You are on page 1of 7

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/338646068

User Authentication for Computer Security Based on Sound-to-Feature


Algorithm and Artificial Neural Network Using Bone Conduction

Article in SSRN Electronic Journal · January 2019


DOI: 10.2139/ssrn.3626988

CITATIONS READS

2 899

10 authors, including:

Jessica Velasco Edmon Fernandez


Technological University of the Philippines Mapúa Institute of Technology
39 PUBLICATIONS 354 CITATIONS 24 PUBLICATIONS 130 CITATIONS

SEE PROFILE SEE PROFILE

Ira Valenzuela Lean Karlo Tolentino


De La Salle University Technological University of the Philippines
48 PUBLICATIONS 434 CITATIONS 103 PUBLICATIONS 597 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Jessica Velasco on 17 January 2020.

The user has requested enhancement of the downloaded file.


User Authentication for Computer Security
Based on Sound-to-Feature Algorithm and
Artificial Neural Network Using Bone
Conduction
J. Velasco, E. Fernandez, G. J. Adiao, E. B. Alarilla, P. C. L. Cano, M. C. De Lara, J. C. Grande, J.
Nuevo,
I. Valenzuela and L. K. Tolentino
Technological University of the Philippines, Manila, Philippines
jessica_velasco@tup.edu.ph

Abstract—A user authentication system for unlocking and a lot of things had gone complex. People demand to
and security of desktop computer using bone conduction own devices which are technology-based. Due to a low
through skull is presented in this study. It is an electronic standard of living and poverty, a wide-ranging number
device that outputs an identity based on the characteristic of crimes transpired. The most common crime is
frequency response gathered. White noise will be used as
an input provided that the external noise is filtered
robbery. Stealing of personal belongings especially
successfully using a laryngophone. It will enter the skull cellphones, computers and money has led to the idea of
through the bone conduction speakers and the microphone having password and authentication in which the first
will pick up the recorded frequency characteristic was applied in Compatible Time-Sharing System
response. This response is then fed to the computer and (CTSS) in 1961 at the Massachusetts Institute of
will be processed by the Sound-to-Feature algorithm for its Technology [1]. The first passwords were simple and
feature extraction. This noise will be classified by the can be easily retained, since complicated hacking
Artificial Neural Network (ANN) algorithm so that it will networks and password-cracking programs are not yet
identify and authenticate the user. For testing and available. Meanwhile, there is a demand for a much-
validation, thirty persons, who were used as subjects, wore
laryngophones which are placed at the back of their heads
sophisticated security.
for gathering of audio samples in WAV format. Each DNA matching, Shape of the ear, Retina recognition,
person has thirty samples with a total of 900 samples. The Face recognition, Fingerprint recognition, Finger
samples were all trained in a database wherein the Geometry Recognition, Behavioral Recognition, Hand
proponents can set who among the thirty people can Geometry Recognition, Olfactory Biometric, Signature
unlock the desktop computer. To be able to test the Recognition, Typing Recognition, Vein Recognition,
significant difference and similarity between the WAV files Speaker Recognition and Speech Recognition [2] are
of each person, the Explore and Score method of the some of biometrics, which are mentioned in literatures,
Sound Analysis Pro (SAP) 2011 was used. A round-robin that are used for authentication but replicas of these
test which is equal to 435 tests was initiated to show that
each subject has a unique wav file. From the 435
biometrics can easily be produced. In addition, the bone
comparisons, all were proven unique and failed the conduction technology of sound through the user’s skull
similarity test achieving an accurate way of identifying the [3] as a user identification and authentication system for
user. biometric systems using eyewear computers was
proposed and constructed. It implemented an algorithm
Index Terms—Artificial Neural Network; Bone called Mel Frequency Cepstral Coefficient (MFCC) for
Conduction; Sound-to-Feature Algorithm; User feature extraction and 1-Nearest Neighbor as classifier
Authentication. to control and analyze the response of user’s skull. It
was tested in a controlled environment.
I. INTRODUCTION A user authentication system using the bone
conduction through skull based on Sound-to-Feature for
During early times, people did not necessarily need to unlocking the computer (PC) is presented in this study.
provide passwords for personal belongings especially The system is tested in an open area wherein external
those expensive ones. As time goes by, technology grew noise is present. White noise will be used as an input to

