RF fingerprint recognition method and system based on convolutional attention mechanism and multi-packet inference

Abstract

The invention relates to the fields of communication networks and artificial intelligence, and provides an RF fingerprint recognition method and system based on a convolutional attention mechanism and multi-packet inference. [Solution] The invention includes step 1 capturing the device transmission signal and preprocessing the transmission signal to obtain a spectrogram; step 2 constructing an RF fingerprint recognition model, inputting the obtained spectrogram into the RF fingerprint recognition model for training to obtain a trained RF fingerprint recognition model; and step 3 performing prediction using the trained RF fingerprint recognition model to obtain a final prediction result. The present invention addresses the problems of conventional model fusion mechanisms, such as high computational complexity and redundant module interactions, by adopting a module separation strategy and independently optimizing the convolutional module and the Transformer module, thereby reducing computational complexity and improving model stability.

Description

[Technical Field]

[0001] The present invention relates to an RF fingerprint recognition method and system based on a convolutional-attention mechanism and multi-packet inference, and belongs to the technical fields of communication networks and artificial intelligence. [Background technology]

[0002] Radio Frequency Fingerprinting (RFF) is a crucial technology for ensuring device authentication and security in wireless communication systems. RFF effectively recognizes the identity of a device by extracting unique features generated by the wireless device during signal transmission. Conventional RF fingerprinting methods typically rely on classifiers based on specific feature engineering and statistical methods. These methods often suffer from poor performance and classification accuracy when faced with complex channel conditions, signal noise interference, and similar characteristics between devices. In recent years, with the advancement of deep learning technology, deep learning-based RF fingerprinting models are gradually becoming mainstream. These models can automatically extract time-frequency domain features of signals, significantly improving the accuracy of device classification.

[0003] Currently, the application of deep learning models in the field of RF fingerprint recognition primarily focuses on classical structures such as convolutional neural networks (CNNs) and long-term short-term memory networks (LSTMs). However, a single model may have limitations when dealing with complex signal features. For example, CNNs are good at capturing local features but lack the ability to model global information in time sequences. Recurrent networks like LSTMs can capture long-term dependencies in time sequences, but they are computationally inefficient and struggle to process high-dimensional signal data.

[0004] In response to the aforementioned challenges, Transformers have gradually gained attention as a hotspot in the study of sequenced signal processing due to their powerful global feature modeling capabilities and parallel computing efficiency. However, when directly applying Transformers to RF fingerprint recognition, several challenges still remain, such as the inability to fully utilize the locality of signal features and the potential limitations of the expressive power of input features relative to global modeling capabilities. To address these challenges, there is an urgent need for an RF fingerprint recognition method that can simultaneously capture both local and global dependencies of signals, thereby improving the accuracy of model classification and its robustness against noisy data. [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] This invention provides an RF fingerprint recognition method and system based on a convolutional-attention mechanism and multi-packet inference to address the shortcomings of the prior art, such as the degraded performance of conventional RF fingerprint recognition models in noisy environments, insufficient feature extraction, and the difficulty in balancing global and local features when relying on a single model for complex data. By analyzing in detail the distribution patterns of local and global features of RF signals, this invention proposes RFF-CAT, a deep learning model architecture combining a convolutional neural network (CNN) and a Transformer encoder. This model enhances local feature extraction capabilities by utilizing convolutional layers, extracting local features from the original RF signal, and effectively removing noise and preserving important signal features through layer-by-layer convolutional operations. Meanwhile, the Transformer encoder module enhances sequence modeling capabilities by capturing global dependencies of RF signals through a multi-head self-attention mechanism and combining it with explicit position encoding.

[0006] Regarding module design, the present invention addresses the problems inherent in conventional model fusion mechanisms, such as high computational complexity and redundant module interactions, by employing a module isolation strategy. This reduces computational complexity and improves model stability by independently optimizing the convolutional module and the Transformer module. To further improve model accuracy, the present invention also introduces a multi-packet adaptive fusion method (MPF). This method divides data into multiple packets and performs weighted fusion of different packets using a weighted summation algorithm, thereby improving recognition performance in low signal-to-noise ratio (SNR) environments. [Means for solving the problem]

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

[0008] A first aspect of the present invention is: Step 1 involves capturing the device transmission signal and preprocessing the transmission signal to obtain a spectrogram, Step 2 involves constructing an RF fingerprint recognition model, inputting the obtained spectrogram into the RF fingerprint recognition model for training, and obtaining a trained RF fingerprint recognition model. The RF fingerprint recognition model includes a convolutional layer, a position encoding layer, and a Transformer encoder. The convolutional layer transforms the input feature map into a higher-order feature representation. The position encoding layer is used to add positional information (absolute or relative position), giving the model the ability to process sequential information. The Transformer encoder includes a multi-head self-attention mechanism and a feedforward neural network (FFN) layer. The multi-head self-attention mechanism captures global dependencies across the entire sequence, and the feedforward neural network layer further enhances feature representation capabilities, as described in step 2. Step 3 of realizing device authentication and recognition is included, which uses a trained RF fingerprint recognition model to make predictions and combines with a multi-packet inference method to obtain a final prediction result. A convolutional-attention mechanism and an RF fingerprint recognition method based on multi-packet inference are provided.

