A radio frequency fingerprint extraction and recognition method based on multi-channel convolutional neural network

By dividing the signal into pre-transient, steady-state, and tail-transient parts to generate a differential constellation trajectory map, and using a multi-channel convolutional neural network for classification, the feature loss problem caused by signal masking in the existing technology is solved, and the recognition accuracy of wireless devices is improved, especially in low signal-to-noise ratio environments.

CN114417914BActive Publication Date: 2024-11-26SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111624298.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-28
Publication Date
2024-11-26
Estimated Expiration
2041-12-28

AI Technical Summary

Technical Problem

Existing technologies, when transforming an entire signal into a differential constellation trajectory diagram, ignore the masking relationship between the transient and steady-state parts of the signal and the difference in the number of sampling points, resulting in poor classification and recognition performance, especially in low signal-to-noise ratio scenarios.

Method used

The signal is divided into pre-transient, steady-state, and tail-transient parts, and differential constellation trajectory maps are generated for each part. A multi-channel convolutional neural network is used for classification and recognition. Local features are extracted through the multi-channel convolutional neural network structure and the feature weights of different parts are balanced.

Benefits of technology

It improves the recognition accuracy of wireless devices, especially significantly enhancing recognition performance in low signal-to-noise ratio scenarios, avoiding feature loss and balancing the influence of features from different parts of the signal.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114417914B_ABST
    Figure CN114417914B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of information security technology, and more specifically, to a method for extracting and identifying radio frequency fingerprints based on a multi-channel convolutional neural network. The method comprises: (1) collecting the transmitted signals of wireless devices and performing preprocessing; (2) for each signal, dividing the signal into pre-transient, steady-state, and tail transient parts, and generating differential constellation trace figures (DCTF) respectively; (3) dividing the obtained DCTF data set into a training set and a test set, and using the training set to train a multi-channel convolutional neural network classifier; (4) for the received signals sent by the wireless device, using the trained multi-channel convolutional neural network to classify and identify them. The present invention uses the DCTF of different parts of the signal to construct a multi-channel convolutional neural network for radio frequency fingerprint extraction and identification, so as to avoid some signal features being submerged in the DCTF and balance the feature weights of different parts of the signal, so as to achieve better recognition effect, especially in low signal-to-noise ratio scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of information security technology, and in particular to a radio frequency fingerprint extraction and recognition method based on a multi-channel convolutional neural network. Background Art

[0002] Wireless communication technology has become an integral part of modern life. However, the broadcast nature of wireless communication transmission makes mobile devices more vulnerable to malicious attacks. RF fingerprinting is a physical layer solution for mobile device identification and authentication. Due to the limitations of the manufacturing process, electronic components will have some small defects and flaws. These small deviations will cause some signal distortion or deformation, which can be used to detect the device transmitting the signal. By extracting and identifying the RF fingerprint of mobile devices, the security of the wireless communication system is enhanced.

[0003] The scheme most similar to the present invention is found in the literature: S. Wang, L. Peng, H. Fu, A. Hu and X. Zhou, "A Convolutional Neural Network-Based RF Fingerprinting Identification Scheme for Mobile Phones," IEEE INFOCOM 2020-IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), 2020, pp. 115-120, doi: 10.1109 / INFOCOMWKSHPS50562.2020.9163058. The main idea is to obtain its statistical characteristics by transforming the entire received signal into a differential constellation trajectory diagram, and then use it to train a convolutional neural network, and finally input the received signal into the network for recognition.

[0004] In the above scheme, in the process of transforming the entire signal into a differential constellation trajectory diagram, the masking relationship between the transient and steady-state pixels of the signal in the differential constellation trajectory diagram is ignored, and the contribution of the transient part is reduced in the differential constellation trajectory diagram due to the difference in the number of sampling points between the transient and steady-state parts of the signal, thereby affecting the classification and recognition effect.

[0005] List of abbreviations, English and definitions of key terms:

[0006] 1. DCTF: Differential Constellation Trace Figure, differential constellation trace figure

[0007] 2. I / Q signal: In-phase / Quadraturesignal, in-phase / quadrature signal

[0008] 3. USRP: Universal Software Radio Peripheral Summary of the invention

[0009] In order to solve the technical problems raised by the above-mentioned background technology, the present invention aims to provide a radio frequency fingerprint extraction and identification method based on a multi-channel convolutional neural network, which divides the pre-transient, steady-state, and tail transient parts of the signal and generates differential constellation trajectory diagrams respectively, and then trains a multi-channel convolutional neural network to classify and identify wireless devices.

