Robust radio frequency fingerprint open set identification method

Abstract

The invention discloses a robust radio frequency fingerprint open set identification method, which comprises the following steps of: receiving and preprocessing a baseband signal sent by a transmitter, and performing feature extraction on the baseband signal; probability scores are obtained through calculation based on the extracted features, and a center sample set is obtained through the probability scores; performing similarity calculation on the to-be-identified features and the features of the registered transmitter and the central sample set, calculating an ROC curve according to the similarity, and setting an optimal threshold value of the registered transmitter; and comparing the similarity of the to-be-identified features with the optimal threshold value of the registered transmitter to judge the identity of the to-be-identified transmitter. According to the method, the radio frequency fingerprint features with higher quality in the data are captured through the feature extraction network with deeper residual network design, and the radio frequency fingerprint features are properly processed through the M-sigmoid classifier to obtain a better classification result, so that the radio frequency fingerprint closed set and open set identification accuracy is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and particularly to a robust radio frequency fingerprint open-set recognition method. Background Art

[0002] Defects at the hardware level during device production will form the radio frequency fingerprint features of each device. Since such defects are unique, the radio frequency fingerprint features of each device are unique. And this radio frequency fingerprint feature is like the "fingerprint" of the device and has uniqueness. Therefore, this physical-level radio frequency fingerprint feature is called the physical fingerprint radio frequency fingerprint feature or radio frequency fingerprint feature. During the communication process, the electromagnetic wave signals sent by the device will introduce the radio frequency fingerprint features of the device. In recent years, research has shown that radio frequency fingerprints can be used as the unique radio frequency fingerprint features of wireless devices, based on which the identification and authentication of electromagnetic radiation sources can be realized. Since radio frequency fingerprints are only related to the physical level of the device, they will not be interfered by changes in signal content and form. In addition, compared with traditional device authentication methods, radio frequency fingerprint features effectively avoid attacks such as illegal acquisition and forgery due to their uniqueness. All the above advantages indicate that radio frequency fingerprint technology has broad application prospects.

[0003] Currently, the technologies for radio frequency fingerprint recognition based on machine learning have been relatively mature, such as SVM, random forest, decision tree, etc. In recent years, relevant models in the field of deep learning have also been introduced into the research of radio frequency fingerprint recognition and have achieved rich results. However, most of the current research still focuses on radio frequency fingerprint recognition in a closed-set environment. In actual authentication applications, unknown devices will inevitably be encountered, thus bringing challenges to open-set recognition. In addition, during the communication process, the channel will have a certain impact on radio frequency fingerprints, resulting in a decrease in the quality of radio frequency fingerprints, which undoubtedly increases the difficulty of radio frequency fingerprint open-set recognition. And currently, there is little research on radio frequency fingerprint open-set recognition. This research gap indicates that it is necessary to conduct in-depth research on a robust radio frequency fingerprint open-set recognition method. Summary of the Invention

[0004] The present invention aims to solve at least to some extent the technical problems existing in the related technologies.

[0005] The object of the present invention is to provide a method for radio frequency fingerprint open-set recognition.

[0006] To achieve the above object, the present invention provides a robust radio frequency fingerprint open-set recognition method, including the following steps:

[0007] S1. Receive and preprocess the baseband signal sent by the transmitter, and then extract the radio frequency fingerprint features from the preprocessed signal;

[0008] S2. Calculate the probability score based on the extracted RF fingerprint features, and obtain the central sample set using the obtained probability score;

[0009] S3. Calculate the similarity between the RF fingerprint features to be recognized and the RF fingerprint features of the registered transmitters respectively with the central sample set, calculate the ROC curve based on the similarity, and then set the optimal similarity threshold for the registered transmitters;

[0010] S4. Determine the identity of the transmitter to be recognized by comparing the similarity of the RF fingerprint features to be recognized with the optimal similarity threshold of the registered transmitters.

