Дослідження біометричної верифікації користувача на основі автокодерів з перевірками руху
In the current research, we continue our previous study regarding motion-based user biometric verification, which consumes sensory data. Sensory-based verification systems empower the continuous authentication narrative – as physiological biometric methods mainly based on photo or video input meet a...
Saved in:
| Date: | 2022 |
|---|---|
| Main Authors: | , |
| Format: | Article |
| Language: | English |
| Published: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2022
|
| Subjects: | |
| Online Access: | https://journal.iasa.kpi.ua/article/view/265654 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Journal Title: | System research and information technologies |
| Download file: |  |
Institution
System research and information technologies| _version_ | 1866302863953625088 |
|---|---|
| author | Havrylovych, Mariia Danylov, Valeriy |
| author_facet | Havrylovych, Mariia Danylov, Valeriy |
| author_sort | Havrylovych, Mariia |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2022-10-17T22:12:39Z |
| description | In the current research, we continue our previous study regarding motion-based user biometric verification, which consumes sensory data. Sensory-based verification systems empower the continuous authentication narrative – as physiological biometric methods mainly based on photo or video input meet a lot of difficulties in implementation. The research aims to analyze how various components of sensor data from an accelerometer affect and contribute to defining the process of unique person motion patterns and understanding how it may express the human behavioral patterns with different activity types. The study used the recurrent long-short-term-memory autoencoder as a baseline model. The choice of model was based on our previous research. The research results have shown that various data components contribute differently to the verification process depending on the type of activity. However, we conclude that a single sensor data source may not be enough for a robust authentication system. The multimodal authentication system should be proposed to utilize and aggregate the input streams from multiple sensors as further research. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2022.2.10 |
| first_indexed | 2025-07-17T10:27:59Z |
| format | Article |
| fulltext |
M.P. Havrylovych, V.Y. Danylov, 2022
128 ISSN 1681–6048 System Research & Information Technologies, 2022, № 2
UDC 004.896
DOI: 10.20535/SRIT.2308-8893.2022.2.10
RESEARCH OF AUTOENCODER-BASED USER BIOMETRIC
VERIFICATION WITH MOTION PATTERNS
M.P. HAVRYLOVYCH, V.Y. DANYLOV
Abstract. In the current research, we continue our previous study regarding motion-
based user biometric verification, which consumes sensory data. Sensory-based veri-
fication systems empower the continuous authentication narrative – as physiological
biometric methods mainly based on photo or video input meet a lot of difficulties in
implementation. The research aims to analyze how various components of sensor
data from an accelerometer affect and contribute to defining the process of unique
person motion patterns and understanding how it may express the human behavioral
patterns with different activity types. The study used the recurrent long-short-term-
memory autoencoder as a baseline model. The choice of model was based on our
previous research. The research results have shown that various data components
contribute differently to the verification process depending on the type of activity.
However, we conclude that a single sensor data source may not be enough for a ro-
bust authentication system. The multimodal authentication system should be pro-
posed to utilize and aggregate the input streams from multiple sensors as further re-
search.
Keywords: motion patterns recognition, biometric verification, recurrent autoen-
coders.
INTRODUCTION
In the modern world, we strive to automate all the systems and processes but to
keep stability, reliability, and trustworthiness of automatization on the highest
possible level. Although there are still a lot of use cases, which are impossible to
make fully automated due to ethical and other artificial intelligence issues (for
example, in the medical area), automated ensuring of the system reliability and
stability is critical in many real-world applications. There is a need for advanced
monitoring and anomaly detection applications for this purpose. The appliance of
such systems is very broad: financial operations, sensors or medical data, security
sector, and many others.
All of such systems meet some challenges during development as there is a
lot of uncertainty. First, you never know what anomaly data should look like,
despite some apparent cases, especially if you have an inlier type of anomaly.
Second, all the processes change over time, so the data – and if new data points
are not similar to the previous ones – it does not mean they are anomalies. As
well, real-world data frequently has a pretty complex structure and is heterogeneous,
leading to the inefficiency of classic statistical and machine learning methods.
