A Radio Frequency Fingerprint Extraction and Recognition Method Based on Cross-Channel Robust Strategy

Abstract

The present invention proposes a method for extracting and identifying radio frequency fingerprints based on a cross-channel robust strategy, including: Step 1: Receive electromagnetic signals; Step 2: Perform frame synchronization and frequency offset compensation on the electromagnetic signals to obtain a time-frequency synchronized signal; Step 3: Perform channel estimation based on the discrete Fourier transform channel estimation algorithm to obtain channel state information; Step 4: Perform channel equalization on the time-frequency synchronized signal after channel estimation to obtain frequency domain symbols independent of the channel; Step 5: Perform Fourier transform and power normalization on the frequency domain symbols independent of the channel; Step 6: Use the Mix-up regularization technique to construct a multi-dimensional convolutional network model; Step 7: Input the power-normalized time-domain complex signal into the multi-dimensional convolutional network model for radio frequency fingerprint extraction and identification. The present invention uses channel knowledge and model regularization constraints, with strong flexibility of the method, fast algorithm response time, strong engineering feasibility, and greatly improves the engineering efficiency of deep learning algorithms.

Description

Technical Field

[0001] The present invention relates to a method for extracting and identifying radio frequency fingerprints based on a cross-channel robust strategy, belonging to the technical field of signal recognition. Background Art

[0002] With the popularization of the Internet of Things (IoT) and wireless communication technologies, more and more devices and systems communicate and interact through radio frequency signals. However, this also provides opportunities for malicious attackers, who can conduct fraud, steal information, or carry out other forms of attacks by impersonating the radio frequency signals of legitimate devices. In this case, radio frequency fingerprint recognition technology can effectively classify and identify unknown signals or devices by analyzing and identifying the characteristics of radio frequency signals, ensuring the security of communication and preventing unauthorized devices from accessing the network or system.

[0003] Currently, deep learning-based radio frequency fingerprint (DL-RFF) recognition has been a research hotspot in recent years. However, deep learning-based radio frequency fingerprint technology has an overfitting problem and overfits the environmental factors and time-varying device noises in the training dataset, resulting in high accuracy in a static environment but being unable to adapt when facing domain changes (such as location or time changes). Therefore, out-of-distribution generalization is crucial for radio frequency fingerprint recognition, which will help solve the existing fingerprint robustness and persistence problems, especially when crossing domains, to achieve training in a specific single domain (environment, location, time) and the effectiveness of fingerprint recognition under any domain change. Summary of the Invention

[0004] In order to solve the technical problem of existing cross-channel robustness, the present invention further proposes a method for extracting and identifying radio frequency fingerprints based on a cross-channel robust strategy.

[0005] The technical solution adopted by the present invention to solve the above problems is as follows: The present invention proposes a method for extracting and identifying radio frequency fingerprints based on a cross-channel robust strategy, and the steps include:

[0006] Step 1: Receive electromagnetic signals;

[0007] Step 2: Perform frame synchronization and frequency offset compensation on the received electromagnetic signals to obtain time-frequency synchronized signals;

[0008] Step 3: Perform channel estimation based on the discrete Fourier transform channel estimation algorithm to obtain channel state information;

[0009] Step 4: Perform channel equalization on the time-frequency synchronized signals after channel estimation to obtain frequency domain symbols independent of the channel;

[0010] Step 5: Perform Fourier transform on the frequency domain symbols independent of the channel and perform power normalization to obtain power-normalized time-domain complex signals;

[0011] Step 6: Construct a multi-dimensional convolutional network model using the Mix-up regularization technique;

[0012] Step 7: Input the time-domain complex signal after power normalization into the multi-dimensional convolutional network model for RF fingerprint extraction and recognition.

