A review of applications in federated learning

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

There have been two mainstream open-source frameworks for FL up to now: TensorFlow Federated (TFF) and Federated AI Technology Enabler (FATE).

“There have been two mainstream open-source frameworks for FL up to now and they are starting to take shape.”

• cites Google, 2019Not checked yet

  Not checked in the free run.

• cites Webanka, 2019Not checked yet

  Not checked in the free run.

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

The primitive framework of FL is FedAvg.

“The primitive framework of FL is FedAvg.”

• cites McMahan et al., 2017Not checked yet

  Not checked in the free run.

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

In FL, data is widespread in tens of thousands edge nodes or mobile devices, and available data in each node may be no more than the total number of nodes.

“In FL, data is widespread in tens of thousands edge node or mobile devices. Available data in each node may no more than the total number of nodes.”

May combine two claims.

Federated transfer learning is suitable when data shares neither sample space nor feature space, and the main problem is lack of data labels with poor data quality.

“Unlike the scenarios in horizontal FL and vertical FL, in most cases, data shares neither sample space nor feature space. Thus, the main problem in this setting is lack of data labels with poor data quality.”

May combine two claims.

FL is concentrated on unbalanced and non-IID data because of heterogeneity among device resources, in contrast to distributed systems which work primarily on balanced and IID data.

“In contrast with distributed system, which works primarily on balanced and IID data distribution, FL is concentrated on unbalanced and non-IID data because of the heterogeneity among device resources.”

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

Horizontal FL is suitable when data spread across nodes have overlapping features but differ in sample space.

“In the case of horizontal FL, there is a certain amount of overlap between the feature of data spread across various nodes, while the data are quite different in sample space.”

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

Federated transfer learning is an effective way to protect both data security and user privacy while breaking the barriers of data islands.

“However, federated transfer learning is an effective way to protect both data security and user privacy while breaking the barriers of data islands.”

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

FL is appropriate for application when data are privacy-sensitive because data are not exposed to a third central server.

“Due to the property that data are not exposed to third central server, FL is appropriate for application when data are privacy-sensitive.”

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

FL largely falls into three groups: horizontal FL, vertical FL, and federated transfer learning.

“Based on paper presented by Yang, Qb et al., (2019), FL largely falls into three groups, respectively, horizontal FL, vertical FL and federated transfer learning.”

• cites Yang, Qb et al., 2019Not checked yet

  Not checked in the free run.

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

Vertical FL is suitable when data is partitioned vertically according to feature dimension, with parties holding homogeneous data that partially overlap on sample ID but differ in feature space.

“Vertical FL is suitable for cases in which data is partitioned in the vertical direction according to feature dimension. All the parties hold homogeneous data which means they have partial overlap on sample ID whereas differ in feature space.”

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

FATE is the first open source industrial-level framework and primarily serves for cross-organizational architecture.

“As the first open source industrial-level framework, it primarily serves for cross-organizational architecture.”

• cites Webanka, 2019Not checked yet

  Not checked in the free run.

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

TFF is the first self-contained framework designed at production level mainly for mobile devices.

“It is the first self-contained framework designed at production level mainly for mobile devices.”

• cites Google, 2019Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Federated Learning (FL) is a collaboratively decentralized privacy-preserving technology to overcome challenges of data silos and data sensibility.

“Federated Learning (FL) is a collaboratively decentralized privacy-preserving technology to overcome challenges of data silos and data sensibility.”

• cites McMahan et al., 2017Not checked yet

  Not checked in the free run.

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

FedAvg is the baseline of FL in many other researches.

“This kind of federated training approach called federated average (FedAvg), which is the baseline of FL in many other researches.”

• cites McMahan et al., 2017Not checked yet

  Not checked in the free run.

May combine two claims.

FL is able to tackle both horizontally partitioned data according to samples and vertically partitioned data according to features in collaborative-learning setting.

“Therefore, FL is able to tackle not only horizontally partitioned data according to samples but also vertically partitioned data according to features in collaborative-learning setting.”

• cites Yang, Qb et al., 2019Not checked yet

  Not checked in the free run.

May combine two claims.

Data governance and data silo are two major challenges in data science development in industrial engineering.

“However, there are two major challenges in this area as data science development. Firstly, data governance is the most significant aspect.”

• cites EU, 2018Not checked yet

  Not checked in the free run.

May combine two claims.

FedAvg can deal with some lightweight Non-IID data but is still faced with challenges of high communication overhead and structural heterogeneity.

“Though it could deal with some lightweight Non-IID data. It is still faced with challenges of high communication overhead and structural heterogeneity.”

• cites McMahan et al., 2017Not checked yet

  Not checked in the free run.

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

FL was first put forward by Google in 2016 to predict user's text input within tens of thousands Android devices while keeping data on devices.

“It was first put forward by Google in 2016 to predict user’s text input within tens of thousands Android devices while keeping data on devices (McMahan et al., 2017).”

• cites McMahan et al., 2017Not checked yet

  Not checked in the free run.

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

FL is a decentralized technology that enables scattered clients or organizations to train a collaborative model autonomously while keeping data localized.

“To sum up, FL is a decentralized technology that enable scattered clients or organizations to train a collaborative model autonomously, while keeping data localized.”

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

These studies fill a part of the theoretical gap in the research of convergence speed of a FL algorithm.

“These studies fill a part of the theoretical gap in the research of convergence speed of a FL algorithm.”

May combine two claims.

Skew characteristic and clinical validation may vary among different equipment version.

“Skew characteristic and clinical validation may vary among different equipment version (Godinho et al., 2016).”

• cites Godinho et al., 2016 sourceNot checked yet

  Not checked in the free run.

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

Wu et al. (2019) made use of a cache structure to store unreliable user update thus alleviates their trustless impact on global model.

“Wu et al., (2019) also fully considered device straggling phenomenon in heterogeneous network. They made use of a cache structure to store those unreliable user update thus alleviates their trustless impact on global model.”

• cites Wu et al., (2019)Not checked yet

  Not checked in the free run.

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

Wang et al. (2019) brought forward an improved adaptive method to reduce loss function within constraints of resource budget.

“and further bring forward an improved adaptive method to reduce loss function within constraints of resource budget.”

• cites Wang et al., (2019)Not checked yet

  Not checked in the free run.

