Radio frequency fingerprint identification method based on memory vector depth generation network

Abstract

The invention relates to a radio frequency fingerprint identification method based on a memory vector depth generation network, which constructs a data set based on same-direction / orthogonal signal data, constructs a deep learning model based on the memory vector depth generation network, does not depend on a specific data processing method, does not need to label the data, and can identify the fingerprint after training on a public data set. The method can be conveniently and rapidly migrated to other application scenes, has a good generalization ability, and can achieve a good recognition effect.

Description

Technical Field

[0001] The present invention relates to the field of communication security technology, and in particular to a radio frequency fingerprint recognition method based on a memory vector deep generation network. Background Art

[0002] RF fingerprint refers to the unique and distinctive subtle features existing in RF signals from different radiation sources. It is a unique, stable and individually different feature set introduced by the physical characteristics of the transmitter hardware (such as transistor characteristics, circuit layout, etc.) in the RF signal. These features are reflected in multiple dimensions such as signal amplitude, phase, and frequency. Even for RF devices of the same model and batch, the amplitude fluctuation characteristics of the signal are different under the same modulation method and transmission frequency, that is, these features still exist in RF devices of the same signal and batch.

[0003] By extracting characteristic fingerprints and analyzing and identifying RF fingerprint features, key tasks such as identification, authentication, and authorization of RF devices can be effectively achieved; an ideal RF fingerprint extraction method should have characteristics such as uniqueness, versatility, independence, and robustness. Currently, RF fingerprint recognition methods are mainly divided into two categories: feature engineering methods and deep learning methods.

[0004] The RF fingerprint recognition method based on feature engineering requires an in-depth understanding of the generation and propagation of RF signals and the characteristics of RF devices in order to design effective features. By utilizing the prior knowledge of experts, transient features and steady-state features are manually designed through data analysis, and then support vector machines, decision trees or shallow neural networks are used to complete subsequent recognition tasks. The RF fingerprint features designed by experts have a clear physical meaning and good interpretability, and can achieve a high recognition accuracy in a specific system. However, the feature engineering method relies on expert knowledge, the fingerprint features are sensitive to noise, and the generalization ability is poor, and the designed features are difficult to apply to other scenarios.

[0005] Deep learning methods represented by supervised learning are driven by massive amounts of labeled data, automatically extract features through deep neural networks, and then complete the identification of RF fingerprints. The main advantage of deep learning methods is that they can have good feature extraction capabilities without the need for domain expert knowledge. However, deep learning methods have a strong dependence on data, and the quality of RF fingerprint extraction is limited by data quality; the collection and simulation of RF data are relatively difficult, and data labeling is time-consuming and labor-intensive, which makes it difficult to meet the needs of deep learning training for massive data; some RF fingerprint extraction methods based on deep learning use time domain and frequency domain transformation methods to convert co-directional / orthogonal signals into images. The conversion process loses a lot of information and it is difficult to meet the needs of real-time signal processing; at the same time, the models trained by deep learning methods on specific data sets are difficult to migrate to other cross-domain scenarios. Summary of the invention

[0006] In view of the defects of the prior art, the present invention proposes a radio frequency fingerprint recognition method based on a memory vector deep generation network, which is based on the co-directional / orthogonal signal data and does not rely on a specific data processing method. The deep generation method is used to train the model, and there is no need to accurately label the data. After training on a public data set, it can be easily and quickly migrated to other application scenarios and has good generalization capabilities.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] A radio frequency fingerprint recognition method based on a memory vector deep generation network includes the following steps: Step 1: using several data sets based on co-directional / orthogonal signals that are publicly available on the Internet or collected and simulated by oneself to construct an original data set; and using a data loader to pre-process the data; Step 2: Build a deep generative model based on a deep generative network, where the deep generative network consists of three parts: encoder, memory, and decoder. Step 3: Use the data loader to extract training data from the dataset to train the deep generative model, and introduce the loss function in the training to perform back-propagation optimization on the network parameters; Step 4: Take the encoder network Encoder as the feature extraction network, encapsulate its network structure and parameters into the application scenario as a RF fingerprint extractor to achieve individual identification and authentication of RF devices or intrusion detection.

[0009] Furthermore, the process of preprocessing data using the data loader specifically includes: randomly selecting data from several data sets according to a certain ratio, performing data augmentation on the selected data, adding noise, jitter, phase shift and frequency shift, etc., and then randomly cropping the data to a fixed length.