[0013] Optionally, the steps to obtain the time-frequency synchronization signal in Step 2 include:

[0014] Step 2.1: Use the method of sliding window and threshold to obtain the rough starting point of the electromagnetic signal, and use the cyclic prefix CP of the orthogonal frequency division multiplexing symbol OFDM to achieve fine synchronization to obtain the frame synchronization signal;

[0015] Step 2.2: Use the position of the known cyclic prefix CP in the frame synchronization signal and the reference signal to estimate the carrier frequency offset, and perform frequency offset compensation on the frame synchronization signal to obtain the time-frequency synchronization signal;

[0016] The expression for the rough starting point of the electromagnetic signal is:

[0017]

[0018] In formula (1), W is the sliding window length, S is the sliding step, T is the threshold, x is the signal sequence, and k is the position of the current sliding window;

[0019] The expression for fine synchronization is:

[0020]

[0021] In formula (2), L is the search length, N G is the sampling point of the cyclic prefix CP, N sub is the effective data length sampling point of the cyclic prefix CP, n0 is the frame synchronization signal, δ is the offset of the signal starting point, and i is the index of the signal sample currently being processed;

[0022] The expression for frequency offset compensation is:

[0023]

[0024] In formula (3), is the fractional multiple frequency offset estimation based on the cyclic prefix, is the integer multiple frequency offset estimation based on the reference signal, x2[n] is the electromagnetic signal after frequency offset compensation, and j is the imaginary unit.

[0025] Optionally, the steps to obtain the channel state information in Step 3 include:

[0026] Utilize the uplink channel reference signal (SRS) symbol X in the received electromagnetic signal SRS Perform channel estimation through the DFT algorithm to obtain channel state information;

[0027] The calculation expression of the signal state information is:

[0028]

[0029]

[0030] In formulas (4) and (5), D SRS is the ideal transmitted SRS symbol by the transmitter, N zc is the length of the SRS symbol, is the channel frequency-domain response estimated by the least squares algorithm, w(n) is the time-domain noise, and win(n) is the window function, is the channel time-domain impulse response, is the DFT channel estimation result.

[0031] Optionally, the expression of the frequency-domain symbol independent of the channel in step 4 is:

[0032]

[0033] In formula (6), D PUSCH is the ideal transmitted PUCSh symbol by the transmitter, is the frequency response of the PUCSh symbol (QPSK) by the transmitter radio frequency front end, is the channel frequency-domain fading of the PUCSh symbol, is the frequency response of the SRS symbol by the transmitter radio frequency front end, is the channel state information obtained in step 2, D is a constant, is the frequency-domain symbol independent of the channel.

[0034] Optionally, the steps to obtain the time-domain complex signal after power normalization in step 5 include:

[0035] Convert the frequency-domain symbol independent of the channel into a time-domain complex signal through Fourier transform, and perform power normalization on the time-domain complex signal to obtain the time-domain complex signal after power normalization;

[0036] The expression of the Fourier transform is:

[0037]

[0038] In formula (7), x′[n] is the time-domain complex signal, is the basic complex exponential function in the Fourier transform, where j is the imaginary unit, and the exponent involves the frequency-domain index k, the time-domain index n, and the total number of points N;

[0039] The expression for power normalization is as follows:

[0040]

[0041] In formula (8), abs(g) takes the absolute value, and are the real part and the imaginary part of the complex signal, respectively.

[0042] Optionally, the multi-dimensional convolutional network model in step 6 includes multiple layers of multi-scale convolutional layers, channel attention components, and residual connections. The features are processed and extracted layer by layer through the multiple layers of multi-scale convolutional layers, channel attention components, and residual connections, and the vertical merging of the feature maps is achieved through adaptive maximum pooling.

[0043] Optionally, the steps for RF fingerprint extraction and recognition in step 7 include:

[0044] Step 7.1: Use the Mix-up regularization technique to mix the features and corresponding labels of any layer in the multi-dimensional attention network model;

[0045] Step 7.2: Input the time-domain complex signal after power normalization into the multi-dimensional convolutional network model after label mixing for RF fingerprint extraction and recognition;

[0046] The expression for label mixing is as follows:

[0047]

[0048] In formula (9), (x i ′, y i ) and (x′ j , y j ) are any two groups of input label pairs, are the input features of x i ′ and x′ j at the k-th layer respectively, λ: Be(α, β) is the mixing ratio coefficient, is the multi-dimensional convolutional network model, is the result of the input signal after label mixing at the k-th layer, is the target label after mixing.

