Radio frequency fingerprint fast channel adaptive method based on low-rank aggregation
By optimizing the RF fingerprint extractor using a low-rank aggregation method and a non-gradient iterative algorithm, the problem of insufficient generalization ability of RF fingerprint technology under variable channels is solved, and rapid adaptation and efficient identification of unknown channels and devices are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-14
- Publication Date
- 2026-03-27
AI Technical Summary
Existing radio frequency fingerprinting technology has weak generalization ability under variable channels, cannot effectively achieve rapid identification of unknown channels and unknown devices, and has high computational consumption and complex model version control.
A low-rank aggregation method is adopted, which weights and aggregates low-rank adaptive modules of existing channels and optimizes the module weights using an iterative algorithm without gradient participation to construct an aggregated radio frequency fingerprint extractor that is suitable for unknown channels and unknown devices.
It achieves rapid generalization capability of the radio frequency fingerprint extraction model on unknown channels and unknown devices, reduces computing costs and memory requirements, and improves the efficiency of device identification and verification.
Smart Images
Figure CN121750362A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of radio frequency fingerprint extraction, in particular to a radio frequency fingerprint fast channel adaptive method based on low-rank aggregation. BACKGROUND
[0002] Radio frequency fingerprint (RFF) is derived from the differences in transmitter circuit hardware design and production process tolerance. Even the parameters of devices in the same production batch have slight differences.
[0003] The radio frequency fingerprint of a device is unique and difficult to be forged, and therefore can be used for identity recognition and authentication of wired or wireless devices. However, due to the influence of the existing radio frequency fingerprint on the propagation channel, the existing public radio frequency fingerprint technology based on deep learning often cannot effectively realize the generalization feature extraction of the variable channel, that is, the propagation channel of the identified sample must match the channel in the known training set, so that an effective radio frequency fingerprint can be extracted from the identified sample by a trained radio frequency fingerprint extractor, and then high-precision device recognition and verification can be realized based on the effective radio frequency fingerprint.
[0004] The existing radio frequency fingerprint technology has weak generalization ability, and the main reason is that the radio frequency fingerprint extractor uses the optimization objective of maximum likelihood estimation during training, and the optimization method based on maximum likelihood estimation is prone to overfitting to the training set data distribution when learning radio frequency fingerprint features. At the same time, although the full-parameter fine-tuning learning method using a small amount of new scene signal can make the radio frequency fingerprint extraction model adapt to the new scene signal, the method often has high computational consumption. Moreover, repeated loading of the exported model is required during actual edge device deployment, which increases the complexity of model version control. On the other hand, when the generalization ability of the radio frequency fingerprint technology is low and cannot quickly adapt to new scene signals, the range of identifiable devices is limited, which cannot meet the growing application requirements of wireless device identification and verification.
[0005] Therefore, it is urgent to quickly improve the generalization ability of the radio frequency fingerprint extraction model so that it can still guarantee the high distinguishability of the fingerprint on the unseen device under unknown channel conditions. SUMMARY
[0006] The purpose of the present application is to provide a radio frequency fingerprint fast channel adaptive method based on low-rank aggregation, which can quickly learn and adapt to new scene signals using a small amount of data and has high rapid generalization ability for unknown devices or unknown channels. The present application uses a non-gradient participating iterative algorithm to optimize the module weight by weighting aggregation of the low-rank adaptive module of the existing channel, so that the model can quickly adapt to unknown channels and unknown devices. Finally, the high-distinguishability radio frequency fingerprint extracted by the aggregated radio frequency fingerprint extractor can realize device recognition and verification only by distance comparison.
