Хмарний сервіс ідентифікації людини за її електрокардіограмою
Telemedicine grows faster with each year. In scope of it new technologies have been created to solve information communication problems. One of them is distance monitoring which requires electrocardiogram analysis. In this case it is important to confirm that the transferred electrocardiogram throug...
Збережено в:
| Дата: | 2023 |
|---|---|
| Автор: | |
| Формат: | Стаття |
| Мова: | English |
| Опубліковано: |
Інститут програмних систем НАН України
2023
|
| Теми: | |
| Онлайн доступ: | https://pp.isofts.kiev.ua/index.php/ojs1/article/view/550 |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
| Назва журналу: | Problems in programming |
| Завантажити файл: |  |
Репозитарії
Problems in programming| id |
pp_isofts_kiev_ua-article-550 |
|---|---|
| record_format |
ojs |
| resource_txt_mv |
ppisoftskievua/60/21fcbf4d24d8ec3617561e82dc8f4f60.pdf |
| spelling |
pp_isofts_kiev_ua-article-5502023-06-25T08:23:36Z Cloud service for authentication of a person based on their electrocardiogram Хмарний сервіс ідентифікації людини за її електрокардіограмою Luhovskyi, Yu.O. electrocardiogram; authentication; classification; cloud infrastructure UDC 004.9 електрокардіограма; автентифікація; класифікація; хмарний сервіс УДК 004.9 Telemedicine grows faster with each year. In scope of it new technologies have been created to solve information communication problems. One of them is distance monitoring which requires electrocardiogram analysis. In this case it is important to confirm that the transferred electrocardiogram through information channels belongs to the patient. Researchers from different countries work on this problem. They suggest different methods of authentication by electrocardiogram. The goal of this work is to suggest a prototype of a service that could be used to authenticate electrocardiograms.The paper describes which algorithms have been used to build authentication technology and how it was implemented. There is a short history of built applications. It shows their structures and the purpose. The recently developed system is a prototype to authentication service. It performs registration of new electrocardiograms which is the most time consuming process in authentication. The paper describes the architecture of the system and shows the result of executed experiments. The results show that there is a performance issue with the machine learning library ML.NET. When a lot of cores are allocated to one machine learning instance the overheads highly decrease the overall experiment time. These experiments confirmed Amdahl’s law. Nevertheless, an architecture was found where experiments took the least time for execution. Knowing the issue with the machine learning library, a new architecture setup was suggested and will be implemented in future works.Besides, attention is paid on how developed service should help researchers to improve the technology. The main idea of the system is using one service for developing and testing the technology. That means we can be sure that the cloud service that runs on production would behave the same as in the research phase.Prombles in programming 2022; 3-4: 485-490 Телемедицина з кожним роком пришвидшує свій розвиток. Відбувається це за рахунок винайдення нових технологій, що розвивають комунікацію та роботу з даними. Одним із напрямків, в телемедицині є дистанційний моніторинг, задачею якого є постійно контролювати стан пацієнта. Це здійснюється перевіркою різних показників, в тому числі, йі електрокардіограми. Важливо, щоб надіслана кардіограма належала саме пацієнтові на моніторингу. Для цього розробляється технологія автенти- фікації по електрокардіограмі. Дослідники з різних країн працюють над цим питанням використовуючи підходи, що можуть сильно відрізнятись.В цій роботі описується прототип хмарного сервісу по автентифікації. Описуються алгоритми, що використовувались для побудови технології автентифікації та додатків, що реалізують цю технологію. Показуються послідовні розробки додатків та описуються їх задачі. Останній розроблений додаток - це хмарний сервіс, побудований на мікросервісах, що виконує реє- страцію електрокардіограм. Це є найбільш витратна щодо часу операція автентифікації. Приводиться архітектура сервісу та результати експериментів. По результатах помітно, що є проблеми з машинним навчанням, а саме бібліотекою ML.NET. Коли виділяється багато ядер для одного екземпляра мікросервіса, що навчає нейромережі, виникають накладні витрати на роз- паралелювання, що сильно впливає на продуктивність. Це підтверджує закон Амдала. За наявної архітектури була знайдена конфігурація, при якій час на проведення експериментів був найменшим. Враховуючи проблему з накладними витратами нова архітектура була запропонована для подальших досліджень.Розглядається, як розроблена архітектура повинна допомогти в розвитку технології. Ідея системи полягає у використанні одного і того ж сервісу для досліджень і в промисловому використанні. Це дасть гарантію, що промисловий сервіс буде поводитись так само як і при дослідженні.Prombles in programming 2022; 3-4: 485-490 Інститут програмних систем НАН України 2023-01-23 Article Article application/pdf https://pp.isofts.kiev.ua/index.php/ojs1/article/view/550 10.15407/pp2022.03-04.485 PROBLEMS IN PROGRAMMING; No 3-4 (2022); 485-490 ПРОБЛЕМЫ ПРОГРАММИРОВАНИЯ; No 3-4 (2022); 485-490 ПРОБЛЕМИ ПРОГРАМУВАННЯ; No 3-4 (2022); 485-490 1727-4907 10.15407/pp2022.03-04 en https://pp.isofts.kiev.ua/index.php/ojs1/article/view/550/603 Copyright (c) 2023 PROBLEMS IN PROGRAMMING |