From the machine learning perspective, it is impossible to apply a classical
supervised approach, as in most cases, only regular data examples are available.
In this study, we will look at the user motion-based verification problem. In this
application, it is not feasible to solve such a problem in a classic supervised man-
Research of autoencoder-based user biometric verification with motion patterns …
Системні дослідження та інформаційні технології, 2022, № 2 129
ner due to the huge population and inability to compare specific users with all
others (as well, there could be personal information concerns). Therefore, only
self-supervised or unsupervised approaches may fit for problem solutions.
The research object is motion-based user biometric verification based on
sensor data.
The purpose is to research and analyze various components, which affect the
result of verification and understanding how sensor data (accelerometer) and its
components may express human behavioral patterns with different activity types.
LITERATURE REVIEW
According to [1], biometric data can be physiological: like iris, face, fingerprint,
and behavioral: like motion or gate patterns, mouse or keyboard movements, or
even both (voice data). For example, in [2], the authors propose interesting ap-
proaches for biometric identification systems based on circular kernel principal
component analysis, Chebyshev transforms hashing, and Bose–Chaudhuri–
Hocquenghem error-correcting codes for ear-based biometrics.
Nevertheless, many biometric-based verification systems are explicit and re-
quire specific actions by the person or from an external supervisor. Many biomet-
ric approaches utilize computer vision approaches like iris, ear, face, or gate-
based, which require a camera sensor. A camera sensor is not convenient for con-
tinuous authentication, and for explicit authentication, somebody, for example,
should take a photo. Usage of the camera for continuous and implicit authentica-
tion may cause problems: it should process a considerable amount of data (video
or image streaming), there are issues with saving and processing personal data,
and a person may feel uncomfortable as they are constantly watched. On the
other hand, the motion-based or mouse/keyboard movement authentication, that
can be implicit and be conducted all the time in the background without user
intervention, such data is more lightweight, and in most cases do not need addi-
tional sensor installation, because most devices already have all the sensors for
other purposes.
We can reformulate the motion-based user verification task as a time series
classification problem. The anomaly detection on time series data has already
been a research topic in many fields. As well, time-series data anomalies, accord-
ing to [3], have their taxonomy: point outliers, subsequent, and time series. In user
verification, the subsequent outliers are of the most significant interest, as we can
reformulate the task from a pattern recognition and classification point of view.
Worth mentioning that the time series classification task has a lot of applications,
besides anomaly detection, especially in the field of motion sensory data.
In [4], we can see the survey on existing machine learning algorithms, both
statistical and deep learning, for anomaly detection on univariate time series. In
the case of univariate data – deep learning methods do not overperform classic
machine learning and statistical methods in the survey. However, for multivariate
and heterogeneous time series anomaly detection, according to [5], deep learning
methods are the ones with the greatest accuracy. In [6], we can review the com-
parative study for anomaly detection on multivariate time series for LSTM and
CNN-based (Temporal Convolutional Network) architectures, and the CNN-
based solution even slightly outperformed the recurrent one. In [7], the CNN-
M.P. Havrylovych, V.Y. Danylov
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 130
based and LSTM model are also compared for biometric motion-based verifica-
tion on various datasets and activity types. The hybrid LSTM-CNN-based neural
network was proposed, as it allows to capture of temporal dependencies on more
complex features extracted with convolutional layers.
In [4], the authors measures not only predictions quality metrics but the
computational time for model training and inference and conclude that amongst
all deep learning approaches, the best combination of quality metrics and compu-
tational time has an autoencoder-based approach, which is very important in the
angle of verification and security to achieve best results with the lowest latency.
In our previous research [8], we conducted a comparative analysis of various
types of recurrent autoencoders for user biometric verification. We compared
them to classical machine learning algorithms such as Isolation Forest and One-
Class SVM. In [9], the authors designed the user personalization and biometric
verification system based on geometric concepts on the convex hull. They men-
tioned the SVM as a state-of-the-art approach (in 2012), but the one which re-
quires a lot of time for training, thus do not fit for on-the-fly verification model
training. In our research, One-Class SVM showed the lowest and most unstable
performance, though. In contrast, the variational autoencoders showed the best
performance; even so, all types of autoencoders provide comparable results.
