A small sample radio frequency fingerprint recognition method

Abstract

The present invention relates to a small-sample radio frequency fingerprint recognition method in the field of artificial intelligence. This method constructs a target recognition model using several labeled samples from the radiation source to be identified and other common samples, and uses the target recognition model to complete the target small-sample radio frequency fingerprint recognition task. The method is characterized in that the target recognition model construction process includes the following steps: pre-training the target recognition model using a large-scale training paradigm based on a base class dataset, initializing a soft label generation model using the pre-trained target recognition model; extracting meta-tasks using the base class dataset that are identical in format to the target small-sample radio frequency fingerprint recognition task, including a meta-support set and a meta-query set; training the target recognition model using the meta-tasks, and fine-tuning the target recognition model on labeled samples of the target small-sample radio frequency fingerprint recognition task to obtain the final target recognition model. The target recognition model can accurately perform classification predictions when faced with new, unseen radiation sources.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a small-sample radio frequency fingerprint recognition method. Background Art

[0002] The basic task of small-sample RF fingerprinting technology is to extract the RF fingerprint characteristics of a small number of observed signals to identify the specific emitter target of the signal, thereby obtaining enemy tactical electronic intelligence. Small-sample RF fingerprinting technology has a wide range of applications in many fields, including electronic intelligence reconnaissance, electronic far-reaching, cognitive radio, and wireless network security.

[0003] Existing RF fingerprint recognition technology usually first uses short-time Fourier transform to process the radiation source signal to obtain a time-frequency diagram, and then uses deep learning models such as convolutional neural networks to extract the RF fingerprint features of the image, and inputs the output features into the subsequent classification layer for recognition.

[0004] The main problem with existing RF fingerprinting technology is that, in RF fingerprinting tasks, only very few labeled samples can be collected for each radiation source category, which is insufficient for model training. To address this issue, existing methods typically use a large-scale training paradigm to pre-train the model on a base class dataset from a similar domain. The model is then fine-tuned on labeled samples from the target small-sample RF fingerprinting task to adapt to the data distribution of the target task. However, the task format of the target small-sample RF fingerprinting task differs significantly from that used during pre-training. Pre-trained models are often unable to adapt to the small-sample task format, resulting in reduced model generalization. Summary of the Invention

[0005] To address the above issues, the present invention provides a small-sample radio frequency fingerprint recognition method. This method constructs a target recognition model using labeled samples from the radiation source to be identified and labeled samples of other common objects. This target recognition model is then used to complete the small-sample radio frequency fingerprint recognition task. The target recognition model construction process includes the following steps:

[0006] Step 1: Prepare the base class dataset and the original support set of the target task Target task original query set And the original support set With the original queryset The signal samples in the sample are preprocessed to obtain the target task support set D in the form of time-frequency graph. S and the target task query set D Q ;

[0007] Step 2: Pre-train the target recognition model using a large-scale training paradigm based on the base class dataset, and use the pre-trained target recognition model to initialize the soft label generation model;

[0008] Step 3: Use the base class dataset to extract the meta-task T with the same form as the target small sample RF fingerprint recognition task Meta , meta-task T Meta Including meta-support set D MetaS With the meta-query set D MetaQ ;

[0009] Step 4: Generate meta-query set D using soft label generation model MetaQ Category soft labels of samples in ;

[0010] Step 5: Use the target recognition model to calculate the meta-query set D MetaQ The predicted category probability distribution of samples in ;

[0011] Step 6: Calculate the meta-query set D MetaQ The predicted category probability distribution of the sample and the loss value of its category soft label are used to train the target recognition model, thus obtaining the target recognition model that completes meta-learning.

[0012] Step 7: Use the meta-learned target recognition model to calculate the target task support set D S The predicted category probability distribution of samples in ;

[0013] Step 8: Calculate the target task support set D S The cross entropy of the predicted category probability distribution of the sample and its label is used to complete the meta-learning target recognition model through back propagation training based on the cross entropy to obtain the fine-tuned target recognition model.

[0014] Furthermore, the base class dataset includes: labeled image samples of other objects, the number of labeled image samples of each object is greater than a set threshold, and the threshold is greater than the target task support set D S 10 times the number of all samples in .

[0015] Furthermore, the target task support set D S Includes: a number of short-time Fourier transform time-frequency diagrams including labeled signal samples generated by all radiation sources to be identified.

[0016] Furthermore, the target task query set D Q Includes: a number of short-time Fourier transform time-frequency diagrams of unlabeled signal samples generated by all radiation sources to be identified.