| institution |
Problems in programming |
| baseUrl_str |
https://pp.isofts.kiev.ua/index.php/ojs1/oai |
| datestamp_date |
2023-06-25T08:23:36Z |
| collection |
OJS |
| language |
English |
| topic |
electrocardiogram authentication classification cloud infrastructure UDC 004.9 електрокардіограма автентифікація класифікація хмарний сервіс УДК 004.9 |
| spellingShingle |
electrocardiogram authentication classification cloud infrastructure UDC 004.9 електрокардіограма автентифікація класифікація хмарний сервіс УДК 004.9 Luhovskyi, Yu.O. Хмарний сервіс ідентифікації людини за її електрокардіограмою |
| topic_facet |
electrocardiogram authentication classification cloud infrastructure UDC 004.9 електрокардіограма автентифікація класифікація хмарний сервіс УДК 004.9 |
| format |
Article |
| author |
Luhovskyi, Yu.O. |
| author_facet |
Luhovskyi, Yu.O. |
| author_sort |
Luhovskyi, Yu.O. |
| title |
Хмарний сервіс ідентифікації людини за її електрокардіограмою |
| title_short |
Хмарний сервіс ідентифікації людини за її електрокардіограмою |
| title_full |
Хмарний сервіс ідентифікації людини за її електрокардіограмою |
| title_fullStr |
Хмарний сервіс ідентифікації людини за її електрокардіограмою |
| title_full_unstemmed |
Хмарний сервіс ідентифікації людини за її електрокардіограмою |
| title_sort |
хмарний сервіс ідентифікації людини за її електрокардіограмою |
| title_alt |
Cloud service for authentication of a person based on their electrocardiogram |
| description |
Telemedicine grows faster with each year. In scope of it new technologies have been created to solve information communication problems. One of them is distance monitoring which requires electrocardiogram analysis. In this case it is important to confirm that the transferred electrocardiogram through information channels belongs to the patient. Researchers from different countries work on this problem. They suggest different methods of authentication by electrocardiogram. The goal of this work is to suggest a prototype of a service that could be used to authenticate electrocardiograms.The paper describes which algorithms have been used to build authentication technology and how it was implemented. There is a short history of built applications. It shows their structures and the purpose. The recently developed system is a prototype to authentication service. It performs registration of new electrocardiograms which is the most time consuming process in authentication. The paper describes the architecture of the system and shows the result of executed experiments. The results show that there is a performance issue with the machine learning library ML.NET. When a lot of cores are allocated to one machine learning instance the overheads highly decrease the overall experiment time. These experiments confirmed Amdahl’s law. Nevertheless, an architecture was found where experiments took the least time for execution. Knowing the issue with the machine learning library, a new architecture setup was suggested and will be implemented in future works.Besides, attention is paid on how developed service should help researchers to improve the technology. The main idea of the system is using one service for developing and testing the technology. That means we can be sure that the cloud service that runs on production would behave the same as in the research phase.Prombles in programming 2022; 3-4: 485-490 |
| publisher |
Інститут програмних систем НАН України |
| publishDate |
2023 |
| url |
https://pp.isofts.kiev.ua/index.php/ojs1/article/view/550 |
| work_keys_str_mv |
AT luhovskyiyuo cloudserviceforauthenticationofapersonbasedontheirelectrocardiogram AT luhovskyiyuo hmarnijservísídentifíkacíílûdinizaííelektrokardíogramoû |
| first_indexed |
2024-09-12T19:29:55Z |
| last_indexed |
2024-09-12T19:29:55Z |
| _version_ |
1811522605534937088 |
| fulltext |
485
Прикладне програмне забезпечення
УДК 004.9 http://doi.org/10.15407/pp2022.03-04.485
CLOUD SERVICE FOR AUTHENTICATION
OF A PERSON BASED
ON THEIR ELECTROCARDIOGRAM
Yurii Luhovskyi
Телемедицина з кожним роком пришвидшує свій розвиток. Відбувається це за рахунок винайдення нових технологій, що
розвивають комунікацію та роботу з даними. Одним із напрямків, в телемедицині є дистанційний моніторинг, задачею якого
є постійно контролювати стан пацієнта. Це здійснюється перевіркою різних показників, в тому числі, йі електрокардіограми.
