1. Introduction
Federated Learning (FL) is a machine learning technology that uses decentralized learning
to prevent user data from being sent to a central server or cloud. A training model
is deployed to all the users’ systems from a server. Each user trains the model using
their own data, produces a new model, and sends it to the central server. Data used
to train the model is never sent from the device, protecting the user’s privacy. The
server collects the locally trained model and aggregates them based on weightage.
The server updates the model along with the previous model, and this updated model
is sent back to the connected users’ devices. This is useful in storing large data.
Various machine learning (ML) algorithms like the Support Vector Machine (SVM), Decision
Tree (DT), logistic regression, and random forest algorithms are used to measure performance
parameters. Advanced clustered FL is also used in healthcare applications to improve
accuracy and performance. Apache Spark is used for data management and advanced FL
to detect and diagnose diseases at an early stage. This helps to reduce the medication
rate.
Blockchain technology is a collection of blocks arranged sequentially that creates
a copy and sends it to all the users within the network. It is immutable since each
user has a copy of every transaction, so tampering with data is not possible. Each
block contains the current block's hash, a previous block's hash, a ``nonce,'' and
a valid hash, which is a unique code that secures the information. Mining is the process
of validating the blocks and adding them to the blockchain. Transactions are stored
in a transaction pool until they are verified, and only after verification, they are
added to the block along with an existing block as a sequence. Smart contracts that
work on the Ethereum platform are a decentralized network that does not require third-party
intervention. This is programmed in a special language called solidity, which improves
data privacy.
There are certain limitations in FL and blockchain technology. In FL, since multiple
devices are connected to the server, and an aggregator takes model updates based on
the network availability of devices, there might be an imbalance of data between users
and the cloud, leading to data privacy issues. To solve this, we can implement smart
contracts between the users in the network so that data cannot be manipulated. Even
in a blockchain, it is limited to handling large volumes of data. In this case, FL
can be implemented so that it distributes nodes and smart contracts are used on these
nodes to prevent heterogeneity and interoperability issues. Hence, combining advanced
FL technologies with a blockchain effectively improves data management, scalability,
and privacy.
The taxonomy of the review based on various technologies is portrayed in Fig. 1.
Fig. 1. Taxonomy of the review.
A scheme is proposed and can be summarized as follows. Blockchain healthcare applications
help to reduce data breaches and data security. FL helps to train models within multiple
users without the exchange of users’ data. Sensitive information can be preserved
and ensures authorized access.
2. Related Work
2.1 Healthcare using Federated Learning
Medical tracking through wearable devices and monitoring their intake using FL has
been analyzed [1-3]. Medication tracking using an Android smartwatch app collects sensor data through
an accelerometer and gyroscope [1]. Cloud-based data utilizes distributed data storage, database management, and computing
frameworks. This system used sensor data extraction, pre-processing, and ML algorithms.
The medical adherence rate was reduced by 50% for chronic diseases compared to acute
diseases. Two main factors for medical non-adherence were stress and the complexity
of tasks.
Using a lightweight wearable device and comfortable user interface (UI), a patient’s
medication can be tracked, and proper feedback can be given on time. Using Internet
of Things (IoT) health monitoring solutions can reduce healthcare costs by 68.3%.
A smartwatch application collects user activity data and sends it to distributed data
storage [Amazon Web Service (AWS)]. Pre-processed data is stored in a distributed
database [MongoDB]. This is then connected to a distributed processing framework [Apache
Spark].
A data mining method utilized Microsoft Kinect Sensors. Inertial sensors and an RGB
depth camera attached to a patient’s wrist were used to model and predict the patient's
medication records. For classification, a decision tree algorithm was used. Smartwatches
are effective in monitoring activities since they have sensors embedded to capture
the motions of a body. They contain a three-axis accelerometer, gyroscope, near-field
communication (NFC), and heart rate monitor. AWS’s simple storage service makes data
accessible from anywhere and can duplicate data in multiple storage across many places.
This can allow only authorized users to access the data. MongoDB stores data in JSON
(Javascript Object Notation) format so that adding or removing fields is easy for
users. It balances data distributed through multiple servers and allows access to
users with valid data points. It uses random forest, gradient-boosted tree, logistic
regression, and support vector machine algorithms to compare performance parameters.
