A radio frequency fingerprint-based unmanned aerial vehicle identity authentication method

Abstract

The application discloses a kind of unmanned plane identity authentication methods based on radio frequency fingerprint, it is related to information security technical field.The present application includes the following steps: antenna receiving array is arranged at base station side, and the radio frequency signal data of unmanned plane collected is preprocessed by server;Synchronous calibration is carried out to the multi-dimensional feature after preprocessing, and feature size chart is generated;The logical identification of unmanned plane is verified, and the server is matched to the radio frequency fingerprint of unmanned plane;After radio frequency fingerprint matching succeeds, the secondary verification model is established in combination with the flight trajectory of unmanned plane;Specific test signal is injected to unmanned plane to detect the camouflage behavior of unmanned plane.The present application generates feature size chart by synchronous calibration to the multi-dimensional feature of the radio frequency information of unmanned plane, and the feature size chart is matched to the radio frequency fingerprint library of unmanned plane of server, and the multiple features are fused into feature size chart to carry out identification, increase the receptive field of network, improve the radio frequency fingerprint identification performance, accuracy and reliability.

Description

Technical Field

[0001] This invention belongs to the field of information security technology, and in particular relates to a method for drone identity authentication based on radio frequency fingerprinting. Background Technology

[0002] Radio frequency fingerprinting technology, a key technology in the field of network security identity authentication, refers to a technology that uses signal processing at the physical layer to extract features from collected wireless signals, thereby enabling individual identification of the radiation source device. Its main idea is to extract unique features from wireless devices to generate unforgeable signatures. Unique fingerprints used to identify wireless devices can prevent spoofing or imitation attacks.

[0003] With the widespread application of drones in logistics, surveying, and other fields, their identity security issues are becoming increasingly prominent. Current mainstream authentication technologies have the following limitations: The shortcomings of traditional drone authentication methods are as follows: (1) MAC address authentication is easily tampered with and forged, and hackers can easily copy it through SDR devices; (2) Digital certificate schemes require pre-set keys, which increases the hardware cost of drones by more than 20%; (3) Biometric technology is affected by ambient light, and the false recognition rate is as high as 15%. The bottlenecks of existing radio frequency fingerprinting technology are as follows: (1) The feature dimensions collected by a single antenna are insufficient (usually <6 dimensions); (2) Static matching threshold leads to an increase in rejection rate under poor channel conditions; (3) There is a lack of active defense mechanism against device cloning; (4) Literature shows that the existing scheme has a discrimination rate of less than 60% for the same type of UAV. Summary of the Invention

[0004] The purpose of this invention is to provide a drone identity authentication method based on radio frequency fingerprints. By collecting multi-dimensional features of drone radio frequency information and synchronously calibrating them to generate a feature size map, the feature size map is matched with the drone radio frequency fingerprint database on the server, thus solving the problems of insufficient feature dimensions and insufficient radio frequency fingerprint recognition performance in existing identity authentication methods.

[0005] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: This invention relates to a method for drone identity authentication based on radio frequency fingerprinting, comprising the following steps: Step S1: Deploy an antenna receiving array on the base station side to collect the UAV radio frequency signal at 0.5ms intervals; Step S2: The collected UAV radio frequency signal data undergoes signal preprocessing via a server; Step S3: Simultaneously calibrate the preprocessed multi-dimensional features to generate a feature size map; Step S4: Verify the drone's logical identifier. If the verification is successful, proceed to step S5. If the verification fails, issue an alarm directly. Step S5: The server performs radio frequency fingerprint matching on the drones verified by the logical identifier; Step S6: After successful RF fingerprint matching, a secondary verification model is established by combining the drone's flight trajectory; Step S7: Inject specific test signals into the abnormal drone to detect the drone's camouflage behavior.

[0006] As a preferred technical solution, in step S1, the antenna receiving array includes an antenna unit, a radio frequency link unit, and a synchronization control unit; the antenna unit is a dual-polarized microstrip patch antenna loaded with an adjustable phase shifter; the dual-polarized microstrip patch antenna supports vertical / horizontal dual-polarization multiplexing, each antenna unit includes a parasitic patch to assist in extending bandwidth, and when the antenna units are spatially deployed, they adopt a double-layer diamond topology arrangement, with the horizontal layer spacing set to 1 / 4 wavelength (12.9 mm for the 5.8 GHz band), and the vertical layer spacing extended to 3 / 8 wavelength to reduce mutual coupling effects; The radio frequency link unit includes a receiving channel module and a phase control module; the receiving channel module is a 16-channel zero-IF receiving link; each channel includes a low-noise amplifier, a bandpass filter, and a variable gain amplifier; the phase control module is an integrated digitally controlled phase shifter (adjustment accuracy 0.5°), and the phase control module implements adaptive beamforming through an FPGA; The synchronization control unit includes a spatiotemporal network module and a timing calibration module; the spatiotemporal network module uses an OCXO temperature-controlled crystal oscillator (frequency stability ±0.1ppb) to distribute clock signals to 16 channels through an H-tree topology; for each channel, a digital delay compensation module (resolution 5ps) is used to automatically align the IQ sampling time deviation (error <1ns).

