Radio frequency fingerprint causal invariance modeling and end side radiation source open set identification method

Abstract

The invention discloses a radio frequency fingerprint causal invariance modeling and end-side radiation source open set identification method, which comprises the following steps: constructing a teacher model, and training the teacher model by adopting original I / Q signals of N radiation sources with known categories to obtain a trained teacher model; according to the radio frequency fingerprint causal invariance modeling and end-side radiation source open set identification method, a preprocessing technology for revealing inherent nonlinear fingerprints of a transmitter and a self-supervised training framework for learning domain invariant causal characterization are combined, so that under the condition that only single source domain training data exists, the single source domain training data can be identified; the method is effectively generalized to a target domain in which channel characteristics, noise levels and the like have significant differences, and the pain point of poor generalization ability of a traditional deep learning model is solved; the output of the classifier and the extreme value theory based on Weibull distribution are calibrated, and a decision boundary with statistical significance is constructed for the known radiation source category, so that the model can accurately classify the known radiation source.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication and machine learning technology, specifically relating to a method for modeling the causal invariance of radio frequency fingerprints and identifying open sets of end-side radiation sources. Background Technology

[0002] Radio frequency (RF) fingerprints of transmitters (radiation sources) are unintentional, unique physical layer features introduced by hardware defects. Like biometric fingerprints, they provide a unique identifier for each transmitter. Radiation source identification (SEI) technology uses RF fingerprints for transmitter authentication and has significant application value in both military (target identification, intrusion detection, etc.) and civilian (IoT security, dynamic spectrum access management, etc.) fields. However, existing technologies face several challenges in achieving efficient and robust radiation source identification. First, traditional RF fingerprint modeling methods heavily rely on predefined, manually designed features by domain experts. These features are typically extracted from the transient or steady-state portions of signals. The extraction process requires complex signal preprocessing steps (such as synchronization, demodulation, carrier tracking, etc.) and is highly dependent on expert knowledge, resulting in poor technological adaptability and long development cycles. Second, these methods typically require a pre-built "feature library" or "database" to store fingerprint information of known transmitters. Such technologies are unable to effectively identify novel or unknown radiation sources not recorded in the database, which is a serious limitation in dynamically changing electromagnetic environments.

[0003] Although deep learning technology has been applied to the field of radiation source identification technology, it still faces several bottlenecks: (1) Insufficient generalization ability of the model: The performance of the model trained in a single environment (source domain) often drops sharply when deployed to a new, unknown actual environment (target domain) with different channel conditions, noise levels or interference; (2) Inaccurate detection of unknown device intrusion: Most existing models are designed based on the "closed-set" assumption, that is, assuming that all signals encountered in the testing phase come from the known categories during training. In the real world, this is obviously not true. When a signal from an unknown emission source is input, the closed-set classifier will incorrectly classify it into a known category, which is unacceptable for applications such as security monitoring. Therefore, how to solve the more realistic "open-set" identification problem, that is, to accurately identify known transmitters while effectively rejecting unknown transmitters, is the key challenge we are currently facing; (3) Limited resources for edge deployment: Although modern deep neural network models have powerful performance, their huge number of parameters and computational complexity make it difficult to deploy on edge transmitters where computing, storage and power consumption resources are strictly limited. Summary of the Invention

[0004] The purpose of this invention is to address the problems raised in the background art by proposing a method for modeling the causal invariance of radio frequency fingerprints and identifying open sets of end-side radiation sources.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] This invention proposes a method for modeling the causal invariance of radio frequency fingerprints and identifying open sets of end-side radiation sources, comprising:

[0007] A teacher model is constructed and trained using the raw I / Q signals of N known categories of radiation sources to obtain a trained teacher model.

[0008] The trained teacher model is then lightweightly optimized to obtain the student model;

[0009] Based on knowledge distillation loss, a lightweight model is obtained by using a pre-trained teacher model to guide the training of the student model.

[0010] The raw I / Q signal of a preset length from the radiation source to be detected is acquired, and after preprocessing, the amplitude spectrum and nonlinear features are obtained. Both the amplitude spectrum and nonlinear features are input into a lightweight model to obtain the initial decision score vector of the radiation source to be detected. Based on N known categories of radiation sources, the initial decision score vector of the radiation source to be detected is corrected, and finally the category of the radiation source to be detected is obtained.

