The number of data samples can vary significantly from device to device, and there may be an underlying statistical structure that captures the relationship among devices and their associated distributions.

“In addition to the variety of data, the number of data samples can vary significantly from device to device, and there may be an underlying statistical structure that captures the relationship among devices and their associated distributions.”

May combine two claims.

Part-Data-Sharing introduces a trade-off between accuracy and centralization and could lead to significant communication overhead, especially for medical data.

“However, despite introducing a trade-off between accuracy and centralization, this could lead to significant communication overhead, especially for medical data.”

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

FAug allows each device to generate missing data samples locally using a conditional GAN to rectify the non-IID training dataset while addressing communication overhead.

“To rectify the non-IID training dataset while addressing the communication overhead, Jeong et al. [91] proposed FAug, in which each device can generate the missing data samples locally using a generative model, i.e., a conditional GAN.”

• cites [91] Jeong, 2018, Communication-efficient on-device machine learning: Federated distillation and augmentation under non-IID private dataNot checked yet

  Not checked in the free run.

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

Part-Data-Sharing addresses the statistical challenge of FL when local data is non-IID by distributing a small amount of globally shared data containing examples from each class.

“It addresses the statistical challenge of FL when local data is non-IID by distributing a small amount of globally shared data containing examples from each class.”

• cites [31] Zhao, 2018, Federated learning with non-IID dataNot checked yet

  Not checked in the free run.

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

In ELFISH, the neurons that have less significant weight parameter updates are randomly masked.

“With ELFISH, the neurons that have less significant weight parameter updates will be randomly masked.”

• cites [107] Xu, 2019, ELFISH: Resource-aware federated learning on heterogeneous edge devices, RatioNot checked yet

  Not checked in the free run.

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

Anh et al. applied a Deep Q-learning algorithm based on the Double Deep Q-Network (DDQN) to optimize resource allocation for model training.

“Furthermore, Anh et al. [109] applied a Deep Q-learning algorithm based on the Double Deep Q-Network (DDQN) to optimize resource allocation for model training.”

• cites [109] sourceNot checked yet

  Not checked in the free run.

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

Training CNN-based models for COVID-19 detection (e.g., ResNet-50) requires storage and memory resources that may not be available on all devices.

“For example, training CNN-based models for COVID-19 detection (e.g., ResNet-50) requires storage and memory resources that may not be available on all devices.”

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

The proposed DDQN-based approach can minimize energy consumption and improve the training latency.

“The simulation results show that the proposed approach can minimize energy consumption and improve the training latency.”

• cites [109] sourceNot checked yet

  Not checked in the free run.

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

Data-driven medicine aiming to improve patient care globally requires federated efforts.

“While machine learning approaches, and especially deep learning, are becoming the de-facto knowledge discovery approach in digital healthcare, it appears that data-driven medicine aiming to improve patient care globally requires federated efforts.”

• AssumptionRests on an assumption: Data-driven medical methods face scarcity and privacy of user data, as well as difficulty of institutions exchanging medical information.
• AssumptionRests on an assumption: With insufficient data, ML is prevented from reaching its full potential, which is only possible if the database consists of the full spectrum of possible anatomies, pathologies, and input data types.

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

Xu et al. proposed a resource-aware federated learning framework called ELFISH.

“Similarly, Xu et al. [107] proposed a resource-aware federated learning framework, called ELFISH.”

• cites [107] Xu, 2019, ELFISH: Resource-aware federated learning on heterogeneous edge devices, RatioNot checked yet

  Not checked in the free run.

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

Jiang et al. proposed a new FL paradigm called PruneFL in order to minimize the model's size.

“In the same direction, Jiang et al. [108] proposed a new FL paradigm called PruneFL in order to minimize the model's size.”

• cites [108] sourceNot checked yet

  Not checked in the free run.

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

Federated Dropout reduces local computations by 1.7x without affecting the accuracy of the model.

“Simulation results demonstrate the superior performance of this solution by reducing the local computations by 1.7× without affecting the accuracy of the model.”