[0011] Preferably, in step S1, receive and preprocess the baseband signal sent by the transmitter, and then extract the RF fingerprint features from the preprocessed signal. Specifically:

[0012] S11. Receive the signal carrying its RF fingerprint information sent by the transmitter. The received signal is:

[0013]

[0014] where represents the RF fingerprint function of the transmitter ; represents the continuous-time variable, represents the ideal baseband signal sent by the transmitter, describes the channel effect, represents the convolution operation, represents the additive white Gaussian noise, represents the number of registered transmitters;

[0015] S12. Preprocess the received signal to obtain the signal ;

[0016] S13. Extract the RF fingerprint features from the preprocessed signal to obtain the RF fingerprint features .

[0017] Preferably, in step S12, preprocess the received signal to obtain the signal , specifically:

[0018] First, the received signal is normalized to obtain:

[0019]

[0020] where represents the total duration of the signal;

[0021] Secondly, perform frequency offset compensation. After coarse frequency offset compensation processing:

[0022]

[0023] Where represents coarse frequency offset compensation, represents the maximum value of the cross-correlation between the received signal and the local preamble, represents the th sampling point, represents the length of, represents the sampling interval, represents the local preamble;

[0024] After coarse frequency offset compensation, the compensated signal is obtained. Then, perform chip matching and signal compression:

[0025]

[0026] Where represents the index of the symbol, represents the chips of each symbol, represents the matching chip sequence;

[0027] The fine frequency offset compensation is expressed as:

[0028]

[0029] Where represents the symbol number used in the fine frequency offset compensation estimation;

[0030] The phase shift is expressed as:

[0031] ;

[0032] The finally frequency offset compensated signal is:

[0033] .

[0034] Preferably, for the preprocessed signal in step S13, perform radio frequency fingerprint feature extraction to obtain the radio frequency fingerprint feature , specifically:

[0035] S131. Input the preprocessed signal into a low-pass filter and a radio frequency fingerprint enhancement model for processing to obtain signals and :

[0036]

[0037]

[0038] wherein represents low - pass filter operation, represents the enhanced processing of auto - encoder radio frequency fingerprint information;

[0039] S132. Based on the signals output after being processed by the low - pass filter and the radio frequency fingerprint enhancement model and , calculate the reconstruction loss and feedback it to the radio frequency fingerprint enhancement model for parameter adjustment. The calculation formula of the reconstruction loss is as follows:

[0040]

[0041] wherein represents the expectation operator, representing the mean value of all samples when calculating the loss, represents the squared Euclidean norm;

[0042] S133. Input the signal after being processed by the radio frequency fingerprint enhancement model into the feature extraction network designed based on the residual network. The process is expressed as:

[0043]

[0044] wherein represents the radio frequency fingerprint feature after the signal is extracted by the network, represents the feature extraction process.

[0045] Preferably, step S2 calculates the probability score based on the extracted radio frequency fingerprint feature and obtains the central sample set by using the obtained probability score. Specifically:

[0046] S21. Based on the extracted radio frequency fingerprint feature, use the M - sigmoid classifier to calculate the probability score and classify it according to the probability score:

[0047]

[0048] wherein represents that the radio frequency fingerprint feature of the th sample is extracted, represents the batch size;

[0049] S22. Calculate the center loss and cross - entropy loss based on the radio frequency fingerprint feature and the above - mentioned probability score:

[0050] Calculate the cross - entropy loss:

[0051]

[0052] where is the true label of the -th sample represented in one-hot encoding form,

[0053] Calculate the center loss:

[0054]

[0055] where represents the centroid of the true class to which the sample belongs, which is dynamically updated. The two losses are jointly fed back to update the parameters of the feature extraction network;

[0056] S23. Select the sample with the highest probability for each class based on the probability scores obtained in step S21 to form a central sample set, denoted as , where represents the -th sample of the -th class in the central sample set.

[0057] Preferably, in step S3, the radio frequency fingerprint features to be recognized and the radio frequency fingerprint features of the registered transmitters are respectively calculated for similarity with the central sample set, and the ROC curve is calculated based on the similarity, and then the optimal similarity threshold of the registered transmitters is set, specifically:

[0058] S31. Calculate the similarity between the extracted radio frequency fingerprint features to be recognized and the radio frequency fingerprint features of the registered transmitters and the central sample set respectively, and then obtain the average similarity with each sample in the central sample set;

[0059] S32. Classify the radio frequency fingerprint features of the registered transmitters and the radio frequency fingerprint features to be recognized based on the similarity;

[0060] S33. Calculate the ROC curve of each registered transmitter based on the similarity, and then set the optimal similarity threshold of each registered transmitter.