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

Above-mentioned algorithms are all based on stochastic gradient descent, but this method could be inefficient if the function is anisotropic.

“The above-mentioned algorithms all based on stochastic gradient descent (SGD), but this method could be inefficient if the function is anisotropic.”

May combine two claims.

Methods to deal with structural heterogeneity mainly focus on resource allocation for heterogeneous devices and fault tolerance for devices prone to be offline.

“Up to now, methods to deal with structural heterogeneity mainly focus on resource allocation for heterogeneous devices and fault tolerance for devices prone to be offline.”

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

Unreliable and unstable network may lead to devices' drop out.

“The unreliable and unstable network may lead to devices’ drop out.”

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

MOCHA introduced multi-task learning to make utilization of shared representation.

“Another idea is to optimize modeling way to achieve personalization for individual devices such as MOCHA, which introduced multi-task learning to make utilization of shared representation (Smith et al., 2017).”

• cites Smith et al., 2017Not checked yet

  Not checked in the free run.

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

Mohri et al. improved global model to cope with any target distribution comprised by a mixture of different clients.

“They improved global model to cope with any target distribution comprised by a mixture of different clients.”

• cites Mohri et al., (2019)Not checked yet

  Not checked in the free run.

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

By separating independent data into different clusters, then processing federated training on each community, the non-IID problem can be resolved.

“By separating independent data into different clusters, then processing federated training on each community, the non-IID problem is thus can be resolved.”

• cites Huanga et al., (2019)Not checked yet

  Not checked in the free run.

May combine two claims.

Li, T. et al. (2020) designed a fairness metrics to measure loss in devices and a q-Fair optimization goal to impel fair resource allocation in FL.

“To address this issue, Li, T. et al., (2020) designed a fairness metrics to measure loss in devices and a q-Fair optimization goal to impel fair resource allocation in FL.”

• cites Li, T. et al., (2020)Not checked yet

  Not checked in the free run.

May combine two claims.

Huanga et al. (2019) introduced clustering thought with FL and constructed a community-based FL method.

“For data pre-processing, Huanga et al., (2019) introduced clustering thought with FL and constructed a community-based FL method.”

• cites Huanga et al., (2019)Not checked yet

  Not checked in the free run.

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

Increasing parallelism by motivating more clients to join training each round is an effective way to improve communication efficiency.

“They also pointed out that increase parallelism which means motivate more clients to join training on each round is an effective way.”

• cites McMahan et al., 2017Not checked yet

  Not checked in the free run.

May combine two claims.

High communication cost, statistical heterogeneity, and structural heterogeneity are major issues faced by researchers in FL optimization.

“In terms of optimization for grounding application, high communication cost, statistical and structural heterogeneity are major issues faced by researchers currently (Li et al., 2020).”

• cites Li et al., 2020Not checked yet

  Not checked in the free run.

May combine two claims.

Yang, Qb. et al. (2019) studied a novel and effective distributed backdoor attack, dividing an attack trigger into many slices and embedding each slice into different attackers.

“On the basis research of Bagdasaryan et al., (2018), Yang, Qb. et al., (2019) studied a novel and effective distributed backdoor attack. They divided an attack trigger into many slices and embedded each slice into different attackers instead of embedding a complete trigger into only one attacker.”

• cites Yang, Qb. et al., (2019)Not checked yet

  Not checked in the free run.

May combine two claims.

Bayesian Nonparametric FL framework can aggregate local models into a federated model without extra parameters, thus avoiding unwanted communication rounds, and can obtain satisfactory accuracy with only one communication round.

“Yurochkin et al., (2019) designed Bayesian Nonparametric FL framework, which is state of the art since it can aggregate local models into a federated model without extra parameters thus avoid unwanted communication rounds. The experiment shows that they can obtain satisfactory accuracy rating with only one communication round.”

• cites Yurochkin et al., 2019Not checked yet

  Not checked in the free run.

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

In hierarchical heterogeneous horizontal framework, projecting each embedding submanifold into a common embedding space overcomes data heterogeneity.

“In hierarchical heterogeneous horizontal framework, it projects each embedding submanifold into a common embedding space to overcome data heterogeneity (Gao et al., 2019).”

• cites Gao et al., 2019Not checked yet

  Not checked in the free run.

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

Mohri et al. (2019) noticed previous work ignore the importance of fairness which may lead to bias centralized model.

“In addition, Mohri et al., (2019) noticed previous work ignore the importance of fairness which may lead to bias centralized model.”

• cites Mohri et al., (2019)Not checked yet

  Not checked in the free run.

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

Huangb et al. (2019) put cross-entropy loss into transmission process and assign different local update times for each client in each round to overcome Non-IID problem.

“Huangb et al., (2019) also gained a good deal of enlightenment from the previous data sharing ideology to overcome Non-IID problem. They put cross-entropy loss into transmission process and assign different local update times for each client in each round.”

• cites Huangb et al., (2019)Not checked yet

  Not checked in the free run.

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

Stragglers and heterogeneity run through FL framework.

“In a nutshell, stragglers and heterogeneity run through FL framework.”

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

Huanga et al.'s community-based FL method is not suitable for massively data training due to high parameter conversion overhead.

“However, one drawback is that it’s not suitable for massively data training due to high parameter conversion overhead.”

May combine two claims.

Huangb et al. (2019) research concerned computation complexity and communication cost as well as accuracy for each client and therefore outperformed baselines.

“This research concerned computation complexity and communication cost as well as accuracy for each client therefore outperform baselines.”

• cites Huangb et al., (2019)Not checked yet

  Not checked in the free run.

May combine two claims.

Xu et al. (2020) devised VerifyNet, the first protocol that can verify correctness of returned model from cloud.

“Thereafter Xu et al., (2020) devised VerifyNet, the first protocol that can verify correctness of returned model from cloud.”

• cites Xu et al., (2020)Not checked yet

  Not checked in the free run.

May combine two claims.

Each medical institute might have a lot of patient data, but that may be far from enough to train their own prediction models.

“Each medical institute might have a lot of patient data, but that may be far from enough to train their own prediction models (Szegedi, Kiss and Horváth,2019).”

• cites Szegedi, Kiss and Horváth,2019Not checked yet

  Not checked in the free run.