[0009] According to the present invention, preferably, the step of capturing a device transmission signal and preprocessing the transmission signal to obtain a spectrogram is First, collect a transmission signal from a LoRa (registered trademark) device using the LoRa (registered trademark) protocol and record it as r(n); Then, as shown in the following formula, it is the step of calculating the carrier frequency offset (CFO),

Equation

Equation

Equation

number

[0010] According to the present invention, preferably, the steps of constructing an RF fingerprint recognition model, inputting the obtained spectrogram into the RF fingerprint recognition model for training, and obtaining a trained RF fingerprint recognition model include the following: The convolutional layers include normalization layers, point convolutional layers, GLU activation function layers, deep separable convolutional layers, batch normalization layers, Swish activation function layers, cascaded convolutional layers, and Dropout layers. First, the obtained spectrogram is input to a convolutional layer, the input spectrogram is normalized to a mean of 0 and a variance of 1 to improve training stability and obtain spectral features with a more uniform distribution, maintaining the normalized output of the original structure, the channel dimension is adjusted using 1x1 convolutions by point convolutional layers, feature information is merged or expanded to obtain a feature map with an adjusted number of channels, important information in the feature map is enhanced by a GLU activation function layer and redundant information in the feature map is suppressed, the number of parameters in the feature map is reduced by a deep separable convolutional layer and spatial features are enhanced, a feature map with a stable distribution is obtained by a batch normalization layer, an activated feature map is obtained by a Swish activation function layer, small gradient information in the negative region is preserved, a final feature representation with an optimized number of channels is obtained by cascaded convolutional layers, and finally, a normalized feature map X is obtained by a Dropout layer. The module extracts low-level local features from the signal, such as spectrogram edges and textures, and these features are useful for subsequent pattern recognition of the signal. Subsequently, to preserve temporal and sequential information, a positional encoding layer is introduced after the convolutional layer operations to obtain a positionally encoded feature map, appropriately adjusting the output of the convolutional layer and capturing the temporal relationships present in the data. Specifically, positional embeddings are added to the output of the convolutional layer as shown in the following equation. PE(X) = X + Embedding(pos) Here, PE(X) represents the position-encoded feature map, X represents the output of the convolutional layer, and Embedding(pos) is the position encoding. This explicit encoding helps the model understand the order and dependencies of features in the sequence, which is a crucial element for time-sequence analysis and signal processing tasks. Finally, the position-encoded feature map is input into the Transformer encoder, and the Transformer encoder uses the multi-head self-attention mechanism to capture global dependencies across the entire sequence and simulate long-range relationships between features, enabling the model to understand the context and global patterns essential for accurate classification. Specifically, the operation is represented by the following equation: head i =Attention(W i Q X1, W i K X1, W i V X1) Here, X1 is the input, i.e., the position-encoded feature map, and W i Q is the query matrix of the i-th attention head, which is used to map the input features to query vectors, and W i K is the key matrix of the i-th attention head, which is used to map the input features to key vectors, and W i V is the value matrix of the i-th attention head, which is used to map the input features to value vectors. h attention heads generate the output of the final multi-head self-attention mechanism, and the output is cascaded and linearly transformed as shown in the following equation: MHA(X1)=W O (head1||... || head h ) Here, MHA(X1) represents the output of the multi-head self-attention mechanism, and W OThis maps the integrated output of the multi-head self-attention mechanism, which is the output weight matrix, and verifies that it conforms to the final output dimension of the Transformer model. The attention mechanism allows the model to selectively focus on important features located at different positions within the sequence, thereby more effectively simulating global context and dependencies, and the feedforward neural network layer further refines the feature representation. After processing by a Transformer encoder, a one-dimensional global mean pooling layer compresses the time dimension of the feature map into a fixed-length feature vector. Finally, the feature vector is transmitted to fully connected and softmax layers for classification and training to obtain a trained RFF-CAT model, an RF fingerprint recognition model. The RFF-CAT model is designed to perform local and global feature extraction independently, with convolutional layers used only to capture local features, and the Transformer encoder excels at simulating global dependencies. By separating these tasks, the model reduces computational complexity, increases flexibility, and improves performance in noisy environments. Compared to conventional Conformer models (which typically fuse CNN and Transformer operations into a single module), this separation simplifies the model architecture and makes it easier to interpret. Here, the training of the RFF-CAT model is as follows: 1) Minimize the cross-entropy loss function using the RMSProp optimization algorithm. 2) Batch processing of training data and performing multiple iterations of training until convergence occurs, 3) In the learning process, adopt a process that includes introducing a learning rate scheduling strategy to improve learning efficiency.