24 individuals performed medication intake activities and non-medication activities.
They wore smartwatches on either the left or right wrist. Each activity was repeated
five times. 80% of the data was used for training ML models, and 20% was used for
validation. Criteria like the dominant hand, sex, age groups, and watch wrist were
tabulated. Another set of medication intake and non-medication activities was also
tabulated.
Gradient boosted tree had the highest F1 score of 0.983 with a training time of 208.784
seconds. Random forest has the second-highest F1 score of 0.977 with a training time
of 13.313 seconds. When implemented with more CPUs, memory, and storage, the Spark
cluster can make the training model faster. In the future, this method can be developed
by improving the patient’s medication adherence by sending notifications of when they
miss or take incorrect medication.
Resolving data islanding and personalization problems have been presented [2]. Data islanding is where user data always exists in the form of isolated islands,
so performing aggregated weightage is difficult. The personalization problems are
that these models fail in personalization since different users have different characteristics,
and the deployed model needs to perform personalized healthcare. To overcome these
challenges, FedHealth is used to provide high accuracy in wearables [1]. Federated transfer learning algorithms and homomorphic encryption were used to achieve
accurate, personalized healthcare.
FedHealth aggregates the data from different organizations and builds a powerful learning
model. Federal transfer learning was used to transfer knowledge from an existing domain
to a new one. Daily activities were monitored using body-worn sensors. RGB-D cameras
were used to record users’ activities. Using a UCI (Unique Caller Identifier) smartphone,
a public human activity recognition dataset was created. Six activities (walking,
walking upstairs, walking downstairs, sitting, standing, and lying down) were measured
with 30 volunteers in the age group of 19-48 years. A convolutional neural network
(CNN) was used for training and prediction between a server and a user with a convolution
size of 1x9 (width x height). It used minibatch Stochastic Gradient Descent (SGD)
for optimization, the batch size was considered as 64, and 80 training epochs were
tabulated. FedHealth achieved better accuracy by 5.3% than a previous method. In the
future, FedHealth can be deployed at a large scale in more healthcare applications
like elderly care, fall detection, and cognitive disease detection.
A comparison of performance measures using four ML algorithms for predicting stroke
has been presented [3]. The Stroke Prediction dataset was created in Apache Spark to handle large amounts
of data, including the MLlib library. Various ML algorithms were used to predict stroke.
Performance measures like accuracy, precision, and recall were compared between random
forest, logistic regression, and support vector algorithms and computed [2]. The dataset consisted of 10 independent variables representing features and one
dependent variable for predicting a disease as a class label, which was either 0,
which denotes no stroke, or 1, which denotes stroke disease.
When applying logistic regression to performance measures like precision, recall,
and F1 score for class 0, the precision had the highest percentage at 81%. The recall
had the lowest percentage at 73%, and for class 1, the F1 score had the highest value
of 79%. Similarly, when applying random forest to performance measures for class 0,
precision had the highest value of 96%, and recall had the lowest value of 85%. For
class 1, the recall had the highest value of 96%.
Applying the decision tree for class 0 showed precision at its highest value of 82%
and recall at its lowest value of 75%. For class 1, the recall was the highest value
of 84%. Finally, when applying the linear support vector machine for class 0, precision
had its highest value of 81%, and recall had its lowest value of 72%. For class 1,
the recall had its highest value of 83%. Applying these algorithms improves accuracy,
and it was deduced that random forest had the highest accuracy of 90%, decision tree
had the next highest at 79%, and support vector and logistic regression had the same
accuracy of 77%.
Implementation of edge computing in medicine and data processing analysis at the edge
using clustered FL (CFL) in diagnosing COVID-19 has been explained [4]. There were two different clusters: one with clients with X-ray facilities in their
hospitals and the other with ultrasound facilities in their hospitals. The updates
from these clusters were sent to a cloud server, where weighted aggregation and updates
were done using CFL. The VGG16 model was used, which consists of 14 convolution layers,
6 pool layers, and 6 dense layers.
The binary cross-entropy loss was used for model optimization. The focal loss was
used for binary classification tasks to improve the model’s performance. The performance
of CFL was compared between two baselines. One was a specialized FL baseline, and
the other was a multi-modal conventional FL. Performance evaluation was done for three
experimental setups. First, it was done to train the baseline model with a batch size
of 16, and then the same model was trained with CFL with five epochs and a batch size
of 16. The comparison results were studied for multi-modal models trained in CFL and
conventional FL using X-ray and ultrasound.