[0007] As a preferred technical solution, in step S2, since the original radio frequency signal is easily affected by environmental noise, multipath effects, and nonlinear characteristics of the equipment, preprocessing can improve feature recognizability and classification accuracy. Experiments show that the signal-to-noise ratio of the unprocessed signal can decrease by more than 15dB, directly affecting the subsequent fingerprint extraction effect. Therefore, the specific process of signal preprocessing on the server is as follows: Step S21, Noise Suppression: Adaptive wavelet threshold denoising is used to eliminate Gaussian white noise, suppress background noise, and improve the signal-to-noise ratio; Step S22, Power Normalization: Dynamically adjust the signal amplitude to a uniform dimension to eliminate the influence of power fluctuations; Step S23, Signal Segmentation: Based on Ye Bayesian variable point detection, the transient / steady-state stages are divided, and the effective feature regions are separated; Step S24, Frequency Domain Reconstruction: Apply variational mode decomposition to separate co-frequency interference components, and use the EMD algorithm to extract intrinsic mode functions to suppress frequency band overlapping interference; Step S25, Phase Compensation: The carrier phase offset is calibrated by Hibert transform to eliminate hardware crystal oscillator offset; Step S26, Data Dimensionality Reduction: Principal component analysis compresses redundant features, improving computational efficiency and classification accuracy.

[0008] As a preferred technical solution, the preprocessed radio frequency signal data undergoes multi-dimensional feature extraction. The information extracted in the multi-dimensional feature extraction includes I / Q quadrature components, transient response characteristics, time-domain overshoot coefficient, symbol clock jitter, spectral envelope profile, spectral asymmetry, out-of-band radiation intensity, phase noise characteristics, modulation error rate, carrier frequency offset, multipath delay spread, and phase trajectory correlation. By analyzing and extracting these radio frequency signals, a unique feature of the UAV is formed as an authentication credential, namely, a radio frequency fingerprint.

[0009] The 12 physical layer features are explained below: The I / Q quadrature components are: a complex signal sequence composed of in-phase and quadrature components is extracted, and the real and imaginary parts are separated by Hilbert transform to form two-dimensional time-domain features; The transient response characteristics are: capturing the duration of overshoot and ringing during the signal activation phase, and measuring the rise time (the time delay corresponding to the 10%-90% amplitude range). The time-domain overshoot coefficient is the percentage by which the maximum amplitude exceeds the steady-state value. The symbol clock jitter is the symbol period offset caused by the analysis clock recovery circuit, and the TIE is measured using eye diagram analysis. The spectral envelope profile is used to extract the amplitude attenuation slope of the main lobe and side lobes of the signal, and a normalized power spectral density distribution is generated by FFT transformation. The spectral asymmetry is calculated by measuring the power asymmetry of the sidebands on both sides of the carrier. The out-of-band radiation intensity is obtained by measuring the leakage power at a distance of ±1.5 times the bandwidth from the carrier center, and a Gaussian window function is used to suppress spectral leakage. The phase noise characteristic is a quantification of the random fluctuation of the carrier phase, and the noise power at a 1kHz offset is measured by a phase noise analyzer. The modulation error rate is calculated by taking the average Euclidean distance between the ideal constellation point and the measured point. The carrier frequency offset is used to detect the deviation between the actual carrier frequency and the nominal value, and the offset value is estimated by the cyclic prefix correlation method. The multipath delay spread is obtained by extracting the maximum delay spread parameter through the channel impulse response and calculating the root mean square delay spread based on the Rayleigh fading model. The phase trajectory correlation analysis reveals the statistical correlation of phase changes between consecutive symbols, and the autocorrelation function is used to quantify the phase continuity characteristics.

[0010] As a preferred technical solution, the multi-dimensional feature synchronous calibration process in step S3 is as follows: Step S31: Configure an independent timestamp marker for each antenna element with an accuracy of 10ns; Step S32: Use a cross-correlation algorithm to align the sampling start points of the 16 channels and eliminate time delay differences; Step S33: Establish a three-dimensional coordinate system (azimuth, elevation, and distance) to map the spatial positions of each antenna; Step S34: Normalize the 12 features respectively; among them, I / Q quadrature components, transient response characteristics, and time-domain overshoot coefficient belong to amplitude-type features, and the normalization of amplitude-type features adopts logarithmic compression processing to directly compress the dynamic range to 60dB; symbol clock jitter, phase noise characteristics, and phase trajectory correlation belong to phase-type features, and cyclic interpolation normalization processing is used for phase-type features to map the features to The interval; spectral envelope profile, spectral asymmetry, out-of-band radiation intensity, modulation error rate, carrier frequency offset, and multipath delay spread belong to frequency domain features. The normalization of frequency domain features is based on bandpass normalization of Butterworth filters.

