Unmanned aerial vehicle radio frequency fingerprint identification method based on time-frequency diagram

Abstract

The invention belongs to the technical field of radio frequency identification, and particularly relates to an unmanned aerial vehicle radio frequency fingerprint identification method based on a time-frequency diagram. The unmanned aerial vehicle radio frequency fingerprint identification method based on the time-frequency diagram is provided on the basis of a traditional unmanned aerial vehicle identification technology based on original signals, and the method can extract the time-frequency diagram information of the unmanned aerial vehicles in a real scene in which a plurality of unmanned aerial vehicles fly; identifying the unmanned aerial vehicle type of a signal source by using a deep learning technology; according to the method, the identification performance and the adaptive capacity of an unmanned aerial vehicle identification system can be improved, and the identification problem of the unmanned aerial vehicle under the influence of a multipath effect and a low signal-to-noise ratio is solved to a certain extent.

Description

Technical Field

[0001] The present invention belongs to the technical field of radio frequency identification, and specifically relates to a method for identifying the radio frequency fingerprint of an unmanned aerial vehicle based on a time-frequency graph. Background Art

[0002] With the rapid development of information technology and computer technology, the number and deployment scale of Internet of Things terminals have increased exponentially. However, due to the openness of the wireless channel, device communication is extremely vulnerable to various malicious attacks, such as Rogue Access Point (RAP), MAC address spoofing, and brute-force cracking and password prediction of the 802.1x protocol. These attacks may not only lead to data leakage, but also pose a serious threat to the personal and property safety of users. Therefore, there is an urgent need for a new security mechanism to improve the access authentication of wireless devices, effectively identify authorized and unauthorized devices, and enhance the security of radio frequency device access.

[0003] Due to the manufacturing tolerance and drift tolerance of electronic components, these "defects" constitute the unique characteristics of the signal transmitting device, namely the radio frequency fingerprint. Since these "defects" are only related to the hardware circuit, they are difficult to tamper with and impossible to imitate. Against this background, the radio frequency fingerprint has become a research hotspot in the field of Internet of Things security due to its uniqueness, short-term stability, and measurability. The radio frequency fingerprint acts directly on the physical layer and is a promising solution to solve the identity recognition problem in the Internet of Things, which can effectively improve the effectiveness and reliability of recognition.

[0004] In actual scenarios, different types of unmanned aerial vehicles will generate unique radio frequency fingerprint characteristics due to differences in hardware structure, communication protocol, working frequency band, and signal modulation method during operation. The present invention proposes a radio frequency fingerprint identification system for unmanned aerial vehicles based on a time-frequency graph. By analyzing the signal characteristics of unmanned aerial vehicles, effective identification of multiple types of unmanned aerial vehicles can be achieved. Using neural network technology, by extracting features from the time-frequency graph of unmanned aerial vehicles, efficient and rapid identification of multiple different types of unmanned aerial vehicles can be carried out in the real flight scenario of unmanned aerial vehicles. This technology can, to a certain extent, solve the problem of identifying unmanned aerial vehicles under the influence of multipath effects and low signal-to-noise ratio. Summary of the Invention

[0005] The objective of the present invention is to propose a UAV radio frequency fingerprint recognition system based on time-frequency diagrams with physical-layer anti-spoofing. In traditional radio frequency fingerprint recognition technologies, different fingerprint features are usually extracted from the original signal, such as higher-order moments, fractal dimensions, phase noise, etc. However, these methods either extract features based on time-domain signals or extract frequency-domain features through Fourier transform, without fully utilizing the information combining the time domain and the frequency domain. To solve this problem, the present invention proposes a radio frequency fingerprint recognition scheme based on the time-frequency diagrams of the short-time Fourier transform (STFT). Different types of UAVs will generate unique time-frequency diagram features during operation due to differences in hardware structure, communication protocols, operating frequency bands, and signal modulation methods. By analyzing these signal features, effective identification of UAV types can be achieved.