[0010] In order to achieve the above technical objectives, the technical solution of the present invention is:

[0011] The radio frequency fingerprint extraction and recognition method based on a multi-channel convolutional neural network includes the following steps:

[0012] (1) Collecting and preprocessing the transmitted signals of wireless devices;

[0013] (2) For each signal, the signal is divided into pre-transient, steady-state, and tail transient parts, and differential constellation trajectory diagrams are generated respectively;

[0014] (3) The obtained differential constellation trajectory map data set is divided into a training set and a test set, and a multi-channel convolutional neural network classifier is trained using the training set;

[0015] (4) For the signals received from the wireless device, the trained multi-channel convolutional neural network is used to classify and identify them.

[0016] Furthermore, when the receiving end collects the baseband signal of the wireless device in step (1), the sampling rate satisfies the Nyquist theorem.

[0017] Furthermore, the preprocessing in step (1) includes: signal detection and synchronization, and energy normalization.

[0018] Furthermore, in step (2), the front transient is the signal portion from when the device starts to send a signal to when it reaches the rated power, the tail transient is the signal portion from when the device power starts to decrease until the device stops working, and the steady-state portion is the signal portion other than the above portions.

[0019] Furthermore, in step (2), the pre-transient, steady-state, and tail transient parts of the obtained signal are respectively subjected to differential processing according to the selected sampling point interval, and the differentially processed I / Q signals are plotted on a two-dimensional plane with the I path and the Q path as coordinate axes, and the grayscale values ​​are mapped according to the sampling point density to obtain a differential constellation trajectory diagram.

[0020] Furthermore, in step (3), the multi-channel convolutional neural network structure consists of the following parts:

[0021] Differential constellation trajectory map input stage: The differential constellation trajectory map reflecting the characteristics of different parts of the signal in step (2) is input into different channels of the convolutional neural network, and each channel corresponds to a differential constellation trajectory map.

[0022] Feature extraction stage: The differential constellation trajectory map of each channel is input into a convolution branch with the same structure as LeNet-5. Each convolution branch consists of two convolution layers and two pooling layers.

[0023] Classification stage: All matrices in different channels are flattened into vectors and connected into a feature vector. This vector is then input into a fully connected layer to integrate the local feature information obtained from each channel for classification. There are three fully connected layers, and the softmax function is used to obtain the probability of each label during classification.

[0024] Furthermore, in step (4), the received wireless device signal is processed according to steps (1) and (2) to obtain the differential constellation trajectory diagram of the pre-transient, steady-state, and tail transient of the signal, and input into the multi-channel convolutional neural network obtained in step (3), and the probability value of each label is output, and the label with the largest probability value is the prediction result.

[0025] The beneficial effects brought by adopting the above technical solution are:

[0026] Compared with the prior art, the present invention has the following significant advantages: it avoids the feature loss caused by the mutual masking of transient partial pixels and steady partial pixels when the signal is converted into a differential constellation trajectory diagram, and balances the weights of the influence of different signal features on the RF fingerprint. Experiments have shown that the present invention can further improve the recognition performance of wireless devices compared with the existing convolutional neural network method based on differential constellation trajectory diagrams, especially in low signal-to-noise ratio scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A schematic diagram of a flow chart of an embodiment of the present invention;

[0028] Figure 2 An example of an experimental platform for an embodiment of the present invention;

[0029] Figure 3 The DCTF diagram of random access signals of different LTE terminals;

[0030] Figure 4 This is a DCTF diagram of the pre-transient, steady-state, and tail-transient of a random access signal of an LTE terminal;

[0031] Figure 5 It is a multi-channel convolutional neural network structure diagram;

[0032] Figure 6 Schematic diagram for improving the performance of RF fingerprint classification and recognition based on multi-channel convolutional neural network. DETAILED DESCRIPTION

[0033] This embodiment provides a radio frequency fingerprint extraction and recognition method based on a multi-channel convolutional neural network. Figure 1 As shown, the following steps are included:

[0034] (1) Collect the transmitted signals of wireless devices and perform preprocessing.

