M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Digital Walkie-Talkie Identification scheme based on Sparse Representation with Multiple features
Presenter : M.S. Student / Kiwon Yang Advisor : Professor / Heung-No Lee
GIST, Dept. of Electrical Engineering and Computer Science INFONET Lab.
1
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Contents I. Introduction II. Contribution III. Research Process IV. Results V. Comparison VI. Appendix VII. Reference
2
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Motivation โ For the efficient support in the electronic warfare, the ability of
exact detection and analysis on the enemyโs transmitter is necessary.
โ In the Internet of Things (IoT) network, the technique for identifying the access of the counterfeit transmitter is needed.
Identification of the radio transmitters using the transmitted signals is called RF (Radio frequency) fingerprinting.
I. Introduction
3
IoT Network
Security system based on RF fingerprinting
Correct transmitter
Counterfeit transmitter
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Proposed system
A feature is a sample vector cultivated from the transmitted RF signals and bears unique information about the pertinent device. Goal โ We want to check if the performance will be increased or not
once multiple features โ rising transient feature, falling transient feature, and sync feature โ are used simultaneously.
I. Introduction
4
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
The feature is occurred by โ Element characteristic โ A part design such as filter, amplifier etc. โ PCB material, soldering etc.
The feature types
I. Introduction
5
Transient signals [1] Steady-state signals [2]
0.9994 0.9996 0.9998 1 1.0002 1.0004 1.0006 1.0008
x 107
0
5
10
15
Frequency
Mag
nitu
de
Hytera Dv1Hytera Dv2Hytera Dv3Hytera Dv4
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Related works โ Merchant et al [3]
โข Convolutional neural network โ Peng et al [4]
โข Differential constellation trace, carrier frequency offset, and 2 features of the error on I/Q domain
โ Patel et al [5] โข Random forest and AdaBoost
I. Introduction
6
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
The combination of rising transient feature, falling transient feature, and sync feature has not been used in the previous studies. โ Falling transient feature has not been used. โ We show that the performance of the proposed scheme is
improved when more feature is included. โ There are no experiments on RF fingerprinting with multiple
features based on sparse representation-based classification algorithm (SRC).
II. Contribution
7
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
The digital walkie-talkie models
III. Research Process โ Signal acquisition
8
Model name SL1M (2014) BD-358 (2017) Manufacturer MOTOROLA HYTERA
Frequency UHF (CH1 : 423.1875MHz)
UHF (CH1 : 423.1875MHz)
# of devices 4 4
Two models follow DMR standard
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Digital Mobile Radio standard [6] โ 2-slot Time-division multiple access (TDMA) method โ 4 level frequency shift-keying modulation
III. Research Process โ Signal acquisition
9
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Procedure for signal acquisition โ In LOS environment, the receiver gets the transmitted signal.
III. Research Process โ Signal acquisition
10
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Threshold method to extract interest signals
III. Research Process โ Feature extraction
11
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x 105
-3
-2
-1
0
1
2
3x 10
-3
Samples
Am
plitu
de
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x 105
-1.5
-1
-0.5
0
0.5
1
1.5x 10
-3
Samples
Am
plitu
de
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x 105
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2x 10
-4
SamplesA
mpl
itude
Rising transient signal Falling transient signal Sync signal
0 0.5 1 1.5 2 2.5 3
x 106
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
Samples
Am
plitu
de
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Main lobe extraction โ Since main lobe occupies most of the energy of each signal part,
the main lobe is used as a feature.
III. Research Process โ Feature extraction
12
Main lobe of rising transient signal
Main lobe of falling transient signal
Main lobe of sync signal
0.98 0.985 0.99 0.995 1 1.005 1.01 1.015 1.02
x 107
0
5
10
15
20
25
Frequency
Mag
nitu
de
0.98 0.985 0.99 0.995 1 1.005 1.01 1.015 1.02
x 107
0
5
10
15
20
25
Frequency
Mag
nitu
de
0.98 0.985 0.99 0.995 1 1.005 1.01 1.015 1.02
x 107
0
5
10
15
20
25
Frequency
Mag
nitu
de
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
The extracted features โ rising transient feature, falling transient feature, and sync feature โ are concatenated. In the system,
๐ฎ๐ฎ = ๐๐๐๐, the concatenated features for training data are arranged to the columns of ๐๐ and the feature for test data is put into ๐ฎ๐ฎ.
III. Research Process โ Feature concatenation
13
0 1000 2000 3000 4000 5000 60000
5
10
15
20
25
Samples
Mag
nitu
de
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Sparse representation-based classification scheme โ In the underdetermined system ๐ฒ๐ฒ = ๐๐๐๐, ๐๐ has infinite cases of
solutions and SRC finds a sparse solution ๐๐. โ The condition to find the sparse solution ๐๐ is sensitive to the
mutual correlation between columns of ๐๐.