[0006] Based on the traditional UAV radio frequency signal fingerprint recognition technology, the present invention proposes a radio frequency fingerprint recognition scheme based on time-frequency diagrams and Siamese networks, which further improves the accuracy and speed of UAV identification. It is a scheme for identifying UAV models at the hardware level that is independent of information matching. The frequency bands of UAV signals are usually in the range of 2.4 GHz, 5.2 GHz, or 5.8 GHz, and the bandwidth is usually 10 MHz, 20 MHz, or 40 MHz. These frequency bands are widely used in various types of UAVs. UAV manufacturers may choose different frequency bands and different channels to reduce interference and conflicts. The UAV identification system based on radio frequency fingerprint recognition technology can realize the function of identifying multiple different types of UAVs in a complex environment. This technology can, to a certain extent, solve the problem of UAV identification under the influence of multipath effects and low signal-to-noise ratios.

[0007] The technical solution of the present invention is as follows:

[0008] A method for UAV radio frequency fingerprint recognition based on time-frequency diagrams, comprising the following steps:

[0009] S1. Obtain training data, specifically: collect the UAV original signal s(t) at different distances at the same time. Define the signals collected at two different distances as s a (t) and s b (t), where t represents time; preprocess s a (t) and s b (t) to obtain data s a ′(t) and s b ′(t) suitable for subsequent training. Perform short-time Fourier transform on s a ′(t) and s b ′(t) to obtain the time-frequency diagram sets X a and X b , and select from X a and X bA sample pair is formed by extracting one picture from each where i∈N a ,j∈N b , N a and N b Respectively represent the number of time-frequency graphs generated at distance a and distance b, using and the corresponding category label y c Construct the first dataset D base As training data:

[0010]

[0011] Where C represents different category indexes;

[0012] S2. Build a drone radio frequency fingerprint recognition model based on the twin network, including a first encoder, a second encoder, a first projector, a second projector, and a classifier;

[0013] Extract samples from the training data and input them into the model. middle Input to the first encoder, and the first encoder extracts the image features to obtain a one-dimensional vector Will Input to the second encoder, and the second encoder extracts the image features to obtain a one-dimensional vector

[0014] The one-dimensional vector Input to the first projector for nonlinear mapping to obtain The one-dimensional vector Input to the second projector for nonlinear mapping to obtain a one-dimensional vector

[0015] The classifier is used to classify the input of the projector and output the predicted category label of the drone;

[0016] S3. Use the training data to pre-train the drone RF fingerprint recognition model. Specifically, the model is trained using supervised contrast loss:

[0017] The input sample is x i The output of the projector is represented by z′ i Indicates that the mutual information Info(x i ,z′ i ) and uses it as the loss function to guide model training, thereby obtaining a pre-trained drone RF fingerprint recognition model;

[0018] S4, obtain the target drone signal, use the same method as S1 to s a (t) and sb (t) is processed to obtain the time-frequency graph set X t , combined with the category label y c Construct the second dataset D target :

[0019]

[0020] S5, based on the pre-trained drone radio frequency fingerprint recognition model, retain the encoder parameters, and in the second dataset D target The parameters of the projector and classifier are retrained, and the cross entropy is used as the loss function to fine-tune the model based on the target data. The fine-tuned model is then used to accurately identify the drone's RF fingerprint.

[0021] Furthermore, the first encoder and the second encoder process the input data by passing the input data through a two-dimensional convolutional layer, a ReLU activation layer, and a maximum pooling layer in sequence, and finally flattening it into a one-dimensional vector.

[0022] Furthermore, the loss function of the supervised contrast loss in S3 is expressed as:

[0023]

[0024] I={1,2,…,|D base |}

[0025] A(i)={j|j≠i,j∈I}

[0026] P(i)={j∈A(i)|y j =y i}

[0027] Where I represents the index set of the first data set, P(i) represents the index set of positive samples with the same category as the i-th sample, A(i) represents the index set of all samples except the i-th sample itself, and τ is the temperature coefficient.

[0028] Furthermore, the cross entropy loss function in S5 is expressed as:

[0029]

[0030] in, Represents the sample x i Predict its true category y i The predicted probability of .

[0031] The beneficial effects of the present invention are as follows. The present invention proposes a method for identifying the radio frequency fingerprint of an unmanned aerial vehicle (UAV) based on a time-frequency map on the basis of the traditional UAV identification technology based on the original signal. This method can extract the time-frequency map information of the UAV in a real scenario where multiple UAVs are flying and use deep learning technology to identify the type of UAV from which the signal source comes. The present invention can improve the identification performance and adaptability of the UAV identification system and, to a certain extent, solve the problem of UAV identification under the influence of multipath effects and in the case of low signal-to-noise ratio. The identification accuracy on the target dataset is as Figure 7 shown. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is the identification flow chart of the UAV radio frequency fingerprint identification system based on the time-frequency map.