[0061] Preferably, the formula for calculating the similarity in step S31 is:

[0062] .

[0063] Preferably, step S32 classifies the radio frequency fingerprint features of the registered transmitters and the radio frequency fingerprint features to be recognized based on the similarity, specifically:

[0064] Obtain the maximum value of the average similarity of the radio frequency fingerprint features to be recognized and the radio frequency fingerprint features of the registered transmitters:

[0065]

[0066] Among them corresponding is the category to which the RF fingerprint feature belongs, that is, the temporary category of the RF fingerprint feature to be recognized or the category to which the RF fingerprint feature of the registered transmitter belongs.

[0067] Preferably, step S33 calculates the ROC curve of each registered transmitter based on the similarity, and then sets the optimal similarity threshold for each registered transmitter, specifically:

[0068] A series of thresholds are set for each category of the similarities obtained for the registered transmitters, and the ROC curve is calculated based on this. Among them, the true positive rate TPR in the ROC curve is expressed as:

[0069]

[0070] where TP represents the positive examples correctly classified, and FN represents the negative examples misclassified;

[0071] The false positive rate FPR is expressed as:

[0072]

[0073] where FP represents the positive examples misclassified, and TN represents the negative examples correctly classified.

[0074] Definition: , when is the largest, set the threshold corresponding to this point as the optimal similarity threshold of this registered transmitter, denoted as .

[0075] Preferably, step S4 determines the identity of the transmitter to be recognized by comparing the similarity of the RF fingerprint feature to be recognized with the optimal similarity threshold of the registered transmitter, specifically:

[0076] Discriminate the maximum average similarity of the RF fingerprint feature to be recognized with the optimal similarity threshold of the registered transmitters temporarily belonging to the same category. The process is as follows:

[0077]

[0078] where is the final discrimination result, class represents the registered transmitter, and unknown represents the unregistered transmitter.

[0079] Beneficial effects: (1) In the feature extraction process of the present invention, an RF fingerprint enhancement model is used to process the signal, reducing signal data noise and reducing the influence of the channel.

[0080] (2) The feature extraction network designed with a deeper residual network in the present invention can capture higher-quality RF fingerprint features in the data.

[0081] (3) The M-sigmoid classifier adopted in the present invention can properly process RF fingerprint features to obtain better classification results.

[0082] (4) The present invention feeds the reconstruction loss back to the RF fingerprint enhancement model for parameter adjustment; at the same time, the cross-entropy loss obtained based on the probability score and the center loss calculated from the RF fingerprint features are jointly calculated and fed back to the feature extraction model for parameter update, so that more accurate and real-time RF fingerprint features can be extracted, thereby effectively avoiding the influence of defects and changes at the physical level of the device on the RF fingerprint recognition result. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 is the flowchart of the method of the present invention;

[0084] Figure 2 is the structural diagram of the FACE RF fingerprint enhancement model of the present invention;

[0085] Figure 3 is the structural diagram of the RF fingerprint open-set recognition model of the present invention;

[0086] Figure 4 is the graph of the change in the RF fingerprint recognition accuracy of the solution of the present invention and the baseline solution in Embodiment 1 of the present invention;

[0087] Figure 5 is the graph of the change in the RF fingerprint recognition accuracy of the solution of the present invention and the baseline solution in Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0088] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention, and they should not be construed as limiting the present invention. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. In the description of the present invention, it should be understood that the terms used are only for the purpose of description and cannot be construed as indicating or implying relative importance.

[0089] The following will describe a robust RF fingerprint open-set recognition method provided by the present invention in conjunction with Figures 1-5 the accompanying drawings.