III. Research Process โ SRC
14
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Sparse representation-based classification scheme โ Principal component analysis (PCA) removes correlations
among columns of ๐๐. โ ๐ฎ๐ฎ = ๐๐๐๐ is changed to ๐ฒ๐ฒ = ๐๐๐๐ by PCA. โ The sparse solution ๐๐ is obtained by basis pursuit algorithm
min๐๐
||๐๐||1 subject to ๐ฒ๐ฒ = ๐๐๐๐.
โ The class of test data is output from SRC Test โ We used 5 cross validation technique. โ Fifty data were captured per a digital walkie-talkie.
III. Research Process โ SRC & Test
15
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
When additional feature is included, the performance of SRC is improved. The accuracy recorded 98.75%. Falling transient signal could have unique information for RF fingerprinting.
IV. Results
16
4 BD-358 4 SL1M 4 BD-358 4 SL1M
Accuracy rate (Minimum number of PC) TR(R) 88% (24) 82% (48) 90.5% (45) TR(F) 87.5% (45) 90% (12) 92.25% (13)
TR(R + F) 93% (49) 92% (20) 95.5% (63) Sync 99% (45) 83.5% (22) 93.75% (86)
TR(R + F) + Sync 99% (44) 98.5% (22) 98.75% (21)
Accuracy rate of the proposed method
R: Rising, F: Falling, PC: Principal components
0 10 20 30 40 50 60 70 80 90 10010
20
30
40
50
60
70
80
90
100
The number of principal components, P
Cla
ssifi
catio
n ac
cura
cy (%
)
TR (Rising)TR (Falling)TR (Rising+Falling)SyncTR (Rising+Falling) + Sync
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
The cluster on each class is distinctly formed when the concatenated feature is used.
IV. Results
17
SL1M transient (rising+falling) features
SL1M sync features
-100
1020
-40
-20
0
20-10
-5
0
5
10
15
Com
pone
nt3
Component2Component1
dv1dv2dv3dv4
-5
0
5
-20
0
20
40-4
-2
0
2
4
6
Com
pone
nt3
Component2Component1
dv1dv2dv3dv4
SL1M transient (rising+falling) + sync features
-20-10
010
20
-20
0
20
40-10
-5
0
5
10
15
20
Com
pone
nt3
Component2Component1
dv1dv2dv3dv4
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
It is noticeable that the highest accuracy rate is recorded even though the less number of training data is used relatively. The comparison experiment is necessary for more accurate comparison.
IV. Results
18
Method Number of Devices Experiment condition Accuracy rate Number of training
data per a device
The proposed method
8 (Digital walkie-talkies) 1m LOS 98.75% 40
Patel et al. [3] 4 (Zigbee devices) 12 dB Higher than 90% 1500
Peng et al. [4] 54 (Zigbee devices) 1-3m LOS 96% 1 (template feature)
Merchant et al. [5] 7 (Zigbee devices) 28 dB 92.29% 900
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
We proposed the RF fingerprinting scheme based on SRC with multiple features. As a feature, the main lobes of rising transient signal, falling transient signal, and sync signal were used simultaneously. When many features were used as concatenation, the accuracy rate was increased. The accuracy rate of the proposed method recorded 98.75%. As a future work, we need to study on RF fingerprinting scheme based on SRC with the various features besides the used features. The paper on this study is under revision.
V. Conclusion
19
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32 20
Thank you
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Deep learning for RF device fingerprinting [3] โ They used the convolutional neural network (CNN) for the RF
fingerprinting. โ They collected 7,000 data from 7 devices. โ Each full dataset of 7,000 transmissions was randomly partitioned into
80% training data, 10% validation data, and 10% testing data. โ The overall correct identification rate is 92.29%
VI. Appendix
21
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Hybrid RF fingerprint extraction and device classification scheme [4] โ They used 4 features simultaneously โ differential constellation trace
figure, carrier frequency offset, and 2 features of the error on I/Q domain. โ The experiment is performed on the total 54 Zigbee devices. โ They did the experiment on the 4 environment.
โข Line of sight / Non line of sight โข Line of sight after 18 month with same receiver and different receivers
VI. Appendix
22
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Improving zigbee device network authentication using ensemble decision tree classifiers [5] โ The used RF DNA features contain information on variance, skewness,
and kurtosis, within a preamble response. โ They showed the result of โRandom Forestโ and โMulti-class AdaBoostโ
for RF fingerprinting. โ The top-ranked 25 variables selected by Variable Importance (VI) metric
built in Random Forest classifier on 4 zigbee devices are used.