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

Combination of FL and disease prediction is one of the good solutions to break down the barriers of analysis throughout different hospitals.

“Combination of FL and disease prediction is one of the good solutions to break down the barriers of analysis throughout different hospitals.”

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

FL has been paid much attention since concept first put forward by Google to predict users' input from Gboard on Android devices.

“FL has been paid much attention to by the researchers since the concept was first put forward by Google to predict users’ input from Gboard on Android devices.”

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

Existing federated training mostly bases on a small set of organizations and is not able to extend to collaborative training for a huge number of devices or institutions.

“Moreover, one problem is existing federal training mostly base on small set of organizations and is not able to extend to collaborative training for huge number of devices or institutions.”

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

LSH would not cause overmuch communication overhead like homomorphic encryption and reduce accuracy like differential privacy.

“Besides, LSH would not cause overmuch communication overhead like homomorphic encryption and reduce accuracy like differential privacy.”

May combine two claims.

Yu et al. (2020) build a federated multi-task learning framework for smart home IOT to automatically learn users' behavior patterns, which could effectively detect physical hazards.

“Similarly, Yu et al., (2020) build a federated multi-task learning framework for smart home IOT to automatically learn users’ behavior patterns, which could effectively detect physical hazards.”

• cites Yu et al., (2020)Not checked yet

  Not checked in the free run.

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

Training on cryptographic models may raise additional communication overhead since more data such as private key should be conveyed.

“Training on these cryptographic models may raise additional communication overhead since more data such as private key should be conveyed.”

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

Haoa et al. (2019) envisioned a more efficient privacy-preserving scheme for FL, integrating differential privacy and lightweight homomorphic encryption technology.

“By this way, Haoa et al., (2019) envisioned a more efficient privacy-preserving scheme for FL, which integrate differential privacy and lightweight homomorphic encryption technology.”

• cites Haoa et al., (2019)Not checked yet

  Not checked in the free run.

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

Currently, FL mainly contributes to horizontally collaborative training for landing applications, which means feature dimensions of each data are similar to each other.

“Currently, FL mainly contributes to horizontally collaborative training for landing applications, which means feature dimensions of each data are similar to each other.”

May combine two claims.

Pfohl et al. (2019) explored differentially private learning for EMR in federated setting and demonstrated performance is comparable with training in a centralized setting.

“Pfohl et al., (2019) explored differentially private learning for EMR in federated setting. And they further demonstrated the performance is comparable with training in a centralized setting.”

• cites Pfohl et al., (2019)Not checked yet

  Not checked in the free run.

May combine two claims.

Hybrid-One scheme combine use of DP with MPC without compromising accuracy rate, which protect communication messages rely on MPC thus introduce less noise than traditional local DP.

“To reduce noise, the Hybrid-One scheme combine the use of DP with MPC without compromising accuracy rate, which protect communication messages rely on MPC thus introduce less noise than traditional local DP (Truex et al.,2019).”

• cites Truex et al.,2019Not checked yet

  Not checked in the free run.

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

FL enable data owner to broaden the scope of data applications and improve model performance through iteration among different entities.

“To summarize, FL enable data owner to broaden the scope of data applications and improve model performance through iteration among different entities.”

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

FL is widely used in the area of biomedical imaging analysis.

“Recently, FL is also widely used in the area of biomedical imaging analysis.”

May combine two claims.

Wang, Y. et al. (2020) exploit an industrial grade federated framework based on Latent Dirichlet Allocation. It has passed assessment on real data for spam filtering and sentiment analysis.

“For text mining, Wang, Y. et al., (2020) exploit an industrial grade federated framework based on Latent Dirichlet Allocation. It has passed the assessment on real data for spam filtering and sentiment analysis.”

• cites Wang, Y. et al., (2020)Not checked yet

  Not checked in the free run.

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

Most privacy-preserving methods in FL rely on a strong assumption that clients are semi-honest, obeying training rules but keeping curious about private data acquisition.

“Most privacy-preserving method in FL rely on a strong assumption that clients are semi-honest which obey training rules but keeping curious about private data acquisition.”

May combine two claims.

Xiao et al. (2020) detected that mutual information is increased, implying correlation between clients, while the distance of parameters is getting greater with increased iteration; this study shows gradient averaging is possibly not a good manner for gradient aggregation.

“Xiao et al., (2020) detect that the mutual information is increased which implies correlation between clients, while the distance of parameters is getting greater with increased iteration. This study shows gradient averaging is possible not a good manner for gradient aggregation.”

• cites Xiao et al., (2020)Not checked yet

  Not checked in the free run.

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

Security guaranteed by blockchain technology has nothing to do with privacy protection; there is no privacy concern in allusion to individual node.

“Although the security of entire architecture can be guaranteed with block chain technology, this security has nothing to do with privacy protection. There is no privacy concern in allusion to individual node.”

• AssumptionRests on an assumption: The security of the entire architecture can be guaranteed with blockchain technology.

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

Stackelberg-based game theory frameworks enjoy wide popularity in motivation mechanism design.

“The frameworks such as Stackelberg-based game theory enjoy wide popularity in motivation mechanism design.”

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

Introducing blockchain may cause more latency when exchanging learning model.

“But introducing blockchain may cause more latency when exchange learning model.”

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

Some lazy clients with high quality or some selfish clients afraid of power consumption may not attend the whole training process with a certain probability.

“In fact, some lazy clients with high quality or some selfish clients afraid of power consumption may not attend the whole training process with a certain probability.”

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

Aforementioned FL frames based on blockchain did not give a specific rewarding mechanism for clients to join training.

“Aforementioned FL frames based on block chain didn’t give specific rewarding mechanism for clients to join training.”

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

The crowdsourcing framework adopted a two-stage Stackelberg model to acquire utility maximization among clients and server.

“The crowdsourcing framework adopted two-stage stackelberg model to acquire utility maximization among clients and server (Pandey et al., 2019).”

• cites Pandey et al., 2019Not checked yet

  Not checked in the free run.

May combine two claims.

FL has been effectively leveraged in neural network, SVM, and linear classifier.

“For instance, FL have been effectively leveraged in neural network (Wang, S. et al., 2019; Haob et al., 2019; Bonawitz et al., 2019) and SVM (Liu et al., 2019), as well as linear classifier (Hardy et al., 2017).”