[0090] Embodiment 1: As Figure 1 and Figure 3As shown in the figure, this embodiment provides a robust radio frequency fingerprint open set recognition method, and the specific steps are as follows:

[0091] Step 1, transmission signal acquisition:

[0092] is the ideal baseband signal transmitted by the transmitter , and the radio frequency fingerprint information of the transmitter is carried in the signal. The signal propagates through the wireless channel and is received by the receiving device to obtain the signal

[0093]

[0094] where represents the radio frequency fingerprint function of the transmitter , represents the continuous time variable, describes the channel effect, represents the convolution operation, represents the additive white Gaussian noise, represents the number of registered devices.

[0095] Step 2, signal preprocessing:

[0096] First, normalize the signal received by the receiving device to obtain :

[0097]

[0098] where represents the total duration of the signal.

[0099] Secondly, perform frequency offset compensation. After rough frequency offset compensation processing:

[0100]

[0101] where represents the rough frequency offset compensation, represents the maximum value of the cross-correlation between the received signal and the local preamble, represents the th sampling point, represents 's length, represents the sampling interval, represents the local preamble;

[0102] After rough frequency offset compensation, the compensated signal can be obtained, and then chip matching and signal compression are performed on it:

[0103]

[0104] wherein represents the index of the symbol, represents the chip of each symbol, represents the matching chip sequence;

[0105] The fine frequency offset compensation can be expressed as:

[0106]

[0107] wherein represents the symbol number used in the fine frequency offset compensation estimation;

[0108] The phase shift can be expressed as:

[0109] ,

[0110] The signal after the final frequency offset compensation is:

[0111] .

[0112] Step 3: Use the FACE model to enhance the RF fingerprint:

[0113] As Figure 2 shown, input the preprocessed signal into the low-pass filter,

[0114]

[0115] wherein represents the low-pass filter operation; at the same time, input the signal into the FACE RF fingerprint enhancement model,

[0116]

[0117] wherein represents the autoencoder processing;

[0118] Calculate the reconstruction loss of the output signals of the two processing methods and feedback it to the FACE RF fingerprint enhancement model for adjusting the parameters of the FACE RF fingerprint enhancement model:

[0119]

[0120] wherein represents the expectation operator, indicating the mean of all samples when calculating the loss, represents the squared Euclidean norm.

[0121] Step 4: Extract RF fingerprint features based on the residual network

[0122] Input the signal processed by the FACE radio frequency fingerprint enhancement model into the feature extraction network, which is designed based on the residual network, and the process is expressed as:

[0123]

[0124] where represents the radio frequency fingerprint features after the signal is extracted by the network, represents the feature extraction process.

[0125] Step 5, obtain the central sample set according to the probability scores output by the M-sigmoid layer:

[0126] Input the extracted radio frequency fingerprint features into the M-sigmoid layer to calculate the probability scores, and then classify all the radio frequency fingerprint features according to the probability scores. The probability score calculation formula is:

[0127]

[0128] where represents the radio frequency fingerprint features of the th sample, represents the batch size.

[0129] Calculate the cross-entropy loss based on the probability scores:

[0130]

[0131] where is the true label of the th sample represented in one-hot encoding form;

[0132] Calculate the center loss based on the radio frequency fingerprint features:

[0133]

[0134] where represents the centroid of the true class to which the sample belongs in the embedding space and is dynamically updated. Jointly feedback the two losses to the feature extraction network to update the feature extraction network parameters.

[0135] In this step, select the samples with the highest probability for each category on the training set to form the central sample set, denoted as , representing the th sample of the th class in the central sample set.

[0136] Step 6, calculate the similarity between the extracted RF fingerprint features and the central sample set:

[0137] Calculate the cosine similarity between the RF fingerprint features to be recognized and the RF fingerprint features of the registered transmitters and the samples in the central sample set respectively, and further obtain the average cosine similarity between the RF fingerprint features to be recognized and the RF fingerprint features of the registered transmitters and each type of sample in the central sample set. The similarity calculation formula is as follows:

[0138]

[0139] Further obtain the maximum value of the average cosine similarity:

[0140]

[0141] Where Corresponding That is, the category to which the RF fingerprint features of the registered transmitter belong or the temporary category to which the RF fingerprint features to be recognized belong.