[0011] Step S35: Divide the time window into 512ms segments, generate 1024 sampling points at 0.5ms intervals, extend the time series to 256 points using cubic spline interpolation, and apply sliding window differential encoding. The specific formula is as follows: ; In the formula, Let F be the differential rate of change of the physical quantity F at time t. Let F represent the observed values ​​of the physical quantity F at times t+1 and t-1, respectively. This represents a fixed time interval between adjacent sampling time points; Step S36: Convert the 4*4 antenna array into an 8*32 spatial coding matrix to generate a 32-channel virtual array, and use spatial convolution kernels for feature enhancement; Step S37: Construct multidimensional tensors and matrices, create a four-dimensional original data volume (16 antennas * 12 features * 256 time points * 8 spatial partitions), and adjust the feature size map to 256 * 256 using bilinear interpolation; Step S38: Load the pre-built channel impulse response library, apply hyperbolic tangent transform to enhance feature contrast, and add artificial perturbation factors to improve the robustness of the model.

[0012] As a preferred technical solution, in step S4, when verifying the logical identifier of the UAV, the validity of the first 24 digits of the UAV manufacturer code is verified through the IEEE standard OUI database. The server deploys a Bloom filter to match the blacklist and whitelist of UAVs in the database and synchronizes the MAC address registration database of the Civil Aviation Administration's UAV supervision system in real time.

[0013] As a preferred technical solution, the matching process for the radio frequency fingerprint in step S5 is as follows: Step S51: Input the feature size map generated by the RF fingerprint into the ResNet50 recognition model; Step S52: Pass through a convolutional layer with a kernel size of 7*7; Step S53: Use a max pooling layer with a window size of 3*3 to change the image's channels and size; Step S54: Connect each layer of the 4 sets of ResNet residual units to all the previous layers, and then replace the 3*3 convolution of Conv4 with three convolutions of 3*3, 3*1 and 3*3 to improve the efficiency and accuracy of the network. Step S55: Extract multi-scale radio frequency fingerprint features from different spaces and perform linear superposition and fusion to enhance the extraction of radio frequency fingerprint features. This is achieved by decomposing high-dimensional convolutions into 32 identical low-dimensional convolutions and fusing them. The network learns different radio frequency fingerprint features. Finally, the extracted radio frequency fingerprint features are input into the average pooling layer, the fully connected layer, and the softmax layer to complete the classification and recognition task of generating feature size maps of radio frequency fingerprints.

[0014] As a preferred technical solution, in step S54, when each layer of the four sets of ResNet residual units is connected to all the preceding layers, the output of the first layer is... Use the quick connection 'connect1' to change the shape so that it meets the input of the third layer, and then connect it to the output of the second layer. After fusion, the input is processed by the third layer and then output after passing through Conv3. Then output the first layer. Second layer output Use quick connections connect2 and connect3 to change the shape to meet the input of the fourth layer, and then connect it to the output of the third layer. After merging, the input is processed by the fourth layer and then output after passing through Conv4. The specific formula is as follows: ; In the formula, These are the residual layers of the network model. The output, Quick links The output of .

[0015] As a preferred technical solution, the specific process for verifying the flight trajectory of the UAV in step S6 is as follows: Step S61: Deploy an ADS-B receiver to capture the latitude, longitude, and altitude information of the UAV. A radar-assisted positioning system can also be used to compensate for GPS spoofing errors. Step S62: Establish a standard flight feature library to verify the smoothness, speed, and spatial decision-making of the UAV flight trajectory; The trajectory smoothness is obtained by calculating the standard deviation of the rate of change of curvature (threshold < 0.15 rad / s²); the velocity distribution is obtained by statistically analyzing the ratio of maximum acceleration to average velocity (threshold range 0.3-1.2); and the spatial decision is obtained by analyzing the correlation between the steering angle and flight altitude (Pearson coefficient > 0.7 is considered valid). Step S63: Real-time parsing of MAC frame structure to establish a baseline model of communication behavior; the parsed MAC frame structure includes data packet interval and payload length; the periodicity of data transmission can be statistically analyzed through data packet interval (standard deviation < 50 μs), and abnormal load can be detected through payload length, triggering an alarm if it exceeds the baseline value ±20%; Step S64: Construct a hidden Markov model to describe the state transition probability; by acquiring six flight phases of the UAV, including takeoff, cruise, hovering, and landing, observe the UAV's communication strength, frequency band switching frequency, and control command density. By comparing these observed parameters with a preset index library, determine whether there are any abnormalities in the UAV's flight behavior.