It is essential to mention that despite the continuous authentication idea
gaining popularity, some research shows that it still underperforms compared to
explicit biometric systems. In the [10], authors mention that the multimodel au-
thentication systems are a better choice. It allows achieving the highest level of
system robustness, increasing flexibility, and mitigating drawbacks of every veri-
fication modality. In [11], authors propose multimodal authentication based on
face, motion patterns, and touch stroke. Motion-based authentication systems can
also be boosted using multiple sensors [10, 13], such as accelerometers and gyro-
scopes. The combination of multiple sensory data increases the accuracy of the
motion-based authentication system.
It is relevant to note that the deep learning model, especially autoencoder-
based, may be used as a single base model for many purposes and may have in-
teresting applications – for example, activity classification, and more exotic
things like detecting smoking events [12] or gender of the person [13]. Overall,
having a single model for a couple of purposes is a tendency in the modern ma-
chine learning and data science field. It allows to reduce the cost and add addi-
tional context for every task-specific downstream model appliance at the same
time.
In this study, we wanted to continue the in-depth analysis of biometric mo-
tion-based user verification and conduct detailed experiments and research re-
garding how accelerometer motion data and its components describe the person
using recurrent autoencoders.
MATERIALS AND METHODS
Based on our previous research [8], we will use the undercomplete autoencoder as
a baseline model architecture with recurrent LSTM layers types, as all types of
autoencoders (undercomplete, variational, contractive) showed comparable results
between each other.
Research of autoencoder-based user biometric verification with motion patterns …
Системні дослідження та інформаційні технології, 2022, № 2 131
Autoencoder architecture consists of an encoder and decoder. Autoencoders
are learned to reconstruct the input data points from some hidden space. There-
fore, the optimization task objective is to minimize reconstruction error:
n
i
ii xedxE
1
))(( , (1)
where nxx ,,1 is data rows, d is the decoder, and e is the encoder with some
parameters.
In the undercomplete autoencoder type, the data is encoded in lower-
dimensional space. Such compression guarantees that the model will not blindly
memorize the train set but will learn the proper feature space and data embedding,
which later as well can be used for various purposes. The decoder’s purpose is to
recreate the sample from an encoded example.
The optimization process of encoder and decoder parameters is done with
classic backpropagation using gradient descent-based algorithms to minimize
(1) [14].
After model training, the decision threshold ε should be set. This threshold
will be used at the model inference step: the data points that reconstruction error
will be higher than this threshold would be considered anomaly or non-self class,
and self otherwise. The threshold setting process is dependent on the use case and
varies for different applications. It can be set manually by an expert or some
knowledge keeper or automatically based on the error distribution on some prede-
fined dataset.
We will use an autoencoder with LSTM layers. The LSTM layer contains
cells with specific internal structures. The memory cell contains three gates: input,
output, and forget. The input controls the input activations when the output con-
trols the output flow. The forget gate controls what information memory cells
should forget and what to pass further through the network, which theoretically
may solve the problems of long sequences, which is a known problem for vanilla
recurrent neural networks, which fail to learn from big sequences [15].
EXPERIMENTS
Dataset: open-source dataset [16] with accelerometer data (52 Hz) from
15 people with seven activity types. In [9], the original paper that presents the
used dataset, the HAR (human activity recognition) task should be solved first, as
training the algorithm on all types of activities will mostly bring additional noise
to data and decrease the metrics due to not enough amount of data points per each
activity. The problem can be understood as detecting the unique and distinguish-
ing patterns of a specific person’s motions.
For understanding accelerometer data and human movement patterns, we
will train the model separately for each axis (x, y, z), axis pairs (xy, yz, xz), and all
three axes — xyz for different types of activities.