[0011] Preferably, the trained teacher model includes an encoder, a fully connected layer and a classifier connected in sequence, wherein the classifier is a multilayer perceptron and the encoder is a ResNet18 model.

[0012] Preferably, the training process of the teacher model includes:

[0013] Collect raw I / Q signals from N known types of radiation sources, slice each raw I / Q signal, use one slice as a training sample to train the teacher model, and a batch includes multiple samples.

[0014] First, the training samples are preprocessed to obtain the amplitude spectrum and nonlinear features. Then, random perturbations are applied to the amplitude spectrum and nonlinear features to enhance them.

[0015] The amplitude spectrum of each sample after random perturbation enhancement and the corresponding nonlinear features are passed together through an encoder and then through a fully connected layer to obtain two output features, based on Barlow Twins loss. Self-supervised learning is performed on the two output features;

[0016] Two output features are concatenated to obtain concatenated features. A first mask is applied to the preset pixels in the concatenated features to obtain a first mask image. A second mask is applied to the pixels in the concatenated features other than the preset pixels to obtain a second mask image. The concatenated features are multiplied by the first mask image and the second mask image respectively to obtain the first feature and the second feature respectively.

[0017] The first feature and the second feature are respectively input into two classifiers;

[0018] For each batch during the training phase, calculate the cross-entropy loss of the output of the classifier after the first and second features pass through it. and Combined with the loss of Barlow Twins used in self-supervised learning The total loss during the training phase can be obtained as follows: ,in These are weight parameters;

[0019] And based on total loss Train the teacher model until the total loss is reached. It converges to the minimum value.

[0020] Preferably, the preprocessing process is as follows: performing a spectral transformation on the original I / Q signal or slice of the radiation source to be detected for a preset length to obtain an amplitude spectrum, and performing a nonlinear intensity transformation on the amplitude spectrum to obtain nonlinear characteristics;

[0021] The spectral transformation process involves performing a Fast Fourier Transform on the original I / Q signal or slice of the radiation source to be detected for a preset length, thereby transforming the input from the time domain to the frequency domain, and extracting the spectrum. The amplitude is used as the amplitude spectrum;

[0022] The nonlinear intensity transformation process is as follows: the amplitude spectrum is remapped based on a cubic Bézier curve using a preset nonlinear mapping function to obtain the nonlinear characteristics. .

[0023] Preferably, the step of performing lightweight optimization on the trained teacher model to obtain the student model includes:

[0024] The encoder in the trained teacher model is designed as a lightweight convolutional neural network to obtain the student model.

[0025] Preferably, the knowledge distillation loss The calculation formula is as follows:

[0026] ;

[0027] in, and The outputs of the trained teacher model and student model are respectively. and These are the output features of the encoder in the trained teacher model and student model, respectively. For preset temperature, For the softmax function, , and All are weighted parameters. The cross-entropy loss used to train the student model Let KL be the divergence function.

[0028] Preferably, the radio frequency fingerprint causal invariance modeling and end-side radiation source open set identification method further includes:

[0029] For each category of radiation source in the training set, obtain the activation vector corresponding to all training samples in the radiation source of the current category, and calculate the mean of all activation vectors to obtain the average activation vector of the radiation source of the current category. The output of the second to last layer of each model is the activation vector.

[0030] Calculate the distance from the activation vector to the average activation vector of all correctly classified training samples of the radiation source of the current category, and select the m largest distances as extreme distances. Fit the m extreme distances to the Weibull distribution to obtain the cumulative distribution function of the Weibull distribution of the radiation source of the current category.

[0031] This leads to the Weibull cumulative distribution function of the m extreme distances corresponding to the N categories of radiation sources in the training set. , and They are the nth in N categories The scale and shape parameters of each type of radiation source are obtained by fitting the m extreme distances corresponding to each type of radiation source, and all of them are greater than zero. It is a random variable.

