A millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method
By constructing multiple physical layer feature classifiers and performing hierarchical optimization, the problem of inconsistent statistical properties of multiple features in millimeter-wave MIMO systems is solved, improving the reliability and real-time performance of identity authentication and reducing computational complexity.
Patent Information
- Application Number
- CN202611314103.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-17
AI Technical Summary
In existing physical layer authentication methods for millimeter-wave MIMO systems, inconsistent statistical characteristics of multiple features, classifier parameters dependent on empirical settings, and difficulty in coordinating the determination of path voting quantity and detection threshold lead to unstable authentication performance and high computational complexity.
Multiple physical layer feature classifiers are constructed, including path decay, departure angle, arrival angle, array gain error, and array phase error. By hierarchically optimizing the number of path votes, the classifier detection threshold, and the fusion weights, an authentication parameter set is formed for online identity determination.
This improves the reliability of millimeter-wave MIMO systems in detecting identity spoofing attacks, reduces computational complexity, and enhances the real-time nature of authentication and the standardization of parameter configuration.
Smart Images

Figure CN123123899A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of wireless communication security and physical layer identity authentication technology, specifically relating to a millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method. Background Technology
[0002] With the development of millimeter-wave communication, very large-scale antenna arrays (VMAs), and beamforming technologies, millimeter-wave MIMO systems have become an important wireless access method for sixth-generation mobile communication, intelligent transportation, satellite communication, and the industrial internet. These systems typically employ narrow-beam directional transmission and rely on beam training, pilot transmission, and channel estimation to establish communication links. Due to the open propagation characteristics of wireless signals, attackers can impersonate legitimate users by forging legitimate user identifiers, copying some communication parameters, or transmitting signals from nearby locations, thereby illegally accessing the network or obtaining communication resources. Therefore, reliable authentication of the transmitting end without significantly increasing communication overhead has become a crucial issue in the security protection of millimeter-wave MIMO systems.
[0003] Physical layer authentication identifies the transmitter by utilizing wireless channel characteristics, device hardware differences, or a combination of both. Channel-based methods typically use path attenuation, angle of arrival, angle of departure, or channel state information as authentication criteria, reflecting the transmitter's location and propagation environment. However, these features change with movement, obstruction, and scattering, resulting in limited stability. Hardware fingerprinting methods utilize features such as gain and phase errors caused by manufacturing and calibration errors in RF devices and antenna arrays, exhibiting good device correlation. However, the authentication performance of a single hardware feature may still degrade under conditions of low signal-to-noise ratio, large estimation errors, or when attackers possess mimicry capabilities. To improve robustness, existing methods have begun to jointly utilize channel and hardware features, extracting path attenuation, spatial angles, and array errors through tensor decomposition, then constructing classifiers separately and performing weighted fusion.
[0004] However, existing multi-feature authentication schemes still have several shortcomings. First, the statistical distribution, dimensions, stability, and discriminative capabilities of features at different physical layers vary significantly. Using fixed weights, average weights, or empirical thresholds makes it difficult to adapt to different signal-to-noise ratios, number of paths, and attack intensities. Second, channel features typically correspond to multiple propagation paths, requiring the simultaneous determination of path-level detection thresholds and the number of votes for multiple paths. These two factors influence each other, and relying on manual settings can easily lead to excessively high false alarm rates or insufficient detection rates. Third, the statistical characteristics of array gain error and phase error differ from those of time-varying channel features, requiring separate determination of detection thresholds. Existing methods lack a unified and calculable parameter configuration process. Furthermore, there is a correlation between local classifier thresholds, fusion weights, and global authentication thresholds. Directly adjusting these online without hierarchical optimization will increase computational complexity and affect authentication real-time performance. Therefore, it is necessary to establish a millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method. This method determines authentication parameters at each level offline based on calculable false alarm probabilities, detection probabilities, and error probabilities, and directly calls the optimized parameters to complete identity determination online. Summary of the Invention
[0005] To address the issues of inconsistent statistical properties of multiple features, empirically dependent classifier parameter settings, difficulty in coordinating the determination of path voting quantity and detection threshold, and insufficient stability of global fusion results in existing millimeter-wave MIMO physical layer authentication methods, this application provides a millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method. This method utilizes path attenuation, departure angle, arrival angle, array gain error, and array phase error in the millimeter-wave MIMO system to construct multiple physical layer feature classifiers, and performs hierarchical optimization of path voting quantity, classifier detection threshold, fusion weight, and global authentication threshold, thereby improving the reliability of millimeter-wave MIMO system in detecting identity spoofing attacks, the standardization of parameter configuration, and the practical deployment capability.
[0006] To achieve the above objectives, this application employs the following technical solution:
[0007] This application presents a millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method, implemented through a millimeter-wave MIMO-OFDM physical layer identity authentication system consisting of a legitimate sender Alice, an authentication receiver Bob, and an attack sender Eve. The legitimate sender Alice possesses… A transmitting array of antenna elements, with Bob certified as the receiver. The receiver array consists of several antenna elements. The authenticated receiver, Bob, performs authentication without needing to embed additional authentication tags in the communication signal. Five types of physical layer authentication features are defined: path attenuation features... , AoD departure angle features Angle of arrival (AoA) features Transmitter array gain error characteristics Transmitter array phase error characteristics The millimeter-wave MIMO multi-classifier optimization authentication method specifically includes the following steps:
[0008] Step 1: In a trusted registration environment, the legitimate transmitter Alice uses different equivalent transmission beams, and the certified receiver Bob uses a receive combining vector. The received observation signal is obtained on the training subcarrier. The received observation signal is divided into statistical modeling data and parameter optimization data. The frequency domain channel matrix is established, and the statistical parameters of the legitimate transmitter required for subsequent classifier construction and parameter optimization are obtained, including the reference mean, feature variance, time correlation coefficient, feature extraction error variance, and statistical parameters of transmission array gain error and phase error.
[0009] Step 2: Based on the transmission and reception array structures of the millimeter-wave MIMO system, divide the transmission and reception arrays into calibrated antenna elements and uncalibrated antenna elements respectively, and construct the transmission array error matrix. and receiver array error matrix ;
[0010] Step 3: Based on the transmit array error matrix obtained in Step 2 and receiver array error matrix Using the equivalent transmit beam and receive combiner vectors from step 1, the received signal matrix is obtained on multiple training subcarriers, and a third-order received signal tensor is constructed by stacking them along the subcarrier dimensions. ;
[0011] Step 4: Apply the third-order received signal tensor obtained in Step 3. Perform CP decomposition, and jointly obtain the path attenuation estimate based on the received spatial factor matrix, transmitted spatial factor matrix, and frequency response factor matrix obtained from the CP decomposition. Departure angle estimate Angle of arrival estimate Transmit array gain error vector and the phase error vector of the transmitting array Five types of physical layer features were obtained for identity authentication;
[0012] Step 5: Based on the path attenuation estimate obtained in Step 4, complete the path association. Departure angle estimate Angle of arrival estimate Calculate the feature difference between adjacent authentication observation times. The path-level differential energy statistics are constructed from the characteristic differences. Establish the assumption of a legitimate sender. Assuming the sender is impersonating the recipient The statistical distribution below;
[0013] Step 6: Based on the path-level differential energy statistics established in Step 5 and its legitimate sender assumption Assuming the sender is impersonating the recipient The statistical distribution is used to determine the path-level detection threshold and perform path-level binary hypothesis testing. Then, the binary outputs of the path decay classifier, AoD classifier, and AoA classifier are obtained through multi-path voting. At the same time, the number of path votes is jointly optimized based on the classifier error probability. and detection threshold ;
[0014] Step 7: Based on the transmit array gain error vector obtained in Step 4 and the phase error vector of the transmitting array Quadratic detection statistics are constructed respectively, binary hypothesis testing is performed, and the detection threshold is optimized according to the corresponding error probability to obtain the binary output of the gain error classifier and the phase error classifier.
[0015] Step 8: Based on the error probabilities of the path attenuation classifier, AoD classifier, and AoA classifier obtained in Step 6, and the gain error classifier and phase error classifier obtained in Step 7, configure the fusion weights corresponding to each of the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier. Construct a global authentication score using the configured binary outputs of the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier, and the fusion weights, and optimize the global authentication threshold. Combine the optimized path vote count, the detection thresholds of each local classifier, the fusion weights, and the global authentication threshold into an authentication parameter set and store it at the authentication receiver Bob.
[0016] Step 9: In the online identity authentication stage, Bob, the authentication receiver, performs steps 3-7 on the unknown sender to obtain the binary outputs of the five classifiers, and calls the authentication parameter set pre-stored in step 8 to calculate the global authentication score. The identity determination is completed by comparing the global authentication score with the global authentication threshold.
[0017] A further improvement to this application is that step 1 specifically includes the following steps:
[0018] Step 1.1: In a trusted registration environment, collect training received signals from the legitimate sender Alice at multiple consecutive authentication observation times for subsequent statistical modeling and parameter optimization.
[0019] Step 1.2: Divide the received observation signals into statistical modeling data and parameter optimization data;
[0020] Step 1.3: Assume there is an existence between the legitimate sender Alice and the authenticated receiver Bob. The main transmission path, in the first The first certified observation time, the first The frequency domain channel matrix corresponding to each training subcarrier is represented as follows:
[0021]
[0022] in: For the first The propagation path is in the first Complex path attenuation coefficient at each certified observation time; For the first The propagation delay of each propagation path; For the first The angle of arrival of each propagation path; For the first The departure angle of the propagation path; The sampling frequency; The number of points in the inverse discrete Fourier transform of OFDM. The ideal array steering vector for authenticating receiver Bob. The ideal array guide vector for the legitimate transmitter Alice.
[0023] A further improvement in this application is that step 2 specifically includes the following steps:
[0024] Step 2.1: Divide the transmitting array and the receiving array into... Each calibrated antenna element and One uncalibrated antenna element, and satisfying The complex error response vector of the transmitting array is represented as:
[0025]
[0026] in, For the first Gain error coefficients of an uncalibrated antenna element For the first Phase error coefficients of an uncalibrated antenna element;
[0027] Step 2.2: Based on the complex error response vector of the transmitting array obtained in Step 2.1 Construct the transmission array error matrix :
[0028] ;
[0029] Step 2.3, the complex error response vector of the receiving array is expressed as:
[0030]
[0031] in, This indicates the total number of receiver array antenna elements in Bob's certified receiver; This indicates the number of calibrated antenna elements in the receiver array; This indicates the number of uncalibrated antenna elements in the receiver array; Indicates the first Gain error coefficient of an uncalibrated receiving antenna element; Indicates the first The phase error of each uncalibrated receiving antenna element, among which ; Indicates the imaginary unit; superscript Indicates transpose;
[0032] Step 2.4: Based on the complex error response vector of the receiving array obtained in Step 2.3 Construct the receiver array error matrix :
[0033] .