For deep learning models, we split data in overlapping on 50 percent win-
dows with a length of 52 (because of accelerometer frequency).
We split the original dataset into a 33% share for the test set and the rest for
the train.
M.P. Havrylovych, V.Y. Danylov
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 132
An accelerometer measures changes in velocity along one axis (Fig.1). The
values reported by the accelerometers are measured
in increments of the gravitational acceleration, with
the value 1,0 representing an acceleration of 9,8 me-
ters per second (per second) in the given direction.
Depending on the direction of the acceleration, the
sensor values may be positive or negative [17].
Autoencoder model architecture is in Table 1
below. As an activation function, the hyperbolic tan-
gent was used. The model was trained in 10 epochs,
with Adam optimizer and mean absolute error loss.
We build autoencoders with python on Keras library
and Tensorflow backend [17].
The dropout rate was 0,4.
T a b l e 1 . The used model architecture with layers’ type, shape, and amount
of params
Type Layer Shape # of Params
LSTM (None, 52, 20) 1760
LSTM (None, 10) 1240
RepeatVector (None, 52, 10) 0
Dropout (None, 52, 10) 0
LSTM (None, 52, 10) 840
LSTM (None, 52, 20) 2480
TimeDistributed (None, 52, 3) 21
The threshold formula was used as in [8]:
)(
1
i
N
i
i MAEstdn
MAET
, ,
where MAE is the mean absolute error between ground truth and predicted
sample; std – standard deviation, and N is the number of samples in the training
dataset.
As model evaluation metrics [19], the EER (equal error rate), FAR (false ac-
cept rate) and FRR (false reject rate), and ROCAUC (area under the curve) were
chosen, which are typical for assessing the biometric system quality:
TNFP
FP
FPRFAR
;
FNTP
FN
FNRFRR
.
Equal error rate (EER) (illustrated in Fig. 2) is obtained by adjusting the sys-
tem’s detection threshold to equalize FAR and FRR. The EER is calculated using
the following formula:
2
FRRFAR
EER
,
where || FRRFAR is the smallest value [20].
Y
Z
X
Fig. 1. Accelerometer axis
demonstration on the general
smartwatch/bracelet [12]
Research of autoencoder-based user biometric verification with motion patterns …
Системні дослідження та інформаційні технології, 2022, № 2 133
RESULTS
For analysis, we considered only 1, 3, 4, and 7 types of activities. We filtered out
2,5, and 6 activities ( Standing Up, Walking and Going updown stairs, Going
UpDown Stairs, Walking and Talking with Someone) as less than 200 data points
were presented in the train set and less than 100 in the median for 15 users
(Table 2).
T a b l e 2 . Amount of samples in train set (median value for 15 users) for
various activities
Activity 1 2 3 4 5 6 7
# in trainset
(median for 15 users) 133 89 301 599 86 62 860
We can review the results for every metric in the tables below: EER and
AUC (Table 3), FAR, and FRR (Table 4). We have 15 users in the dataset there-
fore we report the average performance metric.
