Open set radio frequency fingerprint recognition method based on layered autoencoder

Abstract

The present invention provides an open set radio frequency fingerprint recognition method based on a hierarchical autoencoder, comprising the following steps: collecting radio frequency data, converting the radio frequency data into a time-frequency diagram; classifying the time-frequency diagram to obtain training samples and test samples; constructing a radio frequency fingerprint recognition network model, and using the training samples to train the radio frequency fingerprint recognition network model; and building an open set radio frequency fingerprint recognition network model based on the trained radio frequency fingerprint recognition network model and the test samples for open set radio frequency fingerprint recognition. The present invention utilizes metric learning to reduce the distance between the same category while increasing the distance between different categories, and proposes an open set radio frequency fingerprint intelligent and precise recognition model based on a hierarchical autoencoder, which not only maintains the recognition accuracy of known samples, but also improves the accuracy of open set radio frequency fingerprint recognition.

Description

Technical Field

[0001] The invention belongs to the field of communication technology, and in particular relates to an open set radio frequency fingerprint recognition method based on a layered autoencoder. Background Art

[0002] With the advent of the era of intelligent communications, the security of wireless communications has become increasingly important. However, wireless communication security faces many severe challenges, especially being vulnerable to malicious attacks such as eavesdropping and spoofing. RF fingerprints, as the inherent properties of electromagnetic signals in transmitting devices, are difficult to tamper with and can be used to determine the identity of the transmitting device. Therefore, RF fingerprint recognition is a method that uses the inherent properties of the transmitting device to infer the intrinsic correlation between the RF device and the communication signal containing the "difference" and identify the transmitting device based on the signal it transmits. Traditional methods of manually extracting features rely on experts and domain knowledge, and it is difficult to extract the optimal features.

[0003] Compared with the traditional method of manually extracting features, the use of deep learning methods can extract more advanced features to improve the performance of RF fingerprint recognition. G.Shen, J.Zhang, A.Marshall, et al., in their paper "RadioFrequency Fingerprint Identification for LoRa Using Spectrogram and CNN" (IEEEINFOCOM 2021, pp.1-10, 2021), used the relatively stable average value of the instantaneous carrier frequency offset to propose a hybrid classifier that adjusts the output of CNN according to the instantaneous carrier frequency offset. Combining the two features of three-dimensional Hilbert spectrum and dual spectrum, a multi-feature fusion RF fingerprint recognition model based on deep learning is proposed, which improves the accuracy of model recognition. M.Liu, C.Liu, Y.Chen et al. proposed a radio frequency fingerprint recognition method based on a deformable convolutional network in their paper "Radio Frequency Fingerprint Collaborative Intelligent Blind Identification for Green Radios" (IEEE Transactions on Green Communications and Networking, vol.7, no.2, pp.940-949, 2023), which makes the convolution operation more inclined to the useful information content in the feature map with higher energy and ignores some background noise information. The patent application "A radio frequency fingerprint recognition method for the Internet of Things suitable for low signal-to-noise ratio environment" with publication number CN112926477A proposed by the University of Electronic Science and Technology of China discloses a radio frequency fingerprint recognition method for the Internet of Things suitable for low signal-to-noise ratio environment.

[0004] In summary, deep learning methods can effectively perform feature analysis and identification on RF devices, breaking through the limitations of traditional algorithms on features. However, most existing methods are implemented under closed-set RF fingerprinting, which assumes that all test samples have the same distribution as training samples. In the case of a closed set, the presence of unknown open devices in the communication area will bring the risk of misidentification. At the same time, in the actual open-set electromagnetic spectrum environment, it is impossible for all classes to be covered by the training set. Therefore, RF fingerprinting of unknown open samples is an urgent problem to be solved. Summary of the invention

[0005] The present invention aims at the fact that most of the existing radio frequency fingerprint recognition technologies are implemented in closed set situations, and in closed set situations, there are difficulties such as unknown devices are difficult to correctly identify and it is impossible for all categories to be covered by the training set. The present invention provides an open set radio frequency fingerprint recognition method based on a hierarchical autoencoder. In order to improve the robustness of the open electromagnetic spectrum environment, the present invention uses metric learning to reduce the distance between the same category while increasing the distance between different categories, and proposes an open set radio frequency fingerprint intelligent and accurate recognition model based on a hierarchical autoencoder, which not only maintains the recognition accuracy of known samples, but also improves the accuracy of open set radio frequency fingerprint recognition.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] The open set radio frequency fingerprint recognition method based on layered autoencoder is characterized by comprising the following steps:

[0008] Collect radio frequency data and convert the radio frequency data into a time-frequency diagram;

[0009] Classify the time-frequency graph to obtain training samples and test samples;

[0010] Construct a radio frequency fingerprint recognition network model, and use training samples to train the radio frequency fingerprint recognition network model;

[0011] According to the trained RF fingerprint recognition network model and test samples, an open set RF fingerprint recognition network model is built for open set RF fingerprint recognition.

[0012] To optimize the above technical solutions, the specific measures taken also include:

[0013] Furthermore, the converting of the radio frequency data into the time-frequency diagram is specifically as follows: dividing the radio frequency data into a plurality of sample data, calculating the mean and variance of the sample data, and performing a normalization operation to convert the sample data into the time-frequency diagram.

[0014] Furthermore, the training samples and the test samples do not intersect with each other; Contains only known samples, where N K represents the number of known samples, represents the i-th time-frequency graph of the training sample, Represents the i-th category label; test sample Contains known samples and unknown open samples, where N T represents the number of samples of the test sample, represents the i-th time-frequency diagram of the test sample, Represents the i-th category label.

[0015] Furthermore, the radio frequency fingerprint recognition network model includes an encoder, a potential characterizer, a decoder and a known classifier; the encoder extracts hierarchical coding features of each layer in a hierarchical manner; the potential characterizer extracts the potential representation of each layer from the hierarchical coding features based on a reorganized attention mechanism; the decoder reconstructs the hierarchical coding features using the potential representation and top-down information, and then reconstructs the time-frequency graph; the known classifier performs recognition of known samples based on the hierarchical coding features;

[0016] Among them, the lth layer is expressed as:

[0017] x l+1 =E l (x l )

[0018] Z l =H l (x l )

[0019]

[0020]

[0021] In the formula, x l represents the hierarchical coding features of the lth layer, represents the hierarchical coding features reconstructed at the lth layer, Z l represents the feature map, represents the potential representation of the lth layer, E l (·) represents the autoencoder of the lth layer, H l (·) represents the extraction of the latent representation in the lth layer, To perform the inverse operation, SA(·) represents the reorganized attention mechanism, D l (·) represents the decoder of layer l.

[0022] Furthermore, the potential characterizer transforms the feature map Z l Divided into G group, Z l =[X1, X2, ..., X G ], k∈{1,...,G}, C represents the number of feature channels, H represents the height of the feature map, W represents the width of the feature map, and the sub-feature X k Divided into channel attention branches and channel attention branch

[0023] For the channel attention branch, global average pooling is used to generate channel statistics s1:

[0024]

[0025] In the formula, F gap (·) represents the average pooling function, yes Specific values ​​in the two dimensions H and W;

[0026] Get the output of the channel attention branch

[0027]

[0028] In the formula, σ(·) represents the sigmoid activation function with a gating mechanism, F c (·) represents the feature extraction function, and Respectively represent the weights and biases corresponding to the channel attention branches;

[0029] For the channel attention branch, the group norm is used to obtain the spatial statistics s2:

[0030]

[0031] In the formula, GN(·) represents the group norm;

[0032] Get the output of the channel attention branch

[0033]

[0034] In the formula, and Respectively represent the weights and biases corresponding to the channel attention branches;

[0035] Finally, the channel shuffling operator is used to aggregate and get

[0036] Furthermore, the known classifier obtains a feature vector from the hierarchical coding feature Where E(·) represents the encoder, Cls(·) represents the classifier, and y is calculatedknown and anchor point set C anchor ={c1, c2, ..., c M}The Euclidean distance d between

[0037] d=O(y known , C anchor )=(d1,...,d M )

[0038] =(||y known -c1||2,...,||y known -c M ||2)

[0039] Where O(·) represents the distance function, ||·||2 represents the Euclidean norm;

[0040] The output of the RF fingerprint recognition network model is:

[0041] l known =arg min(O(y known , C anchor ))

[0042] =(||y known -c1||2,...,||y know -c M ||2)

[0043] In the formula, l known Represents the result of a known classification.