[0016] As a preferred technical solution, the specific process of injecting a specific test signal into the drone to detect the drone's camouflage behavior in step S7 is as follows: Step S71: Using OFDM subcarrier masking technology, a pseudo-random QPSK modulated signal is inserted into the normal communication frequency band; the dynamic parameters of the pseudo-random QPSK modulated signal include: center frequency offset (±500kHz, which can be adjusted by programming), power level (15-20dB lower than the main signal) and duration (randomly varying from 50-200μs). Step S72: Embed the drone's unique identifier; the unique identifier also includes an encrypted timestamp and a base station location fingerprint; wherein, the encrypted timestamp is processed by SHA-256 hashing, and the base station location fingerprint is obtained by encoding geographic coordinates latitude and longitude; Step S73: Generate dynamic phase perturbation code, each chip containing 8 bits of device feature information; Step S74: Insert test signals using MAC layer frame gaps; Step S75: Analyze the response characteristics of the UAV and detect camouflage information.

[0017] Legitimate drone characteristics: Information reflection delay fluctuation <50ns, out-of-band radiated harmonic distribution conforms to the equipment model characteristics, and DPPC decoding success rate >98%; Characteristics of camouflaged drones: The time domain response exhibits a fixed pattern (standard deviation <10ns), and the signal nonlinear distortion is abnormal (EVM value offset >8%).

[0018] The present invention has the following beneficial effects: (1) The present invention collects multi-dimensional features of UAV radio frequency information, performs synchronous calibration to generate feature size map, matches the feature size map with the UAV radio frequency fingerprint database of the server, and integrates multiple features into feature size map for recognition, thereby increasing the receptive field of the network and improving the performance, accuracy and reliability of radio frequency fingerprint recognition.

[0019] (2) This invention achieves a two-stage filtering architecture by combining noise suppression and frequency domain reconstruction. First, it eliminates broadband noise through wavelet denoising and then suppresses narrowband co-frequency interference through VND decomposition. Second, it introduces closed-loop feedback in the phase compensation stage and dynamically adjusts the compensation parameters based on the transient signal start-up detection results. Finally, considering the time-varying nature of UAV signals, it uses sliding window local PCA instead of global dimensionality reduction to preserve transient feature details and significantly improve the differentiation of preprocessed signal features, thereby improving the accuracy of legal / illegal UAV classification.

[0020] (3) The present invention adopts a two-factor authentication mechanism to first verify the logical identifiers such as the MAC address of the UAV and then perform radio frequency fingerprint matching. When both the logical identifier and the radio frequency fingerprint of the UAV pass the verification, the flight trajectory, communication mode and other schemes are used to verify the behavior of the UAV in real time, thereby improving the security of UAV identity authentication.

[0021] (4) This invention improves the security of drone identity authentication by designing a security protection mechanism, injecting specific test signals to detect the spoofing behavior of drones, triggering spectrum interference blocking after three consecutive authentication failures, and dynamically associating radio frequency fingerprints with encryption keys.

[0022] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a drone identity authentication method based on radio frequency fingerprinting according to the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0027] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figure 1 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0028] Please see Figure 1 As shown, the present invention is a drone identity authentication method based on radio frequency fingerprinting, comprising the following steps: Step S1: Deploy an antenna receiving array on the base station side to collect the UAV radio frequency signal at 0.5ms intervals; Step S2: The collected UAV radio frequency signal data undergoes signal preprocessing via a server; Step S3: Simultaneously calibrate the preprocessed multi-dimensional features to generate a feature size map; Step S4: Verify the drone's logical identifier. If the verification is successful, proceed to step S5. If the verification fails, issue an alarm directly. Step S5: The server performs radio frequency fingerprint matching on the drones verified by the logical identifier; Step S6: After successful RF fingerprint matching, a secondary verification model is established by combining the drone's flight trajectory; Step S7: Inject specific test signals into the abnormal drone to detect the drone's camouflage behavior.

[0029] In step S1, the antenna receiving array includes antenna elements, radio frequency link elements, and a synchronization control unit; the antenna elements are dual-polarized microstrip patch antennas loaded with adjustable phase shifters, which support ±45° polarization switching; the dual-polarized microstrip patch antennas support vertical / horizontal dual-polarization multiplexing, and each antenna element includes a parasitic patch to assist in extending the bandwidth. When the antenna elements are deployed in space, a double-layer diamond topology is adopted, with the horizontal layer spacing set to 1 / 4 wavelength (12.9 mm for the 5.8 GHz band) and the vertical layer spacing extended to 3 / 8 wavelength to reduce mutual coupling effects; The RF link unit includes a receiving channel module and a phase control module; the receiving channel module is a 16-channel zero-IF receiving link; each channel includes a low-noise amplifier, a bandpass filter, and a variable gain amplifier; the phase control module is an integrated digitally controlled phase shifter (adjustment accuracy 0.5°), and the phase control module implements adaptive beamforming through FPGA, and generates 16 synchronous trigger pulses (jitter ≤ 5ps) through FPGA to achieve precise alignment of each channel; The synchronization control unit includes a spatiotemporal network module and a timing calibration module; the spatiotemporal network module uses an OCXO temperature-controlled crystal oscillator (frequency stability ±0.1ppb) to distribute clock signals to 16 channels through an H-tree topology; each is equipped with a digital delay compensation module (resolution 5ps) and uses automatic alignment IQ sampling time deviation (error <1ns).