[0007] To solve the technical problems described above, the present application provides the following technical solutions:
[0008] A radio frequency fingerprint fast channel adaptive method based on low-rank aggregation, the method comprising:
[0009] S100, collecting baseband signals and several types of signals experienced known channels, and performing energy normalization processing on the signals;
[0010] S200, learning to obtain a basic radio frequency fingerprint extractor using known baseband signal data;
[0011] S300, using a low-rank matrix adaptation method, fine-tuning the basic radio frequency fingerprint extractor according to the signal data of the several types of signals experienced known channels, to obtain low-rank adaptive modules corresponding to the several types of signals;
[0012] S400, weighting and aggregating all low-rank adaptive modules according to a small amount of new scene signals, and merging with the basic radio frequency fingerprint extractor to obtain a final aggregated radio frequency fingerprint extractor;
[0013] The basic radio frequency fingerprint extractor includes first learnable parameters, the low-rank adaptive module includes second learnable parameters, and the weighting coefficient of the low-rank adaptive module is a third learnable parameter. The first learnable parameters and the second learnable parameters are obtained by maximum likelihood estimation, and the third learnable parameter is obtained by an evolutionary strategy of covariance matrix adaptive adjustment.
[0014] Preferably, S100 comprises:
[0015] The receiver receives a radio frequency signal at a sampling rate The radio frequency signal is received , obtaining a discrete signal , the number of sampling points is , all baseband signals are obtained by down-conversion, and the energy of is normalized to 1; on this basis, all collected signals are divided into different signal types;
[0016] wherein, represents a known baseband signal, represents a signal set of a channel. represents a signal set under a new scene, and the number thereof is much smaller than that of the known signal set.
[0017] Preferably, S200 comprises:
[0018] S201, constructing a basic radio frequency fingerprint extractor for the baseband signal ;
[0019] in, The parameters of the basic radio frequency fingerprint extractor are the first learnable parameters; Depend on It consists of several weight matrices, and the specific formula is as follows: ; Indicates the first One basic weight matrix;
[0020] S202, the first baseband signal Input to the basic radio frequency fingerprint extractor The specific formula for radio frequency fingerprint extraction is as follows: ;
[0021] in, Represented as from the first Radio frequency fingerprint vectors extracted from each sample;
[0022] S203, Constructing an auxiliary linear classifier ;
[0023] in, The total number of devices in the training set. For the first The weight vector corresponding to each device;
[0024] S204, Constructing the first Likelihood function of each sample: ;
[0025] Among them, normalized weights and normalized radio frequency fingerprint , For hyperparameters, For the first The device label for each sample;
[0026] S205, Based on Likelihood Function , construct for parameters The optimization objective of log-maximum likelihood estimation : ;
[0027] in, Represents the radio frequency fingerprint vector;
[0028] S206, utilizing including Training set of baseband signal and device identity sample pairs and basic radio frequency fingerprint extractor By solving The optimal solution for the first learnable parameter is obtained. The optimal solution of the auxiliary linear classifier .
[0029] Preferably, S300 comprises:
[0030] S301, based on the optimal solution of the basic radio frequency fingerprint extractor that has been learned Enable low-rank matrix adaptation method on the basic radio frequency fingerprint extractor:
[0031] Keep the original optimal solution Unchanged, for the first Class of signals experienced known channel, build low-rank adaptive module ; wherein, Indicates the learnable parameters of the low-rank adaptive module corresponding to the signal of the class, which is the second learnable parameter;
[0032] Specifically, the low-rank adaptive module estimates the fine-tuned weight by learning two low-rank matrices And , the expression is: ;
[0033] Wherein, Indicates the basic weight, Indicates the optimal weight obtained after fine-tuning;
[0034] In the training process of the low-rank adaptive module, Keep unchanged, only iterate the values of two low-rank matrices And ; if The shape of , then The shape of , The shape of , Indicates the low-rank hyperparameter;
[0035] S302, use the collected first Class Signal experienced known channel As the model input, repeat the above steps to build a new optimization target for this class of signals experienced known channel ; by solving: ; get the second learnable parameter corresponding to the signal of the class .