[0044] Furthermore, in the process of using the training samples to train the radio frequency fingerprint recognition network model, the total loss function L of the radio frequency fingerprint recognition network model is total Set to:

[0045] L total =L tup +λ anc L anc +λ rec L rec

[0046] in,

[0047]

[0048]

[0049]

[0050] Where, L tup represents the loss function based on metric learning, Indicates that the label is The corresponding anchor point is Indicates that the label is When the classifier output and The distance between them, M represents the class of the known sample; L anc represents the anchor point loss; L rec denotes the reconstruction loss and D(·) denotes the decoder.

[0051] Furthermore, the open set RF fingerprint recognition network model is constructed based on the trained RF fingerprint recognition network model and the test samples, specifically:

[0052] Calculate the test sample and anchor point set C anchor Minimum distance between

[0053] Create a test sample The reconstruction loss function between the reconstructed time-frequency graph and

[0054] If d min ≤Θ and μ-3σ≤RLoss≤μ+3σ, where Θ represents the d obtained from the training set. min The threshold, μ represents the mean of RLoss obtained from the training set, σ represents the variance of RLoss obtained from the training set, then the test sample Belongs to a known class, and its label is determined by a known classifier; otherwise, the test sample Belongs to the unknown open category.

[0055] The present invention also proposes a computer-readable storage medium storing a computer program, characterized in that the computer program enables a computer to execute the above-mentioned open set radio frequency fingerprint recognition method based on layered autoencoder.

[0056] The present invention also proposes an electronic device, characterized in that it includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the open set radio frequency fingerprint recognition method based on the layered autoencoder as described above is implemented.

[0057] The beneficial effects of the present invention are as follows: in the open set RF fingerprint identification method based on layered autoencoders proposed by the present invention, a loss function based on metric learning design is adopted, which reduces the distance between the same class while increasing the distance between different classes; a layered autoencoder based on a reorganized attention mechanism is designed, which can better extract the features of known samples and retain the features of unknown open samples. The open set RF fingerprint identification network framework designed by the present invention can effectively improve the performance of open set RFFI without reducing the recognition accuracy of closed set RFFI, and improve the robustness of the open electromagnetic spectrum environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flow chart of an open set radio frequency fingerprint recognition method based on layered autoencoder proposed in the present invention.

[0059] Figure 2 It is a schematic diagram of the radio frequency fingerprint recognition network model of the present invention.

[0060] Figure 3 It is a schematic diagram of the reorganized attention mechanism used in the radio frequency fingerprint recognition network model of the present invention.

[0061] Figure 4 It is a schematic diagram of the open set radio frequency fingerprint recognition network model of the present invention.

[0062] Figure 5 It is a histogram of the minimum distances from the known samples and unknown open samples to the anchor points using the present invention.

[0063] Figure 6a and Figure 6b is a two-dimensional scatter plot using the visualization feature of the present invention, wherein Figure 6a For those without unknown classes, Figure 6b Contains unknown classes. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0065] like Figure 1 As shown, this embodiment proposes an open set radio frequency fingerprint recognition method based on a layered autoencoder, which specifically includes the following steps.

[0066] Step 1: RF data acquisition.

[0067] RF fingerprint recognition is regarded as a K-type hypothesis testing problem. The received signal of the k-th RF device is represented as x k (n) = s k (n)+ω k (n), where s k (n) represents the nth sampling point of the transmitted signal of the kth RF device, x k (n) represents the nth sampling point of the received signal of the kth RF device, ω k (n) means the mean is 0 and the variance is σ 2The received signal is represented as a vector of I / Q components. where x k Represents x k (n) in vector form, I k and Q k Represent the in-phase and quadrature components of the RF signal, respectively. and Represents the real and imaginary part operation of the received signal.

[0068] Step 2: Convert the I / Q data obtained in step 1 into a time-frequency diagram.

[0069] After obtaining the I / Q data in step 1, the data file is divided into multiple sample data. The mean and variance of the sample data are calculated, and normalization is performed to convert the sample data into a time-frequency diagram, laying the foundation for the subsequent training and testing of the RF fingerprint recognition network model.