[0034] A further improvement in this application is that step 3 specifically includes the following steps:
[0035] Step 3.1: At each authentication observation moment, the legitimate sender Alice... Within each training time slot, The equivalent transmission beam is defined as the first... The equivalent transmit beam for each training time slot is:
[0036] ;
[0037] in, For the precoding matrix, For pilot symbol vectors;
[0038] Step 3.2: Assemble all equivalent transmit beams into a transmit beam matrix:
[0039] ;
[0040] Step 3.3, Bob, the authentication receiver, uses the following methods in sequence. Receive and merge vectors, and form a receive and merge matrix:
[0041] ;
[0042] Step 3.4, in the The first certified observation time, the first The received signal matrix corresponding to each training subcarrier is represented as follows:
[0043]
[0044] in:
[0045]
[0046]
[0047]
[0048] in, For the first The angle of arrival of each propagation path, For the first The departure angle of the propagation path, Sampling frequency, The number of points in the inverse discrete Fourier transform of OFDM. To transmit the array error matrix, For the receiver array error matrix, For additive noise matrix, For the first The propagation path is in the first Equivalent complex path attenuation on each training subcarrier; For equivalent receiving spatial response; To send an equivalent spatial response;
[0049] Step 3.5, The received signal matrix corresponding to each training subcarrier Stacking along the subcarrier dimension to construct a third-order received signal tensor :
[0050]
[0051] in, Represents the outer product of vectors. Represents the noise tensor; Indicates the first The frequency response vector of each propagation path on all training subcarriers:
[0052] .
[0053] A further improvement in this application is that step 4 specifically includes the following steps:
[0054] Step 4.1: Utilize the sparse multipath characteristics of the millimeter-wave channel to convert the third-order received signal tensor... Represented as The sum of rank-1 component tensors, and the sum of the rank-1 component tensors of the third-order received signal tensor. Perform CP decomposition to obtain the receiver space factor matrix. Sending spatial factor matrix and frequency response factor matrix :
[0055]
[0056]
[0057]
[0058] The three factor matrices are obtained by solving the following optimization problem:
[0059]
[0060] The alternating least squares method is used to fix two factor matrices and update the third factor matrix, and the receiving space factor matrix is updated cyclically. Sending spatial factor matrix and frequency response factor matrix ;
[0061] Define normalized tensor reconstruction error :
[0062] in, For the first The tensor reconstructed by the current three factor matrices in the next ALS iteration, when the normalized tensor reconstruction error is... The iteration stops when the factor matrix changes below a preset threshold, between two consecutive iterations, or when the maximum number of iterations is reached.
[0063] Step 4.2: Based on the spatial factor matrix sent in Step 4.1 No. The column recovers the transmitted array response, including array errors. :
[0064]
[0065] in, For the transmit beam matrix: For conjugate transpose; before extraction Each calibrated transmitting element responds to the transmitting array. The phase is unwrapped to obtain the phase vector. :
[0066]
[0067] Based on the obtained phase vector Calculate column vectors :
[0068] in, ;
[0069] Based on column vectors The estimated departure angles for each propagation path are obtained. :
[0070] ;
[0071] Similarly, based on the receiving space factor matrix No. Column, recover the receiver array response :
[0072]
[0073] extract For each calibrated receiving array element, phase unwrapping and LS fitting are performed to obtain the estimated angle of arrival for each propagation path. :
[0074] ;
[0075] Step 4.3: Based on the frequency response factor matrix obtained in Step 4.1 Take its first List As the first The estimated frequency response factor vector of each propagation path is searched to find the candidate time delay response vector. The candidate propagation delay with the highest normalized correlation was selected as the first... Propagation delay estimate of each propagation path :
[0076]
[0077] in, ;
[0078] Step 4.4: Determine the estimated propagation delay value Then, the cross-subcarrier delay phase term is eliminated from the frequency response factor using the least squares method to obtain the corresponding path attenuation estimate. :
[0079] ;
[0080] Step 4.5: Based on the ideal transmission array steering vector expression given in Step 1.3, estimate the departure angle of each propagation path obtained in Step 4.2. Substitute and construct the theoretical transmission array steering vector. By eliminating the path scale factor and theoretical angle phase term, the complex error response, i.e., the gain error, of the uncalibrated transmitting array element is obtained. and phase error ;
[0081] Step 4.6: Average, median average, or robust average after removing outliers on the complex error responses obtained from each propagation path to obtain the... Transmit array gain error vector at each certified observation time and the phase error vector of the transmitting array Combined with the path attenuation estimates for each propagation path obtained in steps 4.2 to 4.4 Departure angle estimate and estimated angle of arrival In the Five types of physical layer features for identity authentication are obtained from each authentication observation time, including: path decay estimate. Departure angle estimate Angle of arrival estimate and the transmit array gain error vector and the phase error vector of the transmitting array .
[0082] A further improvement in this application is that step 5 specifically includes the following steps:
[0083] Step 5.1: For any path-level time-varying channel feature obtained in Step 4 Calculate the first Feature estimates of propagation paths :
[0084]
[0085] in, Indicates the feature estimation error;
[0086] when Time-varying channel characteristics Real angle features ;
[0087] when Time-varying channel characteristics Complex path attenuation characteristics ;
[0088] Step 5.2: Use a first-order autoregressive model to describe the time-varying channel characteristics after mean centering:
[0089]
[0090] in, For hypothetical indexes; Corresponding to the assumption of a legitimate sender , Assuming the sender impersonates the corresponding identity , Time-varying channel characteristics In the assumption The time correlation coefficient below, Represents Gaussian or complex Gaussian driving noise, based on the time correlation coefficient. Characteristic variance and estimation error variance Determine the difference between adjacent time points under the assumption Variance below:
[0091]
[0092] in, Time-varying channel characteristics In the assumption The characteristic variance of the driving stochastic process; For the CP feature extraction process Introduced estimation error variance The feature difference between adjacent certified observation times under the assumption The variance below;
[0093] Step 5.3, for the first Propagation path and any time-varying channel characteristics Calculate the characteristic difference of the propagation path at adjacent authentication observation times. :
[0094]
[0095] in, Indicates the first The first certified observation time Feature estimates of each propagation path; Indicates the first time corresponding to the previous certified observation time. Feature estimates of each propagation path;
[0096] Step 5.4: Enter the first [item] in the authentication window. The corresponding propagation path The path-level difference energy statistics corresponding to the propagation path are constructed by summing the squared moduli of the feature differences between adjacent time points. :
[0097] ;
[0098] Step 5.5: Under the assumptions of a legitimate sender and an identity spoofing sender, establish statistical distributions respectively:
[0099] ;
[0100] ;
[0101] in, Based on the feature type, it can be represented as a real Gaussian distribution or a circularly symmetric complex Gaussian distribution.
[0102] A further improvement in this application is that step 6 specifically includes the following steps:
[0103] Step 6.1, based on the assumption and The likelihood ratio test is constructed using the feature difference variance, and a path-level detection threshold is defined. :
[0104]
[0105] in, This is the likelihood ratio threshold parameter;
[0106] when At that time, the path-level decision is represented as
[0107]
[0108] when At time: Path-level decision is represented as
[0109]
[0110] Indicates the first The path supports the assumption of a valid sender. , This indicates that the sender is supported by the identity spoofing assumption. ;
[0111] Step 6.2: For complex path attenuation characteristics, the degrees of freedom of the path-level difference energy statistics are... For real-valued features, the degrees of freedom of the path-level difference energy statistics are: ;
[0112] Step 6.3, based on degrees of freedom , Time-varying channel characteristics are calculated using the incomplete gamma-ratio function. single-path false alarm probability and single-path detection probability :
[0113]
[0114]
[0115] in, Represents an incomplete gamma-ratio function;
[0116] Step 6.4, for the same time-varying channel characteristics ,right Sum of path-level decision results for each propagation path:
[0117]
[0118] Step 6.5: Based on the summation result of Step 6.4, configure the number of path votes. and execute -out-of- The path voting process yields the output of the path decay classifier. , Exit angle classifier output and the output of the angle of arrival classifier :
[0119] ;
[0120] The false alarm probability of the channel feature classifier is expressed as:
[0121] ;
[0122] The detection probability of the channel feature classifier is expressed as:
[0123] ;
[0124] Step 6.6: To simultaneously optimize path-level detection sensitivity and multi-path consistency requirements, the number of path votes is adjusted. and detection threshold Joint optimization is performed to minimize the error probability of the corresponding channel feature classifier. This allows the sender to impersonate the identity. The prior probability is The legitimate sender assumes The prior probability is Define the error probability of the channel feature classifier:
[0125]
[0126] For each time-varying channel characteristic Enumerate the number of votes for candidate paths within the following integer range. For each candidate parameter combination Calculate the single-path false alarm probability sequentially. Single-path detection probability False alarm probability of channel feature classifier Detection probability of channel feature classifier Error probability of channel feature classifier Choose the parameter combination that minimizes the error probability as the optimization parameters for this channel classifier:
[0127]
[0128] In satisfying Among the candidate parameter combinations, choose the one that makes Maximum number of path votes and detection threshold;
[0129] Step 6.7: Obtain the initial candidate values for the number of path votes based on the following parsing expression. :
[0130]
[0131]
[0132]
[0133] in, Initial values for parsed candidate paths representing the number of votes. Indicates the number of propagation paths. The path voting parameters are constructed based on the false alarm probability and the missed detection probability of a single path. This indicates the rounding up operation. This represents the false alarm probability corresponding to a single propagation path. This represents the detection probability corresponding to a single propagation path. The probability of missed detection for a single propagation path; with candidate initial values The candidate path's vote count is represented by its adjacent integers; for each candidate path's vote count... Search for the detection threshold within the preset detection threshold range. And calculate the corresponding channel feature classifier error probability. ; Select to make The minimum number of path votes and the detection threshold are respectively used as the number of votes for optimizing the path. and optimize detection threshold .