[0032] Preferably, in the lightweight model, the amplitude spectrum and nonlinear features are passed together through an encoder and a fully connected layer to obtain two output features. The two output features are concatenated to obtain concatenated features, which are then passed through a classifier to obtain the output of the lightweight model.

[0033] Preferably, the step of correcting the initial decision score vector of the radiation source to be detected based on N known categories of radiation sources, and finally obtaining the category of the radiation source to be detected, includes:

[0034] Acquire the raw I / Q signal of the radiation source to be detected at a preset length. The corresponding activation vector, referred to as the first activation vector, is used to calculate the distance between the first activation vector and the average activation vectors of the known N categories of radiation sources, forming a first distance set. Each first distance in the first distance set is then substituted into the Weibull cumulative distribution function of each category of radiation source to obtain the probability that the original I / Q signal of the radiation source to be detected at a preset length belongs to the corresponding category of radiation source among the N categories, as shown in the following formula:

[0035] ;

[0036] in, The raw I / Q signal of the preset length of the radiation source to be detected Belongs to the N categories The probability of each type of radiation source. for First activation vector With the N categories, the first The first distance between the average activation vectors of each category of radiation sources and They are the nth in N categories The scale and shape parameters of the cumulative distribution function of the Weibull distribution for each category of radiation source are both greater than zero;

[0037] Select from the first distance set The smallest first distance is used to form the second distance set. The radiation sources corresponding to each second distance in the second distance set are recorded and called the class to be corrected. The classes to be corrected are sorted in ascending order according to their corresponding second distances, and the calculation is performed. The correction weights for each class to be corrected are given by the following formula:

[0038] ;

[0039] in, , for The first class to be corrected Corrected weights for each category of radiation sources, The raw I / Q signal of the preset length of the radiation source to be detected belong The first class to be corrected The probability of each type of radiation source. The raw I / Q signal of the preset length of the radiation source to be detected Not belonging to The first class to be corrected The probability of each type of radiation source. for The first class to be corrected The sorting sequence number of each category of radiation source;

[0040] The initial decision score vector has a dimension of N. 1, which contains N elements, each corresponding one-to-one with one of the N known categories of radiation sources. Each element is called a decision score. The correction weights for each class to be corrected are multiplied by the elements corresponding to the radiation sources of the corresponding class in the initial decision score vector, thereby achieving the correction of the radiation sources of each class. Each element is adjusted for a decision score, while the remaining elements remain unchanged, to obtain the original I / Q signal of the radiation source to be detected at a predetermined length. The corresponding decision score vector is called the first decision score vector;

[0041] The modified decision scores in the first decision score vector are compared with the original decision scores. The modified decision scores that have decreased compared to the original scores are selected, and the difference between the selected modified decision scores and the original decision scores is calculated. All differences are summed and concatenated into the first decision score vector to obtain the second decision score vector. The second decision score vector serves as the decision score vector corresponding to the N+1 categories of radiation sources, and its dimension is (N+1). 1;

[0042] From the N+1 decision scores in the second decision score vector, select the maximum value, obtain the row number corresponding to the maximum value, and use the row number corresponding to the maximum value as the original I / Q signal of the preset length of the radiation source to be detected. For the corresponding radiation source category, if the (N+1)th element in the second decision score vector is the maximum value, then the original I / Q signal of the radiation source to be detected is determined to be of a preset length. From an unknown class.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] This method for RF fingerprint causal invariance modeling and open set identification of end-side radiation sources combines preprocessing techniques that reveal the inherent nonlinear fingerprint of the transmitter with a self-supervised training framework that learns domain-invariant causal representations. This invention can effectively generalize to target domains with significant differences in channel characteristics and noise levels under the condition of only single source domain training data, thus solving the pain point of poor generalization ability of traditional deep learning models.

[0045] By calibrating the classifier output with the extreme value theory based on the Weibull distribution, a statistically significant decision boundary is constructed for known radiation source categories. This enables the model to not only accurately classify known radiation sources, but also effectively identify and reject signals from unknown radiation sources, significantly reducing the false alarm rate of unknown radiation source intrusion detection and improving reliability in real open environments.

