192
S. Hu et al.
agate through the wireless channel and finally reach the RX antenna at the
receiver. The arrived signal is expressed as:
A rx = A tx ⊗ h (t) + n (t) ,
(2)
where h (t) represent the channel function and n (t) represent the noise interference.
Based on Eqs. (1) and (2), in addition to the hardware characteristics of the
transmitter, the signal is also affected by the channel. Thus, the characteristics of
the received signal are divided into channel-based features and transmitter-based
features. Assuming the signals go through the same channel, the characteristics of
the identification signals are essentially derived from transmitter-based features,
also known as device fingerprints.
According to the requirements of signal fingerprint extraction, the identification system needs to carry out several preprocesses, such as phase compensation, energy normalization, and discarding the unqualified signal. The fingerprint
recognition system detects the preprocessed signal transformation and extracts
the relevant signal features in the time domain, the frequency domain, or the
wavelet domain. Then the feature vector of the identification signal is composed by the extracted features. Finally, as shown in Fig. 2, it becomes a typical
classification problem consisting of two parts: training and testing. During the
training stage, the transmitted signals of all target transmitters are tagged with
unique tags to form a signal recognition fingerprint library. In addition, during
the testing stage, the relevant received signal features are extracted and compared with the fingerprint database to obtain the recognition results. Through
the above analysis, it is easy to see that there are two main factors affecting the
identification efficiency: the selection of signal characteristics and classifiers.
3 Recognition Methods
Due to the characteristics of the transmitter hardware, the received signal contains various FR fingerprint features. Hence, the appropriate selection of signal
features is an important factor for RF fingerprint classification. From the existing research, it is known that RF fingerprints include time-domain envelopes,
wavelet coefficients, and so on [7]. In this paper, fractal features, phase noise
spectrum, and constellation features are selected as feature vectors.
3.1 Fractal Dimension
In recent years, the fractal theory has been widely concerned because it can
effectively measure the complexity and irregularity of signals [8] and has been
successfully applied in radar radiation source signals. Similar to radar signals,
the characteristics of the wireless communication signals are mainly reflected in
the variation, and distribution of frequency, phase, and amplitude. Therefore,
the signal pulse can be identified by measuring the complexity of the signal
S. Hu et al.
agate through the wireless channel and finally reach the RX antenna at the
receiver. The arrived signal is expressed as:
A rx = A tx ⊗ h (t) + n (t) ,
(2)
where h (t) represent the channel function and n (t) represent the noise interference.
Based on Eqs. (1) and (2), in addition to the hardware characteristics of the
transmitter, the signal is also affected by the channel. Thus, the characteristics of
the received signal are divided into channel-based features and transmitter-based
features. Assuming the signals go through the same channel, the characteristics of
the identification signals are essentially derived from transmitter-based features,
also known as device fingerprints.
According to the requirements of signal fingerprint extraction, the identification system needs to carry out several preprocesses, such as phase compensation, energy normalization, and discarding the unqualified signal. The fingerprint
recognition system detects the preprocessed signal transformation and extracts
the relevant signal features in the time domain, the frequency domain, or the
wavelet domain. Then the feature vector of the identification signal is composed by the extracted features. Finally, as shown in Fig. 2, it becomes a typical
classification problem consisting of two parts: training and testing. During the
training stage, the transmitted signals of all target transmitters are tagged with
unique tags to form a signal recognition fingerprint library. In addition, during
the testing stage, the relevant received signal features are extracted and compared with the fingerprint database to obtain the recognition results. Through
the above analysis, it is easy to see that there are two main factors affecting the
identification efficiency: the selection of signal characteristics and classifiers.
3 Recognition Methods
Due to the characteristics of the transmitter hardware, the received signal contains various FR fingerprint features. Hence, the appropriate selection of signal
features is an important factor for RF fingerprint classification. From the existing research, it is known that RF fingerprints include time-domain envelopes,
wavelet coefficients, and so on [7]. In this paper, fractal features, phase noise
spectrum, and constellation features are selected as feature vectors.
3.1 Fractal Dimension
In recent years, the fractal theory has been widely concerned because it can
effectively measure the complexity and irregularity of signals [8] and has been
successfully applied in radar radiation source signals. Similar to radar signals,
the characteristics of the wireless communication signals are mainly reflected in
the variation, and distribution of frequency, phase, and amplitude. Therefore,
the signal pulse can be identified by measuring the complexity of the signal