[0036] Preferably, S400 comprises:
[0037] S401, based on the optimal solution of the basic radio frequency fingerprint extractor that has been learned And all low-rank adaptive modules The weighted aggregation low-rank adaptive modules are combined with the basic radio frequency fingerprint extractor to construct an aggregated radio frequency fingerprint extractor ;
[0038] wherein, represents a set of a series of low-rank adaptive module weighting coefficients, is a third learnable parameter, and the specific formula is: ;
[0039] The aggregated radio frequency fingerprint extractor estimates the fine-tuned optimal weight by learning the weighting coefficients of a series of low-rank adaptive modules and weighting and aggregating these low-rank adaptive modules:
[0040] ;
[0041] In the training process of the aggregated radio frequency fingerprint extractor, the optimal solution of the basic radio frequency fingerprint extractor and the optimal solution of the low-rank adaptive module remain unchanged, and only the weighting coefficient corresponding to each low-rank adaptive module is iterated . ;
[0042] S402, a small amount of new scene signal is used as the model input, the above steps are repeated to construct a new optimization objective , by solving: , the optimal weight corresponding to the first low-rank adaptive module is obtained .
[0043] Further, the third learnable parameter and the final aggregated radio frequency fingerprint extractor weight are obtained: .
[0044] Preferably, a distance threshold is set to determine the ownership relationship between two radio frequency fingerprints: if the value of the cosine distance between the two radio frequency fingerprints does not exceed , it is determined that they belong to the same device; if the distance exceeds , it is considered that they come from different devices.
[0045] wherein, the threshold is the optimal segmentation point determined based on the minimum equal error rate statistically calculated in the training data set.
[0046] Preferably, the parameterized basic radio frequency fingerprint extractor, the parameterized low-rank adaptive module, and the aggregated radio frequency fingerprint extractor are convolutional neural networks.
[0047] Compared with the prior art, the present application has the beneficial effects that:
[0048] 1、The application uses a low-rank matrix adaptation method to estimate the fine-tuned full-parameter result with two low-rank matrices, so that the calculation cost and memory requirement required for model adaptation are greatly reduced.
[0049] 2、The application proposes a low-rank aggregation-based radio frequency fingerprint fast channel adaptive method, so that the weight of the final aggregated radio frequency fingerprint extractor can be obtained by weighting aggregation of the weight of the known low-quality adaptive module, so that the aggregated radio frequency fingerprint extractor has strong generalization ability to unknown channels or unknown devices.
[0050] 3、The low-quality adaptive module weighting coefficient optimization method used in the application can complete effective iterative update of the weighting coefficient by using a non-gradient participating iterative algorithm, further reducing the calculation amount, and finally obtaining an aggregated radio frequency fingerprint extractor with fast adaptation speed to new scenes. BRIEF DESCRIPTION OF DRAWINGS
[0051] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, which together with the embodiments of the application, are used to explain the application, and do not constitute a limitation on the application. In the drawings:
[0052] Figure 1 is a flowchart of a low-rank aggregation-based radio frequency fingerprint fast channel adaptive method of the application. DETAILED DESCRIPTION
[0053] The technical solutions in the embodiments of the application will be described in detail below with reference to the drawings of the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.
[0054] Please refer to Figure 1 , the application provides a technical solution:
[0055] Embodiment 1: A low-rank aggregation-based radio frequency fingerprint fast channel adaptive method, specifically as Figure 1 shown, comprising the following steps:
[0056] Step 1, the receiver receives a radio frequency signal at a sampling rate , obtains a discrete signal , the number of sampling points is , all baseband signals are obtained by down-conversion, and the energy of is normalized to 1.
[0057] Specifically, in this embodiment, a 10Msps sampling rate is selected, and a 2MHz symbol rate QPSK preamble signal is collected and down-converted to obtain a corresponding baseband signal . Therefore, the receiving end performs 5 times oversampling on the two signals, and the signal segment used contains 1280 sampling points.
[0058] On this basis, all collected signals are divided into different signal types, wherein, represents a known baseband signal, represents a signal set of a kind of channel. represents a signal set in a new scenario, and the number is much smaller than that of the known signal set.
[0059] Step 2, low-rank adaptive module weighted aggregation, specifically including:
[0060] Step 2.1, constructing a basic radio frequency fingerprint extractor for the baseband signal , wherein, represents the parameters of the basic radio frequency fingerprint extractor, and is the first learnable parameter. Specifically, is composed of weight matrices, expressed as: .
[0061] Step 2.2, inputting the first baseband signal to the basic radio frequency fingerprint extractor for radio frequency fingerprint extraction, expressed as: .