Важливо, щоб надіслана кардіограма належала саме пацієнтові на моніторингу. Для цього розробляється технологія автенти-
фікації по електрокардіограмі. Дослідники з різних країн працюють над цим питанням використовуючи підходи, що можуть
сильно відрізнятись.
В цій роботі описується прототип хмарного сервісу по автентифікації. Описуються алгоритми, що використовувались для
побудови технології автентифікації та додатків, що реалізують цю технологію. Показуються послідовні розробки додатків
та описуються їх задачі. Останній розроблений додаток - це хмарний сервіс, побудований на мікросервісах, що виконує реє-
страцію електрокардіограм. Це є найбільш витратна щодо часу операція автентифікації. Приводиться архітектура сервісу та
результати експериментів. По результатах помітно, що є проблеми з машинним навчанням, а саме бібліотекою ML.NET. Коли
виділяється багато ядер для одного екземпляра мікросервіса, що навчає нейромережі, виникають накладні витрати на роз-
паралелювання, що сильно впливає на продуктивність. Це підтверджує закон Амдала. За наявної архітектури була знайдена
конфігурація, при якій час на проведення експериментів був найменшим. Враховуючи проблему з накладними витратами нова
архітектура була запропонована для подальших досліджень.
Розглядається, як розроблена архітектура повинна допомогти в розвитку технології. Ідея системи полягає у використанні
одного і того ж сервісу для досліджень і в промисловому використанні. Це дасть гарантію, що промисловий сервіс буде по-
водитись так само як і при дослідженні.
Ключові слова: електрокардіограма, автентифікація, класифікація, хмарний сервіс
Telemedicine grows faster with each year. In scope of it new technologies have been created to solve information communication
problems. One of them is distance monitoring which requires electrocardiogram analysis. In this case it is important to confirm that the
transferred electrocardiogram through information channels belongs to the patient. Researchers from different countries work on this
problem. They suggest different methods of authentication by electrocardiogram. The goal of this work is to suggest a prototype of a
service that could be used to authenticate electrocardiograms.
The paper describes which algorithms have been used to build authentication technology and how it was implemented. There is a short
history of built applications. It shows their structures and the purpose. The recently developed system is a prototype to authentication
service. It performs registration of new electrocardiograms which is the most time consuming process in authentication. The paper
describes the architecture of the system and shows the result of executed experiments. The results show that there is a performance
issue with the machine learning library ML.NET. When a lot of cores are allocated to one machine learning instance the overheads
highly decrease the overall experiment time. These experiments confirmed Amdahl’s law. Nevertheless, an architecture was found
where experiments took the least time for execution. Knowing the issue with the machine learning library, a new architecture setup
was suggested and will be implemented in future works.
Besides, attention is paid on how developed service should help researchers to improve the technology. The main idea of the system is
using one service for developing and testing the technology. That means we can be sure that the cloud service that runs on production
would behave the same as in the research phase.
Keywords: electrocardiogram, authentication, classification, cloud infrastructure
Information systems
Introduction. Nowadays telemedicine has become much more popular. After the epidemic of COVID19 the
need for distance consultations rapidly grew. That forced creation of new technologies and new equipment in this field.