CFL outperformed the conventional FL environment. The precision, recall, and F1 score
of CFL with specialized models and conventional FL were studied over increasing epochs.
This showed that an increased number of epochs improves efficiency and accuracy. The
main challenge is maintaining a balance between privacy and execution of the FL model
and heterogeneity issues due to data variation between clients in the cluster and
multi-modal collaborative FL networks.
The design of slime mold optimization with E1Gamal encryption in a hybrid deep learning
classification (SMOEGE-HDL) and their processes have been analyzed [5]. It consists of data encryption and classification. Initially, data was encrypted
using the SMOEGE-HDL technique in an IoT environment. The Sequential Minimal Optimization
(SMO) algorithm was used. Then, the High Dimension Learning (HDL) technique was used
for the classification process. A Nadam optimizer was used for better classification.
This technique’s performance was compared with existing techniques and achieved an
accuracy of 98.5%, precision of 98.75%, F1-score of 98.25%, and specificity of 98.5%.
E1Gamal encryption is a public key encryption algorithm that provides confidentiality
and authentication control to protect sensitive information. Integrating this technique
with proper cryptographic techniques and deep learning-based classification ensures
secure communication between IoT devices. Experimental results have shown that the
proposed technique has improved at different file sizes. Healthcare using FL is illustrated
in Fig. 2.
Fig. 2. Healthcare using federated learning.
2.2 Healthcare using Blockchain
Maintenance of health records and monitoring healthcare using blockchain has been
presented [6-13]. Security and privacy of patient records and data through blockchain-based healthcare
systems were also explained [6]. This lets doctors find information about patients’ health through wearable devices.
Here, a permissioned blockchain was used, and the patient had the right to decide
who could access their information. Using cryptography algorithms and HASH, each block
is connected in sequence with the previous block’s hash and has a unique alphanumeric
code for security purposes.
Decentralized Artificial Intelligence (AI) uses ML algorithms to predict diseases
through symptoms at the early stage. Disease symptoms were given as datasets. Using
Bayesian, k-nearest neighbor, and decision tree ML algorithms, training of the model
was done and then checked for accuracy of the trained model. Integrating ML and blockchain
in the healthcare industry results in secure data exchange and achieves accuracy.
A safe and efficient way of sharing Electronic Medical Records (EMRs) using blockchain
has been analyzed [7,9]. To improve scalability, the blocks were stored in a Hadoop database. The hash used
SHA3-256 for transactions in blocks.
There is no universal standard for data to be exchanged in a blockchain. Ethereum
creates smart contracts, which distribute data within the blockchain network. A smart
contract uses the solidity language. Distributed applications contain smart contracts
that have code to perform a particular task. Data masking technology was used to achieve
privacy in data, and the InterPlanetary File System (IPFS) protocol was used to store
elaborate details of medical records. To store the transaction records in sequence,
Merkle-tree structure and Hash encryptions were used.
A consensus algorithm was used to check the stability of the model. A generalization
algorithm was used for age, and a masking algorithm was used for the address, which
was to be stored in DCM (Digital Imaging and Communication in Medicine) format. Electronic
medical data consists of three parts: the public key field stores patients’ data,
and the corresponding doctor’s public key is used as a unique code. Metadata stores
the timestamp generated. The data summary contains the data after masking and the
hash value. Using a delegated proof of stake (DPOS) in the consensus algorithm enabled
data privacy in managing health records while sharing.
Smart contracts and more advanced permission-controlled structures can be used to
protect patients’ medical information. Performance evaluation is conducted using Apache
JMeter, and average execution time, throughput, and average latency were studied.
The execution time increased as the number of transactions increased. Similarly, throughput
was simulated for 100 to 500 users with a period of 10 to 35 ms, and as the number
of users and requests increased, throughput also increased linearly. The highest latency
concerning throughput was 14 ms. The implementation of a payment module for patients
to pay consultation fees to doctors using decentralized blockchain technology can
further be improved.