[0017] Furthermore, the pre-training of the target recognition model using a large-scale training paradigm described in step 2 specifically refers to the pre-training of the target recognition model using one of the DINO method, the small batch gradient descent method or the DEIT method.

[0018] Furthermore, step 4 specifically includes: using the soft label generation model to extract the meta-support set D Metas With the meta-query set D MetaQ The radio frequency fingerprint features of the samples in the calculation element support set D Metas The soft label prototypes of the categories corresponding to all samples in Use the prototype network classifier to output the meta-query set D MetaQ The predicted category probability distribution of samples in the dataset is obtained by using the meta-query set D MetaQ The predicted category probability distribution of the sample in the regularized sample comes with the label, generating the meta query set D MetaQ The soft labels of the samples in .

[0019] Furthermore, step five specifically includes extracting the meta-support set D using the target recognition model MetaS With the meta-query set D MetaQ The radio frequency fingerprint features of the samples in the calculation element support set D MetaS The category recognition prototypes corresponding to all samples in Identifying prototypes based on categories Use the prototype network classifier to output the meta-query set D MetaQ The predicted class probability distribution of samples in .

[0020] Furthermore, step seven specifically includes adding visual cues to the input space of the meta-learned target recognition model and calculating the target task support set D S The individual prototypes c' of each radiation source corresponding to all samples in i , calculate the target task support set D S The RF fingerprint characteristics of each sample and the individual prototype c' of each radiation source i The distance between the target task support set D is obtained. S The predicted class probability distribution of samples in .

[0021] Furthermore, adding visual cues in the input space of the meta-learned target recognition model specifically refers to adding visual cues in the input space of a partially specified Transformer network layer of the meta-learned target recognition model.

[0022] Furthermore, adding visual cues to the input space of the target recognition model that completes meta-learning specifically refers to adding visual cues to the input space of all Transformer network layers of the target recognition model that completes meta-learning.

[0023] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0024] The target recognition model provided in this application introduces for the first time a regularization strategy for calculating sample soft labels using a soft label generation model in the meta-learning stage, solving the problem of inconsistent task formats in the pre-training and fine-tuning stages in existing small-sample radio frequency fingerprint recognition methods, and avoiding the memory overfitting problem existing in existing methods during the meta-learning process, thereby ultimately improving the generalization ability of the target recognition model. The target recognition model can correctly perform classification predictions when faced with new, unseen radiation sources. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Flowchart of a small sample radio frequency fingerprint recognition method provided by an embodiment of the present invention;

[0026] Figure 2 A flowchart of meta-learning of an object recognition model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. Before describing in detail the technical solutions of each embodiment of the present invention, the nouns and terms involved are explained. In this specification, components with the same name or the same number represent similar or identical structures and are for illustrative purposes only.

[0028] The data required by this invention include base class data set, target task original support set Target task original query set Target task support set D S , target task query set D Q The base class dataset includes: labeled image samples of other objects, the number of labeled image samples of each object is greater than the set threshold, and the threshold is greater than the target task support set D S The number of samples in the dataset is 10 times that of all samples in the dataset. A commonly used large-scale image dataset can be selected as the base dataset, such as the Imagenet dataset. The original support set of the target task Includes: a number of labeled signal samples generated by all radiation sources to be identified. The original query set of the target task Includes: a number of unlabeled signal samples generated by all radiation sources to be identified. Target task support set D S is the original support set of the target task After preprocessing, the target task query set D is obtained, including: a number of short-time Fourier transform time-frequency graphs of labeled signal samples generated by all radiation sources to be identified. Q The original query set for the target task After preprocessing, the obtained data include: a number of short-time Fourier transform time-frequency graphs of unlabeled signal samples generated by all radiation sources to be identified.