[0049] The beneficial effects of the present invention are as follows:

[0050] 1. By designing the RFF recognition frame signal, the present invention includes: PUSCH and SRS signals. The former is used for RFFI, and the latter is used to obtain channel prior knowledge. The SRS signal is used for channel estimation, and channel equalization is performed on the PUSCH signal, so that the PUSCH after equalization has the characteristic of being independent of the channel;

[0051] 2. The present invention uses the Mix-up regularization technique to introduce linear constraints and smooth labels into the model, mixes the features of any layer and the corresponding labels, and improves the robustness of the model.

[0052] 3. The core of the present invention lies in designing a cross-channel robust strategy for electromagnetic signals, adopting channel knowledge and model regularization constraints. The method has strong flexibility, fast algorithm response time, and strong engineering feasibility, and greatly improves the engineering efficiency of deep learning algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flowchart of a radio frequency fingerprint extraction and recognition method based on a cross-channel robust strategy provided by the present invention;

[0054] Figure 2 is a frequency offset compensation block diagram provided by the present invention;

[0055] Figure 3 is an RFF identification frame signal diagram provided by the present invention;

[0056] Figure 4 is a comparison diagram of channel estimation methods provided by the present invention;

[0057] Figure 5 is a comparison diagram of waveforms before and after channel equalization processing provided by the present invention;

[0058] Figure 6 is a cross-channel RFFI performance comparison diagram provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] Combined with Figure 1 This embodiment is described. The steps of a radio frequency fingerprint extraction and recognition method based on a cross-channel robust strategy in this embodiment include:

[0060] Step S1: Receive electromagnetic signals;

[0061] Step S2: Perform frame synchronization and frequency offset compensation on the received electromagnetic signals to obtain a time-frequency synchronized signal;

[0062] Further, the steps of obtaining the time-frequency synchronized signal in step S2 include:

[0063] Step S201: Use the method of a sliding window and a threshold to obtain a rough starting point of the electromagnetic signal, and use the cyclic prefix CP of the orthogonal frequency division multiplexing symbol OFDM to achieve fine synchronization to obtain a frame synchronization signal;

[0064] Step S202: Use the position of the known cyclic prefix CP in the frame synchronization signal and the reference signal to estimate the carrier frequency offset, and perform frequency offset compensation on the frame synchronization signal to obtain a time-frequency synchronized signal;

[0065] The expression for the rough starting point of the electromagnetic signal is:

[0066]

[0067] In formula (1), W is the sliding window length, S is the sliding step, T is the threshold, x is the signal sequence, and k is the position of the current sliding window;

[0068] The expression for fine synchronization is:

[0069]

[0070] In formula (2), L is the search length, N G is the sampling point of the cyclic prefix CP, N sub is the sampling point of the effective data length of the cyclic prefix CP, n0 is the frame synchronization signal, δ is the offset of the signal starting point, and i is the index of the signal sample currently being processed;

[0071] The expression for frequency offset compensation is:

[0072]

[0073] In formula (3), is the fractional multiple frequency offset estimation based on the cyclic prefix, is the integer multiple frequency offset estimation based on the reference signal, x2[n] is the electromagnetic signal after frequency offset compensation, and j is the imaginary unit.

[0074] Step S3: Perform channel estimation based on the discrete Fourier transform channel estimation algorithm to obtain the channel state information;

[0075] Furthermore, the steps to obtain the channel state information in step S3 include:

[0076] Utilize the uplink channel reference signal (SRS) symbol X in the received electromagnetic signal SRS Perform channel estimation through the DFT algorithm to obtain the channel state information, and achieve the elimination of noise and radio frequency fingerprint features outside the maximum channel delay;

[0077] The calculation expression for the signal state information is:

[0078]

[0079]

[0080] In formulas (4) and (5), D SRS is the ideal transmitted SRS symbol of the transmitter, N zc is the SRS symbol length, is the channel frequency-domain response estimated by the least-squares algorithm, w(n) is the time-domain noise, and win(n) is the window function. is the channel time-domain impulse response. is the DFT channel estimation result.