[0035] In this embodiment, Figure 2 The experimental platform is built and the signal is collected as shown, specifically:

[0036] Six LTE mobile devices were selected as target wireless devices and numbered. An LTE pseudo base station was built using open source platform software and a USRPB205 device. The uplink center frequency of the pseudo base station was 2565MHz, the number of resource blocks was 50, and the random access signal center frequency set by the pseudo base station was 2561.4MHz. The mobile phone was connected to the pseudo base station, and the flight mode of the mobile phone was manually triggered to turn off to collect the random access signal of the mobile phone.

[0037] A USRPB205 device is used to collect signals. The sampling center frequency is 2561.4MHz, the sampling rate is 16M, and the number of sampling points of each random access signal of the LTE mobile phone collected is 14450. The sampling environment is a line-of-sight environment with a high signal-to-noise ratio. In order to explore the performance of the present invention under different signal-to-noise ratios, Gaussian white noise is added to the collected signals to make their signal-to-noise ratios range from 10dB to 30dB, and each interval is 5dB as a group. All collected signals under each signal-to-noise ratio are simulated once according to steps (2)-(4).

[0038] The preprocessing includes signal detection and synchronization, and energy normalization. The signal spectrum energy is used to detect the presence of the signal, and the cyclic prefix is ​​used to perform coarse synchronization on the received signal, and then the local standard random access signal is used to perform fine synchronization on the received signal. Then the signal is energy normalized to obtain the signal where α c is the channel attenuation factor, s(t) is the baseband signal, Δf is the deviation between the sampling frequency and the signal transmission frequency, φ is the phase deviation, and A is the RMS of the received signal.

[0039] (2) For each signal, the signal is divided into pre-transient, steady-state, and tail-transient parts, and differential constellation trajectory diagrams are generated respectively.

[0040] The signal is divided into the pre-transient state, the steady state, and the tail transient state. Specifically, according to the signal energy characteristics of the LTE random access signal, 50 sampling points before the signal is sent and 50 sampling points after the signal is sent are divided into the pre-transient state, 50 sampling points before the end of the signal and 50 sampling points after the end are divided into the tail transient state, and the middle part of the signal other than the above parts is divided into the steady state. Differential constellation trajectory diagrams are generated for the above three parts respectively. Figure 3 The differential constellation trajectory diagram of the complete LTE random access signal is shown. Figure 4 The differential constellation trajectory diagram of the pre-transient, steady state, and tail transient of an LTE random access signal is shown.

[0041] The method of generating the differential constellation trajectory diagram is:

[0042] Perform a differential operation on the signal Y(t) obtained in step (1):

[0043]

[0044] Among them (g) * is the conjugate operation, t n is the differential interval. After differential processing, a stable phase rotation factor will be obtained. Thus, a stable differential constellation trajectory diagram can be obtained.

[0045] The I / Q signal after differential processing is plotted on a two-dimensional plane with the I path and the Q path as coordinate axes, and the grayscale value is mapped according to the sampling point density to obtain a differential constellation trajectory diagram.

[0046] (3) The obtained differential constellation trajectory map dataset is divided into a training set and a test set, and a multi-channel convolutional neural network classifier is trained using the training set.

[0047] The above differential constellation trajectory map dataset is randomly divided into a training set and a test set in a ratio of 5:1. In order to ensure that the ratio of each label in the training set and the test set is the same as that in the dataset, a stratified sampling method is used. A multi-channel convolutional neural network is trained using the above training set.

[0048] The multi-channel convolutional neural network structure is as follows Figure 5 Its structure consists of the following parts:

[0049] Differential constellation trajectory map input stage: The differential constellation trajectory maps reflecting the characteristics of different parts of the signal in step (2) are input into different channels of the convolutional neural network, and each channel corresponds to a differential constellation trajectory map. The size of each differential constellation trajectory map is 129*129.

[0050] Feature extraction stage: The differential constellation trajectory map of each channel is input into a convolution branch with the same structure as LeNet-5. Each convolution branch consists of two convolution layers and two pooling layers. The convolution layer is used to extract the local features of DCTF, and the size of the convolution kernel is 5*5. The pooling layer is added after the convolution layer to reduce the number of parameters to speed up the operation and prevent overfitting. In this embodiment, maximum pooling is used with a step size of 2 and a pooling size of 2*2.