[0134] A further improvement in this application is that step 7 specifically includes the following steps:
[0135] Step 7.1: Analyze the array error characteristics obtained in Step 4.6. Do not establish a binary hypothesis model; establish the following binary hypothesis models respectively:
[0136]
[0137]
[0138] in ,when hour, ;when hour, :
[0139] Step 7.2: Define scale parameters and integration center :
[0140]
[0141]
[0142] Step 7.3, based on scale parameters and integration center Quadratic detection statistics for constructing array error characteristics:
[0143]
[0144] According to the quadratic statistic With detection threshold The comparison results are used to generate the output of the gain error classifier. and phase error classifier output :
[0145]
[0146] in, For gain error classifier, For phase error classifier;
[0147] Step 7.4, under the assumption Below, the quadratic statistic of array error characteristics It follows a non-central chi-square distribution, and its non-centrality parameter is expressed as:
[0148]
[0149] Using generalized Marcum- The function calculates the false alarm probability of the array error classifier. and detection probability :
[0150]
[0151]
[0152] in, For order is The generalized Marcum-Q function; The number of uncalibrated transmit array elements; and The quadratic statistic is respectively in and The corresponding non-central parameter;
[0153] Define the error probability of the array error classifier:
[0154] ;
[0155] Step 7.5: Within the preset threshold search range, perform one-dimensional numerical optimization on the detection thresholds corresponding to gain error and phase error respectively to determine the gain error classifier. Phase error classifier The optimal detection threshold with the lowest error probability:
[0156]
[0157]
[0158] in, This indicates the characteristics of the transmit array gain error. This indicates the phase error characteristics of the transmitting array; and These represent the error probabilities of the gain error classifier and the phase error classifier, respectively. and These represent the corresponding optimized detection thresholds; in the implementation using target false alarm constraints, when satisfying... Among the candidate thresholds, choose the one that makes The highest threshold.
[0159] A further improvement to this application is that step 8 specifically includes the following steps:
[0160] Step 8.1: Read the path attenuation classifier, AoD classifier, and AoA classifier from Step 6, and the gain error classifier and phase error classifier from Step 7, and calculate the five classifiers respectively. Corresponding false alarm probability Detection probability And calculate the error probability. :
[0161] ;
[0162] Step 8.2: Based on the error probability of each classifier Configure non-negative reliability weights :
[0163]
[0164]
[0165] in, For unnormalized weights, To prevent the default positive number with a denominator of zero
[0166] Therefore, the following is satisfied:
[0167]
[0168] ;
[0169] Step 8.3: Based on the five classifiers binary output and corresponding reliability weights Construct a global authentication score:
[0170]
[0171] Under the modeling condition that the classifier outputs are independent or approximately independent, the global authentication score will be... In the assumption and The following are approximations of Gaussian distributions.
[0172] In the assumption The mean of the global authentication score is:
[0173]
[0174] The variance of the global authentication score is:
[0175]
[0176] In the assumption The mean of the global authentication score is:
[0177]
[0178] The variance of the global authentication score is:
[0179]
[0180] Step 8.4: For a given global authentication threshold Then calculate and optimize the global authentication threshold:
[0181]
[0182] The global false alarm probability is expressed as:
[0183]
[0184] The global detection probability is expressed as:
[0185]
[0186] Based on global false alarm probability and global detection probability Define global error probability :
[0187]
[0188] Global authentication threshold Perform a one-dimensional search to select the option that minimizes the global error probability. Minimum threshold Alternatively, while satisfying the target global false alarm probability constraint Among the candidate thresholds, choose the one that maximizes the global detection probability. Maximum threshold;
[0189] Step 8.5: Combine the legitimate sender channel feature statistical parameters obtained in Step 1, the classifier parameters optimized in Steps 6 and 7, and the global fusion parameters determined in Step 8.4 to form an authentication parameter set. :
[0190]
[0191] in: This refers to the channel feature statistical parameters obtained in step 1 based on the training data from the legitimate sender. This represents the number of votes for the optimized path obtained by the path decay classifier, AoD classifier, and AoA classifier. These represent the optimized detection thresholds obtained by optimizing the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier, respectively. These represent the fusion weights corresponding to the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier, respectively. This indicates an optimization of the global authentication threshold.
[0192] A further improvement in this application is that step 9 specifically includes the following steps:
[0193] Step 9.1: During the online identity authentication phase, collect the received observation signals of multiple equivalent transmit beams, multiple receive merging vectors, and multiple training subcarriers corresponding to the unknown transmitter.
[0194] Step 9.2: Construct a third-order received signal tensor according to steps 3 and 4 and perform CP decomposition to obtain the path attenuation estimate, departure angle estimate, arrival angle estimate, transmit array gain error vector, and transmit array phase error vector of the unknown transmitter.
[0195] Step 9.3: Following the statistical construction and local decision method described in Steps 5-7, call the optimized path voting quantity and local detection threshold pre-stored in the authentication parameter set from Step 8.5 to obtain the binary outputs of the five classifiers. ;
[0196] Step 9.4: Calculate the global authentication score by calling the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier reliability weights stored in Step 8.5.
[0197]
[0198] and optimize global authentication thresholds Compare:
[0199]
[0200] The sender was deemed legitimate and authentication was successful.
[0201]
[0202] If the sender is identified as a potential impersonator, authentication fails and an impersonation alarm is output.
[0203] The beneficial effects of this application are:
[0204] This application combines path attenuation, departure angle, arrival angle, transmit array gain error, and transmit array phase error to combine time-varying channel characteristics reflecting the propagation environment with array error characteristics reflecting equipment differences. This can reduce the performance degradation caused by the influence of a single feature on movement, obstruction, noise, or partial imitation by an attacker, and improve the authentication robustness in complex scenarios.
[0205] This application addresses the characteristic of multiple propagation paths corresponding to channel features by introducing multi-path voting on the basis of path-level detection. It also determines the number of path votes and the detection threshold jointly based on the classifier error probability, which can more reasonably balance detection sensitivity and false alarm suppression capability.
[0206] This application establishes performance models for channel differential energy statistics and array error quadratic statistics, calculates the false alarm probability, detection probability and error probability of each local classifier, and determines the local detection threshold accordingly, so that parameter configuration has a clear statistical basis and reduces performance fluctuations caused by repeated manual experiments and subjective settings.
[0207] This application configures non-negative reliability weights according to the classifier error probability, so that classifiers with strong discriminative ability can obtain greater weight in global decision-making. It also improves the trade-off between false alarm probability and detection probability by optimizing the global authentication threshold, and avoids weak features amplifying decision errors when equal-weighted fusion is used.
[0208] This application separates parameter configuration from online authentication, fixing the path voting quantity, local thresholds, fusion weights, and global thresholds in the offline stage. In the online stage, only feature extraction, local decision-making, and weighted comparison need to be completed, reducing online search overhead and improving authentication real-time performance. Furthermore, this application directly utilizes existing beam training, pilot, and channel estimation data, eliminating the need for additional authentication tag embedding and thus not significantly increasing communication protocol overhead. Attached Figure Description
[0209] Figure 1 This is a schematic diagram of the millimeter-wave MIMO physical layer authentication system model to which this application applies.
[0210] Figure 2 This is a schematic diagram showing the changes in complex path attenuation coefficient, signal arrival angle, and signal departure angle over time for different propagation paths.
[0211] Figure 3 This is a schematic diagram of random radiation pattern distortion caused by antenna array gain error and phase error.
[0212] Figure 4 The chart shows a comparison between the theoretical and simulated ROC curves of path attenuation characteristics, spatial angle characteristics, array gain error characteristics, and multi-feature fusion certification results.
[0213] Figure 5 This is a comparison chart of the authentication performance of the authentication method in this application and existing authentication methods in static and dynamic scenarios.
[0214] Figure 6 This diagram illustrates the impact of the time correlation coefficient of attacker path decay features on single-feature authentication performance and multi-feature fusion authentication performance.
[0215] Figure 7 This diagram illustrates the impact of the number of propagation paths in a millimeter-wave channel on path attenuation characteristic authentication performance and overall authentication performance.
[0216] Figure 8This diagram illustrates the impact of the number of uncalibrated transmit antennas on the array phase error characteristic certification performance and overall certification performance.
[0217] Figure 9 This diagram illustrates the impact of path attenuation feature decision threshold and fusion decision threshold on false alarm probability and missed detection probability.
[0218] Figure 10 This diagram illustrates the optimized selection of the number of path votes in the path decay feature classifier and its relationship with the decision threshold.
[0219] Figure 11 The chart compares the authentication performance of three classifier weight calculation methods: inverse error probability weighting, equal weight weighting, and random weight weighting. Detailed Implementation
[0220] The embodiments of the present invention will be disclosed below with reference to the drawings. For clarity, many practical details will be described in the following description. However, it should be understood that these practical details are not intended to limit the present invention. That is, in some embodiments of the present invention, these practical details are not essential. In addition, for the sake of simplicity, some conventional structures and components will be shown in the drawings in a simple schematic manner.
[0221] like Figure 1 As shown, this application presents a millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method, applicable to scenarios such as intelligent manufacturing, vehicle networking, unmanned systems, satellite communication, and high-reliability low-latency communication. It is implemented through a millimeter-wave MIMO-OFDM physical layer authentication system consisting of a legitimate sender Alice, an authenticated receiver Bob, and an attack sender Eve that may impersonate the legitimate sender. The legitimate sender Alice possesses… A transmitting array of antenna elements, with Bob certified as the receiver. The receiver array consists of several antenna elements. The legitimate transmitter, Alice, sends beam training signals or pilot signals to the authenticated receiver, Bob, via multiple transmit beams. Bob then sequentially receives these signals using multiple receiving directions. Bob performs authentication using existing beam training signals, pilot signals, or channel estimation data from the communication system, without needing to embed additional authentication tags in the communication signals. Five types of physical layer authentication features are defined: path attenuation features... , AoD departure angle features Angle of arrival (AoA) features Transmitter array gain error characteristics Transmitter array phase error characteristics The millimeter-wave MIMO multi-classifier optimization authentication method includes an offline parameter configuration stage and an online identity authentication stage. The offline parameter configuration stage is used to form an authentication parameter set. The authentication parameter set is invoked during the online identity authentication phase. Performing identity verification on an unknown sender involves the following steps:
[0222] Step 1: In a trusted registration environment, the legitimate transmitter Alice uses different equivalent transmit beams, while the authenticated receiver Bob uses a receive combining vector. Received observation signals are obtained on the training subcarriers. These signals are divided into statistical modeling data and parameter optimization data. A frequency domain channel matrix, i.e., a millimeter-wave sparse multipath channel model, is established. The statistical parameters of the legitimate transmitter required for subsequent classifier construction and parameter optimization are obtained, including the reference mean, feature variance, time correlation coefficient, feature extraction error variance, and statistical parameters of transmit array gain error and phase error; specifically, the path attenuation estimate. Departure angle estimate Angle of arrival estimate Transmit array gain error vector and the phase error vector of the transmitting array Reference mean; path decay estimate Departure angle estimate Angle of arrival estimate Transmit array gain error vector and the phase error vector of the transmitting array Characteristic variance and time correlation coefficient in statistical modeling process This refers to the time correlation coefficient of time-varying channel parameters such as path attenuation, angle of arrival, and angle of departure, used to describe the correlation of features over time. Indicates the feature type; The propagation path is indicated; the feature extraction error variance refers to the statistical parameter of the estimation error generated during the corresponding CP decomposition and parameter estimation process, which is used to describe the error between the feature extraction result and the true parameter.