[0033] Figure 2 is the network architecture of the Siamese network based on the supervised contrastive loss.

[0034] Figure 3 is the time-frequency map of the UAV signal collected at distances a and b.

[0035] Figure 4 is the schematic diagram of the encoder.

[0036] Figure 5 is the schematic diagram of the projector.

[0037] Figure 6 is the visualization of the representation in the pre-training stage

[0038] Figure 7 is the UAV identification accuracy of the target dataset. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] The present invention will be described in detail below with reference to the accompanying drawings.

[0040] The overall flow chart of the present invention is as Figure 1 shown. First, it is necessary to collect the same UAV signal at two different distances a and b in advance to form the base class dataset D base ; after preprocessing the base class dataset, the original signal is converted into a set of time-frequency maps X a and X b through the short-time Fourier transform; then the time-frequency maps and are fed into the Siamese network for pre-training so that the network can learn how to extract the time-frequency map features. For the UAV signal to be identified collected in the real scenario, the target dataset D target is formed; then the time-frequency map X t is generated in the same way, and the model is fine-tuned on the above-mentioned pre-trained Siamese network so that the model can quickly adapt to the target dataset, thereby improving the performance and adaptability of the UAV identification system.

[0041] The key to this invention is the design of a twin network that can effectively identify the type of drone. The network architecture of the twin network based on supervised contrast loss is as follows: Figure 2 As shown in the figure, during the pre-training phase, the input base class dataset is passed through the encoder and projector to obtain the corresponding representation vector. Then, supervised contrastive learning is used to guide the model to learn how to extract time-frequency graph features. During the fine-tuning phase, the parameters learned by the encoder during the pre-training phase are retained, and the projector and classifier are fine-tuned on the target dataset. The following is an explanation of the specific steps:

[0042] 1) Data collection and preprocessing

[0043] (1) Collection at different distances

[0044] The original signal s(t) of the drone is collected at different distances at the same time, and the signal collected at two different distances is defined as s a (t) and s b (t), where t represents time.

[0045] (2) Data preprocessing

[0046] 1. Perform preprocessing operations on the collected signal, including clipping, power normalization, and moving to zero frequency, etc., recorded as Preprocess(·), and obtain the preprocessed signal s a ′(t) and s b ′(t).

[0047] s a ′(t)=Preprocess(s a (t))#(1)

[0048] s b′ (t)=preprocess(s b (t))#(2)

[0049] 2. Preprocessed signal s a ′(t) and s b ′(t) is short-time Fourier transformed to obtain the time-frequency graph set X a and X b , where N a and N b represents the number of time-frequency maps generated at distance a and distance b, respectively. Compared to the original signal, time-frequency maps can capture both the time and frequency domain characteristics of the signal and are more resistant to noise. Furthermore, as two-dimensional images, time-frequency maps are more suitable for extracting complex features using convolutional neural networks (CNNs). Figure 3 Shows the time-frequency graph set X a and X b One picture each in xa and x b , we can see that due to the different collection distances, the signal strength and degree of interference of the two are different.

[0050]

[0051] 3. Based on the time-frequency graph set X a and X b Construct base class dataset D base , the base class dataset contains sample pairs and its category label y c , where c=1,2,…,C represents different category indices. By constructing sample pairs collected at different distances It enables the model to learn how to remove noise and frequency deviation and better extract time-frequency graph representation.

[0052]

[0053] 4. For the drone signals to be identified collected in real scenes, construct the target dataset D target First, the same method as above is used to obtain the time-frequency graph set X for the UAV signal t , and according to the category label y c Construct the target dataset D target For this dataset, it is no longer necessary to pair samples.

[0054]

[0055] 2) Pre-training stage

[0056] The drone RF fingerprint recognition model is constructed based on the twin network. Two sub-networks with the same structure extract the two inputs, specifically including the first encoder, the second encoder, the first projector, the second projector and the classifier, such as Figure 2 shown.