[0070] Step 3: Classify the time-frequency diagram obtained in step 2 to obtain training samples and test samples.

[0071] The training of the RF fingerprint recognition network model is completed in the training samples, and the testing is completed in the test samples. The data of the training samples and the test samples do not intersect with each other. The training samples only contain known samples, that is, Where N K represents the number of known samples, represents the i-th time-frequency diagram of the received signal, represents the i-th category label, i∈{1, 2, ..., N k}. Where M is the number of known sample categories. In addition to the known samples of the same category as the training samples, the test samples also contain unknown open samples, that is, Where N T Indicator S T The number of samples in . Represents the class label. Part S T Belongs to the known sample, that is, S T-K , part S T Belong to the unknown open samples in the open space, that is, S T-U .

[0072] Step 4: Build the RF fingerprint recognition network model.

[0073] The RF fingerprint recognition network model SA-HANet consists of three parts, namely, encoder E(·), latent representation, decoder D(·) and known classifier Cls(·), as shown in Figure 2As shown in Figure 1, the encoder consists of a multi-layer encoder that extracts features in layers. The decoder consists of a multi-layer decoder that restores the original features in layers. Different from the traditional autoencoder, each layer in the SA-HANet scheme designed by the present invention consists of a simple autoencoder. The potential representation is based on the reorganized attention mechanism, which extracts features from the hierarchical encoding x. l Extract the potential representation of each layer like Figure 3 Then, the decoder uses the latent representation and top-down information To reconstruct the hierarchical encoding features The advantage of the SA-HANet model structure of the present invention is that it can retain important information that is ignored in the convolution process, making it easy to reconstruct the time-frequency diagram and laying the foundation for the subsequent open-set radio frequency fingerprint recognition.

[0074] Specifically, the lth layer of SA-HANet can be expressed as:

[0075] x l+1 =E l (x l )

[0076] Z l =H l (x l )

[0077]

[0078]

[0079] in, represents the hierarchical coding features reconstructed at the lth layer, x l+1 represents the hierarchical coding features of the l+1th layer, E l (·) represents the encoder in the lth layer. l (·) consists of a convolutional layer and a rectified linear unit (RELU) to extract the latent representation in the lth layer. SA(·) represents the reorganized attention mechanism to extract the important parts of the latent representation. and D l (·) Perform an inverse operation to recover the original information.

[0080] Reorganize the attention mechanism to map the feature map Z l Divided into G groups, to capture specific semantics from sub-features, namely Z l =[X1, X2, ..., X G ], C, H, and W represent X respectively. KIn the three-dimensional parameters, C represents the number of feature channels, H represents the height of the feature map, and W represents the width of the feature map. k There are two branches

[0081] For the channel attention branch, global average pooling (GAP) is used to generate channel statistics, i.e. F gap (·) represents the average pooling function, yes The specific values ​​of the last two dimensions. In order to achieve adaptive selection, the present invention uses a sigmoid activation function σ(.) with a gating mechanism. Therefore, the output result of the channel attention branch is obtained, that is, F c (·) represents the feature extraction function, where and denote the weights and biases corresponding to the channel attention branches, respectively.

[0082] For the channel attention branch, the group norm (GN) is used to obtain the spatial statistics s2, that is, Therefore, the output result of the channel attention branch is obtained, that is, in and They represent the weight and bias corresponding to the channel attention branch. The two branches are connected as

[0083] Finally, the channel shuffling operator is used to aggregate all sub-features to obtain

[0084] The recognition of known samples can be performed by a known classifier. The classifier Cls(.) is required to extract the input time-frequency graph from the Get the feature vector from is the hierarchical encoding feature obtained by the encoder from the time-frequency diagram. The non-trainable parameter C anchor is a set of anchor points, i.e. (c1, c2, ..., c M ). The distance-based function O(y known , C anchor ) Calculate y known and anchor point set C anchor The Euclidean distance between them can be expressed as:

[0085]

[0086] Among them, ||·||2 represents the Euclidean norm. Therefore, the output result of the SA-HANet model for identifying known samples proposed in the present invention can be expressed as:

[0087] l known =arg min(O(y known , C anchor ))

[0088] =(||y known -c1||2,...,||y known -c M ||2)

[0089] Among them, l known Represents the result of a known classification.

