Wearable Biometric Authentication Based on
Human Body Communication
Zedong Nie, Yuhang Liu,Changjiang Duan, Zhongzhou Ruan,Jingzhen Li, Lei Wang
Shenzhen Institutes of Advanced Technology
Chinese Academy of Sciences
Shenzhen, China
zd.nie@siat.ac.cn
Abstract—Human body communication (HBC) is a short-
range, wireless communication in the vicinity of, or inside a
human body. In this paper, biometric authentication based on
capacitive coupled HBC is presented for the wearable devices. In-
situ experiments were conducted with 20 volunteers to investigate
the feasibility. The S21 parameters of the HBC channel from one
palm to the other within the frequency range of 300 KHz – 50
MHz were measured. A total of 2,561,600 data are acquired. The
data are analyzed by the support vector machines (SVM)
including C-SVM and nu-SVM, where 2,241,400 data are used to
train the SVM model and 320,200 data are used to estimate the
authentication rate. Linear, polynomial, and radial basis function
(RBF) are adopted as the kernel functions, respectively. In
addition, to verify whether the features in low frequency band
will affect the performance of HBC authentication, the features in
four frequency bands, i.e., from 300 KHz to 50 MHz, from 3.4
MHz to 50 MHz, from 5.6 MHz to 50 MHz, and from 9.6 MHz to
50 MHz are used as the biometric trait, respectively. The
experiment results show that, in biometric identification mode,
identification rate of 98% is achieved, and in biometric
verification mode, the equal error rate (EER) is 0.24%, the
average area under the curve (AUC) of receiver operating
characteristic (ROC) reaches 0.9993.
Keywords—Biometric authentication; wearable system; human
body communication; support vector machines
I. INTRODUCTION
With the rapid development of wearable technology,
wearable devices are experiencing an exponential growth [1].
Wearable devices are generally small, portable, low power
consumption [2], and worn on the multiple locations on the
user for diverse functions such as video recording, pedometer
and health monitoring [3, 4]. As the information stored in the
wearablCe devices are almost private, such as personal photo,
video and health data measured by the biosensors including
heart rate, blood pressure, and electrocardiogram [4], it is
important to prohibit the unauthorized persons from accessing
the wearable device. Biometric authentication is an excellent
approach to solve this problem[5].
Biometric authentication refers to verifying or identifying
individuals based on the physical or behavioral characteristics
such as face, fingerprint, hand geometry, iris, typing rhythm,
voice, gait [6]. Biometric is inherently more reliable than the
password-based authentication as biometric traits cannot be
This study was financed partially by National Natural Science Foundation of
China (Grant No.61403366), Shenzhen Basic Research Project Fund
(JCYJ20140417113430695).
978-1-4673-7201-5/15/$31.00 ©2015 IEEE
lost or forgotten. It is also more difficult to forge biometrics [5].
A number of biometric characteristics have been in use for
different applications [7]. Several groups have studied the
biometrics in the mobile platforms. Wang et al. and Liu et al.
utilized the finger-vein recognition system for mobile devices
[8, 9]. Klonovs et al. and Tao et al. introduced EEG-based
biometric and face recognition, respectively [10]. Other
researchers combined several biometrics such as face, voice,
and teeth to provide a better performance [11]. However,
considering that the wearable devices may be worn on
different position such as legs [12], the aforementioned
biometric are not suitable for some wearable devices. In
addition, the device of the authentication system must be small
enough to be integrated into wearable devices. For this reason,
some biometrics, like gait [13], typing rhythm [14], hand veins
[15], DNA, hand geometry [16], and iris recognition [17] are
not acceptable for wearable application. Therefore, there is a
need to propose a biometric trait which is small enough and
suitable for wearable systems.
Human body communication (HBC) is a short-range
wireless communication in the vicinity of, or inside a human
body by using the human body as a propagation medium [18,
19]. HBC is divided into two solutions: galvanic coupling and
capacitive coupling [20]. The former requires one pairs of
electrodes in both the transmitter (TX) and the receiver (RX),
whereas the latter only requires a single electrode for the TX
and the RX [21]. The capacitive coupling makes it possible to
miniaturize the size of device, and is more suitable for
applications requiring the devices to be miniature enough.
Since HBC can transfer in high data rates while maintaining
low power consumption [22], and provide high security and
easy integration within body-worn devices [23], HBC shows
great potential for wearable devices. Moreover, as the
proportion of biological tissues such as muscle, fat, and
skeleton is different between individuals, the overall dielectric
constants of human body are diverse, as well as the signal
propagated through human body. The diverse HBC
propagation signal can be utilized as the biometric trait to
authenticate individuals. By means of employing the HBC as
both the authentication and the communication approaches, the
size of wearable devices will be more miniature. Due to the use
of propagation signal between devices, the HBC authentication
is suitable for wearable device regardless of the location.
Therefore, it is of great significance to research HBC-based
authentication.
 Some researchers have attempted to investigate the galvanic
coupled HBC authentication for general applications [24], [25].
However, the white noise used in these studies will increase
the variation between individual biometric samples. Moreover,
it is more appropriate to apply the capacitive coupled HBC
authentication to wearable application in view of the volume
restriction. In this paper, an investigation of capacitive coupled
HBC authentication is presented, employing the sinusoidal
signal rather than white noise to measure the biometric traits.
The obtained data are analyzed by support vector machines
(SVM) including C-SVM and nu-SVM. Linear function,
polynomial, and radial basis function (RBF) are adopted as the
kernel functions, respectively. The performance is evaluated
both in verification mode and identification mode.
2(a) shows five of 20 volunteers’ features measured in one
time, and Fig. 2(b) shows one volunteer’s features measured in
four times. As can be seen from Fig. 2(a), difference exists
between features of different volunteers. In Fig. 2(b), the
features in low frequency bank have a larger variation in the
four curves which may reduce the performance of HBC
authentication.
-10
-15
-20
-25