[0030] In step S2, since the original radio frequency signal is susceptible to interference from environmental noise, multipath effects, and nonlinear characteristics of the equipment, preprocessing can improve feature recognizability and classification accuracy. Experiments show that the signal-to-noise ratio of the unprocessed signal can decrease by more than 15dB, directly affecting the subsequent fingerprint extraction effect. Therefore, the specific process of signal preprocessing on the server is as follows: Step S21, Noise Suppression: Adaptive wavelet threshold denoising is used to eliminate Gaussian white noise, suppress background noise, and improve the signal-to-noise ratio; the specific method consists of two steps: Step 1: Use the sym8 wavelet basis for 5-level decomposition, and apply soft thresholding to the high-frequency coefficients (threshold formula is...). , Step 1: Optimize threshold rotation using Stein unbiased risk estimation (for noise standard deviation) to preserve signal edge features; Step 2: Perform variational mode decomposition to separate the intrinsic modes of the signal, with a preset mode number K=6 and penalty factor a=2000; Filter effective mode components (threshold>0.7) based on correlation coefficient to reconstruct the target signal; Step S22, Power Normalization: Dynamically adjust the signal amplitude to a uniform dimension to avoid interference from differences in UAV transmission power and eliminate the impact of power fluctuations; Step S23, Signal Segmentation: Based on Ye Bayesian variable point detection, the transient / steady-state stages are divided, and the effective signal segments are located through cyclic stationarity analysis to separate the effective feature regions; Step S24, Frequency Domain Reconstruction: Apply variational mode decomposition to separate co-frequency interference components, and use the EMD algorithm to extract intrinsic mode functions to suppress frequency band overlapping interference; Step S25, Phase Compensation: The carrier phase offset is calibrated by Hibert transform, the frequency is dynamically corrected, and the hardware crystal oscillator offset is eliminated; Step S26, Data Dimensionality Reduction: Principal component analysis compresses redundant features, improving computational efficiency and classification accuracy.

[0031] By combining noise suppression and frequency domain reconstruction, a two-stage filtering architecture is formed by first eliminating broadband noise through wavelet denoising and then suppressing narrowband co-frequency interference through VND decomposition. Secondly, closed-loop feedback is introduced in the phase compensation stage, and the compensation parameters are dynamically adjusted in combination with the transient signal start-up detection results. Finally, in view of the time-varying characteristics of UAV signals, sliding window local PCA is used instead of global dimensionality reduction to preserve transient feature details.

[0032] After preprocessing, the radio frequency signal data undergoes multi-dimensional feature extraction. The extracted information includes I / Q quadrature components, transient response characteristics, time-domain overshoot coefficient, symbol clock jitter, spectral envelope profile, spectral asymmetry, out-of-band radiation intensity, phase noise characteristics, modulation error rate, carrier frequency offset, multipath delay spread, and phase trajectory correlation. By analyzing and extracting these radio frequency signals, a unique feature of the UAV is formed as an authentication credential, namely, a radio frequency fingerprint.

[0033] The 12 physical layer features are explained below: I / Q quadrature components: Extract the complex signal sequence composed of in-phase and quadrature components, separate the real and imaginary parts through Hilbert transform, and form two-dimensional time domain features; Transient response characteristics: Capture the duration of overshoot and ringing during the signal activation phase, and measure the rise time (the time delay corresponding to the 10%-90% amplitude range). The time-domain overshoot factor is the percentage by which the calculated maximum amplitude exceeds the steady-state value; Symbol clock jitter is the symbol period offset caused by the analysis clock recovery circuit, and TIE is measured using eye diagram analysis. The spectral envelope profile is used to extract the amplitude attenuation slope of the main lobe and side lobes of the signal, and the normalized power spectral density distribution is generated by FFT transformation. Spectral asymmetry is calculated by assessing the power asymmetry of the sidebands on both sides of the carrier. Out-of-band radiation intensity was measured by measuring the leakage power at a distance of ±1.5 times the bandwidth from the carrier center, and a Gaussian window function was used to suppress spectral leakage. Phase noise characteristics are used to quantify the degree of random fluctuation in the carrier phase, and the noise power at a 1kHz offset is measured using a phase noise analyzer. The modulation error rate is calculated by taking the average Euclidean distance between the ideal constellation point and the measured point. Carrier frequency offset is used to detect the deviation between the actual carrier frequency and the nominal value. The offset value is estimated by the cyclic prefix correlation method. Multipath delay spread is extracted by the maximum delay spread parameter through the channel impulse response, and the root mean square delay spread is calculated based on the Rayleigh fading model; Phase trajectory correlation analysis is used to determine the statistical correlation of phase changes between consecutive symbols, and the autocorrelation function is used to quantify the phase continuity characteristics.