[0090] Step 5: Use the training samples to train the RF fingerprint recognition network model.

[0091] In the first step, the trainable parameters of the network are randomly initialized, the number of network training iterations is initialized to 1, the maximum number of iterations is 100, and the Adam optimization algorithm is used as the network training optimizer.

[0092] In the second step, the training samples are input into the network in batches for training. The training batch size is adjustable, and the training error obtained in each batch is back-propagated to optimize the network parameters. The loss function should not only reduce the distance to the class center, but also increase the distance to other classes. To achieve this, the loss function L based on metric learning is tup Minimize to ground truth anchors The distance to other classes is maximized, which can be expressed as:

[0093]

[0094] In order to make the known class closer to its true anchor, the anchor loss L anc The stimulus input minimizes the distance to its anchor point, i.e. Indicates that the label is The corresponding anchor point is Indicates that the label is When the classifier outputs the result and In order to provide information to distinguish known samples from unknown open samples during the training process, an autoencoder is used to reconstruct the time-frequency map. The reconstruction loss L rec is the distance between the input time-frequency graph and the reconstructed time-frequency graph, that is, Therefore, the total loss function is given by L tup , L anc and L rec Composition, namely L total =L tup +λ anc L anc +λ rec L rec, where the parameter λ is used to balance each loss term. Through the combination of each loss term, not only the distance to the center of the true class is minimized while the distance to other classes is maximized, but also the reconstruction loss is reduced.

[0095] A training epoch ends when backpropagation through all training data is complete.

[0096] Step 6: Determine whether the RF fingerprint recognition network model has been trained.

[0097] Determine whether the current training period has reached the maximum number of training periods. If so, save the radio frequency fingerprint recognition network model and execute step 7. If not, increase the number of training iterations by one and continue to retrain the model.

[0098] Step 7: Build an open set RF fingerprint recognition network model.

[0099] In order to classify known samples and unknown open samples using the SA-HANet network model of the present invention, it is necessary to calculate the minimum distance between the test sample and the anchor point set and the reconstruction loss between the test sample and the reconstructed spectrum map, such as Figure 4 As shown in Figure 2. Since the loss function based on metric learning increases the distance between different classes while clustering the same class, a distance-based function is designed for the open set RF fingerprint recognition algorithm. Test Sample and anchor point set C ancho The distance-based function between r is Test samples The reconstruction loss function between the reconstructed time-frequency graph is Therefore, the open set RF fingerprint recognition algorithm is based on the distance function and the reconstruction loss function. If d min ≤Θ and μ-3σ≤RLoss≤μ+3σ, Θ means d is obtained from the training set min The threshold, μ represents the mean of RLoss obtained from the training set, σ represents the variance of RLoss obtained from the training set, then the test sample belongs to a known class and its label can be determined by the known sample classifier. Otherwise, the test sample Belongs to the unknown open category.

[0100] Step 8: Output the classification results.

[0101] The effect of the present invention is further described below in conjunction with simulation experiments.

[0102] 1. Simulation conditions and parameter settings:

[0103] The simulation experiment of the present invention is carried out on the simulation platform of Python3.6 and Pytorch1.5.1. The computer CPU model is Intel Core i5, equipped with an independent graphics card model NVIDIA T600. The data set uses the public data set drone detect dataset: a radio frequency dataset of unmanned aerial system (UAS) signals for machine learning detection & classification, including seven different drones, namely DJI Mavic Pro, DJI Mavic Pro 2, DJI Mavic Mini, DJI Inspire 2, DJI Phantom 4, DJI Mavic 2Air S and Parrot Disco. The data file of each drone consists of 1.2×10 8 The sample points are divided into 100 parts, that is, each sample is equivalent to 20ms of recording time. Some of them are selected as known samples of closed-set RF fingerprint recognition, and the rest are used as unknown open samples of open-set RF fingerprint recognition.