[0142] Step 7, set the similarity threshold for each type of device based on the ROC curve:

[0143] In this step, a series of thresholds are set for each category for the similarity obtained on the registered transmitter, and the ROC curve is calculated based on this. The true positive rate TPR in the ROC curve is expressed as:

[0144]

[0145] Where TP represents the correctly classified positive examples, and FN represents the misclassified negative examples;

[0146] The false positive rate FPR is expressed as:

[0147]

[0148] Where FP represents the misclassified positive examples, and TN represents the correctly classified negative examples.

[0149] Definition: , when is the largest, set the threshold corresponding to this point as the best similarity threshold of this registered transmitter, denoted as .

[0150] Step 8, determine the device identity through threshold judgment:

[0151] Discriminate the maximum average cosine similarity of the RF fingerprint features to be recognized and the best similarity threshold of the registered transmitters temporarily belonging to the same category. The process is as follows:

[0152]

[0153] where is the final discrimination result, class indicates a registered transmitter, and unknown indicates an unregistered transmitter. When ≥ at that time, the corresponding transmitter is a registered transmitter ; when < at that time, the corresponding transmitter is an unregistered transmitter.

[0154] To verify the recognition effect of this solution in the closed set scenario, in this embodiment, in the composite channel scenario, data is collected in a complex indoor environment under line-of-sight and non-line-of-sight conditions respectively, and the distance range between the sending device and the receiving device is 10 - 40 m. The experiment conducts radio frequency fingerprint recognition research on 1072 frames of signals of 12 actual ZigBee devices collected. To evaluate the performance of this method under various signal quality conditions, a noisy dataset with signal-to-noise ratios of {0, 5, 10, 15, 20, 25, 30} dB is constructed by the method of adding artificial noise, and the dataset is divided into a training set and a test set according to the ratio of 8:2. In this context, this embodiment respectively uses the existing baseline solution and this solution to conduct radio frequency fingerprint recognition.

[0155] This embodiment compares the recognition accuracies of the two solutions on the noisy dataset with signal-to-noise ratios of {0, 5, 10, 15, 20, 25, 30} dB. As Figure 4 shown, the closed set accuracies of the baseline solution on the noisy dataset with signal-to-noise ratios of {0, 5, 10, 15, 20, 25, 30} dB are {9.0, 17.2, 24.7, 40.9, 54.9, 72.1, 80.5}% respectively, while the accuracies of this solution on the noisy dataset with signal-to-noise ratios of {0, 5, 10, 15, 20, 25, 30} dB are {27.8, 43.7, 72.6, 86.0, 92.6, 94.4, 94.9}% respectively. In all ranges, the present invention has brought a significant improvement in accuracy, with the highest improvement exceeding 40%. This shows that the method of the present invention is an effective radio frequency fingerprint feature recognition method in the closed set environment.

[0156] Example 2: This embodiment compares the open set recognition performances of this solution and the baseline solution under different signal-to-noise ratios. Based on Example 1, in this embodiment, 4 devices are selected from 12 actual ZigBee devices as the dataset of unregistered devices, and the rest are used as the dataset of registered devices. The following operations are performed on the two datasets: 20% of the data is taken as the test set, and the remaining data is divided into a training set and a validation set according to the ratio of 8:2 (the validation set of unregistered devices does not participate in the threshold calculation).

[0157] In this embodiment, the recognition accuracies of two schemes are compared on 7 noisy datasets with {0, 5, 10, 15, 20, 25, 30} dB. As Figure 4 shown, the accuracies of the baseline scheme on the 7 noisy datasets with {0, 5, 10, 15, 20, 25, 30} dB are {11.8, 14.8, 19.1, 33.8, 45.4, 63.4, 66.7}% respectively, while the open-set accuracies of this scheme on the 7 noisy datasets with {0, 5, 10, 15, 20, 25, 30} dB are {22.2, 41.2, 62.8, 76.8, 80.6, 84.3, 86.1}% respectively. In the range of signal-to-noise ratio from 10 dB to 15 dB, the accuracy improvement brought by the method of the present invention is more than 40%. It can be seen that the improvement of the open-set performance of the method of the present invention is significant, which indicates that the method of the present invention is a method that can effectively perform open-set recognition of RF fingerprints under the influence of channels.