81
the bone conduction transducer to provide the B. Hardware Development
characteristic frequency response. Since the system will The audio amplifier as shown in Figure 3 strengthens
be tested in an open area, external noises will be filtered low-power audio signals to a strong level which can
using noise reduction. Artificial Neural Networks drive loudspeakers and can make the signal audible to
(ANN) algorithm will be used to identify the user. listeners. Meanwhile, Figure 4 shows the schematic
design which illustrates a low noise pre-amplifier which
II. METHODOLOGY is ideally suited to low impedance (600 ohms nominal)
microphone. This pre-amplifier helps to boost signal
A. Research Design coming from a low amplitude signal. It also allows the
The researcher will input white noise as an audio mic lead to pick up noise with long cable runs or in
signal. The audio frequency response will be picked up hostile environment.
by the laryngophone (as shown in Figure 2) that
contains the bone conduction speaker. The recorded
frequency response will be analyzed and utilized in
MATLAB for sound-to-feature extraction. These
features will be classified by the ANN algorithm and
will be used to authenticate or unlock a computer.

Figure 3: Schematic Diagram of Audio Amplifier

Figure 4: Schematic Diagram of Low Noise Mic Pre-Amplifier

Figure 1: Block Diagram of the System

Figure 5: Schematic Diagram of White Noise Generator

C. Software Development
The first part of the software uses Zelscope to save
Figure 2: Throat Vibration Microphone Transducer or Laryngophone the wav files of the user’s samples. Next, for the

82
database of the system, MATLAB was used to import system.
the wav files acquired from the Zelscope in the To show a great difference with the data gathered, the
program. These samples will be named and trained wav files of each of the samples were analyzed using
meticulously so that samples from different persons will MATLAB and the fast Fourier transform of the wav
not overlap the other samples. For the verification of the files were plotted in a pile.
database, the researchers checked the database
information or even import a sample and then would try
F. Evaluation of Data
to recognize it if it outputs accurately.
For the unlocking system, the user’s frequency The degree of similarity of the trials of each subject is
response will be acquired in the Zelscope through tested using Sound Analysis Pro 2011 (SAP2011),
laryngophone. Artificial Neural Network was used as measurement software of audio similarity [6],
the main algorithm to classify the wav file coming from specifically the Explore and Score method. Thirty
the Zelscope. This frequency response will be loaded in comparisons were made for this test. The hierarchies of
the system then it will unlock the computer if the user is pairing are two consecutive non-repeating numbers (1-
set to be the administrator/owner of the computer. 2, 3-4, 5-6, etc.), two consecutive non-repeating even
Otherwise, it will not unlock the computer. numbers (2-4, 6-8, 10-12, etc.), then two consecutive
non-repeating odd numbers (1-3, 5-7, 9-11, etc.), and
first to last number (1-30).
The degree of similarity is also tested on the means of
the trials of the subjects to show that each subject has a
Figure 6: Block Diagram of Creating the Database of the System unique wav file. It is a round-robin test that is equal to
four hundred thirty-five tests.
D. Device Testing
III. RESULTS AND DISCUSSION
Saving and Storing the Samples in the Database. The
In this part, uniqueness test was performed to show
volume of white noise generator must be adjusted first
the difference with the gathered data. Similarity test was
to its desired amount. The laryngophone must be placed
also done to justify that the degree of similarity of the
at the back of the head. Its push-to-talk button must be
trials done on one subject are identical and the degree of
pressed to lessen the noise from the environment. Then,
similarity of the means of the wav files on each subject
the bone conduction device must be placed in contact
are significantly different. Evaluation on the wav files
with any part of the skull. It will gather thirty (30)
of identical twins is also presented.
audio samples per user that will serve as the database of
the computer system. The thirty samples per user will A. Fast Fourier Transform (FFT) of WAV Files
have their own folder or directory where they were The characteristic frequency responses gathered from
recorded in the MATLAB database. each individual are in wav file form. The Fast Fourier
Process on Using the Device. The that will unlock the Transform (FFT) of each WAV files was obtained via
computer needs to input the wav file which is generated MATLAB.
by the Zelscope. The recorded data by the user will be
uploaded in the Dialog Box of the Graphic User
Interface of the MATLAB Artificial Neural Network. B. Uniqueness Test
The Neural Network will sort the information gathered In order to show difference between the gathered
to identify whether the data was authorized to unlock characteristic frequency responses, the FFT of these
the computer or not. If the system recognizes the user’s wav files undergone a uniqueness test where in the
frequency response, it will unlock the computer. graphs were piled as shown in Figures 7 and 8.
However, if the system does not recognize the user’s The (a) shows the magnitude of the FFT and (b)
frequency response, it will proceed as an unknown shows the phase of the FFT. It can be seen in Figure 7
guest. that the sample of Subject 1 has the widest bandwidth
and highest magnitude compare to other samples. It can
E. Data Gathering Procedure also be seen in Figure 8 that the sample of Subject 4 has
the widest bandwidth and the sample of Subject 5 has
The data were gathered from thirty different people the highest magnitude. The results show that there are
with thirty samples each. These samples were the distinct differences between the wav files.
frequency responses of the person’s skull. The person
will wear the headset which includes the bone
conduction speaker that emits white noise and the
microphone for recording the frequency responses. The
recorded date will be saved to the database of the