[0034] In step S3, the multi-dimensional feature synchronous calibration process is as follows: Step S31: Configure an independent timestamp marker for each antenna element with an accuracy of 10ns; Step S32: Use a cross-correlation algorithm to align the sampling start points of the 16 channels and eliminate time delay differences; Step S33: Establish a three-dimensional coordinate system (azimuth, elevation, and distance) to map the spatial positions of each antenna; Step S34: Normalize the 12 features respectively; among them, I / Q quadrature components, transient response characteristics, and time-domain overshoot coefficient belong to amplitude-type features, and the normalization of amplitude-type features adopts logarithmic compression processing to directly compress the dynamic range to 60dB; symbol clock jitter, phase noise characteristics, and phase trajectory correlation belong to phase-type features, and cyclic interpolation normalization processing is used for phase-type features to map the features to The interval; spectral envelope profile, spectral asymmetry, out-of-band radiation intensity, modulation error rate, carrier frequency offset, and multipath delay spread belong to frequency domain features. The normalization of frequency domain features is based on bandpass normalization of Butterworth filters.

[0035] Step S35: Divide the time window into 512ms segments, generate 1024 sampling points at 0.5ms intervals, extend the time series to 256 points using cubic spline interpolation, and apply sliding window differential encoding. The specific formula is as follows: ; In the formula, Let F be the differential rate of change of the physical quantity F at time t. Let F represent the observed values ​​of the physical quantity F at times t+1 and t-1, respectively. This represents a fixed time interval between adjacent sampling time points; Step S36: Convert the 4*4 antenna array into an 8*32 spatial coding matrix to generate a 32-channel virtual array, and use spatial convolution kernels for feature enhancement; Step S37: Construct multidimensional tensors and matrices, create a four-dimensional original data volume (16 antennas * 12 features * 256 time points * 8 spatial partitions), and adjust the feature size map to 256 * 256 using bilinear interpolation; Step S38: Load the pre-built channel impulse response library, apply hyperbolic tangent transform to enhance feature contrast, and add artificial perturbation factors to improve the robustness of the model.

[0036] In step S4, during the logical identification verification of the drone, the validity of the first 24-bit drone manufacturer code is verified through the IEEE standard OUI database. The verification is performed by parsing the control field of the 802.11 frame. The server deploys a Bloom filter to match the blacklist and whitelist of drones in the database and synchronizes the MAC address registration database of the Civil Aviation Administration's drone supervision system in real time. After the logical identification verification of the drone is passed, the radio frequency fingerprint of the drone is verified.

[0037] In step S5, the matching process for the radio frequency fingerprint is as follows: Step S51: Input the feature size map generated by the RF fingerprint into the ResNet50 recognition model; Step S52: Pass through a convolutional layer with a kernel size of 7*7; Step S53: Use a max pooling layer with a window size of 3*3 to change the image's channels and size; Step S54: Connect each layer of the 4 sets of ResNet residual units to all the previous layers, and then replace the 3*3 convolution of Conv4 with three convolutions of 3*3, 3*1 and 3*3 to improve the efficiency and accuracy of the network. Step S55: Extract multi-scale radio frequency fingerprint features from different spaces and perform linear addition and fusion to enhance the extraction of radio frequency fingerprint features. This is achieved by decomposing a high-dimensional convolution into 32 identical low-dimensional convolutions and fusing them. The network learns different radio frequency fingerprint features. Finally, the extracted radio frequency fingerprint features are input into an average pooling layer, a fully connected layer, and a softmax layer to complete the classification and recognition task of generating feature size maps of radio frequency fingerprints.

[0038] In step S54, when each layer of the four sets of ResNet residual units is connected to all the previous layers, the output of the first layer is... Use the quick connection 'connect1' to change the shape so that it meets the input of the third layer, and then connect it to the output of the second layer. After fusion, the input is processed by the third layer and then output after passing through Conv3. Then output the first layer. Second layer output Use quick connections connect2 and connect3 to change the shape to meet the input of the fourth layer, and then connect it to the output of the third layer. After merging, the input is processed by the fourth layer and then output after passing through Conv4. The specific formula is as follows: ; In the formula, These are the residual layers of the network model. The output, Quick links The output of .