[0046] This method can significantly reduce the number of model parameters, computational complexity, and inference latency while maintaining recognition accuracy. This enables the high-precision SEI algorithm to be successfully deployed on resource-constrained edge transmitters, achieving real-time or near-real-time recognition tasks. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the radio frequency fingerprint causal invariance modeling and end-side radiation source open set identification method of the present invention;

[0048] Figure 2 This is a structural block diagram of the teacher model or student model of the present invention;

[0049] Figure 3 This is a schematic diagram of the teacher model training process of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0052] like Figures 1-3 As shown, a method for modeling the causal invariance of radio frequency fingerprints and identifying open sets of end-side radiation sources is provided, including:

[0053] Step 1: Construct a teacher model and train it using the raw I / Q signals of N known categories of radiation sources to obtain a trained teacher model.

[0054] Step 1.1: Constructing the Teacher Model: The trained teacher model consists of an encoder, a fully connected layer, and a classifier connected in sequence. The classifier is a multilayer perceptron (the penultimate layer of the classifier is a fully connected layer, and the last layer is an activation layer (Sigmoid activation)). The encoder is a ResNet18 model. The input of the trained teacher model is the amplitude spectrum and nonlinear features. The amplitude spectrum and nonlinear features are passed together through the encoder and the fully connected layer (in sequence) to obtain two output features (these two output features correspond to the amplitude spectrum and the nonlinear features, respectively). The two output features are concatenated to obtain the concatenated features. The concatenated features are then passed through the classifier to obtain the output of the teacher model (i.e., the decision score vector).

[0055] Step 1.2, the training process of the teacher model includes:

[0056] Raw I / Q signals from N known types of radiation sources are collected. Each raw I / Q signal is sliced ​​(using a sliding window mechanism with preset length and step size to divide it into a series of overlapping slices). Each slice is used as a training sample to train the teacher model. A batch includes multiple samples.

[0057] First, the training samples are preprocessed to obtain the amplitude spectrum and nonlinear features. (The preprocessing process involves: performing a spectral transform on the slices to obtain the amplitude spectrum, and then performing a nonlinear intensity transform on the amplitude spectrum to obtain the nonlinear features; the spectral transform process involves performing a Fast Fourier Transform on the slices to convert the input from the time domain to the frequency domain, and then extracting the spectrum.) The amplitude is taken as the amplitude spectrum; the nonlinear intensity transformation process is as follows: the amplitude spectrum is remapped based on a cubic Bézier curve through a preset nonlinear mapping function to obtain the nonlinear characteristics. ,and ,and and It follows a Gaussian distribution, α ~ (1, σ1), σ1≈0.1; β ~ (0, σ2), σ2≈0.5; (For cubic Bessel transform), the amplitude spectrum and nonlinear characteristics are enhanced by random perturbation (random scaling and offset perturbation).

[0058] The amplitude spectrum of each sample after random perturbation enhancement and the corresponding nonlinear features are passed together through an encoder and then through a fully connected layer to obtain two output features. , ), and based on Barlow Twins loss Self-supervised learning is performed on the two output features;

[0059] Two output features are concatenated to obtain a concatenated feature. A first mask is applied to the preset pixels in the concatenated feature to obtain a first mask image (the pixels in the concatenated feature other than the first mask portion are blank, forming the first mask image). A second mask is applied to the pixels in the concatenated feature other than the preset pixels to obtain a second mask image (the pixels in the concatenated feature other than the second mask portion are blank, forming the second mask image). The concatenated feature is multiplied by the first mask image and the second mask image respectively to obtain the first feature and the second feature.

[0060] The first feature and the second feature are respectively input into two classifiers;

[0061] For each batch during the training phase, calculate the cross-entropy loss of the output of the classifier after the first and second features pass through it. and Combined with the loss of Barlow Twins used in self-supervised learning The total loss during the training phase can be obtained as follows: ,in These are weight parameters;

[0062] And based on total loss Train the teacher model until the total loss is reached. It converges to the minimum value.

[0063] Step 2: Perform lightweight optimization on the trained teacher model to obtain the student model, including:

[0064] The encoder in the trained teacher model is designed as a lightweight convolutional neural network (such as MobileNetV3) to obtain the student model.