T a b l e 3 . Average EER and AUC (on 15 users) for various activities
1 activity 7 activity 3 activity 4 activity 1 activity 7 activity 3 activity 4 activity Axis data
for train Mean EER Mean AUC
x-axis 0,202 0,350 0,407 0,318 0,837 0,671 0,617 0,714
y-axis 0,168 0,276 0,364 0,345 0,868 0,756 0,657 0,687
z-axis 0,161 0,358 0,394 0,237 0,873 0,669 0,631 0,791
x-axis_y-axis 0,084 0,200 0,346 0,280 0,938 0,833 0,680 0,759
x-axis_z-axis 0,101 0,254 0,358 0,191 0,924 0,777 0,673 0,841
y-axis_z-axis 0,081 0,223 0,349 0,225 0,938 0,807 0,680 0,807
x-axis_
y-axis_z-axis 0,074 0,189 0,334 0,197 0,949 0,846 0,689 0,830
Fig. 2. Illustration of the EER calculation from [19]
ROC w\ EER 0,229,AUROC 0,856, ACC 0,776
1,0
0,8
0,6
0,4
0,2
0,0
0,0 0, 0,4 0,6 0,8 1,0
False Positive Rate
T
ru
e
P
os
it
iv
e
R
at
e
M.P. Havrylovych, V.Y. Danylov
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 134
T a b l e 4 . Average FAR (on 15 users) for various activities
1 activity 7 activity 3 activity 4 activity 1 activity 7 activity 3 activity 4 activity Axis data
for train Mean FAR Mean FRR
x-axis 0,260 0,527 0,589 0,459 0,067 0,132 0,177 0,113
y-axis 0,207 0,388 0,526 0,517 0,058 0,101 0,160 0,108
z-axis 0,182 0,534 0,564 0,307 0,071 0,128 0,174 0,112
x-axis_y-axis 0,066 0,219 0,464 0,374 0,058 0,116 0,176 0,107
x-axis_z-axis 0,076 0,312 0,488 0,196 0,075 0,133 0,166 0,122
y-axis_z-axis 0,059 0,269 0,459 0,259 0,064 0,117 0,181 0,127
x-axis_
y-axis_z-axis 0,030 0,188 0,452 0,222 0,073 0,120 0,171 0,117
DISCUSSION
The obtained results show that all axes contribute to the final results and hold im-
portant information about human patterns. Therefore, training on all three axes
shows the best performance.
Another interesting thing we can notice is that for different types of activities
– different axis brings more value. For example, in walking (4-th activity), we can
say that the model trained on the z-axis only or on another axis in combination
with the z-axis has the best performance if looking at EER and AUC. On the other
side, for the 7th activity (Talking While Standing), the y-axis brings the most value.
As well, if looking at the performance of models trained on the x-axis only,
they always have the lowest performance compared to others. Still, in combina-
tion with the y-axis, the performance increases a lot.
The best results were shown for 1 activity (Working at Computer), but this can
be related, that this type of activity has the highest amount of training samples.
Overall, the performance is highly correlated with the number of training
samples (Table 1), which may point out for need in artificial synthetic data gen-
eration for model training, because for the personalization system is important to
be able to work and be reliable as fast as possible. The continuous training of the
model during the system is alive should be considered to prevent model and data
drift and allow the system to prevent the cold start when there are not that many
available training samples.
CONCLUSIONS
In this research, we conducted an in-depth analysis of different components on
human motion patterns from sensory data (accelerometer in our case) and whether
we can extract distinguishing person patterns from such data and use it for bio-
metric verification systems.
The deep learning approaches have already proved their applicability and
stable performance in such cases. Still, as already mentioned, the motion-based
authentication shows lower accuracy than other biometric verification (e.g.,
physiological), but this does not mean that motion-based verification should not
be used. The solution to overcome problems drawbacks of various types of verifi-
Research of autoencoder-based user biometric verification with motion patterns …
Системні дослідження та інформаційні технології, 2022, № 2 135
cation – is a multimodal authentication system, which increases stability, robust-
ness, and flexibility for customization in different environments.
Looking at the metrics for different accelerometer data components and ac-
tivities, we can see that every axis contributes to the final result not equally. De-
pending on the activity type, different features are important, proving that we
probably need a multi-stage system with preliminary human activity classification
in case of motion-based verification. The advantage of the autoencoder model is
that single model can be used for both tasks without the need to train different
models.
Further research should be done to create a sensory-based authentication sys-
tem that utilizes multiple sensors. Such approach should increase the quality of
the system but keep the continuous option.
Additional analysis of evaluation metrics should also be done, as there are
raising concerns regarding commonly used metrics and biometric evaluation
framework, as they may lead to incorrect decisions and be misleading.
REFERENCES
1. S. Minaee, A. Abdolrashid, H. Su, M. Bennamoun, and D. Zhang, Biometrics Rec-
ognition Using Deep Learning: A Survey. 2021. [Online]. Available:
https://arxiv.org/abs/1912.00271. (Accessed on: Feb 15, 2022).