[0010] Furthermore, in the deep generative network, the encoder is used to compress the original in-phase / orthogonal signal and extract its key features to obtain the latent variable z in the high-dimensional space. The memory is used to store the feature vectors of historical signals. On the basis of ensuring the diversity and sparsity of the features in the memory, the latent variable z obtained by encoding each different signal has better specificity. The decoder is used to reconstruct the original signal from the latent variable z to ensure that the latent variable z obtained by the encoder has better representation ability.

[0011] Furthermore, the encoder is composed of modules that can process one-dimensional time series signals, including but not limited to one-dimensional convolutional neural networks, recurrent neural networks, and Transformers; the decoder is the reverse structure of the encoder, and uses a network structure similar to the encoder to gradually reconstruct the latent variables into signals.

[0012] Furthermore, during model training, a data loader is used to extract N samples from the data set to form a batch of training data. For any data in the batch, an encoder is used to compress it into a high-dimensional space to obtain a latent variable z. The latent variable z is then compared with the feature vector in the memory, and the feature of the latent variable z is used to update the closest feature in the memory. Finally, a decoder is used to reconstruct the signal.

[0013] Furthermore, reconstruction loss, feature similarity loss and feature sparsity loss are introduced in the training to ensure the specificity and representativeness of the generated latent variable z; among them, the reconstruction loss L rec Using L2 distance, feature similarity loss L sim Using cosine similarity, feature sparsity loss L spr Use L2 regularization with parameter λ sim and λ spr To balance the effects of various loss functions, the total loss function is: L = L rec -λ sim L sim +λ spr L spr .

[0014] Furthermore, in step 4, for the access device, during the deployment phase, the RF signal of the access device is collected, and the RF fingerprint is extracted by the RF fingerprint extractor and stored in the database; during the use phase, when the RF signal is detected, its RF fingerprint is extracted and compared with the RF fingerprint in the database. If the RF fingerprint of the device meets the tolerance, the authentication is successful, otherwise the authentication fails.

[0015] Furthermore, the tolerance is obtained in the following manner: constructing an identification data set for an individual radiation source, sampling the signals in the data set and inputting them into a radio frequency fingerprint generator, and calculating the mean similarity measurement D1 of the individual radio frequency fingerprint and the mean similarity measurement D2 of the radio frequency fingerprints of different individuals, and the tolerance is ρ=(D1+min(D2)) / 2.

[0016] Beneficial effects:

[0017] 1. The present invention directly uses the co-directional / orthogonal signals to extract radio frequency fingerprints, does not need to construct positive / negative sample pairs, does not rely on specific data processing methods, and can reduce the hardware and time overhead of preprocessing.

[0018] 2. The present invention trains the model through a deep generation method, does not rely on a specific data set, does not require data labeling, and does not need to use the annotations of the original data set as training labels. Instead, the signal itself is used as supervision for the deep generation model. The trained model has good generalization ability, can adapt to different RF fingerprint application requirements, and can achieve good recognition effects.

[0019] 3. The present invention adopts a memory vector-based deep learning grid to construct a deep learning model. On the basis of ensuring the diversity and sparsity of features in the memory, the latent variable z obtained by encoding each different signal has good specificity, maximizes the difference between the signal latent variable z and the historical feature vector in the memory, and introduces a loss function in training to perform back-propagation optimization on the network parameters. Compared with other existing technologies, it can avoid problems such as model overfitting caused by excessive similarity or lack of diversity of positive / negative samples in some simulation and synthetic data sets, and solve the problem of dependence on negative samples, reduce the demand for video memory in the training stage, avoid possible mode collapse problems, and improve the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the model structure of the memory vector deep generative network in the present invention;

[0021] Figure 2 Schematic diagram of electronic device identity authentication process based on radio frequency fingerprint;

[0022] Figure 3 The invention provides an application scenario of the radio frequency fingerprint recognition method in electronic device identity authentication. DETAILED DESCRIPTION

[0023] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] The radio frequency fingerprint recognition method based on a memory vector deep generative network of the present invention comprises the steps of constructing a co-directional / orthogonal signal data set, preprocessing the data, constructing a deep generative model and training the model, and applying the model to perform radio frequency fingerprint recognition.

[0025] Constructing data sets: Use several data sets based on co-directional / orthogonal signals that are publicly available on the Internet or collected and simulated by yourself to construct the original data set.