[0104] 2. Simulation content:

[0105] Table 1 is a comparison chart of the accuracy and AUROC of the present invention and other existing technologies. Three technical solutions, Softmax, Openmax and CGDL, were selected for comparison. From the perspective of accuracy, it can be seen that the recognition accuracy of the present invention in closed-set RF fingerprint recognition is better than all benchmark solutions. At the same time, in the actual open electromagnetic spectrum environment, the open-set RF fingerprint recognition accuracy of the present invention can reach more than 90%. This shows that the present invention can effectively improve the recognition performance of open-set RF fingerprint without sacrificing the recognition accuracy of closed-set RF fingerprint. In addition, using AUROC, a different performance evaluation method, it can be seen from the results of CGDL that in the actual open electromagnetic spectrum environment, it is difficult for general deep learning methods to distinguish known samples from unknown open samples using only reconstruction loss. The AUROC of the present invention is significantly improved compared with the three technical solutions, which shows that the open-set RF fingerprint recognition network model proposed in the present invention has a strong ability to distinguish between known and unknown open samples. The reason is that the loss function proposed in the present invention not only reduces the distance between the same class, but also increases the distance between different classes.

[0106] Table 1 Comparison of accuracy and AUROC of the present invention and other existing technologies

[0107] method Accuracy (closed set recognition) Accuracy (open set recognition) AUROC (Open Set Recognition) Softmax 0.972 0.773 0.934 Openmax 0.972 0.875 0.939 CGDL 0.980 0.555 0.865 The present invention 0.983 0.914 0.960

[0108] Figure 5 It is a histogram of the minimum distances from the known samples and unknown open samples to the anchor points using the present invention. The horizontal axis represents the minimum distance to the anchor point, and the vertical axis represents the frequency of occurrence. It can be seen that the present invention uses the distance metric in space to make the known samples densely clustered at their anchor points, while increasing the distance between different samples, thereby reducing the overlap with the unknown open samples, achieving better distinction between known samples and unknown open samples, and providing guidance for practical applications.

[0109] In order to further demonstrate the effectiveness of the open set RF fingerprint recognition network model proposed in this invention, the extracted features are visualized by the t-SNE algorithm. Figure 6a-6b is a two-dimensional scatter plot using the visualization feature of the present invention. Figure 6a It can be seen from the figure that the characteristics of the RF fingerprint recognition network model proposed in the present invention are more discriminative between devices. Figure 6b As can be seen in , known samples and unknown open samples do not project to the same region in the feature space. This reduction in overlap enables the open set classifier to effectively identify and reject unknown open samples, thereby improving recognition accuracy.

[0110] Based on the above simulation results and analysis, the open set RF fingerprint intelligent and accurate identification system and method based on hierarchical autoencoder proposed in the present invention can better extract the features of known samples and retain the features of unknown open samples, which not only maintains the recognition accuracy of known samples, but also improves the recognition accuracy of open set RF fingerprint.

[0111] In another embodiment, a computer-readable storage medium is provided, which stores a computer program. The computer program enables a computer to execute the open set radio frequency fingerprint recognition method based on layered autoencoder as described in the first embodiment.

[0112] In another embodiment, an electronic device is proposed, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the open set radio frequency fingerprint recognition method based on a layered autoencoder as described in Embodiment 1 is implemented.