[0081] Step S4: Perform channel equalization on the time-frequency synchronization signal after channel estimation to obtain frequency-domain symbols independent of the channel.

[0082] Furthermore, the expression for the frequency-domain symbols independent of the channel in Step S4 is:

[0083]

[0084] In formula (6), is the frequency-domain symbol independent of the channel, D PUSCH is the ideal transmitted PUC SH symbol of the transmitter. is the frequency response of the transmitter RF front-end to the PUC SH symbol (QPSK). is the channel frequency-domain fading of the PUC SH symbol. is the frequency response of the transmitter RF front-end to the SRS symbol. is the channel state information obtained in Step 2, D = D PUSCH / D SRS , Since the PUSCH and SRS symbol contents remain unchanged and D is a constant, is the frequency-domain symbol independent of the channel, and D is a constant. is the frequency-domain symbol independent of the channel.

[0085] Step S5: Perform Fourier transform on the frequency-domain symbols independent of the channel and perform power normalization to obtain the time-domain complex signal after power normalization.

[0086] Furthermore, the steps to obtain the time-domain complex signal after power normalization in Step S5 include:

[0087] Convert the frequency-domain symbols independent of the channel into a time-domain complex signal through Fourier transform, and perform power normalization on the time-domain complex signal to obtain the time-domain complex signal after power normalization.

[0088] The expression for the Fourier transform is:

[0089]

[0090] In formula (7), x′[n] is the time-domain complex signal. is the basic complex exponential function in the Fourier transform, where j is the imaginary unit, and the exponent involves the frequency-domain index k, the time-domain index n, and the total number of points N.

[0091] The expression for power normalization is as follows:

[0092]

[0093] In formula (8), abs(g) is the absolute value, and are the real part and the imaginary part of the complex signal respectively.

[0094] Step S6: Construct a multi-dimensional convolutional network model using the Mix-up regularization technique;

[0095] Furthermore, the multi-dimensional convolutional network model in step S6 includes multiple layers of multi-scale convolutional layers, channel attention components, and residual connections. The features are processed and extracted layer by layer through the multiple layers of multi-scale convolutional layers, channel attention components, and residual connections, and the vertical merging of the feature maps is achieved through adaptive maximum pooling.

[0096] Step S7: Input the time-domain complex signal after power normalization into the multi-dimensional convolutional network model for RF fingerprint extraction and recognition;

[0097] Furthermore, the steps for RF fingerprint extraction and recognition in step S7 include:

[0098] Step S701: Use the Mix-up regularization technique to mix the features of any layer in the multi-dimensional attention network model with the corresponding labels to improve the generalization performance of the model;

[0099] Step S702: Input the time-domain complex signal after power normalization into the multi-dimensional convolutional network model after label mixing for RF fingerprint extraction and recognition;

[0100] The expression for label mixing is:

[0101]

[0102] In formula (9), (x i ′, y i ) and (x′ j , y j ) are any two sets of input label pairs, are the input features of x i ′, x′ j at the k-th layer respectively, λ:Be(α, β) is the mixing ratio coefficient, is the multi-dimensional convolutional network model, is the result of the input signal after label mixing at the k-th layer, is the target label after mixing.

[0103] Embodiment

[0104] Taking the physical layer security authentication of 5G devices as an example, combined withFigures 2-6 This embodiment will be described. In this embodiment, for scenarios 1 (C1) and 2 (C2), the 5G devices are respectively fixed at positions 0.2 m (LOS) and 1 m (NLOS) away from the receiver. Scenario 3 (C3) is set in an open conference room, where the distance between the 5G device and the receiver is 5 m. The system center frequency is 3.232 GHz, the bandwidth is 20 MHz, the subcarrier spacing is 30 KHz, the sampling rate is 30.72 MHz, the OFDM cyclic prefix of the transmitted signal is set to the normal mode, and the RFF frame structure is as Figure 2 shown.