Data management in healthcare using a canal system to improve security and computational
speed compared to the bitcoin network has been explained [8]. The network traffic decreased as the number of nodes increased. The Head Blockchain
Manager (HBCM) was used as a single authority that regularizes the network like a
miner in bitcoin technology. There are two HBCMs. One is the leader, which generates
blocks, and the other is the follower, which replicates them. The nodes were grouped
into clusters, each with one Blockchain Manager (BCM), which maintains a ledger of
valid and invalid transactions.
To maintain privacy in transactions, canals were used within the network. There were
three canals. Canal 0 was a public blockchain ledger that was accessible by all participants
in the network. Canals 1 and 2 were private blockchain ledgers accessible only by
Canal 1 and Canal 2. The Practical Byzantine Fault Tolerance (PBFT) consensus algorithm
was used for resolving computational issues faced by bitcoin technology.
A performance comparison between lightweight blockchain and bitcoin technology was
simulated using an NS3 simulator. The block size was considered as 1 MB, performance
was evaluated for 100 nodes, and SHA-256 was the hash algorithm. Analysis was done
based on three situations: as the number of blocks increased, as the number of nodes
increased, and to check the time it was taken to replicate data and update. In the
first situation, the number of nodes was kept constant, and the blocks varied. In
the second situation, the nodes were kept constant at 10, and the number of blocks
varied.
The results showed that lightweight blockchain outperforms bitcoin technology regarding
the amount of data transferred and execution time. It showed an improvement of 67%
in replication and execution time. It also showed that network traffic was reduced
by 1/10 that of the bitcoin network. This technique can be further enhanced by using
cluster along with fog computing.
A blockchain technology-based smart ration shop in villages using smart contracts
for immutable transactions has been presented [10]. IoT technology was used to track the assets and stock items when transporting them
from distribution centers to ration shops. Voice-based and text-based interfaces were
used to improve the effectiveness of the tracking system. The performance of blockchains
was evaluated based on their successful transaction rate, time delay, and system robustness.
Blockchain technology applied to overcome the losses incurred in financial, logistic,
human resource, and healthcare industries has been analyzed [11]. Many tracking applications in wearables and IoT techniques are used to prevent virus
transmission. Mobile applications with contact tracing using Bluetooth or NFC were
used to predict disease spreading. Blockchain protects the data through hash encryption
making it immutable.
The COVID-19 pandemic situation has led to a system of online education for students.
A secure Learning Management System (LMS) using a blockchain has been examined [12]. Online assessments and grades were achieved using blockchain’s security properties.
A private blockchain where multiple nodes are controlled by Higher Education Institutions
(HEI) was used. All users’ access to the blockchain is monitored and verified using
a consensus mechanism. This blockchain-enabled LMS can be deployed in several on-campus
systems, making the e-learning work efficiently.
An analytical Hierarchy Process (AHP) for the development of smart cities using the
IoT, big data, AI, and blockchain has been explained [13]. This method was based on the reciprocity axiom, homogeneity axiom, dependency axiom,
and axiom of expectations. Out of 6 groups of criteria and sub-criteria, three groups
were taken. Calculation of a comparison matrix and weights for each of three groups
namely government (G), healthcare (H), and education (E) were taken and tabulated.
Blockchain-based solutions on applying AHP and on the basis of principles of decomposition,
comparative judgements, and hierarchical decomposition ranked the highest out of the
15 groups. The various stages of blockchain technology in healthcare are illustrated
in Fig. 3.
Fig. 3. Healthcare using blockchain.
3. Healthcare using Federated Learning and Blockchain
The implementation of FL using blockchain technology has been analyzed [14-20]. To increase transparency and privacy, FL in blockchain technology was adopted [14]. Chained Distributed ML (C-DISTRIM) was used to train distributed models connected
through intelligent contracts using the solidity programming language within a blockchain
network. Once each participant is trained with the model, communication between the
participant and the cloud is done through smart contracts. Successive iterations are
done until all the participants have finished training. Google Cloud Storage (GCS)
was used to store the shared models.
The Ganache Network utilizes the Ethereum blockchain for testing purposes. The Non-Small
Cell Lung Cancer (NSCLC)-Radiomics dataset consists of 422 patients. The computed
tomography (CT) scan reports had variable segmentations performed by radiologists,
and corresponding survival outcomes after 2 years were observed as classes. Data augmentation
was done to balance the two classes, and the cases were increased to 704 from 422.