[0026] Build a data loader: randomly select data from several data sets in a certain proportion, perform data augmentation on the selected data, add noise, jitter, phase shift, frequency shift, etc., and then randomly crop it to a fixed length.

[0027] Building a deep generative model: Figure 1 As shown in the figure, the deep generative network proposed in the present invention includes three parts: an encoder, a memory and a decoder; the encoder is used to compress the original in-phase / orthogonal signal and extract its key features to obtain latent variables in a high-dimensional space; the memory is used to store the feature vectors of historical signals, so that the latent variables obtained by encoding each different signal have better specificity while ensuring the diversity and sparsity of the features in the memory; the decoder is used to reconstruct the original signal from the latent variables to ensure that the latent variables obtained by the encoder have better representation capabilities.

[0028] The encoder is composed of modules that can process one-dimensional time series signals, including but not limited to one-dimensional convolutional neural networks, recurrent neural networks (such as RNN, LSTM, GRU, etc.), Transformer, etc.; the decoder is the reverse structure of the encoder, and uses a similar network structure to gradually reconstruct the hidden variables into signals.

[0029] Training process: Use the data loader to take N samples from the data set to form a batch of training data. For any data in the batch, use the encoder to compress it into a high-dimensional space to obtain the latent variable z; then compare the latent variable z with the feature vector in the memory, and use the feature of the latent variable z to update the closest feature in the memory; finally, use the decoder to complete the reconstruction of the signal.

[0030] Reconstruction loss, feature similarity loss, and feature sparsity loss are introduced in the training to ensure the specificity and representativeness of the generated latent variables. rec Using L2 distance, feature similarity loss L sim Using cosine similarity, feature sparsity loss L spr Use L2 regularization with parameter λ sim and λ spr To balance the effects of various loss functions, the typical value λ is taken in the experiment sim =0.02,λ spr =0.001, the total loss function is:

[0031] L=L rec -λ sim L sim +λspr L spr ;

[0032] The loss function proposed in the present invention is used to perform back propagation optimization on network parameters, and the optimizer that can be used includes but is not limited to SGD, Adam and other optimizers.

[0033] Testing phase: Construct a radiation source individual identification data set, sample the signals in the data set and input them into the encoder network Encoder of the deep generative model, and calculate the mean similarity measure D1 of the individual's RF fingerprint and the mean similarity measure D2 of the RF fingerprints of different individuals. The tolerance can be taken as ρ = (D1 + min(D2)) / 2. This tolerance value can be used in the subsequent application phase to determine whether two RF fingerprints belong to the same individual.

[0034] The similarity measurement may use Euclidean distance, Mahalanobis distance, correlation distance or information entropy, etc.

[0035] Application stage: Take the encoder network of the deep generative model as the feature extraction network, and encapsulate its network structure and parameters into the application scenario, such as Figure 3 As shown, it can be used as a radio frequency fingerprint extractor to identify and authenticate radio frequency fingerprints by collecting and authenticating related radio frequency signals; the process of collecting and authenticating radio frequency signals by the radio frequency fingerprint extractor is as follows Figure 2 As shown, taking the identity authentication of electronic equipment based on RF fingerprint as an example, in the deployment stage, the RF signals of the device in different states are collected multiple times, input into the RF fingerprint extractor, and the output is stored in the legal fingerprint database as the RF fingerprint of the device; in the device authentication stage, the RF signal of the device to be authenticated is collected, and the RF fingerprint is extracted by the RF fingerprint extractor, and compared with the RF fingerprint in the legal fingerprint database of the access device; if the RF fingerprint of the device meets the tolerance, the authentication is passed, otherwise the authentication fails.

[0036] This method constructs a deep generative model based on a deep generative network of memory vectors. A large-scale data set is required for training. Real data, simulation data, other public data sets of co-directional / orthogonal signals, or mixed multiple data sets can be used for simultaneous training. The method described in the present invention does not need to use the annotations of the original data set as training labels, but uses the signal itself as supervision of the deep generative model. In the forward propagation stage of training, a segment of the co-directional / orthogonal signal is used as input, and the segment is compressed using the encoder network Encoder to obtain the hidden variable of the signal segment in the high-dimensional space as a representation of the signal; then the obtained hidden variable is compared with the historical feature vector memorized in the memory to maximize the difference between the hidden variable and the historical feature vector in the memory; finally, the decoder Decoder is used to gradually reconstruct the information of the signal segment from the hidden variable to obtain a reconstructed signal. In the back propagation stage of training, the Euclidean distance and Mahalanobis distance between the original signal and the reconstructed signal can be used as the reconstruction loss function, and the cosine similarity loss and feature sparsity loss can be used as the memory vector diversity loss function. Under the constraint of the loss function, the weights of the deep generative network are updated using the back propagation algorithm. The model trained by the deep generative method of the present invention has good migration ability, can adapt to different radio frequency fingerprint application requirements, and can achieve good recognition effect.