• cites Wang, S. et al., 2019; Haob et al., 2019; Bonawitz et al., 2019Not checked yet

  Not checked in the free run.

• cites Liu et al., 2019Not checked yet

  Not checked in the free run.

• cites Hardy et al., 2017Not checked yet

  Not checked in the free run.

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

If a malicious clinic or hospital joins collaborative training, it may snoop other participants' privacy information.

“If there is a malicious clinic or hospital join in the collaborative training, it may spare no effort to snoop other participants’ privacy information.”

May combine two claims.

Ilias and Georgios (2019) utilized blockchain smart convention to coordinate all clients and additionally used homomorphic encryption to provide extra privacy guarantee.

“Hence Ilias and Georgios (2019) utilized blockchain smart convention to coordinate all clients and additionally used homomorphic encryption to provide extra privacy guarantee.”

• cites Ilias and Georgios (2019)Not checked yet

  Not checked in the free run.

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

Traditional Byzantine fault-tolerant systems are supported by defense mechanisms rather than malicious client detection.

“Traditional Byzantine fault-tolerant system is supported by some defense mechanism rather than malicious client detection.”

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

Muñoz-González and Lupu (2019) discussed adaptive FL to grub abnormal updates via a Hidden Markov Model to evaluate model quality.

“Muñoz-González and Lupu (2019) discussed adaptive FL to grub abnormal updates via a Hidden Markov Model to evaluate model quality.”

• cites Muñoz-González and Lupu (2019)Not checked yet

  Not checked in the free run.

May combine two claims.

Several blockchain-assisted schemes for industrial data sharing have been proposed, including quality surveillance of 3D-printed articles, consumption monitoring and privacy-preserving energy trading for smart grids, and emergency medical service for pre-hospital care.

“Several blockchain-assisted schemes for industrial data sharing have been proposed, including quality surveillance of 3D-printed articles (Kennedy et al., 2017), consumption monitoring and privacy-preserving energy trading for smart grids (Aitzhan and Svetinovic, 2018) and emergency medical service for pre-hospital care (Hasavari and Song 2019).”

• cites Kennedy et al., 2017Not checked yet

  Not checked in the free run.

• cites Aitzhan and Svetinovic, 2018Not checked yet

  Not checked in the free run.

• cites Hasavari and Song 2019Not checked yet

  Not checked in the free run.

May combine two claims.

Using blockchain, the contribution of each party to optimize the global model could be traced, making an incentive mechanism possible.

“Furthermore, take advantage of block chain, the contribution of each party to optimize global model could be traced, which make it possible for an incentive mechanism.”

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

This kind of Byzantine problem is always encountered in FL.

“This kind of Byzantine problem is always encountered in FL.”

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

Wang et al. (2019) designed a layer-wise aggregation mode to serially generate layer parameters in a neural network for the global model.

“In addition, Wang et al., (2019) designed a layer-wise aggregation mode to serially generate layer parameters in neural network for global model.”

• cites Wang et al., (2019)Not checked yet

  Not checked in the free run.

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

Blockchain and FL are auxiliary to each other.

“Blockchain and FL are auxiliary to each other.”

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

Existing studies based on blockchain mainly focus on innovative medical information sharing systems, but training collaboratively to maximize data utilization has not been implemented.

“Existing studies based on block chain mainly focus on innovate medical information sharing system but training collaboratively to maximize data utilization has not been implemented.”

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

Recent research has proved that blockchain has potential to significantly transform some issues in FL.

“Recent research has proofed that block chain has potential to significantly transform some issues in FL.”

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

Majeed and Hong (2019) envisioned a robust FL-chain that could verify local model update by injecting blockchain technology into FL.

“Injecting block-chain technology into FL, Majeed and Hong (2019) envisioned a robust Fl-chain that could verify local model update.”

• cites Majeed and Hong (2019)Not checked yet

  Not checked in the free run.

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

Considering Non-iid distribution on clients, gradient aggregation in a simply averaging way is not a good choice.

“Considering Non-iid distribution on clients, gradient aggregation in a simply averaging way isn’t a good choice.”

May combine two claims.

Li et al. (2019) considered an autoencoder enabling model parameters to be replaced by a low-dimension vector and to discover irregular weight updates to detect anomalous model update.

“To detect this anomalous model update, Li et al., (2019) considered an autoencoder enable model parameters to be replaced by low-dimension vector as well as discover irregular weights update.”

• cites Li et al., (2019)Not checked yet

  Not checked in the free run.

May combine two claims.

Yao et al. (2019) keep trace of dispatched global parameters in each local training epoch, enabling gradients to be aggregated in an unbiased way.

“To eliminate gradient bias in local training stage, Yao et al., (2019) keep trace of dispatched global parameters in each local training epoch. Since local gradient update is a function of global parameters, then gradients can be aggregated in an unbiased way.”

• cites Yao et al., (2019)Not checked yet

  Not checked in the free run.

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

Chen, Sun and Jin (2019) designed an asynchronous approach for client model update, dividing deep neural network layers into deep layers and shallow layers with different update frequency.

“At the client side, Chen, Sun and Jin (2019) designed an asynchronous approach for client model update. Layers in deep neural network are divided into deep layers and shallow layers with different update frequency.”

• cites Chen, Sun and Jin (2019)Not checked yet

  Not checked in the free run.

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

Blockchain is essentially a distributed ledger derived from Bitcoin, characterized by decentralization, immutability, traceability, collective maintenance, openness and transparency.

“In short, Block-chain is essentially a distributed ledger, derived from Bitcoin (Nakamoto. S, 2008), which is characterized by decentralization, immutability, traceability, collective maintenance, openness and transparency.”

• cites Nakamoto. S, 2008Not checked yet

  Not checked in the free run.

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

Incorporating blockchain with FL is auspicious because it is a decentralized technology and thus does not need a central server to predict the global model anymore; therefore it could overcome the limitation of bandwidth in FL.

“To sum up, incorporate block chain with FL is auspicious since it is a decentralized technology thus doesn’t need central server to predict global model anymore. Therefore, it could overcome the limitation of bandwidth in FL.”

May combine two claims.

Wu et al. (2019) proposed a semi-asynchronous protocol allowing straggling clients not always to go together with the central server, to speed up training by utilizing their slowly updated models.