[0011] According to the present invention, preferably, the step of making a prediction using a trained RF fingerprint recognition model and combining it with a multi-packet inference method to obtain the final prediction result is: In the predictive inference stage, the RF signal data from an unknown device is first preprocessed to convert it into a channel-independent spectrogram, and then the spectrogram is input into a trained RFF-CAT model, wherein the output of the Softmax layer of the RFF-CAT model is a probability vector, where each element represents the confidence level associated with one of the devices, and the probability vector represents the likelihood that a data packet belongs to each device, thereby achieving accurate and reliable device recognition. This includes the step of obtaining the final output probability from the output obtained by the RFF-CAT model using a multi-packet inference method, i.e., an adaptive fusion method (MPF), The Multi-Packet Inference Method (MPF) is, The steps include dividing the received RF signal data into multiple data packets, calculating an initial predicted value for each data packet, i.e., an initial predicted value obtained by the RFF-CAT model, defining an initial weight set, and then performing weighted summation to obtain the fused prediction probability shown in the following formula,

number

number

[0012] A computer device comprising memory and a processor, wherein a computer program is stored in the memory, and the processor, upon executing the computer program, realizes steps of an RF fingerprint recognition method based on a convolutional attention mechanism and multi-packet inference.

[0013] A computer-readable storage medium in which a computer program is stored, wherein, when executed by a processor, the computer program implements steps of an RF fingerprint recognition method based on a convolutional-attention mechanism and multi-packet inference.

[0014] A second aspect of the present invention is: A preprocessing module configured to capture the device transmission signal, preprocess the transmission signal, and obtain a spectrogram, A model training module configured to construct an RF fingerprint recognition model, input the obtained spectrogram into the RF fingerprint recognition model for training, and obtain a trained RF fingerprint recognition model, The RF fingerprint recognition model includes a convolutional layer, a position encoding layer, and a Transformer encoder. The convolutional layer transforms the input feature map into a higher-order feature representation. The location encoding layer is used to add location information. The Transformer encoder includes a multi-head self-attention mechanism and a feedforward neural network layer. The multi-head self-attention mechanism captures global dependencies throughout the sequence, and the feedforward neural network layer further enhances feature representation capabilities. The present invention provides an RF fingerprint recognition system based on a convolutional-attention mechanism and multipacket inference, which includes a multipacket inference module configured to perform predictions using a trained RF fingerprint recognition model and combine them with a multipacket inference method to obtain a final prediction result. [Effects of the Invention]

[0015] The beneficial effects of the present invention are as follows:

[0016] 1. This invention models RF fingerprint recognition as a deep learning problem based on convolutional layers and transformer encoders, and proposes a multi-packet adaptive fusion method. In model training, the feature extraction capabilities of the convolutional layer and transformer are combined to effectively improve the accuracy of feature representation and classification of RF signals.

[0017] 2. RFF-CAT is a deep learning model that exhibits higher accuracy, stability, and robustness when processing complex spatiotemporal data and high-dimensional feature maps, making it promising for a wide range of applications. This invention demonstrates excellent performance even in low signal-to-noise ratio environments, significantly improving the overall performance of distributed RF fingerprint recognition systems and enhancing the recognition accuracy and robustness of the equipment. [Brief explanation of the drawing]

[0018] [Figure 1] This is a schematic diagram of the RF fingerprint recognition system model and workflow of the present invention. [Figure 2] This is a schematic diagram of the RFF-CAT model of the present invention. [Figure 3] These are time-domain plots of the datasets used in the experiments of the present invention. Figure 3(a) is a time-domain plot of the LoRa® device dataset collected when the LoRa® device expansion factor was set to 7. Figure 3(b) is a time-domain plot of the LoRa® device dataset collected when the LoRa® device expansion factor was set to 8, and Figure 3(c) is a time-domain plot of the LoRa® device dataset collected when the LoRa® device expansion factor was set to 9. [Figure 4] This is a spectrogram generated by preprocessing the dataset used in this invention. [Figure 5] This is an overall comparison chart of the predictive performance models of RFF-CAT and the baseline. [Figure 6] This is a schematic diagram comparing the predictive performance of the present invention at different signal-to-noise ratios under MPF conditions, and the inference based on the average value. Figure 6(a) is a schematic diagram of the accuracy with and without MPF when the signal-to-noise ratio is 0, Figure 6(b) is a schematic diagram of the accuracy with and without MPF when the signal-to-noise ratio is 5, Figure 6(c) is a schematic diagram of the accuracy with and without MPF when the signal-to-noise ratio is 10, and Figure 6(d) is a schematic diagram of the accuracy with and without MPF when the signal-to-noise ratio is 20. [Figure 7] This is a schematic diagram illustrating the classification between the MPF method (weighted sum) used in this invention and conventional methods based on average values. [Modes for carrying out the invention]