[0029] The target recognition model is pre-trained using a large-scale training paradigm in the base class dataset. The soft label generation model is initialized using the pre-trained target recognition model. All parameters of the soft label generation model are frozen and cannot be updated, while the target recognition model needs to update parameters in the subsequent meta-learning phase. The base class dataset is used to extract the meta-task T with the same form as the target small sample RF fingerprint recognition task. Meta (including meta-support set D MetaS With the meta-query set D MetaQ ), the same form refers to the meta-task T Meta The number of object categories is the same as that of the target small sample RF fingerprint recognition task, and the meta-support set D MetaS and the target task support set D S The number of samples in each category is the same, and the meta-query set D MetaQ and the target task query set D Q The number of samples is the same; the soft label generation model is used to extract the meta-support set D MetaS With the meta-query set D MetaQ The radio frequency fingerprint features of the samples in the calculation element support set D MetaS The soft label prototypes of the categories corresponding to all samples in Use the prototype network classifier to output the meta-query set D MetaQ The predicted category probability distribution of samples in the dataset is obtained by using the meta-query set D MetaQ The predicted category probability distribution of the sample in the regularized sample comes with the label, generating the meta query set D MetaQ The soft label of the sample category; the meta-support set D is extracted using the target recognition model MetaS With the meta-query set D MetaQ The radio frequency fingerprint features of the samples in the calculation element support set D MetaS The category recognition prototypes corresponding to all samples in Identifying prototypes based on categories Use the prototype network classifier to output the meta-query set D MetaQ The predicted category probability distribution of samples in; calculate the meta-query set D MetaQ The predicted category probability distribution of the samples and the loss value of their soft labels are used to train the target recognition model through back propagation according to the loss value; after the preset target recognition model meta-learning completion conditions are met, the target recognition model that has completed meta-learning is obtained.

[0030] The present invention supports the target task set D S The fine-tuning of the target recognition model is completed by using visual cues to make the target recognition model more accurate in the target task support set D. S The data distribution of the target domain is learned from the samples of the target domain, and visual cues are added to the input space of the target recognition model after completing the meta-learning. SThe individual prototype c' of each radiation source is calculated from the sample i , and then calculate the target task support set D S The RF fingerprint characteristics of each sample and the individual prototype c' of each radiation source i The distance between the target task support set D is obtained. S The predicted category probability distribution of the samples in the target task support set D is calculated. S The cross entropy of the predicted category probability distribution of the sample and its soft label is used to train the target recognition model through back propagation according to the cross entropy to obtain the fine-tuned target recognition model.

[0031] The present invention is based on the target task query set D O Deploy the fine-tuned target recognition model in the target task query set D Q The classification work of the samples in the target task query set D is performed using the fine-tuned target recognition model. Q The radio frequency fingerprint characteristics of the sample are calculated and compared with the individual prototype c' of each radiation source i The distance between the two samples is used to obtain the predicted category probability distribution of the sample to be identified.

[0032] The present invention is explained below with reference to a specific embodiment.

[0033] The small sample RF fingerprint recognition task in this embodiment is the radio station individual identification task. There is a batch of newly collected signals from unknown radio stations, and there are six radio stations transmitting these signals. This task requires identifying the individual radio station from which each signal comes. Because it is difficult to obtain labeled signal samples of individual radio stations, the number of labeled signal samples available for training for each radio station is relatively small. In the radio station individual identification task of this embodiment, each radio station has 5 labeled signal samples available for training. These labeled signal samples constitute the original support set of the target task. There are also several signal samples to be classified to form the original query set of the target task For the original support set With the original queryset The signal samples in are preprocessed to obtain the target task support set D S and the target task query set D Q In this embodiment, the preprocessing method is the short-time Fourier transform method. To solve the problem that the model cannot learn due to a small sample size, an additional base class dataset consisting of a rich variety of common objects in life with sufficient image samples is provided for the target recognition model to learn.

[0034] Figure 1 The main steps of the small sample radio frequency fingerprint recognition method are shown, including:

[0035] 1. Prepare the data

[0036] The embodiment of the present invention uses the classic short-time Fourier transform method to transform the original support set of the target task and the original query set of the target task The signal samples in the image are preprocessed to obtain a time-frequency plot of the samples. The short-time Fourier transform process is as follows: before the signal is Fourier transformed, it is multiplied by a finite time window function. It is assumed that the non-stationary signal is stationary within the short time interval of the analysis window. By moving the window function on the time axis, the signal is analyzed segment by segment to obtain a set of local "spectra" of the signal.

[0037] The calculation formula is as follows:

[0038]

[0039] Wherein, x(n) is the time series signal to be transformed, w(n) is the window function. In this embodiment, Hamming window is used. R is the number of signal sampling points of the adjacent signal window shift. m (ω) is the time-frequency diagram obtained after the signal samples are preprocessed, m represents time (indicating the m-th frame of short-time signal), ω represents frequency, and j represents the imaginary unit.