S21(dB)
The rest of this paper is organized as follows: Section II
details the experimental setup of measurements. Section III -30
introduces the authentication algorithms. The experimental
1
evaluation is discussed in Section IV. The conclusions are -35
2
presented in Section V. 3
-40 4
II. EXPERIMENTAL SETUP OF MEASUREMENT 5

-45
Fig. 1 shows the experimental setup in these experiments, 0 10 20 30 40 50
including a vector network analyzer (VNA) and two self-
Frequency (MHz)
assembled electrodes. VNA and the electrodes were connected
through two coaxial cables. The electrodes, which were (a)
capacitive coupled to human body, were fabricated with copper -10
conductive cloth covered by plastic insulating tape. The
dimension of the electrodes was 6 cm × 4.5 cm, and the -15
distance between them was 70 cm. The VNA was set to
transmit sinusoidal signals within the frequency band range -20
between 300 KHz and 50 MHz, including 1601 discrete
frequency points. In other words, the frequency interval -25
S21(dB)

between two adjacent points was 31.0625 KHz.

-30

-35
1
2
-40 3
4

-45
0 10 20 30 40 50
Frequency (MHz)

Fig. 1. Experimental setup of experiments. Vector network analyzer and two
electrodes were used to measure the biometric feature of HBC.
To investigate the performance of the HBC authentication,
20 volunteers including 4 women and 16 men were employed.
At the time of these experiments, the age range of the
volunteers is 23-34 years old, weight range is 45-100 kilogram,
and height range is 153-180 cm. The volunteers were asked to
put their left palms flat on the left electrode, and right palms on
the right electrode. Sinusoidal signal transmitted from the VNA
passed to the left electrode through coaxial cable, propagated
through human body from left palms to right palms in
capacitive coupling manner, and then received by VNA through
the right electrode. VNA measured the attenuation of the
sinusoidal signal and plotted amplitude-frequency curves. Fig.
(b)
Fig. 2. Features of HBC authentication. (a) Features of five volunteers
measured in one time. (b) Features of one volunteer measured in four times.
The volunteers were measured for 2 times each day, and
total 8 times in 4 days. During each measurement, 10 groups of
data were obtained from each volunteer, and each group of data
contains 1601 points within the frequency range of 300 KHz –
50MHz. In total, 1600 groups of data that amount to 2561600
were obtained.
III. EXPERIMENTAL EVALUATION
Depending on the application, a biometric authentication
system can operate either in the verification or identification
mode [26]. In the biometric verification mode, only one person
who is authorized has the right to enter the system, and the
biometric features are compared with the unique template of