“Considering trade-off between advantages of synchronous update and asynchronous training, Wu et al., (2019) proposed a semi-asynchronous protocol which allow straggling clients don’t always go together with central server. The main idea is that make stragglers join training properly to speed up training process with utilization of their slowly update model.”

• cites Wu et al., (2019)Not checked yet

  Not checked in the free run.

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

Khan et al. (2019) discussed Stackelberg-based incentive mechanism to set local iteration times adaptively to be effective as much as possible.

“Khan et al., (2019) discussed Stackelberg-based incentive mechanism to set local iteration times adaptively to be effective as much as possible.”

• cites Khan et al., (2019)Not checked yet

  Not checked in the free run.

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

Leroy et al. (2019) learned an embedded wake word detector using an adaptive averaging strategy in place of standard weighted model averaging.

“Leroy et al. 2019 Smart phone voice assistant Learn embedded wake word detector using an adaptive averaging strategy in place of standard weighted model averaging Do not show robustness to back- ground noise”

• cites Leroy et al. 2019Not checked yet

  Not checked in the free run.

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

Sozinov et al. (2018) applied Human Activity Recognition, identifying and rejecting erroneous clients.

“Sozinov et al. 2018 Smart devices motion sensors Human Activity Recognition identifies and rejects erroneous clients producing models with slightly worse, accuracy compared to centralized models”

• cites Sozinov et al. 2018Not checked yet

  Not checked in the free run.

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

Feng et al. (2020) used a group optimization strategy for privacy-preserving human mobility prediction.

“Feng et al.2020 Mobile devices motion sensors Privacy-preserving Human Mobility Prediction Using group optimization strategy, reduce the performance degradation only consider the basic mobility model for the simplicity”

• cites Feng et al.2020Not checked yet

  Not checked in the free run.

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

In most cases where labeled data either do not exist or exist little, unsupervised learning methods are supposed to be applied.

“Actually, in most of cases where labeled data either do not exist, or with little existence, unsupervised learning methods are supposed to be applied.”

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

van Berlo et al. (2020) introduced Federated Unsupervised Representation Learning, which is a breakthrough in unsupervised FL.

“Recently, van Berlo et al., (2020) introduced Federated Unsupervised Representation Learning which is a breakthrough in unsupervised FL.”

• cites van Berlo et al., (2020)Not checked yet

  Not checked in the free run.

May combine two claims.

Unsupervised learning has been widely used for image registration and image classification.

“For example, it has been widely used for image registration (Dalca, Balakrishnan, Guttag and Sabuncu, 2019; de Vos et al., 2019) and image classification (Ahn et al., 2019).”

• cites Dalca, Balakrishnan, Guttag and Sabuncu, 2019Not checked yet

  Not checked in the free run.

• cites de Vos et al., 2019 sourceNot checked yet

  Not checked in the free run.

• cites Ahn et al., 2019Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing).

How to distribute the huge computation load on heterogeneous scenarios is still unexplored.

“Wang et al. (2019) Mobile edge computing optimizing MEC, caching and communication discussed the potential of inte-grating the Deep Reinforcement Learning and FL framework with the mobile edge system how to distribute the huge computation load on heterogenous scenarios are still unexplored.”

• cites Wang et al. (2019)Not checked yet

  Not checked in the free run.

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

Aïvodji et al. (2019) designed a sufficiently secure federated smart home setting.

“Aïvodji et al. 2019 Smart home IOT Design a sufficient secure federated smart home setting combines FL with secure data aggregation rather complex architecture to implement”

• cites Aïvodji et al. 2019Not checked yet

  Not checked in the free run.

May combine two claims.

Collaborative and Adversarial Network (CAN), a novel unsupervised domain adaptation approach, shows effectiveness and high performance.

“Such method like Collaborative and Adversarial Network (CAN), a novel unsupervised domain adaptation approach, shows effectiveness and high performance (Zhang et al., 2018).”

• cites Zhang et al., 2018Not checked yet

  Not checked in the free run.

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

Although researchers have made great progress on federated transfer learning to handle dispersive data with few labels, landing applications remain a bottleneck for unsupervised learning in the federated framework.

“Although researchers have made great progress on federated transfer learning to handle with dispersive data with few labels, the landing applications remain a bottleneck for unsupervised learning in federated framework.”

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

Qian et al. (2019)'s scheme is not able to be used for several edge clouds.

“Qian et al. 2019 Mobile edge computing Privacy-aware service placement for mobile edge computing propose a privacy-aware ser- vice placement (PSP) scheme to meet users’ service demands Not able to be used for several edge clouds”

• cites Qian et al. 2019Not checked yet

  Not checked in the free run.

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

Qian et al. (2019) proposed a privacy-aware service placement (PSP) scheme to meet users' service demands for mobile edge computing.

“Qian et al. 2019 Mobile edge computing Privacy-aware service placement for mobile edge computing propose a privacy-aware ser- vice placement (PSP) scheme to meet users’ service demands Not able to be used for several edge clouds”

• cites Qian et al. 2019Not checked yet

  Not checked in the free run.

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

Ramaswamy et al. (2019) predicted emoji from text typed on a keyboard.

“Ramaswamy et al. 2019 Smart phone keyboard predict emoji from text typed on a keyboard achieve better performance than a server trained model client cache contents are different, and metrics cannot be compared across experiments”

• cites Ramaswamy et al. 2019Not checked yet

  Not checked in the free run.

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

Wang et al. (2019) discussed the potential of integrating Deep Reinforcement Learning and FL framework with the mobile edge system for optimizing MEC, caching and communication.

“Wang et al. (2019) Mobile edge computing optimizing MEC, caching and communication discussed the potential of inte-grating the Deep Reinforcement Learning and FL framework with the mobile edge system how to distribute the huge computation load on heterogenous scenarios are still unexplored.”

• cites Wang et al. (2019)Not checked yet

  Not checked in the free run.

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

Hard et al. (2019) achieved recall improvements.

“Hard et al.2019 Smart phone keyboard next-word prediction in a virtual keyboard train an RNN model from scratch in the server and federated environments and achieve recall improvements Still have high communication cost”

• cites Hard et al.2019Not checked yet

  Not checked in the free run.