VI. Appendix
23
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Principal components analysis [7] โ PCA is a method to project the original data onto the new space on the
variance. โ Let ๐ฎ๐ฎ = ๐๐๐๐, where ๐๐ is the training data matrix and ๐ฎ๐ฎ is the test data. โ A covariance matrix of ๐๐ is eigen-decomposed as,
๐๐ โ๐ฆ๐ฆ๐ฆ๐ฆ ๐๐ โ๐ฆ๐ฆ๐ฆ๐ฆ ๐๐ = ๐๐๐ฒ๐ฒ๐๐๐ป๐ป
where ๐ฆ๐ฆ = ๐ฆ๐ฆ๐๐โ ๐๐ th columns of ๐๐๐๐๐๐=1 , ๐ฆ๐ฆ: = [1 1 โฏ 1].
โ The eigen-vectors of the covariance matrix are orthonormal. โ The eigen-value matrix ๐ฒ๐ฒ is proportional to the variance of ๐๐,
๐๐๐๐ ๐๐ โ๐ฆ๐ฆ๐ฆ๐ฆ ๐๐ โ๐ฆ๐ฆ๐ฆ๐ฆ ๐๐๐๐ = ๐ฒ๐ฒ. โ Let the eigen-values ๐๐1, ๐๐2, โฆ ๐๐๐๐ of the eigen-value matrix ๐ฒ๐ฒ be
rearranged in order of the sizes. โ Let the eigen-vectors ๐ฐ๐ฐ1,๐ฐ๐ฐ๐๐, โฆ๐ฐ๐ฐn of the eigen-vector matrix ๐๐ be also
rearranged by the eigen-values.
VI. Appendix
24
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Principal components analysis [7] โ Since the eigen-vectors of the covariance matrix are orthonormal and
the eigen-value matrix ๐ฒ๐ฒ is proportional to the variance of ๐๐, the eigen-vectors can be basis for the creating the new space on the variance of ๐๐.
โ The training data matrix ๐๐ and the test data ๐ฎ๐ฎ is transformed to the new space by ๐๐ = ๐๐๐๐ ๐๐ โ๐ฆ๐ฆ๐ฆ๐ฆ and ๐ฒ๐ฒ = ๐๐๐๐(๐ฎ๐ฎ โ๐ฆ๐ฆ).
โ PCA removes correlations among columns of ๐๐. โ Also, PCA can remove the size of the columns of ๐๐.
VI. Appendix
25
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
๐ฟ๐ฟ๐๐ norms [8]
||๐ฑ๐ฑ||๐๐ = ๏ฟฝ |๐ฅ๐ฅ๐๐|๐๐๐๐
1๐๐
Uniqueness of sparse solution (๐ฟ๐ฟ1) [8] โ Suppose ๐ฒ๐ฒ = ๐๐๐๐0 with
||๐๐0||0 < 12
1 + 1๐๐ ๐๐
,
where ๐๐ ๐๐ = max1โค๐๐,๐๐โค๐๐,๐๐โ ๐๐
|๐๐๐๐๐๐๐๐๐๐|
||๐๐๐๐||2โ||๐๐๐๐||2
Then ๐๐0 is the unique optimal solution to Minimize | ๐๐ |1 subject to ๐ฒ๐ฒ = ๐๐๐๐.
If the function ๐๐ has a second derivative that is non-negative (positive) over an interval, the function is convex (strictly convex) over that interval. [9]
VI. Appendix
26
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
๐ฟ๐ฟ๐๐ norms level sets
Basis pursuit algorithm [10] โ The mathematical optimization problem of the form
min๐๐
||๐๐||1 subject to ๐ฒ๐ฒ = ๐๐๐๐.
โ To solve the problem, โPrimal-Dual Barrier methodโ is used.