978-1-4673-7201-5/15/$31.00 ©2015 IEEE

the authorized person stored in the system database. In the
biometric identification mode, the biometric features are
compared with all templates stored in the system database, and
thus the system could identify the identities of the users. In
some other applications, the system should accomplish both
verification and identification. For example, in a multi-user
system, the system should not only identify the identities of the
users, but also have to verify whether the users have the
authorization to enter this system. In this paper, the
performance of the HBC biometric authentication is evaluated
both in verification and identification mode.
Fig. 3. Flow chart representing the authentication steps by using support
vector machine.
LIBSVM [27], a library for SVM, is employed to estimate
the performance of the HBC authentication. Fig. 3 shows the
process of the authentication by using SVM. In the data
analysis, the data obtained in previous 7 times including 1400
groups of data which amount to 2,241,400 were used as
templates to train SVM classifier, and then, SVM models was
established. Whereas, the 200 groups of data gained in the last
time which add up to 320,200 serve as test data and were
predicted by the SVM models acquired before. The
performance of HBC biometric authentication could be
obtained by comparing the predicted labels with the true labels.
A. Identification Mode
In identification mode, the performance of the
authentication system is evaluated by the identification rate
[26]. The identification rate is computed as
Rate 100% (1)
Applying support vector machines as the authentication
algorithms, the identification becomes a multiclass
classification problem. However, support vector machines are
formulated for two-class problems which could not be applied
to a multiclass problem directly [28]. In view of this problem,
two types of approaches including one-against-all and one-
against-one are introduced to solve it [29].
One-against-all converts multiclass problem into
corresponding amount two-class problems. For each two-class
problem, the remaining classes are regarded as one class. In
other words, the th SVM classifier regards the th class of
samples as positive labels, and all others samples are
categorized to negative labels. For instance, there are 20
978-1-4673-7201-5/15/$31.00 ©2015 IEEE
classes in this experiment. Accordingly, the number of two-
class problem is 20 as well. One-against-one converts an n-
class problem into n n 1 ⁄2 two-class problems. Put it
another way, each class constructs a two-class problem with
every other class. Practically, one-against-one is more suitable
in multiclass problems [29]. In this paper, one-against-one is
used to solve the multiclass problem. Besides, the voting
strategy is applied to decide the identity of the sample.
In order to get an appropriate authentication algorithm for
HBC authentication, both C-SVM and nu-SVM are used.
Additionally, linear, polynomial, and radial basis function are
adopted as the kernel functions, respectively. The results of
experiment I in Table I show the identification rate by directly
using the measured data within the frequency band range from
300 KHz to 50MHz. The highest rate is 89% by using the nu-
SVM with linear kernel function. However, as can be drawn
from Fig. 2(b), the features in the low frequency are not stable
in several measurements. The features in that band may affect
the performance of HBC authentication. To verify whether the
feature in low band would influence the performance of HBC
authentication, the features in other three bands, i.e., from 3.4
MHz to 50MHz, from 5.6 MHz to 50 MHz, and from 9.6 MHz
to 50 MHz are used as the biometric trait, respectively. The
corresponding experiments are labelled from II to IV. The
results are shown in Table I. The highest identification rates in
experiments II, III, and IV are 91%, 92
% and 98%, respectively. Comparing the identification rate in
experiments I, II, III, and IV, it can be found that the
identification rates increase with the features acquired within
the frequency band begin with higher frequency.
In general, the computing speed of the processor in the
wearable systems is slower than the personal computer. It is
necessary to reduce the amount of calculation and time
consumed for wearable applications. Experiments V, VI, and
VII were done with the number of frequency points reduced. In
experiment V, only one frequency point was used in every
eight frequency points. In other words, the dimension of
feature in experiment V is one-eighth of the dimension of the
feature in experiment IV. Similarly, the dimension of the
feature in experiments VI and VII are a quarter and one second
of the dimension in experiment IV. The rate is stable at around
98% by using the C-SVM with the linear function, around 94%
by using the C-SVM with the RBF function, around 91% by
using the nu-SVM with the polynomial function and around
92.5% by using the nu-SVM with RBF function. However, the
rate in experiment VI is higher than the rate in experiment IV
by using the C-SVM with polynomial function, the rate in
experiment V is higher than the rate in experiment IV by using
the nu-SVM with linear function. Therefore, the identification
rate will not decrease much by reducing the dimension of the
feature.
Additionally, the highest identification rate by using C-SVM
is 98%, and the highest identification rate by using nu-SVM is
96%. Moreover, almost every identification rate by using C-
SVM is higher than the corresponding identification rate by
using nu-SVM. Consequently, the C-SVM is more suitable for
the HBC authentication. Furthermore, no matter what the
authentication algorithm is, C-SVM or nu-SVM, the
identification rates with linear kernel function are higher than
 TABLE I. IDENTIFICATION RATE