[0019] The present invention will be described in more detail below with reference to examples and drawings, but the present invention is not limited thereto. Example 1

[0020] The RF fingerprint recognition method based on a convolutional attention mechanism and multi-packet inference is shown in Figure 1, Step 1 involves capturing the device transmission signal and preprocessing the transmission signal to obtain a spectrogram, Step 2 involves constructing an RF fingerprint recognition model, inputting the obtained spectrogram into the RF fingerprint recognition model for training, and obtaining a trained RF fingerprint recognition model. The RF fingerprint recognition model includes a convolutional layer, a position encoding layer, and a Transformer encoder. The convolutional layer transforms the input feature map into a higher-order feature representation. The position encoding layer is used to add positional information (absolute or relative position), giving the model the ability to process sequential information. The Transformer encoder includes a multi-head self-attention mechanism and a feedforward neural network (FFN) layer. The multi-head self-attention mechanism captures global dependencies across the entire sequence, and the feedforward neural network layer further enhances feature representation capabilities, as described in step 2. Step 3 includes performing predictions using a trained RF fingerprint recognition model and combining them with a multi-packet inference method to obtain the final prediction result, thereby achieving device authentication and recognition. Example 2

[0021] The differences between the RF fingerprint recognition method based on the convolution-attention mechanism and multi-packet inference described in Example 1 are as follows:

[0022] The time-domain plots of the transmitted signals are shown in Figure 3, where Figure 3(a) is a time-domain plot of the LoRa® device dataset collected when the LoRa® device expansion factor is set to 7, Figure 3(b) is a time-domain plot of the LoRa® device dataset collected when the LoRa® device expansion factor is set to 8, and Figure 3(c) is a time-domain plot of the LoRa® device dataset collected when the LoRa® device expansion factor is set to 9. The steps of capturing the device transmission signal and preprocessing the transmission signal to obtain a spectrogram are as follows: First, the transmission signal is collected from a LoRa(registered trademark) device using the LoRa(registered trademark) protocol and recorded as r(n), LoRa® devices refer to terminal devices that use LoRa® (Long Range) modulation technology for wireless communication. These devices typically offer low power consumption, long-range, and low-speed data transmission capabilities, and are widely used in IoT scenarios such as smart meter reading, environmental monitoring, agricultural IoT, and smart cities. LoRa® devices include sensor nodes, gateways, and modules, and are suitable for deployment in energy-efficient application environments because they can operate for extended periods without frequent charging. The LoRa® protocol typically refers to the LoRa® WAN (LoRa® Wide Area Network) protocol, an open, low-power wide-area network communication protocol built on LoRa® physical layer technology. LoRa® WAN employs a star topology structure, supports hierarchical device management, security authentication, and end-to-end encryption mechanisms, and offers advantages such as low power consumption, long-distance transmission, and numerous connections, making it particularly suitable for application scenarios involving intermittent small-data transmission in remote locations. Subsequently, the carrier frequency offset (CFO) is calculated as shown in the following formula,