[0113] In the embodiments disclosed in the present application, the computer storage medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. The computer storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the above. More specific examples of computer storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0114] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0115] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. An open set radio frequency fingerprint recognition method based on layered autoencoders, characterized in that: The steps include: Collect radio frequency data and convert the radio frequency data into a time-frequency diagram; Classify the time-frequency graph into categories to obtain training samples and test samples; the training samples and test samples do not intersect with each other; Training samples Contains only known samples, where N K represents the number of known samples, represents the i-th time-frequency graph of the training sample, Represents the i-th category label; test sample Contains known samples and unknown open samples, where N T represents the number of samples of the test sample, represents the i-th time-frequency diagram of the test sample, Represents the i-th category label; A radio frequency fingerprint recognition network model is constructed, and the radio frequency fingerprint recognition network model is trained using training samples; the radio frequency fingerprint recognition network model includes an encoder, a potential characterizer, a decoder and a known classifier; the encoder extracts hierarchical coding features of each layer in a hierarchical manner; the potential characterizer extracts the potential representation of each layer from the hierarchical coding features based on a reorganized attention mechanism; the decoder reconstructs the hierarchical coding features using the potential representation and top-down information, and then reconstructs the time-frequency graph; the known classifier performs recognition of the known samples based on the hierarchical coding features; Among them, the lth layer is expressed as: x l+1 =E l (x l ) Z l =H l (x l ) In the formula, x l represents the hierarchical coding features of the lth layer, represents the hierarchical coding features reconstructed at the lth layer, Z l represents the feature map, represents the potential representation of the lth layer, E l (·) represents the autoencoder of the lth layer, H l (·) represents the extraction of the latent representation in the lth layer, To perform the inverse operation, SA(·) represents the reorganized attention mechanism, D l (·) represents the decoder of layer l; The latent characterizer maps the feature map Z l Divided into G group, Z l =[X1,X2,...,X G ], k∈{1,...,G},C represents the number of feature channels,H represents the height of the feature map,W represents the width of the feature map,and the sub-feature X k Divided into channel attention branches and channel attention branch For the channel attention branch, global average pooling is used to generate channel statistics s1: In the formula, F gap (·) represents the average pooling function, yes Specific values ​​in the two dimensions H and W; Get the output of the channel attention branch In the formula, σ(·) represents the sigmoid activation function with a gating mechanism, F c (·) represents the feature extraction function, and Respectively represent the weights and biases corresponding to the channel attention branches; For the channel attention branch, the group norm is used to obtain the spatial statistics s2: In the formula, GN(·) represents the group norm; Get the output of the channel attention branch In the formula, and Respectively represent the weights and biases corresponding to the channel attention branches; Finally, the channel shuffling operator is used to aggregate and get According to the trained RF fingerprint recognition network model and test samples, an open set RF fingerprint recognition network model is built for open set RF fingerprint recognition.

2. The open set radio frequency fingerprint recognition method based on layered autoencoder according to claim 1, characterized in that: The converting of the radio frequency data into the time-frequency diagram specifically includes: dividing the radio frequency data into a plurality of sample data, calculating the mean and variance of the sample data, and performing a normalization operation to convert the sample data into the time-frequency diagram.

3. The open set radio frequency fingerprint recognition method based on layered autoencoder according to claim 1, characterized in that: The known classifier obtains the feature vector from the hierarchical encoding feature Where E(·) represents the encoder, Cls(·) represents the classifier, and y is calculated known and anchor point set C anchor ={c1,c2,...,c M }The Euclidean distance d between d=O(y known ,C anchor )=(d1,...,d M ) =(||y known -c1||2,...,||y known -c M ||2) Where O(·) represents the distance function, ||·||2 represents the Euclidean norm; The output of the RF fingerprint recognition network model is: l known =argmin(O(y known ,C anchor )) =(||y known -c1||2,...,||y known -c M ||2) In the formula, l known Represents the result of a known classification.

4. The open set radio frequency fingerprint recognition method based on layered autoencoder as claimed in claim 3, characterized in that: In the process of using the training samples to train the radio frequency fingerprint recognition network model, the total loss function L of the radio frequency fingerprint recognition network model total Set to: L total =L tup +λ anc L anc +λ rec L rec in, Where, L tup represents the loss function based on metric learning, Indicates that the label is The corresponding anchor point is Indicates that the label is When the classifier output and The distance between them, M represents the class of known samples; L anc represents the anchor point loss; L rec denotes the reconstruction loss and D(·) denotes the decoder.

5. The open set radio frequency fingerprint recognition method based on layered autoencoder as claimed in claim 3, characterized in that: The open set RF fingerprint recognition network model is constructed based on the trained RF fingerprint recognition network model and the test samples, specifically: Calculate the test sample and anchor point set C anchor Minimum distance between Create a test sample The reconstruction loss function between the reconstructed time-frequency graph and If d min ≤Θ and μ-3σ≤RLoss≤μ+3σ, where Θ represents the d obtained from the training set. min The threshold, μ represents the mean of RLoss obtained from the training set, σ represents the variance of RLoss obtained from the training set, then the test sample Belongs to a known class, and its label is determined by a known classifier; Otherwise, test the sample Belongs to the unknown open category.

6. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables the computer to execute the open set radio frequency fingerprint recognition method based on layered autoencoder according to any one of claims 1 to 5.

7. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the open set radio frequency fingerprint recognition method based on a layered autoencoder as claimed in any one of claims 1 to 5 is implemented.

Images