[0039] In step S6, the specific process for verifying the flight trajectory of the drone is as follows: Step S61: Deploy an ADS-B receiver to capture the latitude, longitude, and altitude information of the UAV. A radar-assisted positioning system can also be used to compensate for GPS spoofing errors. Step S62: Establish a standard flight feature library to verify the smoothness, speed, and spatial decision-making of the UAV flight trajectory; The trajectory smoothness is obtained by calculating the standard deviation of the rate of change of curvature (threshold < 0.15 rad / s²); the velocity distribution is obtained by statistically analyzing the ratio of maximum acceleration to average velocity (threshold range 0.3-1.2); and the spatial decision is obtained by analyzing the correlation between the steering angle and flight altitude (Pearson coefficient > 0.7 is considered valid). Step S63: Real-time parsing of MAC frame structure to establish a baseline model of communication behavior; the parsed MAC frame structure includes data packet interval and payload length; the periodicity of data transmission can be statistically analyzed through data packet interval (standard deviation < 50 μs), and abnormal load can be detected through payload length, triggering an alarm if it exceeds the baseline value ±20%; Step S64: Construct a hidden Markov model to describe the state transition probability; by acquiring six flight phases of the UAV, including takeoff, cruise, hovering, and landing, observe the UAV's communication strength, frequency band switching frequency, and control command density. By comparing these observed parameters with a preset index library, determine whether there are any abnormalities in the UAV's flight behavior.

[0040] In step S7, the specific process for injecting a specific test signal into the drone to detect its camouflage behavior is as follows: Step S71: Using OFDM subcarrier masking technology, a pseudo-random QPSK modulated signal is inserted into the normal communication frequency band; the dynamic parameters of the pseudo-random QPSK modulated signal include: center frequency offset (±500kHz, which can be adjusted by programming), power level (15-20dB lower than the main signal) and duration (randomly varying from 50-200μs). Step S72: Embed the drone's unique identifier; the unique identifier also includes an encrypted timestamp and a base station location fingerprint; wherein, the encrypted timestamp is processed by SHA-256 hashing, and the base station location fingerprint is obtained by encoding geographic coordinates latitude and longitude; Step S73: Generate dynamic phase perturbation code, each chip containing 8 bits of device feature information; Step S74: Insert test signals using MAC layer frame gaps; Step S75: Analyze the response characteristics of the UAV and detect camouflage information.

[0041] Legitimate drone characteristics: Information reflection delay fluctuation <50ns, out-of-band radiated harmonic distribution conforms to the equipment model characteristics, and DPPC decoding success rate >98%; Characteristics of camouflaged drones: The time domain response exhibits a fixed pattern (standard deviation <10ns), and the signal nonlinear distortion is abnormal (EVM value offset >8%).

[0042] It is worth noting that the various units included in the above system embodiments are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of the present invention.

[0043] Furthermore, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware, and the corresponding program can be stored in a computer-readable storage medium.

[0044] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for drone identity authentication based on radio frequency fingerprinting, characterized in that, The steps include the following: Step S1: Deploy an antenna receiving array on the base station side to collect the UAV radio frequency signal at 0.5ms intervals; Step S2: The collected UAV radio frequency signal data undergoes signal preprocessing via a server; Step S3: Simultaneously calibrate the preprocessed multi-dimensional features to generate a feature size map; Step S4: Verify the drone's logical identifier. If the verification is successful, proceed to step S5. If the verification fails, issue an alarm directly. Step S5: The server performs radio frequency fingerprint matching on the drones verified by the logical identifier; Step S6: After successful RF fingerprint matching, a secondary verification model is established by combining the drone's flight trajectory; Step S7: Inject specific test signals into the abnormal drone to detect the drone's camouflage behavior; In step S1, the antenna receiving array includes an antenna unit, a radio frequency link unit, and a synchronization control unit; the antenna unit is a dual-polarized microstrip patch antenna loaded with an adjustable phase shifter; when the antenna unit is deployed in space, it adopts a double-layer diamond topology arrangement, with the horizontal layer spacing set to 1 / 4 wavelength and the vertical layer spacing extended to 3 / 8 wavelength. The radio frequency link unit includes a receiving channel module and a phase control module; the receiving channel module is a 16-channel zero-IF receiving link; each channel includes a low-noise amplifier, a bandpass filter, and a variable gain amplifier; the phase control module is an integrated digitally controlled phase shifter; The synchronization control unit includes a spatiotemporal network module and a timing calibration module; the spatiotemporal network module uses an OCXO temperature-controlled crystal oscillator to distribute clock signals to 16 channels through an H-tree topology. In step S2, the specific process of signal preprocessing by the server is as follows: Step S21, Noise Suppression: Adaptive wavelet threshold denoising is used to eliminate Gaussian white noise; Step S22, Power Normalization: Dynamically adjust the signal amplitude to a uniform dimension; Step S23, Signal Segmentation: Divide the transient / steady-state phases based on Ye Bayesian variable point detection; Step S24, Frequency Domain Reconstruction: Apply variational mode decomposition to separate co-frequency interference components, and use the EMD algorithm to extract intrinsic mode functions; Step S25, Phase Compensation: Calibrate the carrier phase offset using Hibert transform; Step S26, Data Dimensionality Reduction: Principal Component Analysis to Compress Redundant Features; After the preprocessing is completed, the radio frequency signal data is subjected to multi-dimensional feature extraction. The information extracted by multi-dimensional feature extraction includes I / Q quadrature components, transient response characteristics, time-domain overshoot coefficient, symbol clock jitter, spectral envelope profile, spectral asymmetry, out-of-band radiation intensity, phase noise characteristics, modulation error rate, carrier frequency offset, multipath delay spread, and phase trajectory correlation. By analyzing and extracting these radio frequency signals, a unique feature of the UAV is formed as an authentication credential, namely radio frequency fingerprint.