[0105] Frame detection is performed on the received 5G signal. The method of using a sliding window and a threshold is used to detect the rough starting point, and then the cyclic prefix CP of the OFDM symbol is used to achieve fine synchronization, obtaining the frame synchronization signal x1[n] = x[n + n0]. Through the method of sliding correlation of the known CP position and the reference signal, the carrier frequency offset of the frame synchronization signal x1[n] is compensated for frequency offset from both aspects of IFO and FFO, obtaining the time-frequency synchronization signal as Figure 3 shown.

[0106] Utilize the SRS symbol D in the received signal SRS to perform channel estimation through the DFT algorithm. The DFT channel estimation performance is as Figure 4 shown, and the channel state information is:

[0107]

[0108]

[0109] In formulas (1) and (2), D SRS is the ideal transmitted SRS symbol of the transmitter, N zc is the length of the SRS symbol, is the channel frequency domain response estimated by the least squares algorithm, w(n) is the time domain noise, win(n) is the window function, is the channel time domain impulse response, is the DFT channel estimation result.

[0110] As Figure 5 shown, channel equalization is performed on the time-frequency synchronization signal to obtain frequency domain symbols independent of the channel:

[0111]

[0112] In formula (3), D PUSCH is the ideal transmitted PUCSh symbol of the transmitter, is the frequency response of the transmitter radio frequency front end to the PUCSh symbol (QPSK), is the frequency-domain fading of the PUSCH symbol channel, is the frequency response of the SRS symbol by the transmitter RF front end, D = D PUSCH / D SRS , Since the contents of the PUSCH and SRS symbols remain unchanged and D is a constant, is a frequency-domain symbol independent of the channel, and D is a constant., is related to the content of the transmission signal and the spectral quotient of the transmitter for different modulation methods.

[0113] For the frequency-domain symbol independent of the channel perform Fourier transform to convert it into the time domain, and perform power normalization. Input the power-normalized signal into the neural network model trained with the Mix-up regularization technique for RF fingerprint extraction and recognition.

[0114] As Figure 6 shown, the method of the present invention tests the robust performance of the model in the cross-channel for the embodiment. Figure (a) shows the difference in the channel frequency-domain response between the training set and the test set. Figure (b) is the performance evaluation of different methods in the cross-channel. Raw IQ is the power normalization of the x2[n] signal, MMSE is the traditional channel equalization method, CK-RFF is the robust RF fingerprint based on channel knowledge assistance, CKR-RFF is the robust RF fingerprint based on channel knowledge assistance and regularization, and the ordinate is the accuracy rate. It can be seen that the proposed CKR-RFF in this paper achieves the optimal performance.

[0115] The above is only a preferred embodiment of the present invention, and does not impose any form of limitation on the present invention. Although the present invention has been disclosed above with a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to make equivalent embodiments with equivalent changes within the scope of the technical solution of the present invention. However, as long as it does not depart from the technical solution content of the present invention and is based on the technical essence of the present invention, any simple modification, equivalent replacement, and improvement of the above embodiments within the spirit and principle of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A radio frequency fingerprint extraction and recognition method based on a cross-channel robust strategy, characterized in that The steps of the method for extracting and identifying radio frequency fingerprints based on a cross-channel robust strategy include: Step 1: Receive electromagnetic signals; Step 2: Perform frame synchronization and frequency offset compensation on the received electromagnetic signals to obtain time-frequency synchronized signals; Step 3: Perform channel estimation based on the discrete Fourier transform channel estimation algorithm to obtain channel state information; Step 4: Perform channel equalization on the time-frequency synchronized signals after channel estimation to obtain frequency domain symbols independent of the channel; Step 5: Perform Fourier transform on the frequency domain symbols independent of the channel and perform power normalization to obtain power-normalized time-domain complex signals; Step 6: Use the Mix-up regularization technique to construct a multi-dimensional convolutional network model; Step 7: Input the power-normalized time-domain complex signals into the multi-dimensional convolutional network model for radio frequency fingerprint extraction and identification; The steps of performing radio frequency fingerprint extraction and identification in Step 7 include: Step 7.1: Use the Mix-up regularization technique to mix the features and corresponding labels of any layer in the multi-dimensional attention network model; Step 7.2: Input the power-normalized time-domain complex signals into the multi-dimensional convolutional network model after label mixing for radio frequency fingerprint extraction and identification; The expression for label mixing is: （9） In formula (9), and are any two sets of input label pairs, are respectively the input features of the k-th layer, is the mixing ratio coefficient, is the multi-dimensional convolutional network model, is the result after label mixing of the input signal at the k-th layer, is the target label after mixing.