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

In Ramaswamy et al. (2019), client cache contents are different, and metrics cannot be compared across experiments.

“Ramaswamy et al. 2019 Smart phone keyboard predict emoji from text typed on a keyboard achieve better performance than a server trained model client cache contents are different, and metrics cannot be compared across experiments”

• cites Ramaswamy et al. 2019Not checked yet

  Not checked in the free run.

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

Unsupervised learning is reasonable to be used to infer potential information in messy data.

“Thus, unsupervised learning is reasonable to be used to infer potential information in these messy data.”

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

Chen et al. (2019)'s approach strongly relies on a learned probabilistic model.

“Chen, Mathews, Ouyang and Beaufays. 2019 Smart phone keyboard learn out- of-vocabulary words expanding the vocabulary of a keyboard without exporting sensitive text Strongly relies on a learned probabilistic mode”

• cites Chen, Mathews, Ouyang and Beaufays. 2019Not checked yet

  Not checked in the free run.

The quote only partly states this claim.

van Berlo et al. (2020) study shows competitive performance compared with supervised learning and transfer learning.

“This study also shows competitive performance compared with supervised learning and transfer learning.”

• cites van Berlo et al., (2020)Not checked yet

  Not checked in the free run.

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

Feng et al. (2020) only considered the basic mobility model for simplicity.

“Feng et al.2020 Mobile devices motion sensors Privacy-preserving Human Mobility Prediction Using group optimization strategy, reduce the performance degradation only consider the basic mobility model for the simplicity”

• cites Feng et al.2020Not checked yet

  Not checked in the free run.

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

Sozinov et al. (2018) produced models with slightly worse accuracy compared to centralized models.

“Sozinov et al. 2018 Smart devices motion sensors Human Activity Recognition identifies and rejects erroneous clients producing models with slightly worse, accuracy compared to centralized models”

• cites Sozinov et al. 2018Not checked yet

  Not checked in the free run.

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

As a kind of collaborative training approach, FL has great potential in the unsupervised learning area.

“Therefore, as a kind of collaborative training approach, FL has great potential on unsupervised learning area.”

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

Chen et al. (2019)'s approach expands the vocabulary of a keyboard without exporting sensitive text.

“Chen, Mathews, Ouyang and Beaufays. 2019 Smart phone keyboard learn out- of-vocabulary words expanding the vocabulary of a keyboard without exporting sensitive text Strongly relies on a learned probabilistic mode”

• cites Chen, Mathews, Ouyang and Beaufays. 2019Not checked yet

  Not checked in the free run.

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

Ramaswamy et al. (2019) achieved better performance than a server-trained model.

“Ramaswamy et al. 2019 Smart phone keyboard predict emoji from text typed on a keyboard achieve better performance than a server trained model client cache contents are different, and metrics cannot be compared across experiments”

• cites Ramaswamy et al. 2019Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

Feng et al. (2020) reduced performance degradation.

“Feng et al.2020 Mobile devices motion sensors Privacy-preserving Human Mobility Prediction Using group optimization strategy, reduce the performance degradation only consider the basic mobility model for the simplicity”

• cites Feng et al.2020Not checked yet

  Not checked in the free run.

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

Yang et al. (2019) should take more reliable measurements into account to protect privacy.

“Yang et al. 2019 Financial field detect credit card fraud achieves an average of test AUC 10% higher than traditional method. Should take more reliable measurements into account to protect the privacy”

• cites Yang et al. 2019Not checked yet

  Not checked in the free run.

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

In Wang et al. (2020), noise from the perturbing mechanism will slightly influence the overall performance.

“Wang et al. 2020 text mining spam filtering and sentiment analysis Using Random Response with Priori (RRP), which provides theoretical guarantees on both data privacy and model accuracy. noise from our perturbing mechanism will slightly influence the overall performance”

• cites Wang et al. 2020Not checked yet

  Not checked in the free run.

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

Brisimi et al. (2018) yielded classifiers using relatively few features.

“Brisimi et al. 2018 Predict future hospitalizations for patients Cluster Primal Dual Splitting algorithm Yield classifiers using relatively few features Need more iterations for convergence”

• cites Brisimi et al. 2018 sourceNot checked yet

  Not checked in the free run.

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

Han et al. (2019) provided manufactures with service in automated defect inspection.

“Han et al. 2019 Image detection provide manufactures with the service in automated defect inspection solve the problem of lacking defective samples to detect defects need quick model deployment to serve various industries”

• cites Han et al. 2019Not checked yet

  Not checked in the free run.

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

Yu et al. (2020) effectively detect physical hazards.

“Yu et al. 2020 Smart home IOT learn users’ behavior patterns effectively detect physical hazards Not flexible with mapping mechanism for diverse deployment”

• cites Yu et al. 2020Not checked yet

  Not checked in the free run.

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

Liu et al. (2019) is not suitable for small questionable cases.

“Liu et al.2019 Extraction of clinical notes Two-stage federated NLP method Adding pre-processing step to improve accuracy Not suitable for small questionable cases”

• cites Liu et al.2019Not checked yet

  Not checked in the free run.

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

Liu et al. (2020) is more beneficial for the smaller dataset than the larger one in horizontal FL.

“Liu et al. 2020 Image representation obtain various types of image representations from different tasks Be validated on three kinds of FL settings more beneficial for the smaller dataset than the larger one in horizontal FL”

• cites Liu et al. 2020Not checked yet

  Not checked in the free run.

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

Wang et al. (2020)'s RRP provides theoretical guarantees on both data privacy and model accuracy.

“Wang et al. 2020 text mining spam filtering and sentiment analysis Using Random Response with Priori (RRP), which provides theoretical guarantees on both data privacy and model accuracy. noise from our perturbing mechanism will slightly influence the overall performance”

• cites Wang et al. 2020Not checked yet

  Not checked in the free run.

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

Saputra et al. (2019) further improved prediction accuracy.

“Saputra et al. 2019 Electrical vechicles federated energy demand prediction applied the clustering-based energy demand learning method for to further improve the prediction accuracy Need to be more stable and flexible.”

• cites Saputra et al. 2019Not checked yet

  Not checked in the free run.

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

Liu et al. (2019) used a two-stage federated NLP method for extraction of clinical notes.