[0223] Step 1 specifically includes the following steps:
[0224] Step 1.1: In a trusted registration environment, collect training received signals from the legitimate sender Alice at multiple consecutive authentication observation times for subsequent statistical modeling and parameter optimization.
[0225] Step 1.2: Divide the received observed signal into statistical modeling data and parameter optimization data; among which, the statistical modeling data is used to determine the Alice hypothesis for the legitimate transmitter. The following parameters are provided:
[0226] Reference mean values for path decay, AoA, and AoD;
[0227] The characteristic variances of path decay, AoA, and AoD specifically refer to the path-level physical layer features, i.e., path decay estimates, obtained from the CP decomposition in step 4. Departure angle estimate Angle of arrival estimate Transmit array gain error vector and the phase error vector of the transmitting array The feature variance represents the degree of statistical fluctuation of the corresponding physical layer feature under the assumption of a legitimate sender.
[0228] Time correlation coefficients of various time-varying channel characteristics ; : Feature type; : Propagation path;
[0229] The variance of the estimation error generated during the feature extraction process ;
[0230] Transmit array gain error and phase error Statistical mean and variance.
[0231] Parameter optimization data is used to evaluate the false alarm probability, detection probability, and error probability corresponding to candidate authentication parameters. This assumes the identity spoofing sender is... This eliminates the need to obtain the actual feature distribution of the attacking endpoint beforehand. During offline parameter optimization, different attack conditions are constructed based on preset alternative hypothesis parameter value ranges to evaluate authentication performance under different authentication parameter configurations. This indicates that the current sender is a legitimate sender. This indicates that the current sender is a potential impersonator.
[0232] Step 1.3: Assume there is an existence between the legitimate sender Alice and the authenticated receiver Bob. The main transmission path, in the first The first certified observation time, the first The frequency domain channel matrix corresponding to each training subcarrier is represented as follows:
[0233]
[0234] in: For the first The propagation path is in the first Complex path attenuation coefficient at each certified observation time; For the first The propagation delay of each propagation path; For the first The angle of arrival of each propagation path; For the first The departure angle of the propagation path; The sampling frequency; The number of points in the inverse discrete Fourier transform of OFDM. The ideal array steering vector for authenticating receiver Bob. For Alice, the ideal array steering vector of the legitimate transmitter, for any two different propagation paths, at least one of the path attenuation, propagation delay, angle of arrival, and departure angle is different, so that different propagation paths can be distinguished by spatial angle, propagation delay, or a combination of both.
[0235] Step 2: Based on the transmission and reception array structures of the millimeter-wave MIMO system, divide the transmission and reception arrays into calibrated antenna elements and uncalibrated antenna elements respectively, and construct the transmission array error matrix. and receiver array error matrix Specifically, the steps include the following:
[0236] Step 2.1: Divide the transmitting array and the receiving array into... Each calibrated antenna element and One uncalibrated antenna element, and satisfying For calibrated antenna elements, the gain and phase errors of the calibrated antenna elements are set to zero; for uncalibrated antenna elements, the radiation pattern fingerprint of the transmitting device is formed using the gain and phase errors of the uncalibrated antenna elements, and the complex error response vector of the transmitting array is expressed as:
[0237]
[0238] in, For the first Gain error coefficients of an uncalibrated antenna element For the first Phase error coefficients of an uncalibrated antenna element;
[0239] Step 2.2: Based on the complex error response vector of the transmitting array obtained in Step 2.1 Construct the transmission array error matrix :
[0240] ;
[0241] Step 2.3, the complex error response vector of the receiving array is expressed as:
[0242]
[0243] in, This indicates the total number of receiver array antenna elements in Bob's certified receiver; This indicates the number of calibrated antenna elements in the receiver array; This indicates the number of uncalibrated antenna elements in the receiver array; Indicates the first Gain error coefficient of an uncalibrated receiving antenna element; Indicates the first The phase error of each uncalibrated receiving antenna element, among which ; Indicates the imaginary unit; superscript The transpose is indicated; the complex error response of the calibrated antenna element is set to 1, indicating that it serves as a reference element for extracting array error parameters.
[0244] Step 2.4: Based on the complex error response vector of the receiving array obtained in Step 2.3 Construct the receiver array error matrix :
[0245] .
[0246] Step 3: Based on the transmit array error matrix obtained in Step 2 and receiver array error matrix Using the equivalent transmit beam and receive combiner vectors from step 1, the received signal matrix is obtained on multiple training subcarriers, and a third-order received signal tensor is constructed by stacking them along the subcarrier dimensions. Specifically, the steps include the following:
[0247] Step 3.1: At each authentication observation moment, the legitimate sender Alice... Within each training time slot, The equivalent transmission beam is defined as the first... The equivalent transmit beam for each training time slot is:
[0248] ;
[0249] in, For the precoding matrix, For pilot symbol vectors;
[0250] Step 3.2: Assemble all equivalent transmit beams into a transmit beam matrix:
[0251] ;
[0252] Step 3.3, Bob, the authentication receiver, uses the following methods in sequence. Receive and merge vectors, and form a receive and merge matrix:
[0253]
[0254] The transmit beams corresponding to different training time slots should have sufficient excitation diversity, and the different receive merging vectors should have sufficient spatial identifiability to support parameter estimation for different propagation paths.
[0255] Step 3.4, in the The first certified observation time, the first The received signal matrix corresponding to each training subcarrier is represented as follows:
[0256]
[0257] in:
[0258]
[0259]
[0260]
[0261] in, For the first The angle of arrival of each propagation path, For the first The departure angle of the propagation path, Sampling frequency, The number of points in the inverse discrete Fourier transform of OFDM. To transmit the array error matrix, For the receiver array error matrix, For additive noise matrix, For the first The propagation path is in the first Equivalent complex path attenuation on each training subcarrier; This is the equivalent received spatial response after the combined effects of receiver combining and receiver array errors; The equivalent transmission spatial response after the combined effects of transmission beam and transmission array errors;
[0262] Step 3.5, The received signal matrix corresponding to each training subcarrier Stacking along the subcarrier dimension to construct a third-order received signal tensor :
[0263]
[0264] in, Represents the outer product of vectors. Represents the noise tensor; Indicates the first The frequency response vector of each propagation path on all training subcarriers:
[0265] .
[0266] Step 4: Apply the third-order received signal tensor obtained in Step 3. Perform CP decomposition, and jointly obtain the path attenuation estimate based on the received spatial factor matrix, transmitted spatial factor matrix, and frequency response factor matrix obtained from the CP decomposition. Departure angle estimate Angle of arrival estimate Transmit array gain error vector and the phase error vector of the transmitting array Five types of physical layer features for identity authentication are obtained; for the propagation paths of adjacent authentication observation times, the path correspondence is further established. Specifically, the steps include:
[0267] Step 4.1: Utilize the sparse multipath characteristics of the millimeter-wave channel to convert the third-order received signal tensor... Represented as The sum of rank-1 component tensors, and the sum of the rank-1 component tensors of the third-order received signal tensor. Perform CP decomposition to obtain the receiver space factor matrix. Sending spatial factor matrix and frequency response factor matrix :
[0268]
[0269]
[0270]
[0271] The three factor matrices are obtained by solving the following optimization problem:
[0272]
[0273] The alternating least squares method is used to fix two factor matrices and update the third factor matrix, and the receiving space factor matrix is updated cyclically. Sending spatial factor matrix and frequency response factor matrix ;
[0274] Define normalized tensor reconstruction error :
[0275] in, For the first The tensor reconstructed by the current three factor matrices in the next ALS iteration, when the normalized tensor reconstruction error is... Below a preset threshold, in this embodiment, it is taken as follows: When the change in the factor matrix between two adjacent iterations is lower than the preset threshold or the maximum number of iterations is reached, the iteration is stopped; the column normalization of the CP decomposition result is performed, and the scale factor and common phase factor removed from the received spatial factor matrix and the transmitted spatial factor matrix, i.e. the scale uncertainty and phase uncertainty inherent in the CP decomposition, are absorbed into the corresponding column of the frequency response factor matrix to maintain the same column path correspondence of the three factor matrices.