[0062] Wherein represents the radio frequency fingerprint vector extracted from the first sample.
[0063] Step 2.3, constructing an auxiliary linear classifier , wherein, is the total number of devices in the training set, is the weight vector corresponding to the first device;
[0064] Step 2.4, constructing the likelihood function of the first sample, expressed as:
[0065] ;
[0066] Wherein, the normalized weight and the normalized radio frequency fingerprint , is a hyperparameter, is the first The device label for each sample;
[0067] Specifically, in this embodiment, hyperparameters Set it to 10.
[0068] Step 2.5: Based on the likelihood function constructed in Step 2.4, construct a function for the parameters. The optimization objective of log-maximum likelihood estimation The expression is:
[0069] ;
[0070] in, This refers to the radio frequency fingerprint vector extracted in step 2.2.
[0071] Step 2.6, utilizing including Training set of baseband signal and device identity sample pairs and basic radio frequency fingerprint extractor This can be solved by: ;
[0072] Obtain the optimal solution for the first learnable parameter The optimal solution of the auxiliary linear classifier .
[0073] Specifically, in this embodiment, the selected... A custom signal convolutional neural network, which includes:
[0074] 1) Signal to image layer: The 1280 sampling points are truncated into 16 sampling points each and rearranged into a signal image with 80 rows and 16 columns. The real part and imaginary part of each sampling point are used as two channels of the image.
[0075] 2) The 6-layer convolutional neural network has the following (number of kernels / kernel size / convolution interval / image edge augmentation) for each layer as follows: Layer 1 (8 / 3x3 / 1 / 1), Layer 2 (16 / 3x3 / 1 / 1), Layer 3 (32 / 3x3 / 2 / 1), Layer 4 (64 / 3x3 / 1 / 1), Layer 5 (128 / 3x3 / 2 / 1), and Layer 6 (246 / 3x3 / 2 / 1).
[0076] 3) The fully connected layer is used as the output layer, with an output dimension of 512.
[0077] Step 2.7: Optimal solution based on the learned basic RF fingerprint extractor Low-rank matrix adaptation is enabled on the basic RF fingerprint extractor, while maintaining the original optimal solution. Unchanged, targeting the first Build a low-rank adaptive module based on signals that have experienced a known channel. .in, This represents the learnable parameters of the low-rank adaptive module corresponding to this type of signal, and is the second learnable parameter.
[0078] Specifically, the low-rank adaptive module learns two low-rank matrices. and To estimate the fine-tuned weights, The expression is: .
[0079] in, Based on the weights, These are the final weights obtained after fine-tuning. During the training of the low-rank adaptive module... Keeping the same, only iterate over the two low-rank matrices. and The value of . If The shape is ,So The shape is , The shape is , It is a low-rank hyperparameter.
[0080] Step 2.8: Utilize the collected data... kind Signals passing through a known channel Using this as model input, repeat the above steps to construct a new optimization objective for this type of signal experiencing a known channel. This can be solved by: .
[0081] Obtain the second learnable parameter corresponding to this type of signal. .
[0082] Step 2.9: Optimal solution based on the learned basic RF fingerprint extractor and the optimal solution of all low-rank adaptive parameters The low-rank adaptive module is weighted and aggregated, and then merged with the basic RF fingerprint extractor to construct an aggregated RF fingerprint extractor. . Let represent the set of weighted coefficients of a series of low-rank adaptive modules, which is the third learnable parameter, and its expression is: .
[0083] The aggregated RF fingerprint extractor estimates the fine-tuned final weights by learning the weighting coefficients of a series of low-rank adaptive modules and then weighting and summing these low-rank adaptive modules. The expression is as follows:
[0084] .
[0085] During the training of the aggregated RF fingerprint extractor, the optimal solution of the basic RF fingerprint extractor is maintained. and the optimal solution of low-rank adaptive modules The process remains unchanged, only iterating over the weighting coefficients corresponding to each type of low-rank adaptive module. .