Under six testing scenarios, a validation process was done. A 3D binary classifier
and CNN were used. Cropped volumes were resized as 120x160x60 (width x height x depth).
Integrating blockchain helped trace the origin of data and training-process monitoring.
The storing of models in nodes instead of a central server using chain FL in blockchain
technology has been analyzed [15]. The decentralized FL used a private blockchain for model storage, aggregated the
model, and sent updates using smart contracts. In chain FL, each participant had their
own servers, and ML servers trained their models and uploaded them to the miners.
There is at least one miner for each participant in the network. Private Ethereum
was used in smart contracts, and a consensus mechanism prevented attacks. The federated
averaging algorithm helps participants train models sequentially at different times
without waiting for any training to finish.
The MNIST experiment was done with a small model and a large model. The first one
had 25,450 parameters, and the other had 468,874 parameters. Data was divided into
three parts. 20-epoch training was done with the first data, then averaged. 20 epochs
with the second data, then averaged, and 20 epochs with the third data, then averaged.
The results showed that chain FL was effective.
The CIFAR-10 experiment consisted of 60,000 images of 32x32x3 (width x height x depth)
and 10 output classes. The training was done for 200 epochs, and the learning rate
was reduced by 0.1 at epochs 100 and 150 for classic and federated cases. The results
showed that chain FL was still effective in a CNN and showed a decreased accuracy
score of 2.57%.
Collaborative learning using a blockchain to detect diseases through chest CT and
X-ray images has been presented [16]. This method also concentrated on data security in healthcare centers. A data normalization
process must normalize data received from different sources. A deep learning model
is required to recognize COVID-19 patterns from CT scans using a capsule network,
segmentation, and classification. To solve privacy-related problems, FL was used while
aggregating the weights of various models. To preserve privacy, the hospital only
shared gradients of the trained model to the blockchain. The collaborative model was
better and had accurate predictions since it used the latest symptoms of COVID-19.
CT scan image data were collected from 3 different hospitals. In total, there were
34,006 slices scanned by six different scanners. The collected data had CT slices
for 89 subjects, out of which 68 were positive cases, and the other 21 were negative.
The slices had cylindrical bounds. The pixels outside the cylindrical bounds were
discarded so that pre-processing could be done for only pixels within the bounds.
The volume of an adult female lung was 10 to 12% smaller than a male’s lung. Performance
evaluation of the capsule network for detection of CT image accuracy was done, and
high sensitivity was achieved in the capsule network. The result showed that as data
providers increase, the performance also increases.
Privacy and data fraud detection at various levels with low energy consumption and
lag in healthcare applications have been presented [17]. A dynamic detection method based on FL used blockchain for IoT healthcare applications
in a fog cloud. There were two types of learning models. One was the local learning
model, which was connected at each fog node, and the other was the global master node.
The fog-cloud agent (FCA) was the main node that schedules all the requested work
to be shared with the master globe node. The dynamic scheduler schedules all the workloads
based on the blockchain requirement and sends them to the respective fog nodes. Rescheduling
can be done only if there is any failure during execution. All the nodes were sorted
according to the deadlines, and based on the availability of nodes, the workloads
must be executed until the assigned workloads have finished the execution.
The patient information includes the claims filed by patients during admission to
the hospital. It also includes the admission charges and discharged outpatient information,
which had data about hospital visits without admission to the hospital. Resource specifications
with parameters like memory and speed were also considered and tabulated. Three different
workloads were taken as input: ECG heartbeat and videos as images and blood pressure
as numbers. This was processed in the system for execution.
Baseline 1 consisted of ML methods inside a blockchain for fault detection, and baseline
2 consisted of dynamic ML methods. Fraud analysis was done with the ratio of fraud
and delay initially as high as 60 to 80. After the system had undergone adaptive training
and adaptive testing, both ratios were improved with minimum loss. The results showed
that a fully trained ML algorithm (FTMLA) was widely used to train offloaded data
with scalability.
The FL model had a more optimal result than an existing training model with delay.
It also consumed less power. FL-BETS outperformed all the other existing schemes in
Internet of Medical Things (IoMT) applications. The study focused on mobility fraud
and detection in civil marine applications.