[0158] After receiving a signal with channel interference in the present invention, the receiver performs preprocessing and combines deep learning-related methods for RF fingerprint enhancement, effectively reducing data noise and reducing the influence of channels; and the feature extraction network designed based on a deeper residual network can capture higher-quality RF fingerprint features in the data; finally, the M-sigmoid classifier can properly process the RF fingerprint features to obtain better classification results. The above measures provide accurate data support for the final RF fingerprint recognition, thereby significantly improving the accuracy of RF fingerprint recognition.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical RF fingerprint features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A robust radio frequency fingerprint open-set recognition method, characterized in that, It includes the following steps: S1. Receive and preprocess the baseband signal sent by the transmitter, and then extract the radio frequency fingerprint features from the preprocessed signal; S2. Calculate the probability scores based on the extracted radio frequency fingerprint features, and obtain the central sample set by using the obtained probability scores; S3. Calculate the similarity between the radio frequency fingerprint features to be recognized and the radio frequency fingerprint features of the registered transmitters respectively with the central sample set, calculate the ROC curve according to the similarity, and then set the optimal similarity threshold of the registered transmitters; S4. Judge the identity of the transmitter to be recognized by comparing the similarity of the radio frequency fingerprint features to be recognized with the optimal similarity threshold of the registered transmitters.

2. A robust radio frequency fingerprint open-set recognition method according to claim 1, characterized in that, For step S1 of receiving and preprocessing the baseband signal sent by the transmitter, and then extracting the radio frequency fingerprint features from the preprocessed signal, specifically: S11. Receive the signal carrying its radio frequency fingerprint information sent by the transmitter, and the received signal is: ； Among them represents the radio frequency fingerprint function of the transmitter ; represents the continuous-time variable represents the ideal baseband signal transmitted by the transmitter describes the channel effect represents the convolution operation represents the additive white Gaussian noise represents the number of registered transmitters; S12. Preprocess the received signal to obtain the signal ; S13. Perform radio frequency fingerprint feature extraction on the preprocessed signal to obtain radio frequency fingerprint features .

3. A robust radio frequency fingerprint open set recognition method according to claim 2, characterized in that, The signal received in step S12 is preprocessed to obtain a signal , specifically: First, the received signal is obtained through normalization as follows: ； wherein represents the total duration of the signal; Next, frequency offset compensation is performed. After coarse frequency offset compensation processing: ； Among them represents coarse frequency offset compensation represents the maximum value of the cross-correlation between the received signal and the local preamble represents the th sampling point represents length of represents the sampling interval represents the local preamble The compensated signal is obtained after coarse frequency offset compensation , and then chip matching and signal compression are performed: ； wherein represents the index of the symbol, represents the chip of each symbol, represents the matching chip sequence; The fine frequency offset compensation is expressed as: ； wherein represents the symbol number used in the fine frequency offset compensation estimation; The phase shift is expressed as: ； The finally frequency offset compensated signal is: 。 4. A robust radio frequency fingerprint open-set recognition method according to claim 3, characterized in that, The preprocessed signal in step S13 is subjected to radio frequency fingerprint feature extraction to obtain radio frequency fingerprint features , specifically: S131. Input the preprocessed signal into a low-pass filter and a radio frequency fingerprint enhancement model respectively for processing to obtain signals and : ； ； Among them represents a low-pass filter operation represents the enhanced processing of the radio frequency fingerprint information of the autoencoder S132. Based on the signal output after being processed by the low-pass filter and the radio frequency fingerprint enhancement model and , calculate the reconstruction loss and feedback it to the radio frequency fingerprint enhancement model for parameter adjustment. The calculation formula of the reconstruction loss is as follows: ； Among them represents the expectation operator, which represents the mean of all samples when calculating the loss, represents the squared Euclidean norm; S133. The signal processed by the radio frequency fingerprint enhancement model is input into the feature extraction network designed based on the residual network, and the process is expressed as: ； Among them represents the radio frequency fingerprint feature after the signal is extracted by the network represents the feature extraction process 5. A robust radio frequency fingerprint open set recognition method according to claim 4, characterized in that For step S2 of calculating the probability scores based on the extracted radio frequency fingerprint features, and obtaining the central sample set by using the obtained probability scores, specifically: S21. Based on the extracted radio frequency fingerprint features, use the M-sigmoid classifier to calculate the probability scores and classify them according to the probability scores; ； Among them indicates that the extracted one is the radio frequency fingerprint feature of the th sample, indicating the batch size; S22. Calculate the center loss and cross-entropy loss based on the radio frequency fingerprint features and the above probability scores: Calculate the cross-entropy loss: ； wherein is the true label of the th sample represented in one-hot encoding form, Calculate the center loss: ； Among them represents the centroid of the samples in the embedding space whose true category belongs to is dynamically updated. The two losses are jointly fed back to update the parameters of the feature extraction network; S23. Select the sample with the highest probability for each category based on the probability scores obtained in step S21 to form a central sample set, denoted as , where represents the th sample of the th category in the central sample set.