[0086] Step 2.10: Utilize a small number of new scene signals Using this as model input, repeat the above steps to construct a new optimization objective. This can be solved by: ;
[0087] Get the first Optimal weighting coefficients corresponding to low-rank adaptive modules .
[0088] Furthermore, the third learnable parameter can be obtained. The final aggregated RF fingerprint extractor weights are expressed as follows: .
[0089] Specifically, in this embodiment, stochastic gradient descent is used to solve the optimization equations in steps 2.6 and 2.8. More specifically, stochastic gradient descent (SGD) is used as the optimizer with a learning rate of 1e-2 and parameters... The optimization formula in step 2.10 is solved using an evolutionary strategy that adaptively adjusts the covariance matrix, with an iteration step size of 0.7.
[0090] Specifically, in this embodiment, radio frequency fingerprint recognition and verification are achieved by comparing cosine distances: a distance threshold is set. This is used to determine the attribution relationship between two RFID fingerprints. If the value of the cosine distance between the two RFID fingerprints does not exceed... If they are within the same device, then they are considered to belong to the same device; if the distance exceeds [a certain value], then they are considered to belong to the same device. If the value is different, it is considered to originate from a different device. This threshold... The optimal split point is determined based on the minimum equal error rate statistically analyzed in the training dataset.
[0091] In summary, this invention provides a fast channel adaptation method for RF fingerprinting based on low-rank aggregation. This method allows the weights of the desired aggregated RF fingerprint extractor to be obtained by weighted aggregation of the weights of known low-rank adaptive modules, thus enabling the RF fingerprint model to have strong generalization ability for unknown channels or unknown devices. This invention uses a low-rank matrix adaptation method, estimating the fine-tuned full parameter results using two low-rank matrices, thereby significantly reducing the computational cost and memory requirements of low-rank matrix adaptation. This invention employs a non-gradient-participatory iterative algorithm to iteratively update the weighting coefficients corresponding to the low-rank adaptive modules, further reducing the computational load. The resulting aggregated RF fingerprint extractor adapts quickly to new scenarios. All aspects not detailed in this invention are well-known techniques to those skilled in the art.
[0092] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A fast channel adaptation method based on low-rank aggregation radio frequency fingerprinting, characterized in that: The method includes: S100: Acquire baseband signals and several types of signals that have passed through known channels, and perform energy normalization processing on the signals; S200: Using known baseband signal data, a basic radio frequency fingerprint extractor is learned; S300. Using the low-rank matrix adaptation method, the basic radio frequency fingerprint extractor is fine-tuned based on several types of signal data that have experienced known channels, to obtain the low-rank adaptive modules corresponding to these types of signals. S400: Based on a small number of new scene signals, all low-rank adaptive modules are weighted and aggregated, and then merged with the basic RF fingerprint extractor to obtain the final aggregated RF fingerprint extractor. The basic radio frequency fingerprint extractor includes a first learnable parameter, the low-rank adaptive module includes a second learnable parameter, and the weighting coefficient of the low-rank adaptive module is a third learnable parameter. The first and second learnable parameters are obtained through maximum likelihood estimation, and the third learnable parameter is obtained through an evolutionary strategy of adaptive adjustment of the covariance matrix.
2. The fast channel adaptation method for radio frequency fingerprinting based on low-rank aggregation as described in claim 1, characterized in that, The S100 includes: The receiver uses a sampling rate Receive radio frequency signals To obtain discrete signals The number of sampling points is The entire baseband signal is obtained through down-conversion. and will The energy is normalized to 1; based on this, all acquired signals are classified into different signal types; in, Indicates a known baseband signal. express A set of signals for a given channel. This represents the set of signals in the new scenario, and its number is much smaller than the known set of signals.