[0051] Classification stage: All matrices in different channels are flattened into vectors and connected into a feature vector. This vector is then input into a fully connected layer to integrate the local feature information obtained from each channel for classification. There are three fully connected layers, and the softmax function is used to obtain the probability of each label during classification.

[0052] (4) For the signals received from the wireless device, the trained multi-channel convolutional neural network is used to classify and identify them.

[0053] The received wireless device signal is processed according to step (1) and step (2) to obtain the differential constellation trajectory diagram of the pre-transient, steady-state, and tail transient of the signal, and is input into the multi-channel convolutional neural network obtained in step (3) to output the probability value of each label. The label with the largest probability value is the prediction result.

[0054] The method of the present invention avoids the feature loss caused by mutual masking of transient partial pixels and steady partial pixels when the signal is converted into a differential constellation trajectory diagram, and balances the weights of the influence of different partial features of the signal on the radio frequency fingerprint, which can effectively improve the accuracy of wireless device recognition. Figure 6 The test results of using this method and not using this method to directly classify the test set using the convolutional neural network. It can be seen that when using the method of the present invention, the recognition accuracy is improved in all the tested signal-to-noise ratio conditions (10dB-30dB); especially in the low signal-to-noise ratio scenario (less than or equal to 15dB), the improvement effect is more obvious.

[0055] The above disclosure is only a preferred embodiment of the present invention and cannot be used to limit the protection scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

[0056] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A radio frequency fingerprint extraction and recognition method based on a multi-channel convolutional neural network, characterized in that: The steps include: (1) Collecting and preprocessing the transmitted signals of wireless devices; (2) For each signal, the signal is divided into pre-transient, steady-state, and tail transient parts, and differential constellation trajectory diagrams are generated respectively; (3) The obtained differential constellation trajectory map data set is divided into a training set and a test set, and a multi-channel convolutional neural network classifier is trained using the training set; (4) Classify and identify the received signals sent by the wireless device using the trained multi-channel convolutional neural network; In step (2), the front transient is the signal portion from when the device starts to send a signal to when it reaches the rated power, the tail transient is the signal portion from when the device power starts to decrease until the device stops working, and the steady-state portion is the signal portion other than the above portions; In step (2), the pre-transient, steady-state, and tail transient parts of the obtained signal are respectively subjected to differential processing according to the selected sampling point interval, and the I / Q signal after differential processing is plotted on a two-dimensional plane with the I path and the Q path as coordinate axes, and the grayscale value is mapped according to the sampling point density to obtain a differential constellation trajectory diagram; In step (3), the multi-channel convolutional neural network structure consists of the following parts: Differential constellation trajectory map input stage: The differential constellation trajectory maps reflecting the characteristics of different parts of the signal in step (2) are input into different channels of the convolutional neural network, and each channel corresponds to a differential constellation trajectory map; Feature extraction stage: The differential constellation trajectory map of each channel is input into a convolution branch with the same structure as LeNet-5. Each convolution branch consists of two convolution layers and two pooling layers. Classification stage: All matrices in different channels are flattened into vectors and connected into a feature vector, which is then input into a fully connected layer to integrate the local feature information obtained from each channel for classification. The number of fully connected layers is three, and the softmax function is used to obtain the probability of each label during classification.

2. The radio frequency fingerprint extraction and identification method based on a multi-channel convolutional neural network according to claim 1 is characterized in that: When the receiving end collects the baseband signal of the wireless device in step (1), the sampling rate satisfies the Nyquist theorem.

3. The radio frequency fingerprint extraction and identification method based on a multi-channel convolutional neural network according to claim 1 is characterized in that: The preprocessing in step (1) includes: signal detection and synchronization, and energy normalization.

4. The radio frequency fingerprint extraction and identification method based on a multi-channel convolutional neural network according to claim 1 is characterized in that: In step (4), the received wireless device signal is processed according to steps (1) and (2) to obtain the differential constellation trajectory diagram of the pre-transient, steady-state, and tail transient of the signal, and input into the multi-channel convolutional neural network obtained in step (3), and the probability value of each label is output, and the label with the largest probability value is the prediction result.