Data Band C-SVM nu-SVM

Experiment
Segment (MHz) Linear Polynomial RBF Linear Polynomial RBF
I 1-1601 0.3-50 88.5% 79.5% 83.5% 89.0% 88.5% 84.5%
II 101-1601 3.4-50 91.0% 81.0% 91.5% 88.5% 87.5% 88.0%
III 201-1601 6.5-50 92.0% 90.0% 91.0% 90.0% 90.0% 90.5%
IV 301-1601 9.6-50 98.0% 93.5% 94.5% 94.5% 91.0% 92.5%
V 38-201 9.6-50 98.0% 93.5% 93.5% 96.0% 90.5% 92.0%
VI 75-401 9.6-50 97.5% 96.0% 94.0% 91.5% 92.5% 92.5%
VII 150-:801 9.6-50 98.0% 94.0% 94.0% 95.0% 91.0% 92.0%

the rates with polynomial function or RBF function. Therefore, with the nu-SVM. Consequently, C-SVM is superior to nu-
linear kernel function is more suitable for the HBC SVM in the HBC verification. Lastly, the highest AUC and
authentication. the lowest EER achieved 0.9993 and 0.24% by using the C-
SVM with the polynomial function.
B. Verification Mode
In verification mode, only the authorized person has the
authority to enter the system. To confirm whether the visitor is
the authorized person, the features of the visitor should be
compared with the referential feature stored in the database.
The performance of the verification system can be evaluated
by the equal error rate (EER) and the area under the curve
(AUC) [26]. The EER refers to the error rate which the false
accept rate (FAR) equals the false reject rate (FRR), and the
AUC refers to the area under the operator feature curve (ROC).
The EER and AUC range between 0 and 1. The ROC curve is
a function of the matching score, which plots the FAR on the (a) (b)
x-axis and the “1-FRR” on the y-axis. The matching score is
the similarity between two biometric features. A high AUC
value with a low EER means a high performance of the
verification system.
Considering that the feature in the low band would affect
the performance of HBC authentication, only the features
within the frequency band range of 9.6 MHz to 50MHz is used
to evaluate the performance of HBC authentication in the
verification mode. Fig. 4 shows the ROC curves by using
SVM. In these figures, the volunteers are labeled from A to T.
(c) (d)
Fig. 4(a)-(c) are the ROC curves by using C-SVM with linear,
polynomial and RBF kernel function, respectively. Fig. 4(d)-(f)
are the ROC curves by using nu-SVM with linear, polynomial
and RBF kernel function, respectively. In the Fig. 4(a) and Fig.
4(b), only two of the volunteers have the AUC value of less
than one, others have the AUC value of one. In the Fig. 4(d)
and Fig. 4(e), there are three of the volunteers having the AUC
value of less than one. In addition, the numbers of volunteers
that have the AUC value of less than one are five and six in
the Fig. 4(c) and Fig. 4(f). Thus, the linear and polynomial
kernel functions outperform the RBF in the HBC
authentication. Table II and Table III show the AUCs and (e) (f)
EERs. By comparison, it can be found that no matter what
Fig. 4. ROC curves. (a) ROC curves by using C-SVM with linear function.
kind of SVM is used, the lowest average EERs are obtained by (b) ROC curves by using C-SVM with polynomial function. (c) ROC curves
adopting the polynomial function as the kernel, and the linear by using C-SVM with RBF function. (d) ROC curves by using nu-SVM with
function is just slightly inferior to the polynomial function. linear function. (e) ROC curves by using nu-SVM with polynomial function.
Therefore, the polynomial function is the most suitable for the (f) ROC curves by using nu-SVM with RBF function.
HBC verification. Moreover, it is apparent that the average
AUCs with the C-SVM are larger than the AUCs with the nu-
SVM, and the EERs with the C-SVM are less than the EERs

