Transcript
Page 1: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 2: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 3: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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 enemyโ€™s 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

Page 4: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 5: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 6: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

M.S. Dissertation Presentation

GIST INFONET Lab. 13 December 2018 / 32

Related works โ€“ Merchant et al [3]

โ€ข Convolutional neural network โ€“ Peng et al [4]

โ€ข Differential constellation trace, carrier frequency offset, and 2 features of the error on I/Q domain

โ€“ Patel et al [5] โ€ข Random forest and AdaBoost

I. Introduction

6

Page 7: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 8: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 9: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 10: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 11: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 12: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 13: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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,

๐ฎ๐ฎ = ๐€๐€๐€๐€, the concatenated features for training data are arranged to the columns of ๐€๐€ and the feature for test data is put into ๐ฎ๐ฎ.

III. Research Process โ€“ Feature concatenation

13

0 1000 2000 3000 4000 5000 60000

5

10

15

20

25

Samples

Mag

nitu

de

Page 14: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

M.S. Dissertation Presentation

GIST INFONET Lab. 13 December 2018 / 32

Sparse representation-based classification scheme โ€“ In the underdetermined system ๐ฒ๐ฒ = ๐ƒ๐ƒ๐€๐€, ๐€๐€ has infinite cases of

solutions and SRC finds a sparse solution ๐€๐€. โ€“ The condition to find the sparse solution ๐€๐€ is sensitive to the

mutual correlation between columns of ๐ƒ๐ƒ.

III. Research Process โ€“ SRC

14

Page 15: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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 ๐€๐€. โ€“ ๐ฎ๐ฎ = ๐€๐€๐€๐€ is changed to ๐ฒ๐ฒ = ๐ƒ๐ƒ๐€๐€ by PCA. โ€“ The sparse solution ๐€๐€ is obtained by basis pursuit algorithm

min๐€๐€

||๐€๐€||1 subject to ๐ฒ๐ฒ = ๐ƒ๐ƒ๐€๐€.

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

Page 16: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 17: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 18: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 19: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 20: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

M.S. Dissertation Presentation

GIST INFONET Lab. 13 December 2018 / 32 20

Thank you

Page 21: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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

Page 22: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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.

โ€ข Line of sight / Non line of sight โ€ข Line of sight after 18 month with same receiver and different receivers

VI. Appendix

22

Page 23: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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 โ€˜Random Forestโ€™ and โ€˜Multi-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

Page 24: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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 ๐ฎ๐ฎ = ๐€๐€๐€๐€, where ๐€๐€ is the training data matrix and ๐ฎ๐ฎ is the test data. โ€“ A covariance matrix of ๐€๐€ is eigen-decomposed as,

๐€๐€ โˆ’๐ฆ๐ฆ๐ฆ๐ฆ ๐€๐€ โˆ’๐ฆ๐ฆ๐ฆ๐ฆ ๐‘‡๐‘‡ = ๐–๐–๐šฒ๐šฒ๐–๐–๐‘ป๐‘ป

where ๐ฆ๐ฆ = ๐ฆ๐ฆ๐‘๐‘โˆ‘ ๐‘›๐‘› th columns of ๐€๐€๐‘๐‘๐‘›๐‘›=1 , ๐ฆ๐ฆ: = [1 1 โ‹ฏ 1].

โ€“ The eigen-vectors of the covariance matrix are orthonormal. โ€“ The eigen-value matrix ๐šฒ๐šฒ is proportional to the variance of ๐€๐€,

๐–๐–๐‘‡๐‘‡ ๐€๐€ โˆ’๐ฆ๐ฆ๐ฆ๐ฆ ๐€๐€ โˆ’๐ฆ๐ฆ๐ฆ๐ฆ ๐‘‡๐‘‡๐–๐– = ๐šฒ๐šฒ. โ€“ Let the eigen-values ๐œ†๐œ†1, ๐œ†๐œ†2, โ€ฆ ๐œ†๐œ†๐‘›๐‘› of the eigen-value matrix ๐šฒ๐šฒ be

rearranged in order of the sizes. โ€“ Let the eigen-vectors ๐ฐ๐ฐ1,๐ฐ๐ฐ๐Ÿ๐Ÿ, โ€ฆ๐ฐ๐ฐn of the eigen-vector matrix ๐–๐– be also

rearranged by the eigen-values.

VI. Appendix

24

Page 25: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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 ๐šฒ๐šฒ is proportional to the variance of ๐€๐€, the eigen-vectors can be basis for the creating the new space on the variance of ๐€๐€.

โ€“ The training data matrix ๐€๐€ and the test data ๐ฎ๐ฎ is transformed to the new space by ๐ƒ๐ƒ = ๐–๐–๐‘‡๐‘‡ ๐€๐€ โˆ’๐ฆ๐ฆ๐ฆ๐ฆ and ๐ฒ๐ฒ = ๐–๐–๐“๐“(๐ฎ๐ฎ โˆ’๐ฆ๐ฆ).