6. A robust radio frequency fingerprint open set recognition method according to claim 5, characterized in that, For step S3 of calculating the similarity between the radio frequency fingerprint features to be recognized and the radio frequency fingerprint features of the registered transmitters respectively with the central sample set, calculate the ROC curve according to the similarity, and then set the optimal similarity threshold of the registered transmitters, specifically: S31. Calculate the similarity between the extracted radio frequency fingerprint features to be recognized and the radio frequency fingerprint features of the registered transmitters respectively with the central sample set, and then obtain the average similarity with various samples of the central sample set; S32. Classify the radio frequency fingerprint features of the registered transmitters and the radio frequency fingerprint features to be recognized based on the similarity; S33. Calculate the ROC curve of each registered transmitter based on the similarity, and then set the optimal similarity threshold of each registered transmitter.

7. A robust radio frequency fingerprint open set recognition method according to claim 6, characterized in that The calculation formula of the similarity in step S31 is: 。 8. A robust radio frequency fingerprint open set recognition method according to claim 7, characterized in that For step S32 of classifying the radio frequency fingerprint features of the registered transmitters and the radio frequency fingerprint features to be recognized based on the similarity, specifically: Obtain the maximum value of the average similarity of the radio frequency fingerprint features to be recognized and the radio frequency fingerprint features of the registered transmitters; ； Among them corresponding is the category to which the radio frequency fingerprint feature belongs, that is, the temporary category to which the radio frequency fingerprint feature to be recognized belongs or the category to which the radio frequency fingerprint feature of the registered transmitter belongs.

9. A robust radio frequency fingerprint open-set recognition method according to claim 8, wherein For step S33 of calculating the ROC curve of each registered transmitter based on the similarity, and then setting the optimal similarity threshold of each registered transmitter, specifically: Set a series of thresholds for each class for the similarity obtained by the registered transmitters, and calculate the ROC curve based on this, where the true positive rate TPR in the ROC curve is expressed as: ； Where TP represents the correctly classified positive examples, and FN represents the misclassified negative examples; The false positive rate FPR is expressed as: ； Where FP represents the misclassified positive examples, and TN represents the correctly classified negative examples, Definition: When is at its maximum, set the threshold corresponding to this point as the best similarity threshold of the registered transmitter, denoted as .

10. A robust radio frequency fingerprint open-set recognition method according to claim 9, characterized in that, Determining the identity of the transmitter to be recognized by comparing the similarity of the radio frequency fingerprint features to be recognized with the optimal similarity threshold of the registered transmitter in step S4 specifically as follows: Comparing the maximum average similarity of the radio frequency fingerprint features to be recognized with the optimal similarity threshold of the registered transmitters temporarily belonging to the same category, and the process is as follows: ； wherein is the final discrimination result, class indicates a registered transmitter, and unknown indicates an unregistered transmitter.

Images