The distance monitoring that is a part of telemedicine is also improving. The part of it is biometric authentication by
electrocardiogram (ECG). It is created to prove that sending ECG from a patient to a doctor through information chan-
nels really belongs to that person.
Researchers from different countries suggest using authentication by ECG not only in telemedicine but in
other systems where access can be provided with ECG. It can be used on its own or as an additional security to other
authentication methods. To solve authentication problems researchers use different approaches [1, 2, 3, 4, 5, 6]. Some
of them use naked ECG curves and with help of deep learning create neural models which are used to classify other
naked ECG curves [2]. Some other approaches use algorithms that can compare two different curves which are used to
distinguish ECG curves [3]. And others that process ECG, get features and use them in a couple with machine learning
to classify ECGs.
The paper described developed programs that were based on research presented in the articles [7, 8, 9].
There is a patent for this technology [10]. The intention is not to show the efficiency of technology but describe
chosen program decisions required to implement this technology. First a desktop program was developed which
could register a new ECG and make authentication based on it. When the efficiency of technology was confirmed,
© Ю.О. Луговський, 2022
ISSN 1727-4907. Проблеми програмування. 2022. № 3-4. Спеціальний випуск
486
Прикладне програмне забезпечення
a new application was developed. It was able to register a big set of ECGs. After running a couple of experiments
it became clear that developed algorithms take too much time to register one ECG. It is a big problem since
nobody wants to use this technology on live production. To solve the issue a new information system has been
created that is based on microservices and deployed to a cloud. It was shown the revealed issues with the system
and how the calculation was improved using existing resources. The developed system became a prototype of
cloud service for authentication.
Besides, the developed system saves researchers time usually spent for manual data manipulation or other
manual work. It was built in the way to organize work with experiments.
Results. Before looking into programs and information systems it is worth understanding the principles
used for authentication by ECG. Input data for authentication is a file with ECG signal. The signal is loaded from
the file then processed and presented in three dimensional phase space [3]. Then from the presented cardiac cycles
selected one that is the most closest to the average cycle. Using parametric spline found 4 coordinates in three
dimensional space that approximate the selected cardiac cycle the most [2]. Founded coordinates used as input
for machine learning to learn a model and then to classify ECGs. The efficiency of this approach is described in
the article [1].
The authentication approach can be divided logically into two main parts that are presented on Fig 1. The first
is ECG processing and creating coordinates for machine learning. The second is learning neural networks and classify-
ing ECGs by coordinates.
Прикладне програмне забезпечення
Besides, the developed system saves researchers time usually spent for manual data manipulation or other
manual work. It was built in the way to organize work with experiments.
Results. Before looking into programs and information systems it is worth understanding the principles used for
authentication by ECG. Input data for authentication is a file with ECG signal. The signal is loaded from the file then
processed and presented in three dimensional phase space [3]. Then from the presented cardiac cycles selected one that
is the most closest to the average cycle. Using parametric spline found 4 coordinates in three dimensional space that
approximate the selected cardiac cycle the most [2]. Founded coordinates used as input for machine learning to learn a
model and then to classify ECGs. The efficiency of this approach is described in the article [1].
The authentication approach can be divided logically into two main parts that are presented on Fig 1. The first is
ECG processing and creating coordinates for machine learning. The second is learning neural networks and classifying
ECGs by coordinates.
Fig. 1 Simplified schema of authentication technology
Having prommissing result for the technology described in the article [1] a new program has been developed
using C# language. It implements the described approach so the functionality allows to register a new ECG and
authenticate a loaded one. The goal of this program and other solutions is to authenticate only one registered ECG. The
idea is to provide or not provide access to a registered user.
The functionality related to ECG processing was organized and put into a library. The library for machine
learning was created as well. It uses open source framework ML.NET under the hood. The library consumes a list of
coordinates and provides a learned model or classification results. The library made communication with machine
learning much easier since it defines contracts that used to exchange data.
It is worth describing how the algorithm decides if ECG belongs to a registered one or not. The program contains
coordinates of 10 fake ECGs that don’t belong to anybody. They are all used when registering a new ECG. New ECG
requires 10 learned models. The modal is learned based on the pair of coordinates from input ECG and one fake ECG.
Then these 10 learned models are used to identify ECG by average classification results.