[0065] Step 3: Based on knowledge distillation loss, use the trained teacher model to guide the training of the student model to obtain a lightweight model;

[0066] Among them, knowledge distillation loss The calculation formula is as follows:

[0067] ;

[0068] in, and The outputs of the trained teacher model and student model are respectively. and These are the output features of the encoder in the trained teacher model and student model, respectively. For preset temperature, For the softmax function, , and All are weighted parameters. The cross-entropy loss used to train the student model Let KL be the divergence function, where , ∈[0.1,1.0], ∈ [0.1,1.0].

[0069] In this process, singular value decomposition or tensor decomposition (such as CP / SVD / Tucker) is used on large fully connected layers or convolutional kernels to decompose the weight matrix into multiple smaller matrices. After decomposition, the original computational layer is replaced by a series of new computational layers representing these low-rank factors, and the total number of parameters in these new layers is much smaller than that in the original layer.

[0070] Quantization: Perform quantization-aware training on the lightweight model, inserting pseudo-quantization nodes to simulate INT8 operations and minimize accuracy loss. Alternatively, post-training quantization can be used to quickly obtain an INT8 model.

[0071] Deployment optimization: Lightweight models are exported from their native training frameworks (such as PyTorch) to a standard, hardware-independent intermediate representation format, such as the Open Neural Network Exchange Format (ONNX). ONNX models are loaded using inference engines or compilers provided by specific hardware vendors, and hardware-aware graph optimization is performed. For example: for NVIDIA GPUs, TensorRT can be used for layer fusion, accuracy calibration, and automatic operator tuning to maximize throughput; for Android mobile transmitters, NNAPI can be used to distribute computational tasks to underlying hardware accelerators such as GPUs, DSPs, or NPUs; for FPGAs or dedicated SoC platforms (such as XiliNx and ZyNq), models can be compiled into customized hardware instructions or logic circuits to achieve extreme energy efficiency.

[0072] Step 4, the preparatory work for open set recognition includes: for each category of radiation source in the training set, obtain the activation vectors corresponding to all training samples in the radiation source of the current category (i.e., the intermediate features output by the classifier), and calculate the mean of all activation vectors to obtain the average activation vector of the radiation source of the current category, where the output of the second to last layer of each model is the activation vector.

[0073] Calculate the distance (Euclidean distance) between the activation vectors of all correctly classified training samples of the radiation source of the current category and the average activation vector. Select the m largest distances (such as 3%-5% of all correctly classified training samples of the radiation source of the current category) as extreme distances. Fit the m extreme distances to the Weibull distribution to obtain the cumulative distribution function of the Weibull distribution of the radiation source of the current category.

[0074] This leads to the Weibull cumulative distribution function of the m extreme distances corresponding to the N categories of radiation sources in the training set. , and They are the nth in N categories The scale and shape parameters of each type of radiation source are obtained by fitting the m extreme distances corresponding to each type of radiation source, and all of them are greater than zero. It is a random variable.

[0075] Step 5: Acquire the raw I / Q signal from the radiation source of the preset length, and perform preprocessing to obtain the amplitude spectrum and nonlinear characteristics (the preprocessing process is as follows: perform a spectral transformation on the raw I / Q signal of the preset length from the radiation source to obtain the amplitude spectrum, and perform a nonlinear intensity transformation on the amplitude spectrum to obtain the nonlinear characteristics; the spectral transformation process is as follows: perform a fast Fourier transform on the raw I / Q signal of the preset length from the radiation source to obtain the frequency domain, and extract the spectrum). The amplitude is taken as the amplitude spectrum; the nonlinear intensity transformation process is as follows: the amplitude spectrum is remapped based on a cubic Bézier curve through a preset nonlinear mapping function to obtain the nonlinear characteristics. Furthermore, the amplitude spectrum and nonlinear features are both input into the lightweight model to obtain the initial decision score vector of the radiation source to be detected. Based on N known categories of radiation sources, the initial decision score vector of the radiation source to be detected is corrected, and finally the category of the radiation source to be detected is obtained.

[0076] In the lightweight model, the amplitude spectrum and nonlinear features are passed together through an encoder and a fully connected layer to obtain two output features. The two output features are concatenated to obtain concatenated features, which are then passed through a classifier to obtain the output of the lightweight model (i.e., the initial decision score vector).