2. The radio frequency fingerprint extraction and recognition method based on a cross-channel robust strategy according to claim 1, wherein The steps of obtaining the time-frequency synchronized signals in Step 2 include: Step 2.1: Use the method of sliding window and threshold to obtain the rough starting point of the electromagnetic signals, and use the cyclic prefix CP of the orthogonal frequency division multiplexing symbol OFDM to achieve fine synchronization to obtain frame synchronization signals; Step 2.2: Use the position of the known cyclic prefix CP in the frame synchronization signals and the reference signals to estimate the carrier frequency offset, and perform frequency offset compensation on the frame synchronization signals to obtain time-frequency synchronized signals; The expression for the rough starting point of the electromagnetic signals is: （1） In formula (1), is the sliding window length, is the sliding step size, is the threshold value, x is the signal sequence, k is the position of the current sliding window; The expression for fine synchronization is: （2） In formula (2), N G is the total number of sampling points of the cyclic prefix CP, N sub is the effective sampling length of the cyclic prefix CP, n 0 is the frame synchronization signal, is the offset of the signal starting point, i is the index of the signal sample currently being processed; The expression for frequency offset compensation is: （3） In formula (3), is the fractional frequency offset estimation based on the cyclic prefix, is the integer frequency offset estimation based on the reference signal, is the electromagnetic signal after frequency offset compensation, j is the imaginary unit.

3. A radio frequency fingerprint extraction and recognition method based on a cross-channel robust strategy according to claim 1, characterized in that The steps of obtaining the channel state information in Step 3 include: Using the uplink channel reference signal SRS symbols in the received electromagnetic signal Performing channel estimation through the DFT algorithm to obtain channel state information; The calculation expression for the signal state information is: （4） （5） In Formulas (4) and (5), is the ideal transmitted SRS symbol by the transmitter, is the SRS symbol length, is the channel frequency-domain response estimated by the least squares algorithm, is the time-domain noise, is the window function, is the channel time-domain impulse response, is the DFT channel estimation result.

4. A radio frequency fingerprint extraction and recognition method based on a cross-channel robust strategy according to claim 1, characterized in that The expression for the frequency domain symbols independent of the channel in Step 4 is: （6） In formula (6), is the ideal transmitted PUCSh symbol of the transmitter, is the QPSK frequency response of the transmitter RF front-end to the PUCSh symbol, is the channel frequency-domain fading of the PUCSh symbol, is the frequency response of the transmitter RF front-end to the SRS symbol, is the channel state information obtained in step 3, is a constant, is a frequency-domain symbol independent of the channel.

5. The radio frequency fingerprint extraction and recognition method based on a cross-channel robust strategy according to claim 4, characterized in that, The steps of obtaining the power-normalized time-domain complex signals in Step 5 include: Transform the frequency domain symbols independent of the channel into time-domain complex signals through Fourier transform, and perform power normalization on the time-domain complex signals to obtain power-normalized time-domain complex signals; The expression for Fourier transform is: （7） In formula (7), is a complex signal in the time domain, is the basic complex exponential function in the Fourier transform, where j is the imaginary unit, and the exponent involves the frequency-domain index k, the time-domain index n, and the total number of points N; The expression for power normalization is: （8） In formula (8), is to take the absolute value, and are the real part and the imaginary part of the complex signal respectively.

6. The radio frequency fingerprint extraction and recognition method based on a cross-channel robust strategy according to claim 1, wherein The multi-dimensional convolutional network model in Step 6 includes multiple layers of multi-scale convolutional layers, channel attention components, and residual links. Through the multiple layers of multi-scale convolutional layers, channel attention components, and residual links, features are processed and extracted layer by layer, and vertical merging of the feature maps is achieved through adaptive maximum pooling.

Images