The next step in research was to repeat the experiments using a bigger amount of input data. For this purpose a
new desktop application was created. It also included the ECG processing and machine learning libraries. The
application developed to read the amount of input ECG files and save intermediate and output results with learned
models into a specific folder in a file system.
This application can be described with the schema presented on Fig. 2. It realizes the idea of running isolated
experiments. That means that for each experiment creates a folder with input ECGs. On execution all artifacts are saved
to the same experiment folder. That approach allows to have isolated environment for an experiment where saved data
related only to this experiment. With this approach it is good to track changes. Having intermediate results that are
stored in the folders is very useful since it makes possible to validate a part of the application.
Fig. 1 Simplified schema of authentication technology
Having prommissing result for the technology described in the article [1] a new program has been devel-
oped using C# language. It implements the described approach so the functionality allows to register a new ECG
and authenticate a loaded one. The goal of this program and other solutions is to authenticate only one registered
ECG. The idea is to provide or not provide access to a registered user.
The functionality related to ECG processing was organized and put into a library. The library for machine
learning was created as well. It uses open source framework ML.NET under the hood. The library consumes a
list of coordinates and provides a learned model or classification results. The library made communication with
machine learning much easier since it defines contracts that used to exchange data.
It is worth describing how the algorithm decides if ECG belongs to a registered one or not. The program
contains coordinates of 10 fake ECGs that don’t belong to anybody. They are all used when registering a new
ECG. New ECG requires 10 learned models. The modal is learned based on the pair of coordinates from input
ECG and one fake ECG. Then these 10 learned models are used to identify ECG by average classification results.
The next step in research was to repeat the experiments using a bigger amount of input data. For this pur-
pose a new desktop application was created. It also included the ECG processing and machine learning libraries.
The application developed to read the amount of input ECG files and save intermediate and output results with
learned models into a specific folder in a file system.
This application can be described with the schema presented on Fig. 2. It realizes the idea of running
isolated experiments. That means that for each experiment creates a folder with input ECGs. On execution all
artifacts are saved to the same experiment folder. That approach allows to have isolated environment for an ex-
periment where saved data related only to this experiment. With this approach it is good to track changes. Having
intermediate results that are stored in the folders is very useful since it makes possible to validate a part of the
application.
The application starts with a name of an experiment that corresponds to the name of a folder where input
ECGs are stored. Then all files are read from a folder (1) and the first file loaded into memory (2). Next step is
to process ECG signals and find the coordinates of ECG (3). After that the coordinates from 10 faked persons are
loaded (4). Then loaded coordinates and the found one from the input ECG are combined together and saved into
a file that machine learning will use. Created file loaded into a folder (5). Then the file with coordinates for ma-
chine learning download from the folder and the data are formatted to 10 datasets (6). In the next step one model
is learned using prepared data (7). Learning model repeats 10 times. When all models are learned the application
takes the next ECG file from the list and executes all that steps. At the end of experiments each ECG data will
have corresponding 10 learned models.
Described application performs all calculations sequentially. Running the application on a local machine
with 2 cores 2.3 GHz each using 30 ECGs took 12 min 41 sec. On average registration of one ECG took 25 sec.
Most of the time spent learning 10 neural network models, it took 18 sec on average. The second costly part was
to process ECG and find coordinates, it took 5 sec on average. It takes a lot of time to create a record for one
ECG. If there is a need to run a lot of experiments or the amount of ECG data will rapidly increase, most of the
487
Прикладне програмне забезпечення
time would be spent waiting for results. To decrease time spending for registration of one ECG heavy parts of
application should be parallelised.
So the next step after the application that works with a big amount of data was to create an information sys-
tem that would be more efficient. It was created and built on microservices which deployed into the university cloud.
Shortly speaking it is a separated virtual machine on the university server.
The information system uses the same approaches that a desktop application does. The schema on Fig. 3 pres-
ents the way how the application was divided into microservices. Прикладне програмне забезпечення
[Введите текст]
Fig. 2 Execution workflow for the application
The application starts with a name of an experiment that corresponds to the name of a folder where input ECGs
are stored. Then all files are read from a folder (1) and the first file loaded into memory (2). Next step is to process ECG
signals and find the coordinates of ECG (3). After that the coordinates from 10 faked persons are loaded (4). Then
loaded coordinates and the found one from the input ECG are combined together and saved into a file that machine
learning will use. Created file loaded into a folder (5). Then the file with coordinates for machine learning download
from the folder and the data are formatted to 10 datasets (6). In the next step one model is learned using prepared data
(7). Learning model repeats 10 times. When all models are learned the application takes the next ECG file from the list
and executes all that steps. At the end of experiments each ECG data will have corresponding 10 learned models.