[0077] Based on N known categories of radiation sources, the initial decision score vector of the radiation source to be detected is corrected, ultimately yielding the category of the radiation source to be detected, including:

[0078] Acquire the raw I / Q signal of the radiation source to be detected at a preset length. The activation vector corresponding to the length of the slice (which is the output of the penultimate layer of the lightweight model after preprocessing the original I / Q signal of the radiation source of the preset length) is called the first activation vector. The distances between the first activation vector and the average activation vectors of the known N categories of radiation sources are calculated to form a first distance set. Each first distance in the first distance set is substituted into the Weibull cumulative distribution function of each category of radiation source to obtain the probability that the original I / Q signal of the radiation source of the preset length belongs to the corresponding category of radiation source among the N categories. The formula is as follows:

[0079] ;

[0080] in, The raw I / Q signal of the preset length of the radiation source to be detected Belongs to the N categories The probability of each type of radiation source. for First activation vector With the N categories, the first The first distance between the average activation vectors of each category of radiation sources and They are the nth in N categories The scale and shape parameters of the cumulative distribution function of the Weibull distribution for each category of radiation source are both greater than zero. ;

[0081] Select from the first distance set The smallest first distance ( This forms a second distance set, recording the radiation source category corresponding to each second distance in the second distance set (the radiation source category corresponding to this category is obtained by "calculating the distance between the first activation vector and the average activation vector of the known N categories of radiation sources", thus obtaining the radiation source category corresponding to the first distance), and is called the class to be corrected. The classes to be corrected are sorted in ascending order according to their corresponding second distances, and the calculation is performed. The correction weights for each class to be corrected are given by the following formula:

[0082] ;

[0083] in, , for The first class to be corrected Corrected weights for each category of radiation sources, The raw I / Q signal of the preset length of the radiation source to be detected belong The first class to be corrected The probability of each type of radiation source. The raw I / Q signal of the preset length of the radiation source to be detected Not belonging to The first class to be corrected The probability of each type of radiation source. for The first class to be corrected The sorting sequence number of each category of radiation source;

[0084] The initial decision score vector has a dimension of N. 1, which contains N elements, each corresponding one-to-one with one of the N known categories of radiation sources. Each element is called a decision score. The correction weights for each class to be corrected are multiplied by the elements corresponding to the radiation sources of the corresponding class in the initial decision score vector, thereby achieving the correction of the radiation sources of each class. Each element is adjusted for a decision score, while the remaining elements remain unchanged, to obtain the original I / Q signal of the radiation source to be detected at a predetermined length. The corresponding decision score vector is called the first decision score vector;

[0085] The modified decision scores in the first decision score vector are compared with the original decision scores. The modified decision scores that have decreased compared to the original scores are selected, and the difference between the selected modified decision scores and the original decision scores is calculated. All differences are summed and concatenated into the first decision score vector to obtain the second decision score vector. The second decision score vector serves as the decision score vector corresponding to the N+1 categories of radiation sources, and its dimension is (N+1). 1;

[0086] From the N+1 decision scores in the second decision score vector, select the maximum value, obtain the row number corresponding to the maximum value, and use the row number corresponding to the maximum value as the original I / Q signal of the preset length of the radiation source to be detected. For the corresponding radiation source category, if the (N+1)th element in the second decision score vector is the maximum value, then the original I / Q signal of the radiation source to be detected is determined to be of a preset length. From an unknown class.

[0087] This method for RF fingerprint causal invariance modeling and open set identification of end-side radiation sources combines preprocessing techniques that reveal the inherent nonlinear fingerprint of the transmitter with a self-supervised training framework that learns domain-invariant causal representations. This invention can effectively generalize to target domains with significant differences in channel characteristics and noise levels under the condition of only single source domain training data, thus solving the pain point of poor generalization ability of traditional deep learning models.

[0088] By calibrating the classifier output with the extreme value theory based on the Weibull distribution, a statistically significant decision boundary is constructed for known radiation source categories. This enables the model to not only accurately classify known radiation sources, but also effectively identify and reject signals from unknown radiation sources, significantly reducing the false alarm rate of unknown radiation source intrusion detection and improving reliability in real open environments.