978-1-4673-7201-5/15/$31.00 ©2015 IEEE

 TABLE II. AUC AND EER BY USING C-SVM [8] W. Desong, L. Jianping, and G. Memik, "User identification based on
finger-vein patterns for consumer electronics devices," IEEE Trans.
Linear Polynomial RBF Consumer Electronics, vol. 56, pp. 799-804, 2010.
Volunteer [9] L. Zhi and S. Shangling, "An embedded real-time finger-vein
AUC EER AUC EER AUC EER
recognition system for mobile devices," IEEE Trans. Consumer
E 0.9853 0.0263 0.9895 0.0158 0.9858 0.0211 Electronics, vol. 58, pp. 522-527, 2012.
G 1 0 1 0 0.9984 0.0053 [10] J. Klonovs, C. K. Petersen, H. Olesen, and A. Hammershoj, "ID Proof
J 0.9889 0.1 0.9968 0.0316 0.9947 0.0474 on the Go: Development of a Mobile EEG-Based Biometric
S 1 0 1 0 0.9958 0.0316 Authentication System," Vehicular Technology Magazine, IEEE, vol. 8,
T 1 0 1 0 0.9989 0.0105 pp. 81-89, 2013.
Others 1 0 1 0 1 0 [11] K. Dong-Ju and H. Kwang-Seok, "Multimodal biometric authentication
Average 0.9987 0.0063 0.9993 0.0024 0.9987 0.0058 using teeth image and voice in mobile environment," IEEE Trans.
Consumer Electronics, vol. 54, pp. 1790-1797, 2008.
TABLE III. AUC AND EER BY USING NU-SVM [12] P. H. Veltink and D. De Rossi, "Wearable technology for biomechanics:
e-textile or micromechanical sensors? [Conversations in BME],"
Linear Polynomial RBF Engineering in Medicine and Biology Magazine, IEEE, vol. 29, pp. 37-
Volunteer 43, 2010.
AUC EER AUC EER AUC EER [13] W. Chen, Z. Junping, W. Liang, P. Jian, and Y. Xiaoru, "Human
E 0.9863 0.0263 0.9842 0.0263 0.9853 0.0211 Identification Using Temporal Information Preserving Gait Template,"
G 1 0 1 0 0.9979 0.0053 IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 34, pp.
J 0.9895 0.1000 0.9926 0.0737 0.9953 0.0474 2164-2176, 2012.
L 0.9984 0.0053 0.9995 0.0053 0.9726 0.0632 [14] J. Roth, L. Xiaoming, and D. Metaxas, "On Continuous User
S 1 0 1 0 0.9963 0.0105 Authentication via Typing Behavior," IEEE Trans. Image Processing,
T 1 0 1 0 0.9879 0.0316 vol. 23, pp. 4611-4624, 2014.
Others 1 0 1 0 1 0 [15] S. M. Lajevardi, A. Arakala, S. Davis, and K. J. Horadam, "Hand vein
Average 0.9987 0.0066 0.9988 0.0053 0.9968 0.0090 authentication using biometric graph matching," Biometrics, IET, vol. 3,
pp. 302-313, 2014.
[16] Z. Gang, W. Chia-Jiu, and T. E. Boult, "Application of Projective
IV. CONCLUSION Invariants in Hand Geometry Biometrics," IEEE Trans. Information
Forensics and Security, vol. 2, pp. 758-768, 2007.
This paper presents capacitive coupled HBC authentication [17] K. N. Faddis, J. R. Matey, J. R. Maxey, and J. T. Stracener,
for the wearable devices. The HBC authentication is "Performance assessments of iris recognition in tactical biometric
investigated by experiments on 20 volunteers. It is confirmed devices," Biometrics, IET, vol. 2, pp. 1-10, 2013.
that C-SVM is more suitable than the nu-SVM in both the [18] IEEE, "Standard for Local and metropolitan area networks," in Part
15.6: Wireless Body Area Networks vol. 802.15.6, ed, 2012.
identification and verification modes. The linear function [19] H. Baldus, S. Corroy, A. Fazzi, K. Klabunde, and T. Schenk, "Human-
outperforms the polynomial and RBF function in the centric connectivity enabled by body-coupled communications,"
identification mode. Whereas in the verification mode, the Communications Magazine, IEEE, vol. 47, pp. 172-178, 2009.