2. L. Olanrewaju, O. Oyebiyi, S. Misra, R. Maskeliunas, and R. Damasevicius, “Secure
ear biometrics using circular kernel principal component analysis, Chebyshev trans-
form hashing and Bose–Chaudhuri–Hocquenghem error-correcting codes”, Signal,
Image and Video Processing, vol. 14, no. 5, pp. 847–855, 2020. doi:
10.1007/s11760-019-01609-y.
3. A. Blázquez-García, A. Conde, U. Mori, and J.A. Lozano, “A Review on Out-
lier/Anomaly Detection in Time Series Data”, ACM Computing Surveys, vol. 54, no.
3, pp. 1–33, 2021. doi: 10.1145/3444690.
4. M. Braei and S. Wagner, Anomaly Detection in Univariate Time-series: A Survey on
the State-of-the-Art. 2020. [Online]. Available: https://arxiv.org/abs/2004.00433.
(Accessed on: Feb 15, 2022).
5. M. Braei, “Anomaly detection of time series: A comparison of statistical vs classical
machine learning vs deep learning approaches”, Unpublished, 2019. doi:
10.13140/RG.2.2.17687.80801.
6. S. Gopali, F. Abri, S. Siami-Namini, and A. Namin, “A Comparative Study of De-
tecting Anomalies in Time Series Data Using LSTM and TCN Models”, arXiv.org,
2021. [Online]. Available: https://arxiv.org/abs/2112.09293. (Accessed on Feb 15, 2022).
7. S. Mekruksavanich and A. Jitpattanakul, “Deep Learning Approaches for Continu-
ous Authentication Based on Activity Patterns Using Mobile Sensing”, Sensors,
vol. 21, no. 22, pp. 7519, 2021. doi: 10.3390/s21227519.
8. M. Havrylovych, V. Danylov, and A. Gozhyj, “Comparative Analysis of using Re-
current Autoencoders for User Biometric Verification with Wearable Accelerome-
ter”, Proceedings of the 9th International Conference “Information Control Systems
& Technologies”, pp. 358–370, Sept. 2020.
9. P. Casale, O. Pujol, and P. Radeva, “Personalization and user verification in wear-
able systems using biometric walking patterns”, Personal and Ubiquitous Comput-
ing, vol. 16, no. 5, pp. 563–580, 2011. doi: 10.1007/s00779-011-0415-z.
10. M. Abuhamad, A. Abusnaina, D. Nyang, and D. Mohaisen, “Sensor-Based Continu-
ous Authentication of Smartphones’ Users Using Behavioral Biometrics: A Contem-
porary Survey”, IEEE Internet of Things Journal, vol. 8, no. 1, pp. 65–84, 2021. doi:
10.1109/jiot.2020.3020076.
11. Z. Akhtar, A. Buriro, B. Crispo, and T.H. Falk, “Multimodal smartphone user au-
thentication using touchstroke, phone-movement and face patterns”, 2017 IEEE
Global Conference on Signal and Information Processing (GlobalSIP), 2017. doi:
10.1109/globalsip.2017.8309185.
M.P. Havrylovych, V.Y. Danylov
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 136
12. C.A. Cole, D. Anshari, V. Lambert, J.F. Thrasher, and H. Valafar, “Detecting Smok-
ing Events Using Accelerometer Data Collected Via Smartwatch Technology: Vali-
dation Study”, JMIR mHealth and uHealth, vol. 5, no. 12, pp. e189, 2017. doi:
10.2196/mhealth.9035.
13. A. Mostafa, T. Barghash, A. Assaf, and W. Gomaa, “Multi-sensor Gait Analysis for
Gender Recognition”, Proceedings of the 17th International Conference on Informat-
ics in Control, Automation and Robotics, 2020. doi: 10.5220/0009792006290636.
14. I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning Adaptive Computation
and Machine Learning series. 2016. 775 p.