83
3, the first fifteen comparisons are the highest ranked
data and the succeeding fifteen comparisons are the
lowest ranked data. This shows that the range of
similarity is from 3%-90%. In addition, the highest
ranked pair, pair no. 85, happened to be twins. The
evaluation of the data showed that 90% degree of
similarity is still not enough to unlock another subject’s
account.
The degree of similarity is tested on the trials of
Subject 1 as shown in Table 1. Thirty comparisons were
made for this test.
The results show that the highest degree of similarity
is 99% and the lowest degree of similarity is 97%. This
means that the range of similarity for Subject 1 is from
97%-99%. Also, 97% is the least degree of similarity
that is enough to unlock Subject 1’s account.
Figure 7: FFT of the Subject 1’s (red); Subject 2’s (blue); Subject 3’s Meanwhile, the ranges of the degree of similarity of
(yellow) WAV Files each subject are shown in Table 2.
Similarity test was done on the trials of each subject
and thirty comparisons were made per subject. Nine
hundred similarity tests were done and the resulting
range of the degree of similarity is 97%-99%. This
means that a wav file must at least have 97% degree of
similarity to unlock a subject’s account.
Table 1
Similarity Test Results on Trials of Subject 1

Pair %
Trial %Similarity Accuracy
No. Sequential
1 1 2 99 83.54 100
2 3 4 98 100 100
3 5 6 99 79.74 100
4 7 8 99 82.45 100
5 9 10 98 84.39 100
6 11 12 99 81.59 100
7 13 14 99 82.58 100
8 15 16 99 83.99 100
Figure 8: FFT of the Subject 4’s (red); Subject 5’s (blue); Subject 6’s 9 17 18 98 84.72 100
(yellow) WAV Files 10 19 20 99 82.30 100
11 21 22 98 68.71 100
12 23 24 97 73.65 100
C. Similarity Test 13 25 26 98 96.47 100
The evaluation of the data is obtained by using 14 27 28 98 74.29 100
15 29 30 98 73.30 100
SAP2011 to test the degree of similarity of the subject’s 16 1 3 98 79.34 100
wav files. The trials were tested to determine the 17 5 7 98 79.67 100
similarity of the trial samples. The measurements shown 18 9 11 98 83.20 100
below are %Similarity, Accuracy, and %Sequential. 19 13 15 98 79.51 100
20 17 19 98 79.57 100
%Similarity is the global similarity score or the total 21 21 23 99 83.54 100
degree of similarity of the samples being tested. 22 25 27 99 82.56 100
Accuracy is the local similarity score or the degree of 23 2 4 99 77.83 100
similarity on the frame of a sample on at 15ms duration. 24 6 8 98 81.29 100
25 10 12 99 81.65 100
%Sequential is an indicator of how much part of the 26 14 16 99 81.86 100
signals was actually compared. 27 18 20 99 82.96 100
A round-robin similarity test was done on means of 28 22 24 99 82.36 100
different subjects. Four hundred thirty-five tests were 29 26 28 99 82.25 100
done and the resulting range of degree of similarity is 30
30
1 98 83.71 100
3%-90%. The results are ranked with the highest degree
of similarity to the lowest degree of similarity. In Table