[0276] Step 4.2: Based on the received spatial factor matrix in Step 4.1 and the transmission space factor matrix The actual received array response and transmitted array response, including array errors, are recovered. Elements corresponding to calibrated array elements are extracted from the recovered transmitted array response. Phase extraction and phase unwrapping are performed, and least-squares phase fitting is used to obtain the estimated departure angle for each propagation path. Extract the elements corresponding to the calibrated receiver array elements from the recovered receiver array response, and use least-squares phase fitting to obtain the estimated angle of arrival for each propagation path. Specifically;
[0277] According to the sending space factor matrix in step 4.1 No. The column recovers the transmitted array response, including array errors. :
[0278]
[0279] in, For the transmit beam matrix: For conjugate transpose; before extraction Each calibrated transmitting element responds to the transmitting array. The phase is unwrapped to obtain the phase vector. :
[0280]
[0281] Based on the obtained phase vector Calculate column vectors :
[0282] in, ;
[0283] Based on column vectors The estimated departure angles for each propagation path are obtained. :
[0284] ;
[0285] Similarly, based on the receiving space factor matrix No. Column, recover the receiver array response :
[0286]
[0287] extract For each calibrated receiving array element, phase unwrapping and LS fitting are performed to obtain the estimated angle of arrival for each propagation path. :
[0288] ;
[0289] Step 4.3: Based on the frequency response factor matrix obtained in Step 4.1 Take its first List As the first The estimated frequency response factor vector of each propagation path is searched to find the candidate time delay response vector. The candidate propagation delay with the highest normalized correlation was selected as the first... Propagation delay estimate of each propagation path :
[0290]
[0291] in, ;
[0292] Step 4.4: Determine the estimated propagation delay value Then, the cross-subcarrier delay phase term is eliminated from the frequency response factor using the least squares method to obtain the corresponding path attenuation estimate. :
[0293] ;
[0294] Step 4.5: Based on the ideal transmission array steering vector expression given in Step 1.3, estimate the departure angle of each propagation path obtained in Step 4.2. Substitute and construct the theoretical transmission array steering vector. By eliminating the path scale factor and theoretical angle phase term, the complex error response, i.e., the gain error, of the uncalibrated transmitting array element is obtained. and phase error Specifically:
[0295] First, the calibrated array elements are used to estimate the... Common scale coefficients corresponding to each path:
[0296]
[0297] Then based on the common scale coefficient Calculate the first Complex error of each uncalibrated transmitting element :
[0298]
[0299] Thus, the first Gain error of each uncalibrated transmitting element and phase error :
[0300]
[0301] ;
[0302] Step 4.6: Obtain the complex error response, i.e., the gain error, for each propagation path. and phase error By performing an average, median average, or robust average after removing outliers, we obtain the first... Transmit array gain error vector at each certified observation time and the phase error vector of the transmitting array Combined with the path attenuation estimates for each propagation path obtained in steps 4.2 to 4.4 Departure angle estimate and estimated angle of arrival In the Five types of physical layer features for identity authentication are obtained from each authentication observation time, including: path decay estimate. Departure angle estimate Angle of arrival estimate and the transmit array gain error vector and the phase error vector of the transmitting array Among them, the transmit array gain error vector and the phase error vector of the transmitting array The complex error response of the uncalibrated transmitting array element obtained in step 4.5 is further decomposed to obtain the result.
[0303] Step 5: Based on the path attenuation estimate obtained in Step 4, complete the path association. Departure angle estimate Angle of arrival estimate Calculate the feature difference between adjacent authentication observation times. The path-level differential energy statistics are constructed from the characteristic differences. Establish the assumption of a legitimate sender. Assuming the sender is impersonating the recipient The statistical distribution is as follows. Specifically, it includes the following steps:
[0304] Step 5.1: For any path-level time-varying channel feature obtained in Step 4 Calculate the first Feature estimates of propagation paths :
[0305]
[0306] in, Indicates the feature estimation error;
[0307] when Time-varying channel characteristics Real angle features ;
[0308] when Time-varying channel characteristics Complex path attenuation characteristics ;
[0309] Step 5.2: Use a first-order autoregressive model to describe the time-varying channel characteristics after mean centering:
[0310]
[0311] in, For hypothetical indexes; Corresponding to the assumption of a legitimate sender , Assuming the sender impersonates the corresponding identity , Time-varying channel characteristics In the assumption The time correlation coefficient below, Represents Gaussian or complex Gaussian driving noise, based on the time correlation coefficient. Characteristic variance and estimation error variance Determine the difference between adjacent time points under the assumption Variance below:
[0312]
[0313] in, Time-varying channel characteristics In the assumption The characteristic variance of the driving stochastic process; For the CP feature extraction process Introduced estimation error variance The feature difference between adjacent certified observation times under the assumption The variance below;
[0314] Step 5.3, for the first Propagation path and any time-varying channel characteristics Calculate the characteristic difference of the propagation path at adjacent authentication observation times. :
[0315]
[0316] in, Indicates the first The first certified observation time Feature estimates of each propagation path; Indicates the first time corresponding to the previous certified observation time. Feature estimates of each propagation path;
[0317] Step 5.4: Enter the first [item] in the authentication window. The corresponding propagation path The path-level difference energy statistics corresponding to the propagation path are constructed by summing the squared moduli of the feature differences between adjacent time points. :
[0318] ;
[0319] Step 5.5: Under the assumptions of a legitimate sender and an identity spoofing sender, establish statistical distributions respectively:
[0320] ;
[0321] ;
[0322] in, Based on the feature type, it can be represented as a real Gaussian distribution or a circularly symmetric complex Gaussian distribution.
[0323] Step 6: Based on the path-level differential energy statistics established in Step 5 and its legitimate sender assumption Assuming the sender is impersonating the recipient The statistical distribution is used to determine the path-level detection threshold and perform path-level binary hypothesis testing. Then, the binary outputs of the path decay classifier, AoD classifier, and AoA classifier are obtained through multi-path voting. At the same time, the number of path votes is jointly optimized based on the classifier error probability. and detection threshold Specifically, the steps include the following:
[0324] Step 6.1: Perform path-level binary hypothesis testing of channel characteristics: based on the hypothesis... and The likelihood ratio test is constructed using the feature difference variance, and a path-level detection threshold is defined. :
[0325]
[0326] in, The likelihood ratio threshold parameter is determined based on prior probability, error cost, or target false alarm probability.
[0327] when At that time, the path-level decision is represented as
[0328]
[0329] when At time: Path-level decision is represented as
[0330]
[0331] Indicates the first The path supports the assumption of a valid sender. , This indicates that the sender is supported by the identity spoofing assumption. ;
[0332] Step 6.2, Calculate the single-path detection performance of channel characteristics: For complex path attenuation characteristics, the degrees of freedom of the path-level differential energy statistics are... For real-valued features, the degrees of freedom of the path-level difference energy statistics are: ;
[0333] Step 6.3, based on degrees of freedom , Time-varying channel characteristics are calculated using the incomplete gamma-ratio function. single-path false alarm probability and single-path detection probability :
[0334]
[0335]
[0336] in, Represents an incomplete gamma-ratio function;
[0337] Step 6.4: Use path voting to form three channel feature classifiers: for the same time-varying channel feature ,right Sum of path-level decision results for each propagation path:
[0338]
[0339] Step 6.5: Based on the summation result of Step 6.4, configure the number of path votes. and execute -out-of- The path voting process yields the output of the path decay classifier. , Exit angle classifier output and the output of the angle of arrival classifier :
[0340] ;
[0341] Under the condition that the decision results of different propagation paths are independent or approximately independent, and that each path has the same or approximately the same statistical characteristics, the false alarm probability of the channel feature classifier is expressed as:
[0342] ;
[0343] The detection probability of the channel feature classifier is expressed as:
[0344] ;
[0345] Step 6.6: Jointly optimize the path voting number and detection threshold of the channel classifier: To simultaneously optimize path-level detection sensitivity and multi-path consistency requirements, the path voting number is adjusted. and detection threshold Joint optimization is performed to minimize the error probability of the corresponding channel feature classifier. This allows the sender to impersonate the identity. The prior probability is The legitimate sender assumes The prior probability is Define the error probability of the channel feature classifier:
[0346]
[0347] For each time-varying channel characteristic Enumerate the number of votes for candidate paths within the following integer range. For each candidate path, the number of votes Within the preset threshold range, the detection threshold is set. Perform grid search, golden section search, binary search, or other one-dimensional numerical search to minimize the error probability of the channel feature classifier;
[0348] For each candidate parameter combination Calculate the single-path false alarm probability sequentially. Single-path detection probability False alarm probability of channel feature classifier Detection probability of channel feature classifier Error probability of channel feature classifier Choose the parameter combination that minimizes the error probability as the optimization parameters for this channel classifier:
[0349]
[0350] In the implementation method constrained by the target false alarm probability, when the following conditions are met... Among the candidate parameter combinations, choose the one that makes Maximum number of path votes and detection threshold;
[0351] Step 6.7: Under the conditions of equal prior knowledge, fixed number of paths, and homogeneous path statistics, obtain the candidate initial values of path vote counts according to the following analytical expression. :
[0352]
[0353]
[0354]
[0355] in, Initial values for parsed candidate paths representing the number of votes. Indicates the number of propagation paths. The path voting parameters are constructed based on the false alarm probability and the missed detection probability of a single path. This indicates the rounding up operation. This represents the false alarm probability corresponding to a single propagation path. This represents the detection probability corresponding to a single propagation path. The probability of missed detection for a single propagation path; with candidate initial values The candidate path's vote count is represented by its adjacent integers; for each candidate path's vote count... Search for the detection threshold within the preset detection threshold range. And calculate the corresponding channel feature classifier error probability. ; Select to make The minimum number of path votes and the detection threshold are respectively used as the number of votes for optimizing the path. and optimize detection threshold .
[0356] Step 7: Based on the transmit array gain error vector obtained in Step 4 and the phase error vector of the transmitting array The quadratic detection statistics are constructed, binary hypothesis testing is performed, and the detection threshold is optimized based on the corresponding error probabilities to obtain the binary outputs of the gain error classifier and the phase error classifier. Specifically, the steps are as follows:
[0357] Step 7.1: Construct array gain error and phase error classifiers: For the array error features obtained in Step 4.6... Do not establish a binary hypothesis model; establish the following binary hypothesis models respectively:
[0358]
[0359]
[0360] in ,when hour, ;when hour, :
[0361] Step 7.2: Define scale parameters and integration center :
[0362]
[0363]
[0364] Step 7.3, based on scale parameters and integration center Quadratic detection statistics for constructing array error characteristics:
[0365]
[0366] According to the quadratic statistic With detection threshold The comparison results are used to generate the output of the gain error classifier. and phase error classifier output :
[0367]
[0368] in, For gain error classifier, For phase error classifiers; the detection threshold of channel feature classifiers is uniformly represented as ;
[0369] Step 7.4, Optimize the detection threshold of the array error classifier: Under the assumption Below, the quadratic statistic of array error characteristics It follows a non-central chi-square distribution, and its non-centrality parameter is expressed as:
[0370]
[0371] Using generalized Marcum- The function calculates the false alarm probability of the array error classifier. and detection probability :
[0372]
[0373]
[0374] in, For order is The generalized Marcum-Q function; The number of uncalibrated transmit array elements; and The quadratic statistic is respectively in and The corresponding non-central parameter;
[0375] Define the error probability of the array error classifier:
[0376] ;
[0377] Step 7.5: Within the preset threshold search range, perform one-dimensional numerical optimization on the detection thresholds corresponding to gain error and phase error respectively to determine the gain error classifier. Phase error classifier The optimal detection threshold with the lowest error probability:
[0378]
[0379]
[0380] in, This indicates the characteristics of the transmit array gain error. This indicates the phase error characteristics of the transmitting array; and These represent the error probabilities of the gain error classifier and the phase error classifier, respectively. and These represent the corresponding optimized detection thresholds; in the implementation using target false alarm constraints, when satisfying... Among the candidate thresholds, choose the one that makes The highest threshold.