polynomial is better, and the linear function is slightly inferior [20] B. Kibret, M. Seyedi, D. T. H. Lai, and M. Faulkner, "Investigation of
to the polynomial. It is also proved that the features in the Galvanic-Coupled Intrabody Communication Using the Human Body
Circuit Model," Biomedical and Health Informatics, IEEE Journal of,
band lower than 9.6MHz will affect the performance of HBC vol. 18, pp. 1196-1206, 2014.
authentication. In result, in identification mode the [21] Z. Nie, J. Ma, Z. Li, H. Chen, and L. Wang, "Dynamic Propagation
identification rate achieves 98%, and in verification mode the Channel Characterization and Modeling for Human Body
average EER is 0.24%, the average AUC reaches 0.9993. Communication," Sensors, vol. 12, pp. 17569-17587, 2012.
[22] B. Joonsung, S. Kiseok, L. Hyungwoo, C. Hyunwoo, and Y. Hoi-Jun,
"A 0.24-nJ/b Wireless Body-Area-Network Transceiver With Scalable
Double-FSK Modulation," Solid-State Circuits, IEEE Journal of, vol.
REFERENCES 47, pp. 310-322, 2012.
[23] Z. Nie, J. Ma, K. Ivanov, and L. Wang, "An Investigation on Dynamic
[1] S. Kirk, "The Wearables Revolution: Is Standardization a Help or a
Human Body Communication Channel Characteristics at 45 MHz in
Hindrance?: Mainstream technology or just a passing phase?,"
Different Surrounding Environments," Antennas and Wireless
Consumer Electronics Magazine, IEEE, vol. 3, pp. 45-50, 2014.
Propagation Letters, IEEE, vol. 13, pp. 309-312, 2014.
[2] T. Starner, "How Wearables Worked their Way into the Mainstream,"
[24] I. Nakanishi, T. Inada, Y. Sodani, and L. Shigang, "Performance
Pervasive Computing, IEEE, vol. 13, pp. 10-15, 2014.
Evaluation of Intra-palm Propagation Signals as Biometrics," in
[3] K. Lyons and H. Profita, "The Multiple Dispositions of On-Body and
International Conference on Biometrics and Kansei Engineering, 2013,
Wearable Devices," Pervasive Computing, IEEE, vol. 13, pp. 24-31,
pp. 91-94.
2014.
[25] I. Nakanishi and Y. Sodani, "SVM-Based Biometric Authentication
[4] A. Pantelopoulos and N. G. Bourbakis, "A Survey on Wearable Sensor-
Using Intra-Body Propagation Signals," in IEEE International
Based Systems for Health Monitoring and Prognosis," IEEE Trans.
Conference on Advanced Video and Signal Based Surveillance, 2010,
Systems, Man, and Cybernetics, Part C: Applications and Reviews, vol.
pp. 561-566.
40, pp. 1-12, 2010.
[26] M. Faundez-Zanuy, "Biometric security technology," IEEE Aerospace
[5] S. Rane, W. Ye, S. C. Draper, and P. Ishwar, "Secure Biometrics:
and Electronic Systems Magazine, vol. 21, pp. 15-26, 2006.
Concepts, Authentication Architectures, and Challenges," Signal
[27] C.-C. Chang and C.-J. Lin, "LIBSVM : a library for support vector
Processing Magazine, IEEE, vol. 30, pp. 51-64, 2013.
machines," ACM Transactions on Intelligent Systems and Technology,
[6] A. K. Jain, A. Ross, and S. Pankanti, "Biometrics: a tool for information
vol. 2, p. 27, 2011.
security," IEEE Trans. Information Forensics and Security, vol. 1, pp.
[28] S. Abe, Support Vector Machines for Pattern Classification, 2nd ed.
125-143, 2006.
London: Springer, 2010.
[7] A. K. Jain, P. Flynn, and A. A. Ross, Handbook of Biometrics. New
[29] H. Chih-Wei and L. Chih-Jen, "A comparison of methods for multiclass
York: Springer US, 2008.
support vector machines," IEEE Trans. Neural Networks, vol. 13, pp.
415-425, 2002.

978-1-4673-7201-5/15/$31.00 ©2015 IEEE