[0057] 1. The first encoder and the second encoder: extract samples from the training data and input them into the model. middle Input to the first encoder, and the first encoder extracts the image features to obtain a one-dimensional vector Will Input to the second encoder, and the second encoder extracts the image features to obtain a one-dimensional vector For each encoder, a convolutional neural network (CNN) is used as the encoder to extract image features, such as Figure 4 Specifically, the input data passes through a two-dimensional convolutional layer, a ReLU activation layer, and a maximum pooling layer in sequence, and is finally flattened into a one-dimensional vector.

[0058] 2. First projector, second projector: transform one-dimensional vector Input to the first projector for nonlinear mapping to obtain The one-dimensional vector Input to the second projector for nonlinear mapping to obtain a one-dimensional vector like Figure 5 As shown in the figure, the low-dimensional simple feature representation in the source space is transformed into a high-dimensional complex feature representation through the linear layer and the ReLU activation layer. This can remove redundant features, thereby retaining more useful information and increasing the expressive power of the model.

[0059] 3. Classifier: Define a fully connected layer as the model classifier, and the input is the representation z′ obtained after the projector i , output the predicted category label y of the drone c The classifier is mainly used in the subsequent fine-tuning stage to train the classifier based on the predicted label information and guide the model to output the correct classification results.

[0060] 3. Supervised contrast loss: In the pre-training stage, the supervised contrast loss is defined based on the characteristics of the twin network to train the first encoder, the second encoder, the first projector, and the second projector. Specifically, for the representation obtained after the encoder and projector and According to the information maximization (InfoMax) principle, the input time-frequency x is maximized i Instead of representing z′ i The mutual information Info(x i ,z′ i ), thus guiding the representation extracted by the model to retain more original information. i ,z′ i ) is a theoretical information quantity index and is difficult to calculate directly. Therefore, the mutual information Info(x i ,z′ i ) and uses it as the loss function to guide model training.

[0061] The present invention uses a version based on supervised contrast loss. Where I represents the index set of the base class dataset, P(i) represents the index set of positive samples with the same category as the i-th sample, A(i) represents the index set of all samples except the i-th sample itself, and τ is the temperature coefficient. Compared with the unsupervised contrast loss, the supervised contrast loss considers the label information and can better align similar samples in the embedding space. The pre-trained representation z′ is reduced by the t-SNE method. i Visualize, such as Figure 6As shown in Figure 2, we can see that the supervised contrast loss effectively brings samples of the same type closer together in the feature space, while moving samples of different types further away from each other.

[0062]

[0063] I={1,2,…,|D base |}#(9)

[0064] A(i)={j|j≠i,j∈I}#(10)

[0065] P(i)={j∈A(i)|y j =y i}#(11)

[0066] 3) Fine-tuning stage

[0067] In the fine-tuning phase, the parameters learned by the projector in the pre-training phase are no longer retained, only the encoder parameters are retained. target , fix the encoder parameters, use the encoder to extract the time-frequency graph representation. Then use the classifier in D target Retrain the parameters of the projector and classifier using cross entropy as the loss function. Represents the sample x i Predict its true category y i The prediction probability of the target UAV signal can be identified.

[0068]

[0069] For the second data set D to be identified target , the present invention only needs to perform simple fine-tuning training on the classifier to obtain better recognition results. The recognition accuracy of drones is as follows Figure 7 shown.

[0070] This paper proposes a UAV radio frequency fingerprint recognition system based on time-frequency graph, the core of which is to accurately identify the UAV model through a twin network.

[0071] The system's implementation begins with data acquisition. RF signals from the same drone are collected at different distances to construct a base dataset. This dataset provides diverse signal samples for training, enabling the model to learn feature variations under varying conditions. After acquisition, the one-dimensional time-domain signal is converted into a two-dimensional time-frequency plot using a short-time Fourier transform (STFT). The generated time-frequency plot preserves both the time and frequency domain information of the signal, providing a rich representation for feature extraction.

[0072] During the pre-training phase, after the time-frequency graph data is generated, it is input into the twin network for pre-training. The twin network design consists of an encoder and a projector, where the encoder is responsible for converting the input time-frequency graph into a high-dimensional representation vector, and the projector further processes the representation vector to extract deeper features. During the pre-training phase, a supervised contrastive learning method is used to guide the model to learn the similarities and differences in signal features by optimizing the contrastive loss function. The focus of model training is to ensure that it can accurately distinguish the characteristics of signals from different types of drones, while identifying the common characteristics of signals of the same model. Through this training process, the twin network learns how to extract effective features from the base class dataset, providing a solid foundation for subsequent model fine-tuning.