[0381] Step 8: Based on the error probabilities of the path attenuation classifier, AoD classifier, and AoA classifier obtained in Step 6, and the gain error classifier and phase error classifier obtained in Step 7, configure the fusion weights corresponding to each of the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier. Construct a global authentication score using the configured binary outputs of the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier, and the fusion weights, and optimize the global authentication threshold. Combine the optimized path vote count, the detection thresholds of each local classifier, the fusion weights, and the global authentication threshold into an authentication parameter set and store it at the authentication receiver Bob. Specifically, this includes the following steps:
[0382] Step 8.1: Read the path attenuation classifier, AoD classifier, and AoA classifier from Step 6, and the gain error classifier and phase error classifier from Step 7, and calculate the five classifiers respectively. Corresponding false alarm probability Detection probability And calculate the error probability. :
[0383] ;
[0384] Step 8.2: Based on the error probability of each classifier Configure non-negative reliability weights :
[0385]
[0386]
[0387] in, For unnormalized weights, Normalization yields , To prevent the default positive number with a denominator of zero
[0388] Therefore, the following is satisfied:
[0389]
[0390]
[0391] After the above normalization, the reliability weights of each classifier are all non-negative and the sum of the weights is 1. This makes the global authentication score a normalized weighted combination of the outputs of the five local classifiers, thus avoiding drift in the range of the global authentication score caused by changes in the weight scale. Classifiers with lower error probabilities and higher authentication reliability receive larger fusion weights; classifiers with higher error probabilities and lower authentication reliability receive smaller fusion weights. The weights are heuristic reliability weights determined based on the classifier error probabilities.
[0392] Step 8.3: Determine the global authentication threshold and form the authentication parameter set: based on the five classifiers. binary output and corresponding reliability weights Construct a global authentication score:
[0393]
[0394] Under the modeling condition that the classifier outputs are independent or approximately independent, the global authentication score will be... In the assumption and The following are approximations of Gaussian distributions.
[0395] In the assumption The mean of the global authentication score is:
[0396]
[0397] The variance of the global authentication score is:
[0398]
[0399] In the assumption The mean of the global authentication score is:
[0400]
[0401] The variance of the global authentication score is:
[0402]
[0403] Step 8.4: For a given global authentication threshold , For use in weighted fusion of global authentication scores The global authentication threshold for identity determination differs from the local channel feature detection threshold in step 6. and the array error local detection threshold in step 7 .
[0404] Then calculate and optimize the global authentication threshold:
[0405]
[0406] The global false alarm probability is expressed as:
[0407]
[0408] The global detection probability is expressed as:
[0409]
[0410] Based on global false alarm probability and global detection probability Define global error probability :
[0411]
[0412] Global authentication threshold Perform a one-dimensional search to select the option that minimizes the global error probability. Minimum threshold Alternatively, while satisfying the target global false alarm probability constraint Among the candidate thresholds, choose the one that maximizes the global detection probability. Maximum threshold;
[0413] Step 8.5: Combine the legitimate sender channel feature statistical parameters obtained in Step 1, the classifier parameters optimized in Steps 6 and 7, and the global fusion parameters determined in Step 8.4 to form an authentication parameter set. :
[0414]
[0415] in: This refers to the channel feature statistical parameters obtained in step 1 based on the training data from the legitimate transmitter, including the statistical mean, variance, and time correlation coefficient of path attenuation, angle of arrival, and angle of departure features; This represents the number of votes for the optimized path obtained by the path decay classifier, AoD classifier, and AoA classifier. These represent the optimized detection thresholds obtained by optimizing the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier, respectively. These represent the fusion weights corresponding to the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier, respectively. This indicates an optimization of the global authentication threshold.
[0416] Step 9: In the online identity authentication phase, the authentication receiver Bob executes steps 3-7 on the unknown sender to obtain the binary outputs of the five classifiers, and calls the pre-stored authentication parameter set from step 8 to calculate the global authentication score. Identity determination is then completed by comparing the global authentication score with the global authentication threshold. This improves the reliability, parameter configuration standardization, and practical deployment capability of the millimeter-wave MIMO system against identity spoofing attacks without embedding additional authentication tags or excessively increasing online computation. Specifically, the steps are as follows:
[0417] Step 9.1: During the online identity authentication phase, collect the received observation signals of multiple equivalent transmit beams, multiple receive merging vectors, and multiple training subcarriers corresponding to the unknown transmitter.
[0418] Step 9.2: Construct a third-order received signal tensor according to steps 3 and 4 and perform CP decomposition to obtain the path attenuation estimate, departure angle estimate, arrival angle estimate, transmit array gain error vector, and transmit array phase error vector of the unknown transmitter.
[0419] Step 9.3: Following the statistical construction and local decision method described in Steps 5-7, call the optimized path voting quantity and local detection threshold pre-stored in the authentication parameter set from Step 8.5 to obtain the binary outputs of the five classifiers. ;
[0420] Step 9.4: Calculate the global authentication score by calling the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier reliability weights stored in Step 8.5.
[0421]
[0422] and optimize global authentication thresholds Compare:
[0423]
[0424] The sender was deemed legitimate and authentication was successful.
[0425]
[0426] If the sender is identified as a potential impersonator, authentication fails and an impersonation alarm is output.
[0427] The authentication receiver can also output the global authentication score, the local classifier that triggered the exception, and the corresponding exception path or array error characteristics for security auditing, secondary authentication, or subsequent access control.
[0428] Figure 2The results show the changes in complex path attenuation, angle of arrival (AoA), and angle of departure (AoD) with the certified observation time for different propagation paths. The complex path attenuation includes both real and imaginary parts, therefore six curves are displayed for the three main propagation paths. AoA and AoD are real-valued angular features, so curves corresponding to the three paths are shown separately. Figure 2 It is evident that path attenuation, AoA, and AoD of the same path exhibit significant temporal correlation between adjacent observation times. Therefore, this application employs a first-order autoregressive model to describe the temporal evolution of the channel characteristics and further utilizes the difference between adjacent time points to construct authentication statistics.
[0429] Figure 3 The diagram illustrates the radiation pattern distortion caused by random gain and phase errors in an antenna array. Different random array errors can alter the main lobe gain, side lobe levels, and pattern shape. Since these gain and phase errors primarily originate from array manufacturing errors, component variations, and calibration errors, and are device-dependent, they can be used as hardware fingerprint features for transmitter authentication.
[0430] Figure 4 This is used to verify the consistency between the theoretical analysis results and simulation results of each classifier and fusion authentication model in this application. Among them, Figure 4 In the figure, (a) shows the theoretical ROC curve and the simulated ROC curve of the path decay feature classifier. Figure 4 (b) shows the theoretical and simulated ROC curves of the spatial angle feature classifier. Figure 4 (c) shows the theoretical and simulated ROC curves of the array gain error classifier. Figure 4 In the figure, (d) represents the global authentication ROC curve after weighted fusion of five feature classes. Figure 4 As can be seen from (a) to 4(c) in the figure, the theoretical analysis results of each single feature classifier are basically consistent with the simulation results; Figure 4 The theoretical results in (d) show a slight deviation from the simulation results, mainly due to approximating the weighted fusion score as a Gaussian distribution and the influence of residual correlations between classifiers under finite sample conditions. Overall, the established false alarm probability and detection probability models can effectively characterize authentication performance.
[0431] Figure 5 This is used to compare the authentication performance of the method in this application with existing authentication methods based on channel gain and phase noise in static and dynamic scenarios. Figure 5 Figure (a) shows the ROC curves for different methods. Figure 5 (b) in the table provides the corresponding AUC metric. From... Figure 5As can be seen, the proposed method exhibits high detection performance in static scenarios. However, in dynamic scenarios involving transmitter movement and rapid channel changes, the detection performance of existing methods deteriorates more significantly, while the proposed method maintains a high detection probability and AUC. This result demonstrates that jointly utilizing the spatiotemporal channel characteristics that vary with the propagation environment and the array error characteristics that possess device stability is beneficial for improving authentication robustness in dynamic scenarios.
[0432] Figure 6 The impact of the time correlation coefficient of the attack sender's path attenuation characteristics on authentication performance is shown. Among them, Figure 6 In the diagram (a), the effect of the time correlation coefficient on the ROC performance of the path decay classifier is shown. Figure 6 (b) shows its impact on the performance of multi-feature fusion authentication. As the time correlation coefficient increases, the time variation law of path attenuation at the attacking end gradually approaches that of the legitimate end, and the distinguishability of the path attenuation feature decreases, thus reducing the detection performance of the single-feature classifier. In contrast, since global authentication utilizes other channel features and array error features simultaneously, its performance decline is smaller, indicating that multi-feature fusion can mitigate the impact of reduced distinguishability of a single feature.
[0433] Figure 7 The number of distinguishable propagation paths is shown. The impact on authentication performance. Among them, Figure 7 (a) in the text shows different ROC curve of the path decay classifier Figure 7 Figure (b) shows the corresponding global authentication ROC curve. As the number of distinguishable propagation paths increases, the authentication receiver can obtain more complementary path-level observation information, and the spatial and temporal information utilized by multi-path voting becomes richer. Therefore, the detection performance of both the path decay classifier and the global fusion classifier is improved.
[0434] Figure 8 The number of uncalibrated transmit elements is shown. The impact on array phase error certification performance. Among them, Figure 8 (a) in the text represents different ROC curve of the lower phase error classifier Figure 8 (b) in the figure represents the corresponding global fusion ROC curve. With... The array error dimension, which can be used to characterize differences in the transmitting device hardware, is increased, and the hardware fingerprint has a higher feature resolution, thus improving the performance of the phase error classifier and global authentication.