number

number

number

number

[0023] The steps of constructing an RF fingerprint recognition model, inputting the obtained spectrogram into the RF fingerprint recognition model for training, and obtaining a trained RF fingerprint recognition model include the following, as shown in Figure 2: The convolutional layers include normalization layers, point convolutional layers, GLU activation function layers, deep separable convolutional layers, batch normalization layers, Swish activation function layers, cascaded convolutional layers, and Dropout layers. First, the obtained spectrogram is input to a convolutional layer, the input spectrogram is normalized to a mean of 0 and a variance of 1 to improve training stability and obtain spectral features with a more uniform distribution, maintaining the normalized output of the original structure, the channel dimension is adjusted using 1x1 convolutions by point convolutional layers, feature information is merged or expanded to obtain a feature map with an adjusted number of channels, important information in the feature map is enhanced by a GLU activation function layer and redundant information in the feature map is suppressed, the number of parameters in the feature map is reduced by a deep separable convolutional layer and spatial features are enhanced, a feature map with a stable distribution is obtained by a batch normalization layer, an activated feature map is obtained by a Swish activation function layer, small gradient information in the negative region is preserved, a final feature representation with an optimized number of channels is obtained by cascaded convolutional layers, and finally, a normalized feature map X is obtained by a Dropout layer. The module extracts low-level local features from the signal, such as spectrogram edges and textures, and these features are useful for subsequent pattern recognition of the signal. Subsequently, to preserve temporal and sequential information, a positional encoding layer is introduced after the convolutional layer operations to obtain a positionally encoded feature map, appropriately adjusting the output of the convolutional layer and capturing the temporal relationships present in the data. Specifically, positional embeddings are added to the output of the convolutional layer as shown in the following equation. PE(X) = X + Embedding(pos) Here, PE(X) represents the position-encoded feature map, X represents the output of the convolutional layer, and Embedding(pos) is the position encoding. This explicit encoding helps the model understand the order and dependencies of features in the sequence, which is a crucial element for time-sequence analysis and signal processing tasks. Finally, the position-encoded feature map is input to a Transformer encoder, i.e., a stacked multi-module, which uses a multi-head self-attention mechanism to capture global dependencies across the entire sequence and simulate long-range relationships between features, enabling the model to understand the context and global patterns essential for accurate classification. Specifically, the operation is shown by the following equation: head i =Attention(W i Q X1, W i K X1, W i V X1) Here, X1 is the input, i.e., the location-encoded feature map, and W i Q is the query matrix of the i-th attention head, used to map input features to query vectors; W i K is the key matrix of the i-th attention head, used to map the input features to the key vector; W i V is the value matrix of the i-th attention head, used to map the input features to a value vector, and h attention heads generate the output of the final multi-head self-attention mechanism, which is cascaded and linearly transformed as shown in the following equation. MHA(X1) = W O (head1|| ... || head h ) Here, MHA(X1) represents the output of the multi-head self-attention mechanism, WO This maps the integrated output of the multi-head self-attention mechanism, which is the output weight matrix, and verifies that it conforms to the final output dimension of the Transformer model. The attention mechanism allows the model to selectively focus on important features located at different positions within the sequence, thereby more effectively simulating global context and dependencies, and the feedforward neural network layer further refines the feature representation. After processing by a Transformer encoder, a one-dimensional global mean pooling layer compresses the time dimension of the feature map into a fixed-length feature vector. Finally, the feature vector is transmitted to fully connected and softmax layers for classification and training to obtain a trained RFF-CAT model, an RF fingerprint recognition model. The RFF-CAT model is designed to perform local and global feature extraction independently, with convolutional layers used only to capture local features, and the Transformer encoder excels at simulating global dependencies. By separating these tasks, the model reduces computational complexity, increases flexibility, and improves performance in noisy environments. Compared to conventional Conformer models (which typically fuse CNN and Transformer operations into a single module), this separation simplifies the model architecture and makes it easier to interpret. Here, the training of the RFF-CAT model is as follows: 1) Minimize the cross-entropy loss function using the RMSProp optimization algorithm. 2) Batch processing of training data and performing multiple iterations of training until convergence occurs, 3) In the learning process, the following processes will be adopted, including the introduction of a learning rate scheduling strategy, in order to improve learning efficiency.

[0024] The steps involve using a trained RF fingerprint recognition model to make predictions and combining them with a multi-packet inference method to obtain the final prediction result. In the predictive inference stage, the RF signal data from an unknown device is first preprocessed to convert it into a channel-independent spectrogram, as shown in Figure 4. Next, the spectrogram is input to a trained RFF-CAT model, where the output of the Softmax layer of the RFF-CAT model is a probability vector, where each element represents the confidence level associated with one of the devices. The probability vector represents the likelihood that a data packet belongs to each device, thereby achieving accurate and reliable device recognition. The step includes obtaining the final output probability from the output obtained by the RFF-CAT model using a multipacket inference method, i.e., an adaptive fusion method (MPF), wherein the multipacket inference method (MPF) is: The steps include dividing the received RF signal data into multiple data packets, calculating an initial predicted value for each data packet, i.e., an initial predicted value obtained by the RFF-CAT model, defining an initial weight set, and then performing weighted summation to obtain the fused prediction probability shown in the following formula,

number

number

[0025] After the training of the entire model is complete, the final step of using the trained model to classify signal samples from an unknown instrument is further included; the step of evaluating the model's accuracy, recall, and F1 score based on the classification results is also included; and the step of storing the evaluation results and periodically updating the model to adapt to changes in the characteristics of new instruments.