[0040] 2. Pre-training of target recognition model

[0041] This paper uses a visual Transformer as the model backbone of the target recognition model and pre-trains the target recognition model using the DINO method, a large-scale self-supervised learning paradigm. Pre-training the target recognition model on a large-scale base class dataset enables the target recognition model to obtain a feature extractor capable of extracting rich, universal features. After pre-training the target recognition model, only the feature extractor portion of the target recognition model is retained, and the classification head of the target recognition model is removed for subsequent target recognition tasks.

[0042] The DINO method can be replaced by other large-scale pre-training methods, such as the mini-batch gradient descent method and the DEIT method.

[0043] 3. Meta-learning of object recognition models

[0044] The meta-learning steps of the target recognition model include: sampling and constructing a meta-task T with the same specifications as the small sample RF fingerprint recognition task in the base class dataset Meta , meta-task T Meta The meta-support set D obtained by sampling MetaS With the meta-query set D MetaQ Composition; Use the pre-trained target recognition model to create a meta-query set D MetaQ The soft labels of the samples are obtained; the target recognition model is optimized using the soft labels.

[0045] In the base class dataset, sample and construct a meta-task T with the same specifications as the small sample RF fingerprint recognition task. Meta , aims to enable the target recognition model to further learn in the form of small sample tasks, thereby having stronger generalization ability for small sample RF fingerprint recognition tasks. In this embodiment, the small sample RF fingerprint recognition task is in the form of 6-class-5 samples, that is, 6 radio station individuals, each radio station individual has 5 labeled signal samples for training, a total of 30 labeled signal samples. The time-frequency graph samples obtained after preprocessing constitute the target task support set D S ; In addition, several samples to be classified from these six radio stations are preprocessed to obtain time-frequency graph samples to form the target task query set D Q The target recognition model needs to be in the target task support set D S The specific distribution of radio station data in this task is learned from the 30 samples of the target task query set D Q In order to adapt the target recognition model to the target small sample radio frequency fingerprint recognition task form, the meta-task T extracted in this embodiment is Meta The same 6-class-5 sample format is used, with a total of 30 samples as meta-task T Meta The meta-support set D in MetaS , and extract a number of test samples (e.g., 15 per category, 90 in total) from these 6 categories as the meta-task T Meta The meta-query set D in MetaQ .

[0046] Next, the soft label generation model is used to generate the meta-query set D MetaQ The category soft label of the sample, the specific process is as follows Figure 2 As shown, the soft label generation model is used to extract the meta-support set D MetaS With the meta-query set D MetaQ RF fingerprint features of samples in; calculate the element support set D MetaS The soft label prototypes of the categories corresponding to all samples in Use the prototype network classifier to output the meta-query set D MetaQ The predicted category probability distribution of samples in ; using the meta-query set D MetaQ The predicted category probability distribution of the sample in the regularized sample comes with the label, generating the meta query set D MetaQ The soft labels of the samples in .

[0047] Extracting meta-support set D using soft label generation model MetaS With the meta-query set D MetaQ The expression of the radio frequency fingerprint feature of the sample in is:

[0048]

[0049]

[0050] Among them, f S is a soft label generation model, It is the meta-support set D MetaS The i′th sample of the i-th class in is the meta-query set D MetaQ The kth sample of for Generate model f from soft labels S The extracted radio frequency fingerprint features, for Generate model f from soft labels S The extracted radio frequency fingerprint features;

[0051] Computational element support set D MetaS The soft label prototypes of the categories corresponding to all samples in The expression is:

[0052]

[0053] in, is the soft label prototype of the i-th category.

[0054] Use the prototype network classifier to output the meta-query set D MetaQ The predicted category probability distribution of samples in:

[0055]

[0056] in, Representative meta-query set D MetaQ The predicted category of the kth sample in , Representative meta-query set D MetaQ The conditional distribution of the predicted category probability obtained by the soft label generation model for the k-th sample in represents the category soft label prototype of the k′th category, exp represents the exponential operation, d(·,·) is the metric function used to calculate the similarity, and in this embodiment, the cosine distance or Euclidean distance is used as the metric function.

[0057] Using the predicted category probability distribution p k，s Regularization samples Own label Get the sample Category soft label

[0058]

[0059] Wherein, t is the temperature factor, and in this embodiment, the value of t is 0.2.

[0060] The target recognition model is optimized by using category soft labels, aiming to introduce exogenous knowledge (stored in the meta-query set D MetaQ The soft labels of the samples in the dataset can effectively resist the problem of memory overfitting. The specific process is: using the target recognition model to extract the meta-support set D MetaS With the meta-query set D MetaQ RF fingerprint features of samples in; calculate the element support set D MetaS The category recognition prototypes corresponding to all samples in Use the prototype network classifier to output the meta-query set D MetaQ The predicted category probability distribution of the samples in is used to calculate the loss value output by the current target recognition model using cross entropy, and the target recognition model is trained based on the loss value.