[0435] Figure 9 The local detection threshold for path attenuation features is shown. and global authentication threshold The impact on the system's false alarm probability and false negative probability. Among them, Figure 9 (a) shows the two threshold pairs for false alarm probabilities. The impact, Figure 9 (b) shows the two thresholds for the probability of missed detection. The impact of this is that as the detection threshold increases, the statistical evidence required for the system to determine an identity spoofing sender becomes stronger, thus reducing the probability of false alarms, but simultaneously increasing the probability of missed detections. This result indicates that there is a trade-off between detection sensitivity and false alarm suppression capability between the local detection threshold and the global authentication threshold, requiring optimization through error probability or target false alarm constraints.
[0436] Figure 10 Used to verify the number of path votes The optimized selection results. Among them, Figure 10 (a) in the figure gives the error probability of the path decay classifier under different path vote counts. With detection threshold The relationship between the changes shows that different paths correspond to different optimal detection thresholds and minimum error probabilities. Figure 10 Figure (b) further illustrates the optimal selection of the number of path votes under different signal-to-noise ratios and detection thresholds. This result demonstrates the optimal selection of the number of path votes. With local detection threshold There is a coupling relationship between them, so it is not advisable to rely on experience to set them separately, but rather to perform joint optimization based on the error probability of the classifier.
[0437] Figure 11 This is used to compare the authentication performance of different classifier weight configurations. Figure 11 Figure (a) shows the ROC curves for three methods: inverse error probability reliability weighting, equal weighting, and random weighting. Figure 11 (b) in the table gives different global authentication thresholds. The total error probability is given by... Figure 11 It is evident that configuring reliability weights based on classifier error probability enables classifiers with lower error probability and stronger discriminative ability to contribute more to global decision-making, obtain higher detection probability in areas with lower false alarms, and obtain a lower total error probability near the optimal global authentication threshold, thereby verifying the effectiveness of the reliability weighting method in this application.
[0438] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.
Claims
1. A millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method, implemented through a millimeter-wave MIMO-OFDM physical layer identity authentication system consisting of a legitimate sender Alice, an authentication receiver Bob, and an attack sender Eve. The legitimate sender Alice possesses… A transmitting array of antenna elements, with Bob certified as the receiver. The receiver array of one antenna element, with Bob performing authentication at the certified receiver, does not require additional embedding of authentication tags in the communication signal. Its characteristic is: The millimeter-wave MIMO multi-classifier optimization and authentication method specifically includes the following steps: Step 1: In a trusted registration environment, the legitimate transmitter Alice uses different equivalent transmission beams, and the certified receiver Bob uses a receive combining vector. The received observation signal is obtained on the training subcarrier. The received observation signal is divided into statistical modeling data and parameter optimization data. The frequency domain channel matrix is established, and the statistical parameters of the legitimate transmitter required for subsequent classifier construction and parameter optimization are obtained, including the reference mean, feature variance, time correlation coefficient, feature extraction error variance, and statistical parameters of transmission array gain error and phase error. Step 2: Based on the transmission and reception array structures of the millimeter-wave MIMO system, divide the transmission and reception arrays into calibrated antenna elements and uncalibrated antenna elements respectively, and construct the transmission array error matrix. and receiver array error matrix ; Step 3: Based on the transmit array error matrix obtained in Step 2 and receiver array error matrix Using the equivalent transmit beam and receive combiner vectors from step 1, the received signal matrix is obtained on multiple training subcarriers, and a third-order received signal tensor is constructed by stacking them along the subcarrier dimensions. ; Step 4: Apply the third-order received signal tensor obtained in Step 3. Perform CP decomposition, and jointly obtain the path attenuation estimate based on the received spatial factor matrix, transmitted spatial factor matrix, and frequency response factor matrix obtained from the CP decomposition. Departure angle estimate Angle of arrival estimate Transmit array gain error vector and the phase error vector of the transmitting array Five types of physical layer features were obtained for identity authentication; Step 5: Based on the path attenuation estimate obtained in Step 4, complete the path association. Departure angle estimate Angle of arrival estimate Calculate the feature difference between adjacent authentication observation times. The path-level differential energy statistics are constructed from the characteristic differences. Establish the assumption of a legitimate sender. Assuming the sender is impersonating the recipient The statistical distribution below; Step 6: Based on the path-level differential energy statistics established in Step 5 and its legitimate sender assumption Assuming the sender is impersonating the recipient The statistical distribution is used to determine the path-level detection threshold and perform path-level binary hypothesis testing. Then, the binary outputs of the path decay classifier, AoD classifier, and AoA classifier are obtained through multi-path voting. At the same time, the number of path votes is jointly optimized based on the classifier error probability. and detection threshold ; Step 7: Based on the transmit array gain error vector obtained in Step 4 and the phase error vector of the transmitting array Quadratic detection statistics are constructed respectively, binary hypothesis testing is performed, and the detection threshold is optimized according to the corresponding error probability to obtain the binary output of the gain error classifier and the phase error classifier. Step 8: Based on the error probabilities of the path attenuation classifier, AoD classifier, AoA classifier obtained in Step 6 and the gain error classifier and phase error classifier obtained in Step 7, configure the fusion weights corresponding to the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier respectively. Construct a global authentication score using the configured binary outputs of the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier and the fusion weights, and optimize the global authentication threshold. The optimized path voting number, the detection threshold of each local classifier, the fusion weight, and the global authentication threshold are combined into an authentication parameter set and stored at the authentication receiver Bob. Step 9: In the online identity authentication stage, Bob, the authentication receiver, performs steps 3-7 on the unknown sender to obtain the binary outputs of the five classifiers, and calls the authentication parameter set pre-stored in step 8 to calculate the global authentication score. The identity determination is completed by comparing the global authentication score with the global authentication threshold.
2. The millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method according to claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1.1: In a trusted registration environment, collect training received signals from the legitimate sender Alice at multiple consecutive authentication observation times for subsequent statistical modeling and parameter optimization. Step 1.2: Divide the received observation signals into statistical modeling data and parameter optimization data; Step 1.3: Assume there is an existence between the legitimate sender Alice and the authenticated receiver Bob. The main transmission path, in the first The first certified observation time, the first The frequency domain channel matrix corresponding to each training subcarrier is represented as follows: in: For the first The propagation path is in the first Complex path attenuation coefficient at each certified observation time; For the first The propagation delay of each propagation path; For the first The angle of arrival of each propagation path; For the first The departure angle of the propagation path; The sampling frequency; The number of points in the inverse discrete Fourier transform of OFDM. The ideal array steering vector for authenticating receiver Bob. The ideal array guide vector for the legitimate transmitter Alice.
3. The millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 2.1: Divide the transmitting array and the receiving array into... Each calibrated antenna element and One uncalibrated antenna element, and satisfying The complex error response vector of the transmitting array is represented as: in, For the first Gain error coefficients of an uncalibrated antenna element For the first Phase error coefficients of an uncalibrated antenna element; Step 2.2: Based on the complex error response vector of the transmitting array obtained in Step 2.1 Construct the transmission array error matrix : ; Step 2.3, the complex error response vector of the receiving array is expressed as: in, This indicates the total number of receiver array antenna elements in Bob's certified receiver; This indicates the number of calibrated antenna elements in the receiver array; This indicates the number of uncalibrated antenna elements in the receiver array; Indicates the first Gain error coefficient of an uncalibrated receiving antenna element; Indicates the first The phase error of each uncalibrated receiving antenna element, among which ; Indicates the imaginary unit; superscript Indicates transpose; Step 2.4: Based on the complex error response vector of the receiving array obtained in Step 2.3 Construct the receiver array error matrix : 。 4. The millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method according to claim 3, characterized in that: Step 3 specifically includes the following steps: Step 3.1: At each authentication observation moment, the legitimate sender Alice... Within each training time slot, The equivalent transmission beam is defined as the first... The equivalent transmit beam for each training time slot is: ; in, For the precoding matrix, For pilot symbol vectors; Step 3.2: Assemble all equivalent transmit beams into a transmit beam matrix: ; Step 3.3, Bob, the authentication receiver, uses the following methods in sequence. Receive and merge vectors, and form a receive and merge matrix: ; Step 3.4, in the The first certified observation time, the first The received signal matrix corresponding to each training subcarrier is represented as follows: in: in, For the first The angle of arrival of each propagation path, For the first The departure angle of the propagation path, Sampling frequency, The number of points in the inverse discrete Fourier transform of OFDM. To transmit the array error matrix, For the receiver array error matrix, For additive noise matrix, For the first The propagation path is in the first Equivalent complex path attenuation on each training subcarrier; For equivalent receiving spatial response; To send an equivalent spatial response; Step 3.5, The received signal matrix corresponding to each training subcarrier Stacking along the subcarrier dimension to construct a third-order received signal tensor : in, Represents the outer product of vectors. Represents the noise tensor; Indicates the first The frequency response vector of each propagation path on all training subcarriers: 。 5. The millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method according to claim 4, characterized in that: Step 4 specifically includes the following steps: Step 4.1: Utilize the sparse multipath characteristics of the millimeter-wave channel to convert the third-order received signal tensor... Represented as The sum of rank-1 component tensors, and the sum of the rank-1 component tensors of the third-order received signal tensor. Perform CP decomposition to obtain the receiver space factor matrix. Sending spatial factor matrix and frequency response factor matrix : The three factor matrices are obtained by solving the following optimization problem: The alternating least squares method is used to fix two factor matrices and update the third factor matrix, and the receiving space factor matrix is updated cyclically. Sending spatial factor matrix and frequency response factor matrix ; Define normalized tensor reconstruction error : in, For the first The tensor reconstructed by the current three factor matrices in the next ALS iteration, when the normalized tensor reconstruction error is... The iteration stops when the factor matrix changes below a preset threshold, between two consecutive iterations, or when the maximum number of iterations is reached. Step 4.2: Based on the spatial factor matrix sent in Step 4.1 No. The column recovers the transmitted array response, including array errors. : in, For the transmit beam matrix: For conjugate transpose; before extraction Each calibrated transmitting element responds to the transmitting array. The phase is unwrapped to obtain the phase vector. : Based on the obtained phase vector Calculate column vectors : in, ; Based on column vectors The estimated departure angles for each propagation path are obtained. : ; Similarly, based on the receiving space factor matrix No. Column, recover the receiver array response : extract For each calibrated receiving array element, phase unwrapping and LS fitting are performed to obtain the estimated angle of arrival for each propagation path. : ; Step 4.3: Based on the frequency response factor matrix obtained in Step 4.1 Take its first List As the first The estimated frequency response factor vector of each propagation path is searched to find the candidate time delay response vector. The candidate propagation delay with the highest normalized correlation was selected as the first... Propagation delay estimate of each propagation path : in, ; Step 4.4: Determine the estimated propagation delay value Then, the cross-subcarrier delay phase term is eliminated from the frequency response factor using the least squares method to obtain the corresponding path attenuation estimate. : ; Step 4.5: Based on the ideal transmission array steering vector expression given in Step 1.3, estimate the departure angle of each propagation path obtained in Step 4.