Described application performs all calculations sequentially. Running the application on a local machine with 2
cores 2.3 GHz each using 30 ECGs took 12 min 41 sec. On average registration of one ECG took 25 sec. Most of the
time spent learning 10 neural network models, it took 18 sec on average. The second costly part was to process ECG
and find coordinates, it took 5 sec on average. It takes a lot of time to create a record for one ECG. If there is a need to
run a lot of experiments or the amount of ECG data will rapidly increase, most of the time would be spent waiting for
results. To decrease time spending for registration of one ECG heavy parts of application should be parallelised.
So the next step after the application that works with a big amount of data was to create an information system
that would be more efficient. It was created and built on microservices which deployed into the university cloud.
Shortly speaking it is a separated virtual machine on the university server.
The information system uses the same approaches that a desktop application does. The schema on Fig. 3 presents
the way how the application was divided into microservices.
Fig. 2 Execution workflow for the applicationПрикладне програмне забезпечення
Fig. 3 Application breakdown to different microservices
Four microservices were created: ExperimentRunnerApi, EcgProcessingApi, LearningRunnerApi,
MachineLearning.EngineApi. The idea of partitioning was to move the time consuming calculations to a separate
microservices. The MachineLearning.EngineApi is responsible for learning models and EcgProcessingApi processes
ECG and prepares coordinates for machine learning. ExperimentRunnerApi is an orchestrator that starts the process for
selected ECG and coordinates microservices. LearningRunnerApi prepares data for machine learning and manages
work with neural models.
The communication with ExperimentRunnerApi and between microservices themself happen using HTTP
protocol. The information system uses high performance object storage MinIO to store input ECG files and all artifacts.
The system saved the same principle of isolated experiment. The schema on Fig. 4 presents the flow of requests when
the experiment is executed.
Fig. 4 Requests workflow between microservices and storage
The workflow is the following. Numbers present a sequence of requests or responses.
1) Start the experiment. The request contains the name of the experiment.
Fig. 3 Application breakdown to different microservices
488
Прикладне програмне забезпечення
Four microservices were created: ExperimentRunnerApi, EcgProcessingApi, LearningRunnerApi, Machine-
Learning.EngineApi. The idea of partitioning was to move the time consuming calculations to a separate microser-
vices. The MachineLearning.EngineApi is responsible for learning models and EcgProcessingApi processes ECG and
prepares coordinates for machine learning. ExperimentRunnerApi is an orchestrator that starts the process for selected
ECG and coordinates microservices. LearningRunnerApi prepares data for machine learning and manages work with
neural models.
The communication with ExperimentRunnerApi and between microservices themself happen using HTTP
protocol. The information system uses high performance object storage MinIO to store input ECG files and all artifacts.
The system saved the same principle of isolated experiment. The schema on Fig. 4 presents the flow of requests when
the experiment is executed.
Прикладне програмне забезпечення
Fig. 3 Application breakdown to different microservices
Four microservices were created: ExperimentRunnerApi, EcgProcessingApi, LearningRunnerApi,
MachineLearning.EngineApi. The idea of partitioning was to move the time consuming calculations to a separate
microservices. The MachineLearning.EngineApi is responsible for learning models and EcgProcessingApi processes
ECG and prepares coordinates for machine learning. ExperimentRunnerApi is an orchestrator that starts the process for
selected ECG and coordinates microservices. LearningRunnerApi prepares data for machine learning and manages
work with neural models.
The communication with ExperimentRunnerApi and between microservices themself happen using HTTP
protocol. The information system uses high performance object storage MinIO to store input ECG files and all artifacts.
The system saved the same principle of isolated experiment. The schema on Fig. 4 presents the flow of requests when
the experiment is executed.
Fig. 4 Requests workflow between microservices and storage
The workflow is the following. Numbers present a sequence of requests or responses.
1) Start the experiment. The request contains the name of the experiment.
Fig. 4 Requests workflow between microservices and storage
The workflow is the following. Numbers present a sequence of requests or responses.