โ€“ PCA removes correlations among columns of ๐€๐€. โ€“ Also, PCA can remove the size of the columns of ๐€๐€.

VI. Appendix

25

Page 26: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

M.S. Dissertation Presentation

GIST INFONET Lab. 13 December 2018 / 32

๐ฟ๐ฟ๐‘๐‘ norms [8]

||๐ฑ๐ฑ||๐‘๐‘ = ๏ฟฝ |๐‘ฅ๐‘ฅ๐‘–๐‘–|๐‘๐‘๐‘–๐‘–

1๐‘๐‘

Uniqueness of sparse solution (๐ฟ๐ฟ1) [8] โ€“ Suppose ๐ฒ๐ฒ = ๐ƒ๐ƒ๐€๐€0 with

||๐€๐€0||0 < 12

1 + 1๐œ‡๐œ‡ ๐ƒ๐ƒ

,

where ๐œ‡๐œ‡ ๐ƒ๐ƒ = max1โ‰ค๐‘˜๐‘˜,๐‘—๐‘—โ‰ค๐‘š๐‘š,๐‘˜๐‘˜โ‰ ๐‘—๐‘—

|๐‘Ž๐‘Ž๐‘˜๐‘˜๐‘‡๐‘‡๐‘Ž๐‘Ž๐‘—๐‘—|

||๐‘Ž๐‘Ž๐‘˜๐‘˜||2โˆ™||๐‘Ž๐‘Ž๐‘—๐‘—||2

Then ๐€๐€0 is the unique optimal solution to Minimize | ๐€๐€ |1 subject to ๐ฒ๐ฒ = ๐ƒ๐ƒ๐€๐€.

If the function ๐‘“๐‘“ 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

Page 27: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

M.S. Dissertation Presentation

GIST INFONET Lab. 13 December 2018 / 32

๐ฟ๐ฟ๐‘๐‘ norms level sets

Basis pursuit algorithm [10] โ€“ The mathematical optimization problem of the form

min๐€๐€

||๐€๐€||1 subject to ๐ฒ๐ฒ = ๐ƒ๐ƒ๐€๐€.

โ€“ To solve the problem, โ€˜Primal-Dual Barrier methodโ€™ is used.

VI. Appendix

27

Strictly convex when p>1

Convex p>1 Nonconvex p<1

Page 28: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

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 ๐‘๐‘๐‘‡๐‘‡๐‘ฅ๐‘ฅ ๐‘ ๐‘ . ๐‘ก๐‘ก.๐ด๐ด๐‘ฅ๐‘ฅ = ๐‘๐‘, ๐‘ฅ๐‘ฅ โ‰ฅ 0, min ๐‘๐‘๐‘‡๐‘‡๐œ†๐œ† ๐‘ ๐‘ . ๐‘ก๐‘ก.๐ด๐ด๐‘‡๐‘‡๐œ†๐œ† + ๐‘ ๐‘  = ๐‘๐‘, ๐‘ ๐‘  โ‰ฅ 0 โ€“ The Karush-Kuhn-Tucker conditions for both equation are

๐ด๐ด๐‘‡๐‘‡๐œ†๐œ† + ๐‘ ๐‘  = ๐‘๐‘๐ด๐ด๐‘ฅ๐‘ฅ = ๐‘๐‘๐‘ฅ๐‘ฅ โ‰ฅ 0๐‘ ๐‘  โ‰ฅ 0

๐‘ฅ๐‘ฅ ๐‘–๐‘– ๐‘ ๐‘  ๐‘–๐‘– = 0, โ€‰1 โ‰ค ๐‘–๐‘– โ‰ค ๐‘›๐‘›

โ€“ Let ๐‘ ๐‘  = (๐‘ ๐‘  1 , ๐‘ ๐‘ (2) ,โ‹ฏ , ๐‘ ๐‘  ๐‘›๐‘› ), ๐‘†๐‘† = diag(๐‘ ๐‘ ) , and ๐‘’๐‘’ = (1,1,โ‹ฏ , 1) . We can rewrite the constraints into

๐น๐น๏ฟฝ(๐‘ฅ๐‘ฅ, ๐œ†๐œ†, ๐‘ ๐‘ ) =๐ด๐ด๐‘‡๐‘‡๐œ†๐œ† + ๐‘ ๐‘  โˆ’ ๐‘๐‘๐ด๐ด๐‘ฅ๐‘ฅ โˆ’ ๐‘๐‘๐‘‹๐‘‹๐‘†๐‘†๐‘’๐‘’