“Liu et al.2019 Extraction of clinical notes Two-stage federated NLP method Adding pre-processing step to improve accuracy Not suitable for small questionable cases”

• cites Liu et al.2019Not checked yet

  Not checked in the free run.

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

Gao et al. (2019) designed a hierarchical heterogeneous horizontal FL framework for EEG classification.

“Gao et al.2019 EEG Classification Design a hierarchical heterogeneous horizontal FL framework The first EEG classifier over heterogeneous EEG data Only work on 3 different datasets”

• cites Gao et al.2019Not checked yet

  Not checked in the free run.

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

Yang et al. (2019) achieved an average of test AUC 10% higher than traditional method.

“Yang et al. 2019 Financial field detect credit card fraud achieves an average of test AUC 10% higher than traditional method. Should take more reliable measurements into account to protect the privacy”

• cites Yang et al. 2019Not checked yet

  Not checked in the free run.

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

Yu et al. (2020)'s approach is not flexible with mapping mechanism for diverse deployment.

“Yu et al. 2020 Smart home IOT learn users’ behavior patterns effectively detect physical hazards Not flexible with mapping mechanism for diverse deployment”

• cites Yu et al. 2020Not checked yet

  Not checked in the free run.

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

Liu et al. (2020) was validated on three kinds of FL settings.

“Liu et al. 2020 Image representation obtain various types of image representations from different tasks Be validated on three kinds of FL settings more beneficial for the smaller dataset than the larger one in horizontal FL”

• cites Liu et al. 2020Not checked yet

  Not checked in the free run.

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

Han et al. (2019) solved the problem of lacking defective samples to detect defects.

“Han et al. 2019 Image detection provide manufactures with the service in automated defect inspection solve the problem of lacking defective samples to detect defects need quick model deployment to serve various industries”

• cites Han et al. 2019Not checked yet

  Not checked in the free run.

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

Silva et al. (2019) was only tested in limited dataset.

“Silva et al.2019 MRI Analysis Provide federated analysis framework compatible with the standard ENIGMA pipelines Deal with variability of high dimensional features efficiently Only tested in limited dataset”

• cites Silva et al.2019Not checked yet

  Not checked in the free run.

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

Hu et al. (2018) proposed an environmental monitoring frame based on federated region learning.

“Hu et al. 2018 environment protection environmental monitoring frame based on federated region learning considers the regional characteristics during the distribution of training samples to improve the inference accuracy Need to be extended to multi-layer structures instead of two-layer structure”

• cites Hu et al. 2018Not checked yet

  Not checked in the free run.

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

Hu et al. (2018)'s frame considers regional characteristics during distribution of training samples to improve inference accuracy.

“Hu et al. 2018 environment protection environmental monitoring frame based on federated region learning considers the regional characteristics during the distribution of training samples to improve the inference accuracy Need to be extended to multi-layer structures instead of two-layer structure”

• cites Hu et al. 2018Not checked yet

  Not checked in the free run.

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

Liu et al. (2020) obtained various types of image representations from different tasks.

“Liu et al. 2020 Image representation obtain various types of image representations from different tasks Be validated on three kinds of FL settings more beneficial for the smaller dataset than the larger one in horizontal FL”

• cites Liu et al. 2020Not checked yet

  Not checked in the free run.

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

Brisimi et al. (2018) need more iterations for convergence.

“Brisimi et al. 2018 Predict future hospitalizations for patients Cluster Primal Dual Splitting algorithm Yield classifiers using relatively few features Need more iterations for convergence”

• cites Brisimi et al. 2018 sourceNot checked yet

  Not checked in the free run.

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

Aïvodji et al. (2019)'s architecture is rather complex to implement.

“Aïvodji et al. 2019 Smart home IOT Design a sufficient secure federated smart home setting combines FL with secure data aggregation rather complex architecture to implement”

• cites Aïvodji et al. 2019Not checked yet

  Not checked in the free run.

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

Brisimi et al. (2018) used Cluster Primal Dual Splitting algorithm to predict future hospitalizations for patients.

“Brisimi et al. 2018 Predict future hospitalizations for patients Cluster Primal Dual Splitting algorithm Yield classifiers using relatively few features Need more iterations for convergence”

• cites Brisimi et al. 2018 sourceNot checked yet

  Not checked in the free run.

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

Liu et al. (2019) added pre-processing step to improve accuracy.

“Liu et al.2019 Extraction of clinical notes Two-stage federated NLP method Adding pre-processing step to improve accuracy Not suitable for small questionable cases”

• cites Liu et al.2019Not checked yet

  Not checked in the free run.

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

Yang et al. (2019) detected credit card fraud.

“Yang et al. 2019 Financial field detect credit card fraud achieves an average of test AUC 10% higher than traditional method. Should take more reliable measurements into account to protect the privacy”

• cites Yang et al. 2019Not checked yet

  Not checked in the free run.

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

Yu et al. (2020) learned users' behavior patterns.

“Yu et al. 2020 Smart home IOT learn users’ behavior patterns effectively detect physical hazards Not flexible with mapping mechanism for diverse deployment”

• cites Yu et al. 2020Not checked yet

  Not checked in the free run.

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

Silva et al. (2019) dealt with variability of high dimensional features efficiently.

“Silva et al.2019 MRI Analysis Provide federated analysis framework compatible with the standard ENIGMA pipelines Deal with variability of high dimensional features efficiently Only tested in limited dataset”

• cites Silva et al.2019Not checked yet

  Not checked in the free run.

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

Wang et al. (2020) used Random Response with Priori (RRP) for spam filtering and sentiment analysis.

“Wang et al. 2020 text mining spam filtering and sentiment analysis Using Random Response with Priori (RRP), which provides theoretical guarantees on both data privacy and model accuracy. noise from our perturbing mechanism will slightly influence the overall performance”

• cites Wang et al. 2020Not checked yet

  Not checked in the free run.

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

Liu et al. (2020) used imitation learning to increase imitation learning efficiency of local robots in cloud robotic systems.

“Liu et al. 2020 Robot network robots imitation learning increases imitation learning efficiency of local robots in cloud robotic systems Need to further work on convergence justification of the fusion process”

• cites Liu et al. 2020Not 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.

Saputra et al. (2019) need to be more stable and flexible.