Blockchain-based FL using edge nodes for improving secure healthcare has been explained
[18]. The benefits and limitations of combining both technologies to achieve better performance
were described. Various applications that implement blockchain-enabled ML include
IoMT, electronic health records (EHRs), EMR management, and digital healthcare systems.
The various challenges of combining these two technologies were highlighted.
Blockchain technology helped to develop a framework for the COVID-19 crisis through
hospital end-to-end delivery systems providing data privacy and access. Model View
Control (MVC) makes the system more scalable and accessible. Proof of Work (PoW) is
the primary mechanism that checks the validity and approves a transaction to be added.
FL uses model training and interference, where the model is transferred between FL
devices, but the data does not leave the local server of individual devices. This
helps in model training across a variety of distributed datasets.
Integration of a blockchain and FL increases the training rate and ensures security.
However, this method still has challenges like security weakness, the need for additional
privacy-preserving mechanisms, and data quality. A more accurate model can decrease
costs for providers and insurers.
In developing an architecture for blockchain-enabled FL, implementation of a decentralized
identifier (DID)-based access system is the key component [19]. A decentralized unique ID was introduced, allowing users to manage their IDs and
gain access to FL through their DIDs stored in the blockchain. The system administrators
have control over access and permission for the execution of smart contracts. This
ensures data security.
The method consisted of two phases with six stages. The initial phase comprises two
stages, and the second comprises four. Then, once the evaluation is done, it undergoes
cross-device and cross-silo scenarios. The aggregator aggregates the primary model
from the trainers. Authentication is done using DID and verified credentials (VCs).
The results were evaluated using a primary deep learning method on a Non-Independent
Identically Distributed (Non-IID) dataset. The FEMNIST dataset was used for training.
The results showed a difference between a real environment and a development environment.
A survey on how blockchain is integrated with FL (BlockFed), challenges, solutions,
and future scope has been presented [20]. Initially, the focus was on significant issues in FL, the role of a blockchain,
and its potential to overcome these issues. Then, based on how FL is integrated with
a blockchain, systems were classified as coupled, decoupled, and overlapped systems.
Finally, the advantages and disadvantages of this system and solutions were analyzed.
Some of the challenging issues in FL were the lack of incentive mechanisms, statistical
heterogeneity, system heterogeneity, model security, data privacy, and communication
overhead. To solve these issues, blockchain provides appropriate solutions like incentive
mechanisms where the users are rewarded, implementing blockchain with various distributed
optimization algorithms, and applying consensus mechanisms for data security and privacy.
Fig. 4 portrays the characteristics of healthcare using FL and blockchain.
Fig. 4. Healthcare using FL and blockchain.
Thus, by integrating FL and a blockchain, we can achieve data security and privacy,
improve the prediction and accuracy of medical data, reduce communication delay and
computational and processing time, and analyze the performance effects under various
attack scenarios. A summary of existing and proposed technology is tabulated in Table 1.
Table 1. Summary of review.
|
Reference Number
|
Method
|
Algorithm Used
|
Result
|
Limitation/Enhancements
|
|
Existing Method: Federated Learning
|
|
[1]
|
Sensors for data extraction and processing through ML Algorithms
|
Random Forest
Gradient Boosted Tree
Logistic Regression
Support Vector Machine
|
Gradient Boosted Tree F1 score-0.977
Training time-13.313s
|
Requires more training and testing time on adding features
|
|
[2]
|
Federated Transfer Learning and Homomorphic encryption
|
K-Nearest Neighbors, Support Vector Machine, Random Forest
|
Fedhealth has 5.3% more than conventional
|
Activity recognition for cognitive disease not evaluated.
|
|
[3]
|
Performance using Federated ML Algorithms for the prediction of disease
|
Decision Tree
Random Forest
Support Vector Machine
Logistic Regression
|
Random Forest has a 90% accuracy
|
Requires more Computation time
|
|
[4]
|
Edge Computing and Clustered Federated Learning (CFL)
|
Specialized FL
Multimodal model
CFL
|
CFL outperforms conventional FL
|
Privacy imbalance due to Heterogeneity needs attention
|
|
[5]
|
Slime Mold Optimization with E1Gamal encryption in Hybrid Deep Learning classification
(SMOEGE-HDL)
|
Sequential Minimal Optimization (SMO)
Nadam Optimizer
|
Accuracy-98.5%
Precision-98.75%
F1 score-98.25%
Specificity-98.5%
|
Privacy-enhancing approaches for various techniques are to be explored.