VI. Appendix
27
Strictly convex when p>1
Convex p>1 Nonconvex p<1
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Primal-Dual Barrier method [11] โ A certain class of algorithms that solve linear and nonlinear convex
optimization problems. โ Consider the dual pair for Linear programming problem
min ๐๐๐๐๐ฅ๐ฅ ๐ ๐ . ๐ก๐ก.๐ด๐ด๐ฅ๐ฅ = ๐๐, ๐ฅ๐ฅ โฅ 0, min ๐๐๐๐๐๐ ๐ ๐ . ๐ก๐ก.๐ด๐ด๐๐๐๐ + ๐ ๐ = ๐๐, ๐ ๐ โฅ 0 โ The Karush-Kuhn-Tucker conditions for both equation are
๐ด๐ด๐๐๐๐ + ๐ ๐ = ๐๐๐ด๐ด๐ฅ๐ฅ = ๐๐๐ฅ๐ฅ โฅ 0๐ ๐ โฅ 0
๐ฅ๐ฅ ๐๐ ๐ ๐ ๐๐ = 0, โ1 โค ๐๐ โค ๐๐
โ Let ๐ ๐ = (๐ ๐ 1 , ๐ ๐ (2) ,โฏ , ๐ ๐ ๐๐ ), ๐๐ = diag(๐ ๐ ) , and ๐๐ = (1,1,โฏ , 1) . We can rewrite the constraints into
๐น๐น๏ฟฝ(๐ฅ๐ฅ, ๐๐, ๐ ๐ ) =๐ด๐ด๐๐๐๐ + ๐ ๐ โ ๐๐๐ด๐ด๐ฅ๐ฅ โ ๐๐๐๐๐๐๐๐
= 0
VI. Appendix
28
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Primal-Dual Barrier method [11] โ We relax the last constraint ๐ฅ๐ฅ(๐๐)๐ ๐ (๐๐) = 0 to ๐ฅ๐ฅ(๐๐)๐ ๐ (๐๐) = ๐๐ and obtain
๐น๐น(๐ฅ๐ฅ, ๐๐, ๐ ๐ ) =๐ด๐ด๐๐๐๐ + ๐ ๐ โ ๐๐๐ด๐ด๐ฅ๐ฅ โ ๐๐๐๐๐๐๐๐ โ ๐๐๐๐
= 0
โ The Jacobian will be
๐ฝ๐ฝ =0 ๐ด๐ด๐๐ ๐ผ๐ผ๐ด๐ด 0 0๐๐ 0 ๐๐
and the Newtonโs method read 0 ๐ด๐ด๐๐ ๐ผ๐ผ๐ด๐ด 0 0๐๐ 0 ๐๐
๐๐๐๐๐๐๐๐๐๐๐ ๐
=โ๐ด๐ด๐๐๐๐ โ ๐ ๐ + ๐๐
๐๐ โ ๐ด๐ด๐ฅ๐ฅโ๐๐๐๐๐๐ + ๐๐๐๐
โ Solve
min ๐ต๐ต(๐ฅ๐ฅ๐๐ , ๐๐๐๐) = ๐๐๐๐๐ฅ๐ฅ โ ๐๐๏ฟฝ log๐ฅ๐ฅ๐๐
๐๐
๐๐=1
, ๐๐ > 0
VI. Appendix
29
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
Primal-Dual Barrier method [11]
VI. Appendix
30
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
[1] S. Ur Rehman, K. Sowerby, and C. Coghill, โโRF fingerprint extraction from the energy envelope of an instantaneous transient signal,โโ in Proc. Austral. Commun. Theory Workshop (AusCTW), 2012, pp. 90โ95.
[2] I. Kennedy, P. Scanlon, F. Mullany, M. Buddhikot, K. Nolan, and T. Rondeau, โRadio transmitter fingerprinting: A steady state frequency domain approach,โ in Vehicular Technology Conference, 2008. VTC 2008-Fall. IEEE 68th. IEEE, 2008, pp. 1โ5.
[3] K. Merchant, S. Revay, G. Stantchev, and B. Nousain, โDeep learning for Rf device fingerprinting in cognitive communication network,โ IEEE, J.Sel.Topics Signal Process., vol. 12, no. 1, pp. 160-167, Feb. 2018.
[4] L. Peng, A. Hu, J. Zhang, Y. Jiang, J. Yu, and Y. Yan, โDesign of a Hybrid RF Fingerprint Extraction and Device Classification Scheme,โ IEEE Trans. Internet Things., May, 2018 (in press)
[5] H. J. Patel, M. A. Temple, and R. O. Baldwin, โImproving ZigBee device network authentication using ensemble decision tree classifiers with radio frequency distinct native attribute fingerprinting,โ IEEE Trans. Rel., vol. 64, no. 1, pp. 221โ233, Mar. 2015.
[6] ETSI TS 102 361-1 v2.4.1. โElectromagnetic compatibility and Radio spectrum Matters (ERM); Digital Mobile Radio (DMR) Systems; Part 1: DMR Air Interface (AI) protocol,โ European Telecommunications Standards Institute, 2016.
VII. Reference
31
M.S. Dissertation Presentation
GIST INFONET Lab. 13 December 2018 / 32
[7] H. Abdi and L. Williams, โPrincipal component analysis,โ Wiley Interdiscipl. Rev., Comput. Statist., vol. 2, no. 4, pp. 433โ459, 2010.
[8] A. M. Bruckstein, D. L. Donoho, and M. Elad, โFrom sparse solutions of systems of equations to sparse modeling of signals and images,โ SIAM Rev., vol. 51, no. 1, pp. 34โ81, Feb. 2009.
[9] T. M. Cover and J. A. Thomas, Elements of Information Theory. New York: Wiley, 1991.
[10] S. S. Chen, D. L. Donoho, and M. A. Saunders, โAtomic decomposition by basis pursuit,โ SIAM Rev., vol. 43, no. 1, pp. 129โ159, 2001.
[11] KailaiXu, โInterior point method for Linear programmingโ, Standford.edu, 2017.
VII. Reference
32