84
Table 2 133 5 28 3 63.15 100
Similarity Test Results on Trials of Subject 1 111 5 6 3 59.03 100
26 1 27 21 57.39 100
Range of Degree of
Subject Number Remarks
Similarity
1 97%-99% IDENTICAL Meanwhile, the degree of Similarity is tested on the
2 97%-99% IDENTICAL
3 97%-99% IDENTICAL
means of the wav files of different subject. A round-
4 97%-99% IDENTICAL robin comparison was made for this test. Thirty
5 97%-99% IDENTICAL comparisons are shown in Table 3.
6 97%-99% IDENTICAL
7 97%-99% IDENTICAL
8 97%-99% IDENTICAL
IV. CONCLUSION
9 97%-99% IDENTICAL
10 97%-99% IDENTICAL Based on the data gathered, the graph of the wav files
11 97%-99% IDENTICAL from one subject showed similarity from each other.
12 97%-99% IDENTICAL
13 97%-99% IDENTICAL
The graphs from different users also showed differences
14 97%-99% IDENTICAL from each other.
15 97%-99% IDENTICAL A subject who is assigned to unlock the device tried
16 97%-99% IDENTICAL unlocking the device 100 times and it was a success. A
17 97%-99% IDENTICAL
18 97%-99% IDENTICAL
guest subject (anyone except the user who is set to
19 97%-99% IDENTICAL unlock the device) tried unlocking the device 100 times
20 97%-99% IDENTICAL and failed. This test revealed the effectiveness of the
21 97%-99% IDENTICAL device to recognize the uniqueness of a user’s identity.
22 97%-99% IDENTICAL
23 97%-99% IDENTICAL
The proposed security system using on bone conduction
24 97%-99% IDENTICAL based on sound-to-feature algorithm and artificial neural
25 97%-99% IDENTICAL network which uses the characteristic frequency
26 97%-99% IDENTICAL response of the user’s skull as its passcode in computers
27 97%-99% IDENTICAL
28 97%-99% IDENTICAL
was evaluated having a 100% unlocking accuracy and
29 97%-99% IDENTICAL was successfully accomplished.
30 97%-99% IDENTICAL The study can be a way in proving that a person’s
skull can be acknowledged as a new method of
biometrics which is more unique and secure in all
Table 3
Similarity Test Results on Means of Different Subjects aspects. One cannot steal nor copy the frequency
response of another’s skull at all because it includes its
Pair % shape, the brain, the veins and all other aspects that can
Subjects %Similarity Accuracy
No. Sequential contribute to uniqueness of it. Therefore, using this
85 4 5 90 87.30 100 study helps in aiding a more secure and sophisticated
4 1 5 88 76.99 100 biometric in desktop computers.
200 8 26 87 70.55 100
186 8 12 87 69.11 100
195 8 21 87 68.80 100 ACKNOWLEDGMENT
190 8 16 87 67.91 100
191 8 17 87 66.92 100 The authors would like to acknowledge our mentors
188 8 14 87 65.54 100
75 3 21 87 59.14 100 and colleagues from the Electronics Engineering
70 3 16 86 77.78 100 Department, College of Engineering, Technological
406 22 29 85 80.89 100 University of the Philippines - Manila for the
419 24 29 85 72.08 100 encouragement and support in the whole process of
420 24 30 85 72.08 100
431 27 29 85 70.17 100 conducting this study.
413 23 29 85 64.97 100
38 2 11 19 58.97 100 REFERENCES
91 4 11 16 87.30 100
285 13 16 15 67.24 100
[1] D. C. Walden and T. Van Vleck, eds., 2015. The compatible
318 15 18 15 64.53 100
time sharing system (1961-1973): Fiftieth anniversary
67 3 13 14 61.50 100
commemorative overview. IEEE Computer Society, 2015.
22 1 23 12 59.64 100
[2] P. Kanakam, K. C. B. Rao, and S. M. Hussain. "Olfactory
7 1 8 12 59.64 100
Biometric Technique: An Emerging Technology." Journal of
289 13 20 11 61.00 100
Advancement in Robotics 1, vol. 1, pp. 1-11, Dec. 2015.
244 10 29 10 68.22 100
[3] Lindasalwa Muda, Mumtaj Begam, and Irraivan Elamvazuthi.
292 13 23 8 69.52 100
"Voice recognition algorithms using mel frequency cepstral
283 13 14 6 64.15 100
coefficient (MFCC) and dynamic time warping (DTW)
1 1 2 6 60.65 100

85
techniques." arXiv preprint arXiv:1003.4083, 2010. [Online].
Available from https://arxiv.org/abs/1003.4083
[4] S. Schneegass, Y. Oualil, and A. Bulling. "SkullConduct:
Biometric user identification on eyewear computers using bone
conduction through the skull." in Proceedings of the 34th ACM
SIGCHI Conference on Human Factors in Computing Systems
(CHI 2016), pp. 1379-1384.
[5] M. A. Nielsen. Neural networks and deep learning. 2015.
[Online]. Available from
http://neuralnetworksanddeeplearning.com.
[6] O. Tchernichovski, F. Nottebohm, C. E. Ho, B. Pesaran, and P.
P. Mitra. "A procedure for an automated measurement of song
similarity." Animal Behaviour vol. 59 (6), pp.1167-1176, 2000.

86

View publication stats

You might also like