[0026] The results of the proposed RF fingerprint recognition and multi-packet fusion methods based on convolutional neural networks and Transformer encoders are shown in Figures 5, 6, and 7. Figure 5 compares the classification accuracy of the RFF-CAT model, Transformer model, convolutional neural network model, and Transformer encoder model at different signal-to-noise ratio levels. These classification accuracies were obtained without the use of MPF. As shown in the figure, at low SNR levels, the classification accuracy obtained using the RFF-CAT model is approximately 10% higher than that obtained using the Transformer model. At high signal-to-noise ratio levels, the classification accuracy of the RFF-CAT model is approximately 4-5% higher than that of the Transformer model. Furthermore, the classification accuracy obtained using the RFF-CAT model is generally approximately 10% higher than that obtained using the Transformer model. RFF-CAT is particularly useful at low signal-to-noise ratio levels because it uses convolutional operations to capture local patterns and features. Experimental results obtained from simulations using the proposed RFF-CAT trained model demonstrate the excellent performance of the model. Figure 6(a) is a schematic diagram of the accuracy with and without MPF when the signal-to-noise ratio is 0, Figure 6(b) is a schematic diagram of the accuracy with and without MPF when the signal-to-noise ratio is 5, Figure 6(c) is a schematic diagram of the accuracy with and without MPF when the signal-to-noise ratio is 10, and Figure 6(d) is a schematic diagram of the accuracy with and without MPF when the signal-to-noise ratio is 20. As can be seen from Figures 6 and 7, the MPF method of the present invention can significantly improve the performance of RFF, especially in scenarios with a low signal-to-noise ratio. Example 3

[0027] The computer device includes memory and a processor, the memory storing a computer program, and the processor, upon executing the computer program, realizes the steps of the RF fingerprint recognition method based on the convolutional-attention mechanism and multi-packet inference described in Embodiment 1 or 2. Example 4

[0028] A computer-readable storage medium stores a computer program, which, when executed by a processor, realizes the steps of the RF fingerprint recognition method based on the convolution-attention mechanism and multi-packet inference described in Embodiment 1 or 2. Example 5

[0029] The preprocessing module is configured to capture the instrument transmission signal, preprocess the transmission signal, and obtain a spectrogram. The model training module is configured to construct an RF fingerprint recognition model, input the resulting spectrogram into the RF fingerprint recognition model for training, and obtain a trained RF fingerprint recognition model. The RF fingerprint recognition model includes a convolutional layer, a position encoding layer, and a Transformer encoder. The convolutional layer transforms the input feature map into a higher-order feature representation. The location encoding layer is used to add location information. The Transformer encoder includes a multi-head self-attention mechanism and a feedforward neural network layer. The multi-head self-attention mechanism captures global dependencies across the entire sequence, and the feedforward neural network layer further enhances feature representation capabilities. The multi-packet inference module is configured to perform predictions using a trained RF fingerprint recognition model and combine them with a multi-packet inference method to obtain the final prediction result.

Claims

1. A convolution-attention mechanism and multi-packet inference-based RF fingerprint recognition method, Step 1 involves capturing the device transmission signal and preprocessing the transmission signal to obtain a spectrogram, Step 2 involves constructing an RF fingerprint recognition model, inputting the obtained spectrogram into the RF fingerprint recognition model for training, and obtaining a trained RF fingerprint recognition model. The RF fingerprint recognition model includes a convolutional layer, a position encoding layer, and a Transformer encoder. The convolutional layer transforms the input feature map into a higher-order feature representation. The location encoding layer is used to add location information. The Transformer encoder includes a multi-head self-attention mechanism and a feedforward neural network layer. The multi-head self-attention mechanism captures global dependencies across the entire sequence, and the feedforward network layer enhances feature representation by adding global dependencies to local features extracted by the convolutional layer. The convolutional layer includes a normalization layer, a point convolutional layer, a GLU activation function layer, a deep separable convolutional layer, a batch normalization layer, a Swish activation function layer, a cascade convolutional layer, and a Dropout layer. First, the obtained spectrogram is input to a convolutional layer, the input spectrogram is normalized, the channel dimension is adjusted using 1x1 convolution by a point convolutional layer to obtain a feature map with an adjusted number of channels, redundant information in the feature map is suppressed by a GLU activation function layer, the number of parameters in the feature map is reduced by a deep separable convolutional layer and spatial features are emphasized, a feature map with a normalized mean and variance is obtained by a batch normalization layer to obtain a feature map with a uniform distribution, an activated feature map is obtained by a Swish activation function layer, small gradient information in the negative region is preserved, a feature representation with an optimized number of channels is obtained by a cascaded convolutional layer, and finally, a normalized feature map X is obtained by a Dropout layer. Subsequently, after the convolutional layer operations, a position encoding layer is introduced to obtain a position-encoded feature map, capturing the temporal relationships present in the data. Finally, the position-encoded feature map is input to the Transformer encoder, which uses a multi-head self-attention mechanism to capture global dependencies across the entire sequence. Step 2 involves processing with a Transformer encoder, compressing the time dimension of the feature map into a fixed-length feature vector using a one-dimensional global average pooling layer, and finally transmitting the feature vector to a fully connected layer and a softmax layer for classification and training to obtain an RF fingerprint recognition model, which is an RFF-CAT model, a deep learning model architecture combining a convolutional layer and a Transformer encoder. Step 3 includes making predictions using a trained RF fingerprint recognition model and combining them with a multi-packet inference method to obtain the final prediction result, The multi-packet inference method is, The process involves dividing the received RF signal data into multiple data packets, calculating an initial predicted value for each data packet (i.e., the initial predicted value obtained by the RFF-CAT model), defining an initial weight set, and then performing weighted summation to obtain the fused prediction probability. The steps include: calculating the error between the initial predicted value and the fusion predicted probability, and updating the weights to match the predicted value of each data packet; The process includes the step of obtaining a new fused prediction probability by re-adding the weighted predictions of each data packet using the updated weights, thereby achieving device authentication and recognition, i.e., recognizing the RF fingerprint of the corresponding device. A method for recognizing RF fingerprints based on a convolutional attention mechanism and multi-packet inference, characterized by the above.