= 0

VI. Appendix

28

Page 29: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

M.S. Dissertation Presentation

GIST INFONET Lab. 13 December 2018 / 32

Primal-Dual Barrier method [11] โ€“ We relax the last constraint ๐‘ฅ๐‘ฅ(๐‘–๐‘–)๐‘ ๐‘ (๐‘–๐‘–) = 0 to ๐‘ฅ๐‘ฅ(๐‘–๐‘–)๐‘ ๐‘ (๐‘–๐‘–) = ๐œ‡๐œ‡ and obtain

๐น๐น(๐‘ฅ๐‘ฅ, ๐œ†๐œ†, ๐‘ ๐‘ ) =๐ด๐ด๐‘‡๐‘‡๐œ†๐œ† + ๐‘ ๐‘  โˆ’ ๐‘๐‘๐ด๐ด๐‘ฅ๐‘ฅ โˆ’ ๐‘๐‘๐‘‹๐‘‹๐‘†๐‘†๐‘’๐‘’ โˆ’ ๐œ‡๐œ‡๐‘’๐‘’

= 0

โ€“ The Jacobian will be

๐ฝ๐ฝ =0 ๐ด๐ด๐‘‡๐‘‡ ๐ผ๐ผ๐ด๐ด 0 0๐‘†๐‘† 0 ๐‘‹๐‘‹

and the Newtonโ€™s method read 0 ๐ด๐ด๐‘‡๐‘‡ ๐ผ๐ผ๐ด๐ด 0 0๐‘†๐‘† 0 ๐‘‹๐‘‹

๐‘‘๐‘‘๐‘‹๐‘‹๐‘‘๐‘‘๐œ†๐œ†๐‘‘๐‘‘๐‘ ๐‘ 

=โˆ’๐ด๐ด๐‘‡๐‘‡๐œ†๐œ† โˆ’ ๐‘ ๐‘  + ๐‘๐‘

๐‘๐‘ โˆ’ ๐ด๐ด๐‘ฅ๐‘ฅโˆ’๐‘‹๐‘‹๐‘†๐‘†๐‘’๐‘’ + ๐œ‡๐œ‡๐‘’๐‘’

โ€“ Solve

min ๐ต๐ต(๐‘ฅ๐‘ฅ๐‘˜๐‘˜ , ๐œ‡๐œ‡๐‘˜๐‘˜) = ๐‘๐‘๐‘‡๐‘‡๐‘ฅ๐‘ฅ โˆ’ ๐œ‡๐œ‡๏ฟฝ log๐‘ฅ๐‘ฅ๐‘–๐‘–

๐‘›๐‘›

๐‘–๐‘–=1

, ๐œ‡๐œ‡ > 0

VI. Appendix

29

Page 30: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

M.S. Dissertation Presentation

GIST INFONET Lab. 13 December 2018 / 32

Primal-Dual Barrier method [11]

VI. Appendix

30

Page 31: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

M.S. Dissertation Presentation

GIST INFONET Lab. 13 December 2018 / 32

[1] S. Ur Rehman, K. Sowerby, and C. Coghill, โ€˜โ€˜RF fingerprint extraction from the energy envelope of an instantaneous transient signal,โ€™โ€™ in Proc. Austral. Commun. Theory Workshop (AusCTW), 2012, pp. 90โ€“95.

[2] I. Kennedy, P. Scanlon, F. Mullany, M. Buddhikot, K. Nolan, and T. Rondeau, โ€œRadio transmitter fingerprinting: A steady state frequency domain approach,โ€ in Vehicular Technology Conference, 2008. VTC 2008-Fall. IEEE 68th. IEEE, 2008, pp. 1โ€“5.

[3] K. Merchant, S. Revay, G. Stantchev, and B. Nousain, โ€œDeep 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, โ€œDesign 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, โ€œImproving 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. 221โ€“233, Mar. 2015.

[6] ETSI TS 102 361-1 v2.4.1. โ€œElectromagnetic 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

Page 32: Digital Walkie-Talkie Identification scheme based on ......SL1M transient ( rising+falling) features. SL1M sync features -10 0 10 20-40-20 0 20-10-5 0 5 10 15. Component3 Component1

M.S. Dissertation Presentation

GIST INFONET Lab. 13 December 2018 / 32

[7] H. Abdi and L. Williams, โ€œPrincipal component analysis,โ€ Wiley Interdiscipl. Rev., Comput. Statist., vol. 2, no. 4, pp. 433โ€“459, 2010.

[8] A. M. Bruckstein, D. L. Donoho, and M. Elad, โ€œFrom sparse solutions of systems of equations to sparse modeling of signals and images,โ€ SIAM Rev., vol. 51, no. 1, pp. 34โ€“81, 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, โ€œAtomic decomposition by basis pursuit,โ€ SIAM Rev., vol. 43, no. 1, pp. 129โ€“159, 2001.

[11] KailaiXu, โ€œInterior point method for Linear programmingโ€, Standford.edu, 2017.

VII. Reference

32


Recommended