1) Start the experiment. The request contains the name of the experiment.
2) Read all file names with ECGs
3) Send request to process ECG and find coordinates
4) Download ECG file content
5) Save prepared data for machine learning model
6) Success response that data for machine learning model is created
7) Send request to learn models for previously selected ECG
8) Read prepared data for learning
9) Send request to learn model for one pair of data that consists from input ECG coordinates dataset and one
fake ECG coordinates dataset.
10) Save learned model.
11) Success response that model has been learned
12) Success response that models for input ECG were learned
Microservices have the ability to process a couple of requests simultaneously so it should help to reduce
the overall execution time. Microservices deployed to a virtual machine which consumes a dedicated number of
server cores. A microservice has access to all virtual machine cores. Configuration and deployment of microser-
vices are performed with Kubernetes. Each microservice presented in one instance and there was no horizontal
scaling for loaded instances.
All communications between microservices and storage were sequential by default. Only the machine learning
instance performs learning neural models with parallel computations since ML.NET does it by default. The machine
learning instance will use all allocated cores to the virtual machine. To reduce total experiment time the requests to
learn neural models were sent in the way to learn the specific amount of models simultaneously. The idea was to find
the best configuration for a system that took minimum time for the experiment.
The experiments included registration of 30 ECGs. Table 1 presented the time results of experiments and re-
lated configurations to them.
489
Прикладне програмне забезпечення
Table 1. Experiment results
Allocated cores Simultaneously
learned models
Total experiment
time for 30 ECGs
Average time
to learn one model
2 1 12 min 8 sec 2.1 sec
2 2 10 min 31 sec 3.2 sec
2 3 9 min 2 sec 4 sec
2 4 8 min 59 sec 5.2 sec
2 5 8 min 47 sec 6.4 sec
4 1 12 min 31 sec 2.2 sec
4 2 8 min 15 sec 2.7 sec
4 3 7 min 51 sec 3.5 sec
4 4 7 min 19 sec 4.1 sec
4 5 6 min 59 sec 5 sec
6 1 15 min 28 sec 2.8 sec
6 2 9 min 39 sec 3.2 sec
6 3 8 min 52 sec 3.9 sec
6 4 8 min 48 sec 4.7 sec
8 1 20 min 3 sec 3.7 sec
8 2 12 min 18 sec 4.2 sec
8 3 10 min 29 sec 4.7 sec
8 4 9 min 33 sec 5.8 sec
Learning 5 and more neural models simultaneously didn’t give valuable improvements so the experiments
contained maximum 5 simultaneously learned models.
Analyzing the results we can see a weird pattern when with an increasing number of allocated cores
the execution time increased. There is Amdahl’s law that could be helpful here [11]. It says that performance
for the application that has some parallelizable part of the program would be increasing with adding new cores
to some specific amount of cores. This happen because execution time for all parallel work can’t be less than
longest sequential work that parralized. But in these experiments we see that efficiency dropping down with
adding new cores. This is very visible in experiments when only one model is learning simultaneously. Am-
dahl’s law also mentions this issue and says that it is related to overheads. All parallel work contains some
sequential part which performs parallelization. So sometimes with adding a lot of cores a lot of sequential work
is also added which can decrease the performance. Comparing experiment time when experiment runs on 2 and
8 cores where only one neural model learned simultaneously we see that efficiency dropped for 8 minutes. This
is the issue with learning neural models since the average time to learn one increased by 1.6 seconds. Because
learning neural models took 86% of all experiment time the parallelization delays influence a lot. The ML.NET
didn’t work efficiently in this case.
Better performance was achieved when learning a couple of models simultaneously. For this architecture
where presented only one instance of microservice the best configuration was allocation 4 cores to virtual machine and
learning 5 models simultaneously.
Taking into account the drawback of ML.NET, the better architecture should have up to 10 instances for
machine learning microservice where each instance has its own core. This helps to mitigate the issue with over-
heads.
Future research. The next step in research is to deploy and test architecture with 10 instances created for
machine learning microservice. Then configure horizontal scaling for all microservices and parallelize others.
Also expand functionality for the system by making it possible to measure the efficiency of authentication on a
big set of data. It will help to track all changes in the technology and see how new changes influence the authentication
result. It will help to improve technology iteratively.