2. The steps of capturing the device transmission signal and preprocessing the transmission signal to obtain a spectrogram are: First, the transmission signal is collected from a LoRa® device using the LoRa® protocol and recorded as r(n). Subsequently, the carrier frequency offset is calculated as shown in the following formula, [Math 1] Here, arg(•) represents the phase of the transmission signal, n is the time variable, △f represents the carrier frequency offset, and the step is... As shown in the following equation, the step of compensating for the carrier frequency offset is as follows: [Math 2] 【number】 Here, r(n) is the received signal, 【number】 This is the compensated signal, where j represents the imaginary unit, Δf represents the CFO carrier frequency offset, and t represents the time variable, with a step of... The step of normalizing the compensated signal to a unit amplitude by data normalization, as shown in the following equation, [Math 3] Here, r ｎｏｒｍ (n) represents the normalized received signal, with a step and Finally, the step is to perform a short-time Fourier transform to convert it into a channel-independent spectrogram, as shown in the following equation, [Math 4] Here, S(f, t) represents the spectrogram, w(n-t) represents a matrix window of length N, where a portion of the signal is selected and transformed from time n to each time t, j represents the imaginary unit, and f represents the frequency, including a step and RF fingerprint recognition method based on the convolution-attention mechanism and multi-packet inference described in claim 1.

3. The steps of constructing an RF fingerprint recognition model, inputting the resulting spectrogram into the RF fingerprint recognition model for training, and obtaining a trained RF fingerprint recognition model are as follows: The step involves introducing a positional encoding layer after the convolutional layer operation to obtain a positionally encoded feature map and capture the temporal relationships present in the data, the formula for which is shown below. PE(X)=X+Embedding(pos) Here, PE(X) represents the position-encoded feature map, X represents the output of the convolutional layer, and Embedding(pos) is the position encoding, step, The step involves inputting a position-encoded feature map into a Transformer encoder, the operation of which is shown below: head ｉ =Attention(W ｉ Ｑ X １ 、W ｉ Ｋ X １ 、W ｉ Ｖ X １ ) Here, X １ is the input, i.e., the location-encoded feature map, and W ｉ Ｑ is the query matrix of the i-th attention head, W ｉ Ｋ is the key matrix of the i-th attention head, W ｉ Ｖ is the value matrix of the i-th attention head, and h attention heads generate the output of the final multi-head self-attention mechanism, which is cascaded and linearly transformed as shown in the following equation. MHA(X １ )=W Ｏ (head １ || ... || head ｈ ) Here, MHA(X １ ) represents the output of the multi-head self-attention mechanism, W Ｏ This is the output weight matrix, This includes a step of mapping the integrated output of a multi-head self-attention mechanism and further refining the feature representation using a feedforward neural network layer. RF fingerprint recognition method based on the convolution-attention mechanism and multi-packet inference described in claim 2.