3. The fast channel adaptation method based on low-rank aggregation radio frequency fingerprinting as described in claim 1, characterized in that, The S200 includes: S201, Construct a basic radio frequency fingerprint extractor for baseband signals ; in, The parameters of the basic radio frequency fingerprint extractor are the first learnable parameters; Depend on It consists of several weight matrices, and the specific formula is as follows: ; Indicates the first One basic weight matrix; S202, the first baseband signal Input to the basic radio frequency fingerprint extractor The specific formula for radio frequency fingerprint extraction is as follows: ; in, Represented as from the first Radio frequency fingerprint vectors extracted from each sample; S203, Constructing an auxiliary linear classifier ; in, The total number of devices in the training set. For the first The weight vector corresponding to each device; S204, Constructing the first Likelihood function of each sample: ; Among them, normalized weights and normalized radio frequency fingerprint , For hyperparameters, For the first The device label for each sample; S205, Based on Likelihood Function , construct for parameters The optimization objective of log-maximum likelihood estimation : ; in, Represents the radio frequency fingerprint vector; S206, utilizing including Training set of baseband signal and device identity sample pairs and basic radio frequency fingerprint extractor By solving The optimal solution for the first learnable parameter is obtained. The optimal solution of the auxiliary linear classifier .
4. The fast channel adaptation method for radio frequency fingerprinting based on low-rank aggregation as described in claim 1, characterized in that, The S300 includes: S301, Optimal solution based on the learned basic radio frequency fingerprint extractor Enable low-rank matrix adaptation on the basic RF fingerprint extractor: Maintain the original optimal solution Unchanged, targeting the first Build a low-rank adaptive module based on signals that have experienced a known channel. ;in, This represents the learnable parameters of the low-rank adaptive module corresponding to this type of signal, and is the second learnable parameter; Specifically, the low-rank adaptive module learns two low-rank matrices. and To estimate the fine-tuned weights, The expression is: ; in, Indicates the basic weights, This represents the final weight obtained after fine-tuning; During the training of the low-rank adaptive module Keeping the same, only iterate over the two low-rank matrices. and The value; if The shape is ,So The shape is , The shape is , Indicates low-rank hyperparameters; S302, using the collected data... kind Signals passing through a known channel Using this as model input, repeat the above steps to construct a new optimization objective for this type of signal experiencing a known channel. By solving: ; Obtain the second learnable parameter corresponding to this type of signal. .
5. The fast channel adaptation method for radio frequency fingerprinting based on low-rank aggregation as described in claim 1, characterized in that, The S400 includes: S401, Optimal solution based on the learned basic RF fingerprint extractor and all low-rank adaptive modules The low-rank adaptive module is weighted and aggregated, and then merged with the basic RF fingerprint extractor to construct an aggregated RF fingerprint extractor. ; in, Let represent the set of weighted coefficients of a series of low-rank adaptive modules, which is the third learnable parameter. The specific formula is as follows: ; The aggregated RF fingerprint extractor estimates the fine-tuned final weights by learning the weighting coefficients of a series of low-rank adaptive modules and then weighting and summing these low-rank adaptive modules. ； During the training of the aggregated RF fingerprint extractor, the optimal solution of the basic RF fingerprint extractor is maintained. and the optimal solution of low-rank adaptive modules The process remains unchanged; only the weighting coefficients corresponding to each low-rank adaptive module are iterated. ; S402, Utilizing a small number of new scene signals Using this as model input, repeat the above steps to construct a new optimization objective. By solving: , obtained the Optimal weights corresponding to low-rank adaptive modules ; Furthermore, the third learnable parameter is obtained. And the final aggregated RF fingerprint extractor weights: .
6. A fast channel adaptation method for radio frequency fingerprinting based on low-rank aggregation according to any one of claims 1-5, characterized in that, Set a distance threshold This is used to determine the attribution relationship between two RFID fingerprints: if the value of the cosine distance between the two RFID fingerprints does not exceed [the specified value]. If they are within the same device, then they are considered to belong to the same device; if the distance exceeds [a certain value], then they are considered to belong to the same device. If so, it is assumed that they originated from different devices; Wherein, the threshold The optimal split point is determined based on the minimum equal error rate statistically analyzed in the training dataset.
7. A fast channel adaptation method for radio frequency fingerprinting based on low-rank aggregation according to any one of claims 1-5, characterized in that, The parameterized basic RF fingerprint extractor, the parameterized low-rank adaptive module, and the aggregated RF fingerprint extractor are all convolutional neural networks.