2. Substitute and construct the theoretical transmission array steering vector. By eliminating the path scale factor and theoretical angle phase term, the complex error response, i.e., the gain error, of the uncalibrated transmitting array element is obtained. and phase error ; Step 4.6: Average, median average, or robust average after removing outliers on the complex error responses obtained from each propagation path to obtain the... Transmit array gain error vector at each certified observation time and the phase error vector of the transmitting array Combined with the path attenuation estimates for each propagation path obtained in steps 4.2 to 4.4 Departure angle estimate and estimated angle of arrival In the Five types of physical layer features for identity authentication are obtained from each authentication observation time, including: path decay estimate. Departure angle estimate Angle of arrival estimate and the transmit array gain error vector and the phase error vector of the transmitting array .
6. The millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method according to claim 5, characterized in that: Step 5 specifically includes the following steps: Step 5.1: For any path-level time-varying channel feature obtained in Step 4 Calculate the first Feature estimates of propagation paths : in, Indicates the feature estimation error; when Time-varying channel characteristics Real angle features ; when Time-varying channel characteristics Complex path attenuation characteristics ; Step 5.2: Use a first-order autoregressive model to describe the time-varying channel characteristics after mean centering: in, For hypothetical indexes; Corresponding to the assumption of a legitimate sender , Assuming the sender impersonates the corresponding identity , Time-varying channel characteristics In the assumption The time correlation coefficient below, Represents Gaussian or complex Gaussian driving noise, based on the time correlation coefficient. Characteristic variance and estimation error variance Determine the difference between adjacent time points under the assumption Variance below: in, Time-varying channel characteristics In the assumption The characteristic variance of the driving stochastic process; For the CP feature extraction process Introduced estimation error variance The feature difference between adjacent certified observation times under the assumption The variance below; Step 5.3, for the first Propagation path and any time-varying channel characteristics Calculate the characteristic difference of the propagation path at adjacent authentication observation times. : in, Indicates the first The first certified observation time Feature estimates of each propagation path; Indicates the first time corresponding to the previous certified observation time. Feature estimates of each propagation path; Step 5.4: Enter the first [item] in the authentication window. The corresponding propagation path The path-level difference energy statistics corresponding to the propagation path are constructed by summing the squared moduli of the feature differences between adjacent time points. : ; Step 5.5: Under the assumptions of a legitimate sender and an identity spoofing sender, establish statistical distributions respectively: ; ; in, Based on the feature type, it can be represented as a real Gaussian distribution or a circularly symmetric complex Gaussian distribution.
7. The millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method according to claim 6, characterized in that: Step 6 specifically includes the following steps: Step 6.1, based on the assumption and The likelihood ratio test is constructed using the feature difference variance, and a path-level detection threshold is defined. : in, This is the likelihood ratio threshold parameter; when At that time, the path-level decision is represented as when At time: Path-level decision is represented as Indicates the first The path supports the assumption of a valid sender. , This indicates that the sender is supported by the identity spoofing assumption. ; Step 6.2: For complex path attenuation characteristics, the degrees of freedom of the path-level difference energy statistics are... For real-valued features, the degrees of freedom of the path-level difference energy statistics are: ; Step 6.3, based on degrees of freedom , Time-varying channel characteristics are calculated using the incomplete gamma-ratio function. single-path false alarm probability and single-path detection probability : in, Represents an incomplete gamma-ratio function; Step 6.4, for the same time-varying channel characteristics ,right Sum of path-level decision results for each propagation path: Step 6.5: Based on the summation result of Step 6.4, configure the number of path votes. and execute -out-of- The path voting process yields the output of the path decay classifier. , Exit angle classifier output and the output of the angle of arrival classifier : ; The false alarm probability of the channel feature classifier is expressed as: ; The detection probability of the channel feature classifier is expressed as: ; Step 6.6: To simultaneously optimize path-level detection sensitivity and multi-path consistency requirements, the number of path votes is adjusted. and detection threshold Joint optimization is performed to minimize the error probability of the corresponding channel feature classifier. This allows the sender to impersonate the identity. The prior probability is The legitimate sender assumes The prior probability is Define the error probability of the channel feature classifier: For each time-varying channel characteristic Enumerate the number of votes for candidate paths within the following integer range. For each candidate parameter combination Calculate the single-path false alarm probability sequentially. Single-path detection probability False alarm probability of channel feature classifier Detection probability of channel feature classifier Error probability of channel feature classifier Choose the parameter combination that minimizes the error probability as the optimization parameters for this channel classifier: In satisfying Among the candidate parameter combinations, choose the one that makes Maximum number of path votes and detection threshold; Step 6.7: Obtain the initial candidate values for the number of path votes based on the following parsing expression. : in, Initial values for parsed candidate paths representing the number of votes. Indicates the number of propagation paths. The path voting parameters are constructed based on the false alarm probability and the missed detection probability of a single path. This indicates the rounding up operation. This represents the false alarm probability corresponding to a single propagation path. This represents the detection probability corresponding to a single propagation path. The probability of missed detection for a single propagation path; with candidate initial values The candidate path's vote count is represented by its adjacent integers; for each candidate path's vote count... Search for the detection threshold within the preset detection threshold range. And calculate the corresponding channel feature classifier error probability. ; Select to make The minimum number of path votes and the detection threshold are respectively used as the number of votes for optimizing the path. and optimize detection threshold .
8. The millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method according to claim 7, characterized in that: Step 7 specifically includes the following steps: Step 7.1: Analyze the array error characteristics obtained in Step 4.
6. Do not establish a binary hypothesis model; Establish the following binary hypothesis models respectively: in ,when hour, ;when hour, : Step 7.2: Define scale parameters and integration center : Step 7.3, based on scale parameters and integration center Quadratic detection statistics for constructing array error characteristics: According to the quadratic statistic With detection threshold The comparison results are used to generate the output of the gain error classifier. and phase error classifier output : in, For gain error classifier, For phase error classifier; Step 7.4, under the assumption Below, the quadratic statistic of array error characteristics It follows a non-central chi-square distribution, and its non-centrality parameter is expressed as: Using generalized Marcum- The function calculates the false alarm probability of the array error classifier. and detection probability : in, For order is The generalized Marcum-Q function; The number of uncalibrated transmit array elements; and The quadratic statistic is respectively in and The corresponding non-central parameter; Define the error probability of the array error classifier: ; Step 7.5: Within the preset threshold search range, perform one-dimensional numerical optimization on the detection thresholds corresponding to gain error and phase error respectively to determine the gain error classifier. Phase error classifier The optimal detection threshold with the lowest error probability: in, This indicates the characteristics of the transmit array gain error. This indicates the phase error characteristics of the transmitting array; and These represent the error probabilities of the gain error classifier and the phase error classifier, respectively. and These represent the corresponding optimized detection thresholds; in the implementation using target false alarm constraints, when satisfying... Among the candidate thresholds, choose the one that makes The highest threshold.
9. The millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method according to claim 8, characterized in that: Step 8 specifically includes the following steps: Step 8.1: Read the path attenuation classifier, AoD classifier, and AoA classifier from Step 6, and the gain error classifier and phase error classifier from Step 7, and calculate the five classifiers respectively. Corresponding false alarm probability Detection probability And calculate the error probability. : ; Step 8.2: Based on the error probability of each classifier Configure non-negative reliability weights : in, For unnormalized weights, To prevent the default positive number with a denominator of zero Therefore, the following is satisfied: ; Step 8.3: Based on the five classifiers binary output and corresponding reliability weights Construct a global authentication score: Under the modeling condition that the classifier outputs are independent or approximately independent, the global authentication score will be... In the assumption and The following are approximations of Gaussian distributions. In the assumption The mean of the global authentication score is: The variance of the global authentication score is: In the assumption The mean of the global authentication score is: The variance of the global authentication score is: Step 8.4: For a given global authentication threshold Then calculate and optimize the global authentication threshold: The global false alarm probability is expressed as: The global detection probability is expressed as: Based on global false alarm probability and global detection probability Define global error probability : Global authentication threshold Perform a one-dimensional search to select the option that minimizes the global error probability. Minimum threshold Alternatively, while satisfying the target global false alarm probability constraint Among the candidate thresholds, choose the one that maximizes the global detection probability. Maximum threshold; Step 8.5: Combine the legitimate sender channel feature statistical parameters obtained in Step 1, the classifier parameters optimized in Steps 6 and 7, and the global fusion parameters determined in Step 8.4 to form an authentication parameter set. : in: This refers to the channel feature statistical parameters obtained in step 1 based on the training data from the legitimate sender. This represents the number of votes for the optimized path obtained by the path decay classifier, AoD classifier, and AoA classifier. These represent the optimized detection thresholds obtained by optimizing the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier, respectively. These represent the fusion weights corresponding to the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier, respectively. This indicates an optimization of the global authentication threshold.
10. The millimeter-wave MIMO heterogeneous feature multi-classifier fusion authentication method according to claim 9, characterized in that: Step 9 specifically includes the following steps: Step 9.1: During the online identity authentication phase, collect the received observation signals of multiple equivalent transmit beams, multiple receive merging vectors, and multiple training subcarriers corresponding to the unknown transmitter. Step 9.2: Construct a third-order received signal tensor according to steps 3 and 4 and perform CP decomposition to obtain the path attenuation estimate, departure angle estimate, arrival angle estimate, transmit array gain error vector, and transmit array phase error vector of the unknown transmitter. Step 9.3: Following the statistical construction and local decision method described in Steps 5-7, call the optimized path voting quantity and local detection threshold pre-stored in the authentication parameter set from Step 8.5 to obtain the binary outputs of the five classifiers. ; Step 9.4: Calculate the global authentication score by calling the path attenuation classifier, AoD classifier, AoA classifier, gain error classifier, and phase error classifier reliability weights stored in Step 8.
5. and optimize global authentication thresholds Compare: The sender was deemed legitimate and authentication was successful. If the sender is identified as a potential impersonator, authentication fails and an impersonation alarm is output.
