Machine Learning for RF Fingerprinting Extraction . . .
199
For the classifier, weighted KNN achieves a better performance than bagged tree
and fine Gaussian SVM in a single feature classification.
In order to more intuitively reflect the classification effect of different parameter features, multiple features are considered as the classification features for
the purpose of visual comparison. Figure 5 illustrates the dot plot of 300 sample
points per terminal with QPSK modulation when using box dimension, fractal
dimension, and phase noise spectrum as input features. It can be seen that when
those features are used as input features, except for some points overlapped
with each other, others can be clearly distinguished. Figure 7 is the front view
of Fig. 5, which also presents the dot plot when using box dimension and phase
noise spectrum as input features. Figure 8 is the top view of Fig. 5, which presents
the dot plot when using box dimension and fractal dimension as input features.
Figure 6 shows the dot plot using constellation side length, constellation diagonal length, and constellation angle as input features. Figure 9 is the front view
of Fig. 6, which presents the dot plot when using constellation diagonal length
60
80
100
3.81
120
140
160
Phase noise
spectrum
180
200
3.8
Fractal
dimansion
3.79
3.78
1.72
1.7
1.68
1.66
3.77
1.64
1.62
Box dimention
TX1
TX2
TX3
TX4
Fig. 5. View of box dimension, fractal dimension, and phase noise spectrum with SNR
= 9.07 dB
1
1.01
1.02
1.03
126
128
130
132
134
1.05
1.1
1.15
1.2
1.25
1.3
Constellation
side length
Constellation
angle
TX1
TX2
TX3
TX4
Constellation
diagonal length
Fig. 6. View of constellation side length, constellation diagonal length, and constellation angle with SNR = 9.07 dB