[0061] Extract meta-support set D using target recognition model MetaS With the meta-query set D MetaQ Sample RF fingerprint features:

[0062]

[0063]

[0064] Among them, f R It is a target recognition model that continuously learns and updates after pre-training. for By the target recognition model f R The extracted radio frequency fingerprint features, for By the target recognition model f R The extracted radio frequency fingerprint features.

[0065] Computational element support set D MetaS The category recognition prototypes corresponding to all samples in

[0066]

[0067] Use the prototype network classifier to output the meta-query set D MetaQ The predicted category probability distribution of samples:

[0068]

[0069] in, Representative meta-query set D MetaQ The conditional distribution of the predicted category probability obtained by the target recognition model for the k-th sample in The category recognition prototype representing the k′th category;

[0070] Use cross entropy to calculate the loss value output by the target recognition model:

[0071]

[0072] in, Representative meta-query set D MetaQ The soft label of the category of the k-th sample in, Represents the calculated loss value, which is used to measure the similarity between the soft label and the target recognition model output predicted category probability distribution.

[0073] After reaching the preset target recognition model meta-learning completion condition, the target recognition model that has completed meta-learning is obtained. In this embodiment, the meta-learning completion condition is that the training reaches 1000 times.

[0074] 4. Fine-tuning the object recognition model

[0075] By adding visual cues to the input space of the meta-learned target recognition model, the backbone parameters of the target recognition model can be frozen in subsequent fine-tuning. Only the parameters within the visual cues can be fine-tuned to complete the target recognition model's learning of the data distribution of the target small sample RF fingerprint recognition task. This can significantly reduce the parameters that need to be fine-tuned to resist overfitting, while ensuring that the backbone parameters of the model do not change to prevent the model from forgetting prior knowledge during the fine-tuning stage. The specific expression for adding visual cues is:

[0076] [x l ,·,E l ]=L l ([x l-1 , P l-1 , E n-1 ]), l=1,2,...,N l ;

[0077]

[0078] where x l is the classification token inherent in the existing Transformer model framework, E i is the partial output of each layer of the model (where E0 is the serialized representation of the input image), N l is the total number of network layers, P l The visual cue blocks added by the present invention have the same specifications in each layer, but different initialization parameters. The specific specifications are [number of cue blocks, dimension of cue blocks]. In this embodiment, the number of cue blocks is 10, and the dimension of cue blocks is 384. “·” is the partial output of the previous layer, and its size is the same as that of the P added in the input space of the next layer. l Same, will be P in the input of the next layer l Instead. Ll y represents the lth Transformer network layer with frozen parameters. head represents the classification head. In this embodiment, the prototype network classifier is used to replace the original classification head in the Transformer model framework. y represents the final output of the model.

[0079] Visual cues can be added to the input space of some or all Transformer network layers.

[0080] Afterwards, the target recognition model and the target task support set D that have completed meta-learning are used S Recalculate the individual prototype c′ of each radio station using the samples in i , using the recalculated individual prototype c' of each station i Recalculate the target task support set D S The RF fingerprint characteristics of each sample and the individual prototype c' of each radio station i The distance between the target task support set D is obtained. S The predicted category probability distribution of the samples in the dataset is obtained, and the cross entropy function is used to recalculate the loss value between the model output and the true label of the sample, complete the back propagation, and obtain the fine-tuned target recognition model.

[0081] 5. Solve the task of identifying individual target radio stations.

[0082] Use the fine-tuned target recognition model to complete the target task query set D Q The specific method is to use the fine-tuned target recognition model and the target task support set D S Recalculate the individual prototype c' of each radio station in the sample i , use the fine-tuned model to recalculate the target task query set D Q The individual prototypes c' of each sample and each radio station i The distance between and is used to obtain the predicted category probability distribution of the sample to be classified.

[0083] It should be noted that the small-sample RF fingerprint recognition method proposed in this invention processes images as the data type and solves a classification problem. The concepts of this invention can also be applied to other tasks with similar data types and tasks, such as small-sample image classification, small-sample face recognition, small-sample image retrieval, and small-sample remote sensing image classification.