[0089] This method can significantly reduce the number of model parameters, computational complexity, and inference latency while maintaining recognition accuracy. This enables the high-precision SEI algorithm to be successfully deployed on resource-constrained edge transmitters, achieving real-time or near-real-time recognition tasks.

[0090] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0091] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0092] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for radio frequency fingerprint causality invariance modeling and end-side radiation source open set identification, characterized in that: The radio frequency fingerprint causal invariance modeling and end-side radiation source open set identification method comprises: A teacher model is constructed, and the teacher model is trained by using original I / Q signals of N radiation sources of known categories to obtain a trained teacher model; The trained teacher model is subjected to lightweight optimization to obtain a student model; The trained teacher model is used to guide the student model training based on a knowledge distillation loss to obtain a lightweight model; Original I / Q signals of a preset length from a to-be-detected radiation source are obtained, amplitude spectrum and nonlinear features are obtained after preprocessing, and the amplitude spectrum and the nonlinear features are input into the lightweight model to obtain an initial decision score vector of the to-be-detected radiation source, the initial decision score vector of the to-be-detected radiation source is corrected based on N radiation sources of known categories, and finally the category of the to-be-detected radiation source is obtained.

2. The method of claim 1, wherein the method further comprises: The trained teacher model comprises an encoder, a full connection layer and a classifier connected in sequence, and the classifier is a multilayer perceptron, and the encoder is a ResNet18 model.

3. The method of claim 2, wherein: The training process of the teacher model comprises: N original I / Q signals of known categories of radiation sources are collected, each original I / Q signal is sliced, one slice is taken as a training sample, the teacher model is trained, and one batch comprises multiple samples; First, the training sample is preprocessed to obtain amplitude spectrum and nonlinear features, and the amplitude spectrum and the nonlinear features are respectively subjected to random disturbance enhancement; The amplitude spectrum of each sample after random disturbance enhancement and the corresponding nonlinear feature are input into an encoder and then a full connection layer to obtain two output features, and based on Barlow Twins loss Self-supervised learning is performed on the two output features; Two output features are spliced to obtain spliced features, a preset pixel in the spliced features is subjected to first masking to obtain a first mask image, pixels other than the preset pixel in the spliced features are subjected to second masking to obtain a second mask image, and the spliced features are multiplied by the first mask image and the second mask image respectively to obtain first features and second features respectively; The first features and the second features are input into two classifiers respectively; For each batch in the training phase, the cross-entropy loss of the output of the first feature and the second feature passing through the classifier is calculated respectively and the loss of Barlow Twins used in conjunction with self-supervised learning The total loss of the training phase can be obtained: where is the weight parameter; and based on the total loss training the teacher model until the total loss converges to a minimum.

4. The method of claim 3, wherein: The preprocessing process is that the original I / Q signal or the slice of the to-be-detected radiation source of a preset length is subjected to frequency spectrum transformation to obtain amplitude spectrum, and the amplitude spectrum is subjected to nonlinear intensity transformation to obtain nonlinear features; wherein the spectrum transforming process is: performing a fast Fourier transform on the raw I / Q signal or slice of pre-set length of the radiation source to be detected, so that the input is converted from time domain to frequency domain, and extracting the amplitudes of the spectrum as the amplitude spectrum ; The process of the nonlinear intensity transformation is: performing a cubic Bezier curve-based remapping on the amplitude spectrum through a preset nonlinear mapping function to obtain a nonlinear feature .

5. The method of claim 1, wherein: The trained teacher model is subjected to lightweight optimization to obtain a student model, comprising: The encoder in the trained teacher model is designed as a lightweight convolutional neural network to obtain the student model.

6. The method of claim 1, wherein: The knowledge distillation loss The calculation formula is as follows: ； wherein, and are outputs of the trained teacher model and student model, respectively, and are output features of the encoders in the trained teacher model and student model, respectively, is a preset temperature, is a softmax function, , and are weight parameters, is a cross-entropy loss used for training the student model, is a KL divergence function.