Conclusions. While testing cloud authentication service revealed a problem with decreasing performance
when increasing the number of allocated cores to the virtual machine. The issue was in learning neural models and
it is easy to see on experiments results. The explanation to them is found in Amdahl’s law. The performance dropped
since learning models happened with overheads. It highly impacted overall experiment time since learning models take
86% of time for one ECG registering process. This forces the creation of another architecture where each instance of
machine learning microservices has its own core.
Using this authentication service in research and testing provides us confidence that the service on live produc-
tion will behave the same as on research.
490
Прикладне програмне забезпечення
References
1. Yogendra N. S., Gupta P. Biometrics Method for Human Identification Using Electrocardiogram. Advances in Biometrics: Third International
Conference, ICB 2009. – Italy, 2019. – P. 1270-1279.
2. Pei-Lun Hong, Jyun-Ya Hsiao, Chi-Hsun Chung, Yao-Min Feng, and Shun-Chi Wu. Ecg biometric recognition: template-free approaches based
on deep learning. In 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pages
2633– 2636. IEEE, 2019.
3. Sairul I Safie, John J Soraghan, and Lykourgos Petropoulakis. Electrocardiogram (ECG) Biometric Authentication Using Pulse Active Ratio
(PAR). IEEE Transactions on Information Forensics and Security. – 2011. – Vol. 6. – №4. – P. 1315 – 1322.
4. Adrian D. C. Chan, Mohyeldin M. Hamdy, Armin Badre, Vesal Badee Wavelet Distance Measure for Person Identification Using Electrocar-
diograms. IEEE Transactions on Instrumentation and Measurement. – 2008. – Vol.57. – P. 248-253.
5. Fainzilberg L.S., Potapova T.P. Computer Analysis and Recognition of Cognitive Phase Space Electrocardiographic Image. Proceeding of 6th
International Conference on Computer analysis of Images and Patterns (CAIP-95). – Prague, 1995. – P. 668-673.
6. Shih-Chin Fang, Hsiao-Lung Chan. QRS detection-free electrocardiogram biometrics in the reconstructed phase space. Pattern Recognition
Letters. – 2013. – Vol. 34. – P. 595-602.
7. Вишневский В.В., Романенко Т.Н., Кизуб Л.А. Биометрическая идентификация человека по его электрокардиограмме. Математичні
машини і системи. 2018. № 2. С. 88–95.
8. Вишневский В.В., Калмыков В.Г., Романенко Т.Н. Аппроксимация одно-, дву- и трехмерных дуг кривых параметрическими сплайнами.
Математичні машини і системи. 2015. № 4. С. 57–64.
9. Вишневський В.В., Романенко Т.М. Застосування метрики Хаусдорфа для визначення нетипових кардіоциклів у тривимірному
фазовому просторі координат вектор-кардіограми. Медична інформатика і інженерія. 2019. № 3. С. 31–36.
10. Вишневський В.В. Спосіб автоматичної автентифікації людини за її електрокардіограмою: пат. України на винахід № 117713; заявл.
15.02.17; опубл. 10.09.18, Бюл. № 17. 3.66 с.
11. Gene Amdahl. Validity of the single processor approach to achieving large scale computing capabilities, reprinted from the afips conference
proceedings, vol. 30 (atlantic city, n.j., apr. 18–20). Solid-State Circuits Newsletter, IEEE , 12:19 – 20, 02 2007.
Received 02.08.2022
About the author:
Luhovskyi Yurii Olexandrovich
postgraduate student,
Ukraine, Kyiv, Balzaka street 58, flat 43;
publications in domestic magazines - 3,
h-index 0,
https://orcid.org/0000-0002-0195-2770
Place of work:
The Institute Of Mathematical Machines and Systems Problems
National Academy Of Science Of Ukraine (IMMSP NASU).
42 Academician Glushkov Avenue,
Kyiv 03187, Ukraine. (044) 526-24-97
Прізвища та ініціали авторів і назва доповіді українською мовою:
Луговський Ю. О.
Хмарний сервіс ідентифікації людини за її електрокардіограмою
Прізвища та ініціали авторів і назва доповіді англійською мовою:
Yurii Luhovskyi
Cloud service for authentication of a person based on their electrocardiogram
|