[0084] The above-described embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A small sample radio frequency fingerprint recognition method, which uses a number of labeled samples from the radiation source to be identified and labeled samples of other common objects to build a target recognition model, and uses the target recognition model to complete the target small sample radio frequency fingerprint recognition task, characterized by: The process of constructing the target recognition model includes the following steps: Step 1: Prepare the base class dataset and the original support set of the target task Target task original query set Original support set With the original queryset The signal samples in are preprocessed to obtain the target task support set D S and the target task query set D Q ; Step 2: Pre-train the target recognition model using a large-scale training paradigm based on the base class dataset, and use the pre-trained target recognition model to initialize the soft label generation model; Step 3: Use the base class dataset to extract the meta-task T with the same form as the target small sample RF fingerprint recognition task Meta , meta-task T Meta Includes meta-support set D MetaS With the meta-query set D MetaQ ; Step 4: Generate meta-query set D using soft label generation model MetaQ Soft labels of the categories of the samples in ; Step 5: Use the target recognition model to calculate the meta-query set D MetaQ The predicted category probability distribution of samples in ; Step 6: Calculate the meta-query set D MetaQ The predicted category probability distribution of the sample and the loss value of its category soft label are used to train the target recognition model, thus obtaining the target recognition model that completes meta-learning. Step 7: Use the meta-learned target recognition model to calculate the target task support set D S The predicted category probability distribution of samples in ; Step 8: Calculate the target task support set D S The cross entropy of the predicted category probability distribution of the sample and its label is used to complete the meta-learning target recognition model through back propagation training based on the cross entropy to obtain the fine-tuned target recognition model.

2. The small sample radio frequency fingerprint recognition method according to claim 1, characterized in that: The base class data set includes: labeled image samples of other objects, the number of labeled image samples of each object is greater than a set threshold, and the threshold is greater than the target task support set D S 10 times the number of all samples in .

3. The small sample radio frequency fingerprint recognition method according to claim 1, characterized in that: The target task support set D S Includes: a number of short-time Fourier transform time-frequency diagrams including labeled signal samples generated by all radiation sources to be identified.

4. The small sample radio frequency fingerprint recognition method according to claim 1, characterized in that: The target task query set D Q Includes: a number of short-time Fourier transform time-frequency diagrams of unlabeled signal samples generated by all radiation sources to be identified.

5. The small sample radio frequency fingerprint recognition method according to claim 1, characterized in that: The pre-training of the target recognition model using a large-scale training paradigm as described in step 2 specifically refers to the pre-training of the target recognition model using one of the DINO method, the small batch gradient descent method, or the DEIT method.

6. The small sample radio frequency fingerprint recognition method according to claim 5, characterized in that: Step 4 specifically includes: using the soft label generation model to extract the meta-support set D MetaS With the meta-query set D MetaQ The radio frequency fingerprint features of the samples in the calculation element support set D MetaS The soft label prototypes of the categories corresponding to all samples in Use the prototype network classifier to output the meta-query set D MetaQ The predicted category probability distribution of samples in the dataset is obtained by using the meta-query set D MetaQ The predicted category probability distribution of the sample in the regularized sample comes with the label, generating the meta query set D MetaQ The soft labels of the samples in .

7. The small sample radio frequency fingerprint recognition method according to claim 5 or 6, characterized in that: Step 5 specifically includes extracting the meta-support set D using the target recognition model MetaS With the meta-query set D MetaQ The radio frequency fingerprint features of the samples in the calculation element support set D MetaS The category recognition prototypes corresponding to all samples in Identifying prototypes based on categories Use the prototype network classifier to output the meta-query set D MetaQ The predicted class probability distribution of samples in .

8. The small sample radio frequency fingerprint recognition method according to claim 5 or 6, characterized in that: Step 7 specifically includes adding visual cues to the input space of the meta-learned target recognition model and calculating the target task support set D S The individual prototypes c' of each radiation source corresponding to all samples in i , calculate the target task support set D S The RF fingerprint characteristics of each sample and the individual prototype c' of each radiation source i The distance between the target task support set D is obtained. S The predicted class probability distribution of samples in .

9. The small sample radio frequency fingerprint recognition method according to claim 8, characterized in that: Adding visual cues to the input space of the meta-learned target recognition model specifically refers to adding visual cues to the input space of a partially specified Transformer network layer of the meta-learned target recognition model.

10. The small sample radio frequency fingerprint recognition method according to claim 8, characterized in that: Adding visual cues to the input space of the meta-learned target recognition model specifically refers to adding visual cues to the input space of all Transformer network layers of the meta-learned target recognition model.

Images