[0073] For drone signals to be identified, collected in real-world scenarios, the system further constructs a target dataset and generates time-frequency plots using the same method as the base dataset. To quickly adapt to the new signal environment, the system fine-tunes the pre-trained twin network. During fine-tuning, the learned parameters of the encoder are retained and used to directly transfer and extract the basic features of the target data. The projector and classifier are optimized based on the target dataset, enabling the model to quickly adjust to drone signals in new scenarios, significantly improving recognition accuracy and adaptability. Ultimately, the classifier's output determines the type of drone the signal originated from.

Claims

1. A method for identifying the radio frequency fingerprint of an unmanned aerial vehicle based on a time-frequency diagram, characterized in that, Including the following steps: S1. Obtain training data, specifically: collect the original signal s(t) of the drone at different distances at the same time, and define the signals collected at two different distances as s a (t) and s b (t), where t represents time; For s a (t) and s b (t), preprocess them to obtain data s a′ (t) and s b′ (t) applicable to subsequent training. Perform short-time Fourier transform on s a′ (t) and s b′ (t) to obtain a set of time-frequency diagrams X a and X b . Extract one diagram from each of X a and X b to form a sample pair where i ∈ N a , j ∈ N b , N a and N b represent the number of time-frequency diagrams generated at distances a and b respectively. Use and the corresponding class label y c to construct the first dataset D base as training data: Where C represents different class indices; S2. Construct a UAV radio frequency fingerprint recognition model based on a Siamese network, including a first encoder, a second encoder, a first projector, a second projector, and a classifier; Extract samples from the training data and input them into the model. Specifically, input the sample pair in into the first encoder. After extracting the image features by the first encoder, a one-dimensional vector is obtained. Input into the second encoder. After extracting the image features by the second encoder, a one-dimensional vector Input the one-dimensional vector into the first projector for non-linear mapping to obtain Input the one-dimensional vector into the second projector for non-linear mapping to obtain the one-dimensional vector The classifier is used to classify according to the input of the projector and output the predicted class label of the UAV; S3. Pre-train the UAV radio frequency fingerprint recognition model using training data, specifically by training the model with supervised contrastive loss: Denote the input sample as x i and denote the output of the projector as z′ i Based on the noise contrastive estimation technique, estimate the lower bound of the mutual information Info(x i , z′ i ), and use it as the loss function to guide the model training, so as to obtain the pre-trained UAV radio frequency fingerprint recognition model; S4. Obtain the target UAV signal, and use the processing method for s a (t) and s b (t) to obtain the time-frequency diagram set X t , and then combine with the class label y c to construct the second data set D target : S5. Based on the pre-trained UAV radio frequency fingerprint recognition model, retain the encoder parameters, and retrain the parameters of the projector and classifier on the second dataset D target to fine-tune the model based on the target data using cross-entropy as the loss function, and then use the fine-tuned model to accurately identify the UAV radio frequency fingerprints.

2. The method for identifying the radio frequency fingerprint of an unmanned aerial vehicle based on a time-frequency diagram according to claim 1, wherein The processing process of the first encoder and the second encoder for the input data is to make the input data pass through a two-dimensional convolutional layer, a Relu activation layer, and a max pooling layer in sequence, and finally flatten it into a one-dimensional vector.

3. The method for identifying the radio frequency fingerprint of an unmanned aerial vehicle based on a time-frequency diagram according to claim 1, wherein The loss function of the supervised contrastive loss in S3 is expressed as: I = {1, 2, …, |D base |} A(i) = {j|j≠i, j∈I} P(i) = {j ∈ A(i) | y j = y i} Where I represents the index set of the first data set, P(i) represents the positive sample index set of the same class representation as the i-th sample, A(i) represents the index set of all samples except the i-th sample itself, and τ is the temperature coefficient.

4. The method for identifying the radio frequency fingerprint of an unmanned aerial vehicle based on a time-frequency diagram according to claim 1, characterized in that The cross-entropy loss function in S5 is expressed as: Among them, represents the prediction probability of predicting the sample x i as its true class y i .

Images