4. The steps involve using a trained RF fingerprint recognition model to make predictions and combining them with a multi-packet inference method to obtain the final prediction result. In the predictive inference stage, the RF signal data from an unknown device is first preprocessed to convert it into a channel-independent spectrogram, and then the spectrogram is input into a trained RFF-CAT model. The step includes obtaining the final output probability from the output obtained by the RFF-CAT model using a multi-packet inference method, i.e., an adaptive fusion method. The fusion prediction probability is given by the following formula: [Math 5] Here, p represents the predicted fusion probability, 【number】 represents the predicted probability of the nth data packet, i.e., the initial predicted value, W ｎ is the weight of the predicted probability of the nth data packet, The weight update rule in each iteration is given by the following formula: [Math 6] Here, W ｎ This represents the predicted probability weight array for the data packet at index n, and W ｎ＋１ represents the weight array after updating the weight at index n, μ represents the learning rate, which controls the range of weight adjustment in each iteration, and e ｎ This represents the initial predicted probability, i.e., the error between the initial predicted value and the weighted fusion probability, and p ｎ This is an array containing the initial predicted probabilities of the first N data packets, starting from index n. RF fingerprint recognition method based on the convolution-attention mechanism and multi-packet inference described in claim 3.

5. A computer device comprising memory and a processor, wherein a computer program is stored in the memory, and the processor, upon executing the computer program, realizes the steps of the RF fingerprint recognition method based on a convolutional-attention mechanism and multi-packet inference as described in any one of claims 1 to 4. Computer equipment characterized by the following features.

6. A computer-readable storage medium in which a computer program is stored, wherein, when executed by a processor, the computer program realizes the steps of an RF fingerprint recognition method based on a convolutional-attention mechanism and multi-packet inference as described in any one of claims 1 to 4. A computer-readable storage medium characterized by the following features.

7. A convolutional-attention mechanism and multi-packet inference-based RF fingerprint recognition system, A preprocessing module configured to capture the device transmission signal, preprocess the transmission signal, and obtain a spectrogram, A model training module configured to construct an RF fingerprint recognition model, input the obtained spectrogram into the RF fingerprint recognition model for training, and obtain a trained RF fingerprint recognition model, The RF fingerprint recognition model includes a convolutional layer, a position encoding layer, and a Transformer encoder. The convolutional layer transforms the input feature map into a higher-order feature representation. The location encoding layer is used to add location information. The Transformer encoder includes a multi-head self-attention mechanism and a feedforward neural network layer. The multi-head self-attention mechanism captures global dependencies across the entire sequence, and the feedforward network layer enhances feature representation by adding global dependencies to local features extracted by the convolutional layer. The convolutional layer includes a normalization layer, a point convolutional layer, a GLU activation function layer, a deep separable convolutional layer, a batch normalization layer, a Swish activation function layer, a cascade convolutional layer, and a Dropout layer. First, the obtained spectrogram is input to a convolutional layer, the input spectrogram is normalized, the channel dimension is adjusted using 1x1 convolution by a point convolutional layer to obtain a feature map with an adjusted number of channels, redundant information in the feature map is suppressed by a GLU activation function layer, the number of parameters in the feature map is reduced by a deep separable convolutional layer and spatial features are emphasized, a feature map with a normalized mean and variance is obtained by a batch normalization layer to obtain a feature map with a uniform distribution, an activated feature map is obtained by a Swish activation function layer, small gradient information in the negative region is preserved, a feature representation with an optimized number of channels is obtained by a cascaded convolutional layer, and finally, a normalized feature map X is obtained by a Dropout layer. Subsequently, after the convolutional layer operations, a position encoding layer is introduced to obtain a position-encoded feature map, capturing the temporal relationships present in the data. Finally, the position-encoded feature map is input to the Transformer encoder, which uses a multi-head self-attention mechanism to capture global dependencies across the entire sequence. After processing with a Transformer encoder, a one-dimensional global average pooling layer compresses the time dimension of the feature map into a fixed-length feature vector. Finally, the feature vector is transmitted to a fully connected layer and a softmax layer for classification and training to obtain an RF fingerprint recognition model, which is an RFF-CAT model, a deep learning model architecture combining a convolutional layer and a Transformer encoder. This is the model training module. Includes a multi-packet inference module configured to perform predictions using a trained RF fingerprint recognition model and combine them with a multi-packet inference method to obtain a final prediction result, The multi-packet inference method is, The process involves dividing the received RF signal data into multiple data packets, calculating an initial predicted value for each data packet (i.e., the initial predicted value obtained by the RFF-CAT model), defining an initial weight set, and then performing weighted summation to obtain the fused prediction probability. The steps include: calculating the error between the initial predicted value and the fusion predicted probability, and updating the weights to match the predicted value of each data packet; The process includes the step of obtaining a new fused prediction probability by re-adding the weighted predictions of each data packet using the updated weights, thereby achieving device authentication and recognition, i.e., recognizing the RF fingerprint of the corresponding device. An RF fingerprint recognition system characterized by a convolutional attention mechanism and multi-packet inference.

Images