[0037] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A radio frequency fingerprint recognition method based on a memory vector deep generation network, characterized in that: The following steps are involved: Step 1: Use several data sets based on co-directional / orthogonal signals that are publicly available on the Internet or collected and simulated by yourself to construct the original data set; And use the data loader to preprocess the data; Step 2: Build a deep generative model based on a deep generative network, where the deep generative network includes three parts: encoder, memory, and decoder. Step 3: Use the data loader to extract training data from the dataset to train the deep generative model, and introduce the loss function in the training to perform back-propagation optimization on the network parameters; Step 4: Take the encoder network Encoder as the feature extraction network, encapsulate its network structure and parameters into the application scenario as a RF fingerprint extractor to achieve individual identification and authentication of RF devices or intrusion detection.

2. According to claim 1, a radio frequency fingerprint recognition method based on a memory vector deep generation network is characterized in that: The process of preprocessing data using the data loader specifically includes: randomly selecting data from several data sets in a certain proportion, augmenting the selected data, adding noise, jitter, phase shift and frequency shift, etc., and then randomly cropping it to a fixed length.

3. The radio frequency fingerprint recognition method based on a memory vector deep generation network according to claim 1 is characterized in that: In the deep generative network, the encoder is used to compress the original in-phase / orthogonal signal and extract its key features to obtain the latent variable z in the high-dimensional space. The memory is used to store the feature vectors of historical signals. On the basis of ensuring the diversity and sparsity of the features in the memory, the latent variable z obtained by encoding each different signal has good specificity. The decoder is used to reconstruct the original signal from the latent variable z to ensure that the latent variable z obtained by the encoder has good representation ability.

4. The method for radio frequency fingerprint recognition based on a memory vector deep generation network according to claim 1 is characterized in that: The encoder is composed of modules that can process one-dimensional time series signals, including but not limited to one-dimensional convolutional neural networks, recurrent neural networks, and Transformers; the decoder is the reverse structure of the encoder, and uses a network structure similar to the encoder to gradually reconstruct the hidden variables into signals.

5. The method for radio frequency fingerprint recognition based on a memory vector deep generation network according to claim 1 is characterized in that: During model training in step 3, a data loader is used to extract N samples from the data set to form a batch of training data. For any data in the batch, an encoder is used to compress it into a high-dimensional space to obtain a latent variable z. The latent variable z is then compared with the feature vector in the memory, and the feature of the latent variable z is used to update the feature in the memory that is closest to it. Finally, a decoder is used to reconstruct the signal.

6. The method for radio frequency fingerprint recognition based on a memory vector deep generation network according to claim 5 is characterized in that: Reconstruction loss, feature similarity loss and feature sparsity loss are introduced in the training to ensure the specificity and representativeness of the generated latent variable z; Among them, the reconstruction loss L rec Using L2 distance, feature similarity loss L sim Using cosine similarity, feature sparsity loss L spr Use L2 regularization with parameter λ sim and λ spr To balance the effects of various loss functions, the total loss function is: L = L rec -λ sim L sim +λ spr L spr .

7. The method for radio frequency fingerprint recognition based on a memory vector deep generation network according to claim 1 is characterized in that: In step 4, for the access device, during the deployment phase, the RF signal of the access device is collected, and the RF fingerprint is extracted by the RF fingerprint extractor and stored in the database; during the use phase, when the RF signal is detected, its RF fingerprint is extracted and compared with the RF fingerprint in the database. If the RF fingerprint of the device meets the tolerance, the authentication is successful, otherwise the authentication fails.

8. The radio frequency fingerprint recognition method based on a memory vector deep generation network according to claim 7 is characterized in that: The tolerance is obtained in the following manner: constructing an identification data set for an individual radiation source, sampling the signals in the data set and inputting them into a radio frequency fingerprint generator, and calculating the mean similarity measurement D1 of the individual radio frequency fingerprint and the mean similarity measurement D2 of the radio frequency fingerprints of different individuals, and the tolerance is ρ=(D1+min(D2)) / 2.

Images