|
|
Existing Method: Blockchain
|
|
[6]
|
Permissioned Blockchain and ML Algorithms
|
Hash
Consensus
Naïve Bayesian
K-Nearest Neighbor
Decision Tree
|
Secure exchange of data and accuracy achieved
|
Various blockchains can be integrated with advanced ML.
|
|
[7]
|
Blockchain using Data Masking technology and Interplanetary File System (IPFS)
|
SHA-256
Consensus mechanism
Generalization algorithm
|
Data privacy achieved
|
Lacks On-chain storage proofs
|
|
[8]
|
Data Management of Healthcare using the Canal System
|
Consensus mechanism
Hash
|
Performance improved by 67%, and data traffic was 1/10th of conventional
|
Processing time on increasing the no. of nodes
|
|
[9]
|
Smart contracts and IPFS
|
SHA-256
Consensus mechanism
Generalization algorithm
|
Throughput and latency of 14ms improved
|
Lacks On-chain storage proofs
|
|
[10]
|
Smart contracts and IoT technology
|
Proof of Authority (PoA)
Digital Signature Algorithm
|
Response time - around 130 ms
Average satisfaction -90%
|
Computational complexity by varying the number of transaction requests per second.
|
|
[11]
|
Smart contracts and IoT wearable devices
|
Consensus mechanism
The lightweight digital signature algorithm
|
Wearables readings during self-isolation
Transactions are sent to the blockchain network when the alarm is set (monitored by
authorities)
Health alert to patient and authority
|
Blockchain paired with remote monitoring devices and IoT during an emergency for various
scenarios needs to be explored.
|
|
[12]
|
Private blockchain and Learning Management System (LMS)
|
Consensus mechanism
|
In-house developed LMS connected to multiple nodes of HEI.
Traceable timestamp
|
Implementing blockchain technology in e-learning systems
|
|
[13]
|
IoT, AI, and Blockchain
|
Analytic Hierarchy Process (AHP) algorithm
Comparison matrix and ranking of 15 groups
|
Government, Healthcare and Education were taken for ranking
Sub-criteria showed governance blockchain-based ICT as the highest.
G1- 1.78 times greater than G2
|
All six groups of criteria and sub-criteria are to be identified in detail.
Use of Analytical Network Process (ANP)
|
|
Proposed Method: Integration of FL with Blockchain
|
|
[14]
|
Chain Distributed Machine Learning with smart contracts
|
CNN
3D binary classifier
|
Tracing the origin achieved in diagnosis. p>0.05
|
Performance, when intruders enter, should be analyzed.
|
|
[15]
|
Chain FL along with blockchain using federated averaging
|
Private Blockchain
Federated Averaging
|
Chain FL is 2.57% more effective than conventional
|
Using many smart contracts for storing large amounts of data consumes more time
|
|
[16]
|
Collaborative Learning using Blockchain to detect infections
|
Capsule network
Segmentation
Classification
|
Sensitivity of 0.987
and the performance of 98% was achieved
|
Receiving data from more hospitals consumes more time
|
|
[17]
|
Federated Learning using blockchain in fog-cloud
|
Fog node
Global master node
Scheduler
Fully trained ML
|
FL-BETS outperforms with minimum energy consumption of 41% and delay of 28% than existing
schemes
|
Dynamic run-time attacks are to be considered.
|
|
[18]
|
Blockchain-enabled ML for Medical things
|
Model View Control
Proof of Work
|
A more accurate model reduces cost.
|
Performance evaluation can be done when the current architecture has practical Proof
of functionality.
|
|
[19]
|
Blockchain-enabled FL using Decentralized Identifier (DID)
|
DID
Verified credentials
|
Difference between real and developed environment
|
Blockchain-empowered FL for IoMT to resolve privacy issues.
|
|
[20]
|
Blockchain empowered in FL
|
Consensus mechanism
Coupled
Decoupled
overlapped
|
Blockchain solves FL issues like incentives, heterogeneity, data privacy, and communication
overhead.
|
Implementation of Various distributed optimization algorithms to enhance security
|