2. The method for drone identity authentication based on radio frequency fingerprinting according to claim 1, characterized in that, In step S3, the multi-dimensional feature synchronous calibration process is as follows: Step S31: Configure an independent timestamp marker for each antenna element; Step S32: Align the sampling start points of the 16 channels using a cross-correlation algorithm; Step S33: Establish a three-dimensional coordinate system to map the spatial positions of each antenna; Step S34: Normalize the 12 features respectively; Step S35: Divide the time window into 512ms, generate 1024 sampling points at 0.5ms intervals, and extend the time series to 256 points through cubic spline interpolation; Step S36: Convert the 4*4 antenna array into an 8*32 spatial coding matrix; Step S37: Construct multidimensional tensors and matrices, create a four-dimensional original data volume, and adjust the feature size map to 256*256 using bilinear interpolation; Step S38: Load the pre-built channel impulse response library and apply hyperbolic tangent transform to enhance feature contrast.

3. The method for drone identity authentication based on radio frequency fingerprinting according to claim 1, characterized in that, In step S4, when verifying the logical identifier of the drone, the validity of the first 24 digits of the drone manufacturer code is verified through the IEEE standard OUI database. The server deploys a Bloom filter to match the blacklist and whitelist of drones in the database and synchronizes the MAC address registration database of the Civil Aviation Administration's drone supervision system in real time.

4. The method for drone identity authentication based on radio frequency fingerprinting according to claim 1, characterized in that, In step S5, the matching process for the radio frequency fingerprint is as follows: Step S51: Input the feature size map generated by the RF fingerprint into the ResNet50 recognition model; Step S52: Pass through a convolutional layer with a kernel size of 7*7; Step S53: Use a max pooling layer with a window size of 3*3 to change the image's channels and size; Step S54: Connect each layer of the 4 sets of ResNet residual units to all the previous layers, and then replace the 3*3 convolution of Conv4 with three convolutions of 3*3, 3*1 and 3*3. Step S55: Extract multi-scale radio frequency fingerprint features from different spaces and then perform linear addition and fusion.

5. The method for drone identity authentication based on radio frequency fingerprinting according to claim 4, characterized in that, In step S54, when each layer of the four sets of ResNet residual units is connected to all the preceding layers, the first layer output is... Use the quick connection 'connect1' to change the shape so that it meets the input of the third layer, and then connect it to the output of the second layer. After fusion, the input is processed by the third layer and then output after passing through Conv3. Then output the first layer. Second layer output Use quick connections connect2 and connect3 to change the shape to meet the input of the fourth layer, and then connect it to the output of the third layer. After merging, the input is processed by the fourth layer and then output after passing through Conv4. .

6. The method for drone identity authentication based on radio frequency fingerprinting according to claim 1, characterized in that, In step S6, the specific process for verifying the flight trajectory of the UAV is as follows: Step S61: Deploy the ADS-B receiver to capture the UAV's latitude, longitude, and altitude information; Step S62: Establish a standard flight feature library to verify the smoothness, speed, and spatial decision-making of the UAV flight trajectory; Step S63: Real-time parsing of MAC frame structure to establish a baseline model of communication behavior; Step S64: Construct a hidden Markov model to describe the state transition probabilities.

7. The method for drone identity authentication based on radio frequency fingerprinting according to claim 1, characterized in that, In step S7, the specific process for injecting a specific test signal into the drone to detect its camouflage behavior is as follows: Step S71: Using OFDM subcarrier masking technology, a pseudo-random QPSK modulated signal is inserted into the normal communication frequency band; Step S72: Embed the drone's unique identifier; Step S73: Generate dynamic phase perturbation code, each chip containing 8 bits of device feature information; Step S74: Insert test signals using MAC layer frame gaps; Step S75: Analyze the response characteristics of the UAV and detect camouflage information.

Images