“Saputra et al. 2019 Electrical vechicles federated energy demand prediction applied the clustering-based energy demand learning method for to further improve the prediction accuracy Need to be more stable and flexible.”

• cites Saputra et al. 2019Not checked yet

  Not checked in the free run.

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

Aïvodji et al. (2019) combined FL with secure data aggregation.

“Aïvodji et al. 2019 Smart home IOT Design a sufficient secure federated smart home setting combines FL with secure data aggregation rather complex architecture to implement”

• cites Aïvodji et al. 2019Not checked yet

  Not checked in the free run.

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

Lee et al. (2018) used a federated patient hashing framework for similar patient matching.

“Lee et al. 2018 Similar patient matching Federated patient hashing framework Avoid security attack from reverse engineering Inevitable computational complexity”

• cites Lee et al. 2018 sourceNot checked yet

  Not checked in the free run.

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

Pfohl et al. (2019) established efficacy of FL over centralized and local learning for clinical prediction.

“Pfohl et al.2019 Clinical prediction Establish efficacy of FL over centralized and local learning Perform FL in a differentially private manner Underestimate of privacy cost”

• cites Pfohl et al.2019Not checked yet

  Not checked in the free run.

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

Huangb et al. (2019) training on iid data outperform non-iid data.

“Huangb et al.2019 Mortality prediction over drug utilization data Adaptive boosting method Alleviate non-iid by introducing data-sharing technology Training on iid data outperform non-iid data”

• cites Huangb et al.2019Not checked yet

  Not checked in the free run.

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

Lee et al. (2018) has inevitable computational complexity.

“Lee et al. 2018 Similar patient matching Federated patient hashing framework Avoid security attack from reverse engineering Inevitable computational complexity”

• cites Lee et al. 2018 sourceNot checked yet

  Not checked in the free run.

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

Huangb et al. (2019) alleviated non-iid by introducing data-sharing technology.

“Huangb et al.2019 Mortality prediction over drug utilization data Adaptive boosting method Alleviate non-iid by introducing data-sharing technology Training on iid data outperform non-iid data”

• cites Huangb et al.2019Not checked yet

  Not checked in the free run.

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

Kim et al. (2017) is only accurate with small or skewly distributed data.

“Kim et al.2017 Analysis of computational phenotypes Federated tensor factorization for privacy preserving computational phenotyping Summarized information does not disclose the patient data Only Accurate with small or skewly distributed data”

• cites Kim et al.2017Not checked yet

  Not checked in the free run.

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

Kim et al. (2017) used federated tensor factorization for privacy preserving computational phenotyping.

“Kim et al.2017 Analysis of computational phenotypes Federated tensor factorization for privacy preserving computational phenotyping Summarized information does not disclose the patient data Only Accurate with small or skewly distributed data”

• cites Kim et al.2017Not checked yet

  Not checked in the free run.

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

Gao et al. (2019) is the first EEG classifier over heterogeneous EEG data.

“Gao et al.2019 EEG Classification Design a hierarchical heterogeneous horizontal FL framework The first EEG classifier over heterogeneous EEG data Only work on 3 different datasets”

• cites Gao et al.2019Not checked yet

  Not checked in the free run.

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

Huangb et al. (2019) used adaptive boosting method for mortality prediction over drug utilization data.

“Huangb et al.2019 Mortality prediction over drug utilization data Adaptive boosting method Alleviate non-iid by introducing data-sharing technology Training on iid data outperform non-iid data”

• cites Huangb et al.2019Not checked yet

  Not checked in the free run.

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

Pfohl et al. (2019) performed FL in a differentially private manner.

“Pfohl et al.2019 Clinical prediction Establish efficacy of FL over centralized and local learning Perform FL in a differentially private manner Underestimate of privacy cost”

• cites Pfohl et al.2019Not checked yet

  Not checked in the free run.

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

Gao et al. (2019) only work on 3 different datasets.

“Gao et al.2019 EEG Classification Design a hierarchical heterogeneous horizontal FL framework The first EEG classifier over heterogeneous EEG data Only work on 3 different datasets”

• cites Gao et al.2019Not checked yet

  Not checked in the free run.

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

Li and Liu (2019) model parameters of community will lead to extra communication overhead.

“Li and Liu 2019 Predict mortality and hospital stay time Introduce community-based FL and evaluate it on non-iid icu EMRs converged to higher predictive accuracy in less communication rounds than the baseline FL model Model parameters of community will lead to extra communication overhead”

• cites Li and Liu 2019Not checked yet

  Not checked in the free run.

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

Kim et al. (2017) summarized information does not disclose patient data.

“Kim et al.2017 Analysis of computational phenotypes Federated tensor factorization for privacy preserving computational phenotyping Summarized information does not disclose the patient data Only Accurate with small or skewly distributed data”

• cites Kim et al.2017Not checked yet

  Not checked in the free run.

May combine two claims.

Li and Liu (2019) introduced community-based FL and evaluated it on non-iid ICU EMRs.

“Li and Liu 2019 Predict mortality and hospital stay time Introduce community-based FL and evaluate it on non-iid icu EMRs converged to higher predictive accuracy in less communication rounds than the baseline FL model Model parameters of community will lead to extra communication overhead”

• cites Li and Liu 2019Not checked yet

  Not checked in the free run.

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

Li and Liu (2019) converged to higher predictive accuracy in less communication rounds than the baseline FL model.

“Li and Liu 2019 Predict mortality and hospital stay time Introduce community-based FL and evaluate it on non-iid icu EMRs converged to higher predictive accuracy in less communication rounds than the baseline FL model Model parameters of community will lead to extra communication overhead”

• cites Li and Liu 2019Not checked yet

  Not checked in the free run.

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

Lee et al. (2018) avoided security attack from reverse engineering.

“Lee et al. 2018 Similar patient matching Federated patient hashing framework Avoid security attack from reverse engineering Inevitable computational complexity”

• cites Lee et al. 2018 sourceNot checked yet

  Not checked in the free run.

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

Pfohl et al. (2019) underestimate of privacy cost.

“Pfohl et al.2019 Clinical prediction Establish efficacy of FL over centralized and local learning Perform FL in a differentially private manner Underestimate of privacy cost”

• cites Pfohl et al.2019Not checked yet

  Not checked in the free run.