15. H. Sak, A. Senior, and F. Beaufays, “Long short-term memory recurrent neural net-
work architectures for large scale acoustic modeling”, Interspeech 2014, 2014. doi:
10.21437/interspeech.2014-80.
16. UCI machine learning repository. [Online]. Available: https://archive.ics.uci.edu/
ml/datasets/Activity+Recognition+from+Single+Chest-Mounted+Accelerometer.
(Accessed on: Feb 15, 2022).
17. F. Chollet, Building Autoencoders in Keras. 2016. [Online]. Available:
https://blog.keras.io/building-autoencoders-in-keras.html. (Accessed on: Feb 15, 2022).
18. Getting Raw Accelerometer Events. [Online]. Available: https://developer.apple.
com/documentation/coremotion/getting_raw_accelerometer_events\
19. S. Sugrim, C. Liu, M. McLean, and J. Lindqvist, “Robust Performance Metrics for
Authentication Systems”, Proceedings 2019 Network and Distributed System Secu-
rity Symposium, 2019. doi: 10.14722/ndss.2019.23351.
20. Y. Hong and R. Kumar, Performance Evaluation Metrics for Biometrics-based Au-
thentication Systems. 2021. [Online]. Available: http://hdl.handle.net/10066/23535.
(Accessed on: Feb 15, 2022).
Received 16.02.2022
INFORMATION ON THE ARTICLE
Mariia P. Havrylovych, ORCID: 0000-0002-9797-2863, Institute for Applied System
Analysis of the National Technical University of Ukraine “Igor Sikorsky Kyiv Polytech-
nic Institute”, Ukraine, e-mail: mariia.havrylovych@gmail.com
Valeriy Ya. Danylov, ORCID: 0000-0003-3389-3661, Institute for Applied System
Analysis of the National Technical University of Ukraine “Igor Sikorsky Kyiv Polytech-
nic Institute”, Ukraine, e-mail: danilov1950@ukr.net
ДОСЛІДЖЕННЯ БІОМЕТРИЧНОЇ ВЕРИФІКАЦІЇ КОРИСТУВАЧА НА ОСНОВІ
АВТОКОДЕРІВ З ПЕРЕВІРКАМИ РУХУ / М.П. Гаврилович, В.Я. Данилов
Анотація. Продовжено попереднє дослідження щодо біометричної перевірки
користувача на основі руху з використанням сенсорних даних. Системи сенсор-
ної верифікації розширюють можливості неперервної автентифікації, оскільки
фізіологічні біометричні методи, в основному засновані на фото- або відеода-
них, стикаються з багатьма труднощами під час реалізації. Мета дослідження —
проаналізувати як різні компоненти сенсорних даних від акселерометра впли-
вають і сприяють визначенню процесу унікальних моделей руху людини та
розуміння того, як вони можуть виражати моделі поведінки людини з різними
видами активності. Як базову модель використано рекурентний автокодуваль-
ник довгої-короткої пам’яті. Вибір моделі ґрунтується на попередніх дослі-
дженнях. Результати дослідження показали, що залежно від виду діяльності
різноманітні компоненти даних мають різний внесок. Зроблено висновок, що
одного джерела даних датчика може бути недостатньо для надійної системи
автентифікації. Для подальших досліджень слід запропонувати мультимодаль-
ну систему автентифікації, яка повинна використовувати та об’єднувати вхідні
потоки від кількох датчиків.
Ключові слова: розпізнавання образів руху, біометрична верифікація, рекуре-
нтні автокодувальники.