7. The method of claim 3, wherein: The radio frequency fingerprint causal invariance modeling and end-side radiation source open set identification method further comprises: For each category of radiation source in the training set, the activation vectors corresponding to all training samples in the current category of radiation source are obtained, and the mean value of all activation vectors is calculated to obtain the average activation vector of the current category of radiation source, and the output of the penultimate layer of each model is an activation vector; The distances from the activation vectors corresponding to all correctly classified training samples of the current category of radiation source to the average activation vector are calculated, and the maximum m distances are selected as extreme distances, and the Weibull distribution of the m extreme distances is fitted to obtain the Weibull distribution cumulative distribution function of the current category of radiation source; This leads to the Weibull cumulative distribution function of the m extreme distances corresponding to the N categories of radiation sources in the training set. , and They are the nth in N categories The scale and shape parameters of each type of radiation source are obtained by fitting the m extreme distances corresponding to each type of radiation source, and all of them are greater than zero. It is a random variable.

8. The method of claim 1, wherein: In the lightweight model, the amplitude spectrum and the nonlinear feature are together input into an encoder and a full connection layer to obtain two output features, the two output features are spliced to obtain a spliced feature, and the spliced feature is input into a classifier to obtain an output of the lightweight model.

9. The method of claim 7, wherein: The method for correcting the initial decision score vector of the to-be-detected radiation source based on the N known categories of radiation sources and finally obtaining the category of the to-be-detected radiation source comprises the steps of: Obtaining original i / q signals of a preset length of a radiation source to be detected The corresponding activation vector is called the first activation vector. The distance between the first activation vector and each average activation vector of the known N categories of radiation sources is calculated respectively to form a first distance set. Each first distance in the first distance set is substituted into the Weibull distribution cumulative distribution function of each category of radiation source to obtain the probability that the original I / Q signals of the preset length of the radiation source to be detected belong to the corresponding category of radiation source in the N categories, which is as follows: ； in, The raw I / Q signal of the preset length of the radiation source to be detected Belonging to the N categories The probability of each type of radiation source for First activation vector With the N categories, the first The first distance between the average activation vectors of each category of radiation sources and They are the nth in N categories The scale and shape parameters of the cumulative distribution function of the Weibull distribution for each category of radiation source are both greater than zero; Select from the first distance set The smallest first distance is used to form the second distance set. The radiation sources corresponding to each second distance in the second distance set are recorded and called the class to be corrected. The classes to be corrected are sorted in ascending order according to their corresponding second distances, and the calculation is performed. The correction weights for each class to be corrected are given by the following formula: ； wherein, , is a correction weight of a radiation source of a category in the original I / Q signal of a preset length of the radiation source to be detected belongs to a probability that a radiation source of a category in the original I / Q signal of a preset length of the radiation source to be detected does not belong to a probability that a radiation source of a category in the is a ranking serial number of a radiation source of a category in the The dimension of the initial decision score vector is N 1, that is, contains N elements corresponding to N known categories of radiation sources one by one, each element is called a decision score, and The correction weight of the class to be corrected is multiplied by the element corresponding to the radiation source of the corresponding class in the initial decision score vector, so that The decision score of the element is corrected, and the remaining elements remain unchanged, obtaining the original I / Q signal of the preset length of the to-be-detected radiation source The corresponding decision score vector is called the first decision score vector. The modified decision scores in the first decision score vector are compared with the unmodified decision scores, the modified decision scores that are smaller than the unmodified decision scores are screened out, the difference between the screened out modified decision scores and the unmodified decision scores is calculated, all the differences are accumulated and spliced into the first decision score vector to obtain a second decision score vector, and the second decision score vector is used as a decision score vector corresponding to the N+1 categories of radiation sources, and the dimension of the second decision score vector is (N+1) 1; Select a maximum value from N+1 decision scores in the second decision score vector, obtain a row number corresponding to the maximum value, and take the row number corresponding to the maximum value as the original I / Q signal of the preset length of the to-be-detected radiation source Corresponding to the radiation source category, if the N+1th element in the second decision score vector is the maximum value, it is determined that the original I / Q signal of the preset length of the to-be-detected radiation source From the unknown class.

Images