• cites [89] Caldas, 2018, Expanding the reach of federated learning by reducing client resource requirementsNot checked yet

  Not checked in the free run.

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

Data pre-processing is an important step to remove redundancy in the training set and to find the relevant features to be used during model training.

“In addition, data pre-processing is an important step to remove the redundancy in the training set as well as to find the relevant features to be used during the model training [110], [111].”

• cites [110] sourceNot checked yet

  Not checked in the free run.

• cites [111] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Using edge computing to provide complementary resources for local model training would require time-consuming data transmission between nano-devices and edge nodes and could lead to network congestion.

“A possible solution in which edge computing provides complementary resources in support of the local model training [105], would require time-consuming data transmission between the nano-devices and the edge nodes and could lead to network congestion, especially given the huge amount of data generated by these devices every second.”

• cites [105] sourceNot checked yet

  Not checked in the free run.

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

Caldas et al. introduce Federated Dropout to reduce the computation load of local training.

“Caldas et al. [89] introduce Federated Dropout to reduce the computation load of local training.”

• cites [89] Caldas, 2018, Expanding the reach of federated learning by reducing client resource requirementsNot checked yet

  Not checked in the free run.

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

PruneFL always converges to similar accuracy achieved by classical FL.

“The experimental results show that PruneFL always converges to similar accuracy achieved by classical FL.”

• cites [108] sourceNot checked yet

  Not checked in the free run.

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

FL is optimizing the drug discovery process by reducing costs and time-to-market for pharmaceutical companies.

“Already today, FL is improving medical image analysis by providing clinicians with better diagnostic tools, helping find similar patients for true precision medicine, and optimizing the drug discovery process by reducing costs and time-to-market for pharmaceutical companies.”

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

FL is improving medical image analysis by providing clinicians with better diagnostic tools.

“Already today, FL is improving medical image analysis by providing clinicians with better diagnostic tools, helping find similar patients for true precision medicine, and optimizing the drug discovery process by reducing costs and time-to-market for pharmaceutical companies.”

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

FL is still unable to deliver all its promises and meet the more stringent requirements (e.g., latency, security) of a healthcare system based on multiple IoMT.

“However, FL is still unable to deliver all its promises and meets the more stringent requirements (e.g., latency, security) of a healthcare system based on multiple Internet of Medical Things (IoMT).”

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

FL is helping find similar patients for true precision medicine.

“Already today, FL is improving medical image analysis by providing clinicians with better diagnostic tools, helping find similar patients for true precision medicine, and optimizing the drug discovery process by reducing costs and time-to-market for pharmaceutical companies.”

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

The current settings of FL have not answered all the technical questions.

“However, the current settings of FL have not answered all the technical questions”

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

Although no data are shared among the participants by definition in FL systems, some security risks are still present and can be considered as vulnerabilities from multiple aspects.

“For example, although no data are shared among the participants by definition in FL systems, some security risks are still present and can be considered as vulnerabilities from multiple aspects.”

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

Federated learning appeared as a valuable approach in the medical field, allowing patient data to stay where it is generated.

“To solve these issues, Federated Learning (FL) appeared as a valuable approach in the medical field, allowing patient data to stay where it is generated.”

• AssumptionRests on an assumption: Data-driven medical methods face scarcity and privacy of user data, as well as difficulty of institutions exchanging medical information.
• AssumptionRests on an assumption: With insufficient data, ML is prevented from reaching its full potential, which is only possible if the database consists of the full spectrum of possible anatomies, pathologies, and input data types.

May combine two claims.

Federated learning can protect sensitive medical data while collaboratively training a learning model.

“With the promise of powerful, accurate, safe, robust, and unbiased models, federated learning can neatly protect sensitive medical data while collaboratively training a learning model.”

• Also stated in: § VI. CONCLUSION · plain

  “Since an FL setting allows many clients to collaboratively train a model while keeping training data decentralized, it can protect privacy-sensitive medical data.”