|
| id | journaliasakpiua-article-265654 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2025-07-17T10:27:59Z |
| publishDate | 2022 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/5f/9b7f8f999090a8c232fa28a2c4b3f25f.pdf |
| spelling | journaliasakpiua-article-2656542022-10-17T22:12:39Z Research of autoencoder-based user biometric verification with motion patterns Дослідження біометричної верифікації користувача на основі автокодерів з перевірками руху Havrylovych, Mariia Danylov, Valeriy motion patterns recognition biometric verification recurrent autoencoders розпізнавання образів руху біометрична верифікація рекурентні автокодувальники In the current research, we continue our previous study regarding motion-based user biometric verification, which consumes sensory data. Sensory-based verification systems empower the continuous authentication narrative – as physiological biometric methods mainly based on photo or video input meet a lot of difficulties in implementation. The research aims to analyze how various components of sensor data from an accelerometer affect and contribute to defining the process of unique person motion patterns and understanding how it may express the human behavioral patterns with different activity types. The study used the recurrent long-short-term-memory autoencoder as a baseline model. The choice of model was based on our previous research. The research results have shown that various data components contribute differently to the verification process depending on the type of activity. However, we conclude that a single sensor data source may not be enough for a robust authentication system. The multimodal authentication system should be proposed to utilize and aggregate the input streams from multiple sensors as further research. Продовжено попереднє дослідження щодо біометричної перевірки користувача на основі руху з використанням сенсорних даних. Системи сенсорної верифікації розширюють можливості неперервної автентифікації, оскільки фізіологічні біометричні методи, в основному засновані на фото- або відеоданих, стикаються з багатьма труднощами під час реалізації. Мета дослідження — проаналізувати як різні компоненти сенсорних даних від акселерометра впливають і сприяють визначенню процесу унікальних моделей руху людини та розуміння того, як вони можуть виражати моделі поведінки людини з різними видами активності. Як базову модель використано рекурентний автокодувальник довгої-короткої пам’яті. Вибір моделі ґрунтується на попередніх дослідженнях. Результати дослідження показали, що залежно від виду діяльності різноманітні компоненти даних мають різний внесок. Зроблено висновок, що одного джерела даних датчика може бути недостатньо для надійної системи автентифікації. Для подальших досліджень слід запропонувати мультимодальну систему автентифікації, яка повинна використовувати та об’єднувати вхідні потоки від кількох датчиків. The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2022-08-30 Article Article application/pdf https://journal.iasa.kpi.ua/article/view/265654 10.20535/SRIT.2308-8893.2022.2.10 System research and information technologies; No. 2 (2022); 128-136 Системные исследования и информационные технологии; № 2 (2022); 128-136 Системні дослідження та інформаційні технології; № 2 (2022); 128-136 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/265654/261714 |
| spellingShingle | розпізнавання образів руху біометрична верифікація рекурентні автокодувальники Havrylovych, Mariia Danylov, Valeriy Дослідження біометричної верифікації користувача на основі автокодерів з перевірками руху |
| title | Дослідження біометричної верифікації користувача на основі автокодерів з перевірками руху |
| title_alt | Research of autoencoder-based user biometric verification with motion patterns |
| title_full | Дослідження біометричної верифікації користувача на основі автокодерів з перевірками руху |
| title_fullStr | Дослідження біометричної верифікації користувача на основі автокодерів з перевірками руху |
| title_full_unstemmed | Дослідження біометричної верифікації користувача на основі автокодерів з перевірками руху |
| title_short | Дослідження біометричної верифікації користувача на основі автокодерів з перевірками руху |
| title_sort | дослідження біометричної верифікації користувача на основі автокодерів з перевірками руху |
| topic | розпізнавання образів руху біометрична верифікація рекурентні автокодувальники |
| topic_facet | motion patterns recognition biometric verification recurrent autoencoders розпізнавання образів руху біометрична верифікація рекурентні автокодувальники |
| url | https://journal.iasa.kpi.ua/article/view/265654 |
| work_keys_str_mv | AT havrylovychmariia researchofautoencoderbaseduserbiometricverificationwithmotionpatterns AT danylovvaleriy researchofautoencoderbaseduserbiometricverificationwithmotionpatterns AT havrylovychmariia doslídžennâbíometričnoíverifíkacííkoristuvačanaosnovíavtokoderívzperevírkamiruhu AT danylovvaleriy doslídžennâbíometričnoíverifíkacííkoristuvačanaosnovíavtokoderívzperevírkamiruhu |