Radio frequency fingerprint identification method and device based on enhanced complex convolutional neural network

Abstract

The invention discloses a radio frequency fingerprint identification method, device and equipment based on an enhanced complex convolutional neural network, and relates to the technical field of Internet of Things security, and the method comprises the steps: converting an I / Q signal transmitted by a to-be-identified radio frequency transmitter into a complex value radio frequency signal; and inputting the complex value radio frequency signal and the conjugate form thereof into the enhanced complex convolutional neural network to obtain an identification result of the radio frequency transmitter to be identified. A convolutional layer of a complex-valued convolutional block in the enhanced complex convolutional neural network is a wide linear complex-valued convolutional layer; and the wide linear complex value convolution layer performs convolution operation on the complex value radio frequency signal and the conjugate form thereof through two convolution filters with consistent parameters, and the processing results of the two convolution filters are added as the output of the wide linear complex value convolution layer. According to the method, the characteristics of the complex-valued radio-frequency signal are fully extracted through the wide linear complex-valued convolutional layer; and an SE attention mechanism module is introduced, so that the identification performance of the network model on the radio frequency fingerprint signal is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) security technology, and in particular to a radio frequency fingerprinting method, apparatus, and device based on an enhanced complex convolutional neural network. Background Technology

[0002] Radio frequency (RF) fingerprints refer to the unique signal characteristics generated during the manufacturing process of wireless devices due to differences in hardware technology. These characteristics are reflected in the device's transmitted signals and are non-replicable. Therefore, RF fingerprint recognition technology can achieve highly secure device authentication by extracting and analyzing device characteristics in wireless communication signals, and has significant application value in fields such as IoT security and spectrum management.

[0003] With the rapid development of deep learning, especially the widespread application of convolutional neural networks (CNNs), real-valued CNNs (RV-CNNs), while widely used in image recognition and speech processing, still have significant limitations when processing complex-valued data (such as radar echoes, radio frequency signals, and MRI images). The complex-valued features of radio frequency signals (such as phase noise and frequency shift) are core to device identification. RV-CNNs cannot model the phase modulation characteristics of signals and cannot fully extract these device features, leading to insufficient feature extraction and reduced recognition accuracy.

[0004] In contrast, complex-valued convolutional neural networks (CV-CNNs) can more accurately capture the amplitude and phase features of complex-valued data, thus exhibiting stronger feature extraction capabilities and better recognition performance in radio frequency signal processing tasks. However, because traditional CV-CNNs typically use linear convolutional structures in their convolutional layers, they still cannot fully utilize the complete feature information of complex signals when processing non-circular complex-valued data, thereby limiting the mining of deep features and the network's feature recognition capabilities. Non-circular complex-valued data is non-circular, and its amplitude and phase distributions have specific characteristics. Traditional linear convolutional structures struggle to effectively capture and utilize the second-order statistical properties of non-circular complex-valued data, easily losing feature information related to non-circularity, thus failing to fully extract the features of complex signals. Furthermore, the features in non-circular complex-valued data often have complex nonlinear relationships, while linear convolutional structures are essentially linear operations, with limited ability to express complex nonlinear feature relationships, making it difficult to mine deep, nonlinear features in non-circular complex-valued data, thus limiting the network's ability to learn and recognize deep features of complex-valued data.

[0005] In summary, considering the high complexity of radio frequency fingerprint signals and the subtle feature differences between devices, and the difficulty of traditional CV-CNN in deeply mining the deep features in the signals, the accuracy of radio frequency fingerprint recognition methods based on traditional CV-CNN needs to be improved. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the limitation of the feature recognition capability and accuracy of the existing radio frequency fingerprint recognition network in the prior art because the linear convolution result in the traditional CV-CNN cannot fully utilize the complete feature information of the complex signal.

[0007] To address the aforementioned technical problems, this invention provides a radio frequency fingerprinting method based on an enhanced complex convolutional neural network, comprising: acquiring an I / Q signal transmitted by a radio frequency transmitter to be identified and converting it into a complex-valued radio frequency signal; inputting the complex-valued radio frequency signal and its conjugate form into an enhanced complex convolutional neural network, wherein the enhanced complex convolutional neural network includes multiple complex-valued convolutional blocks, a fully connected layer, and an output layer; the convolutional layer of the complex-valued convolutional blocks is a wide linear complex-valued convolutional layer; the wide linear complex-valued convolutional layer performs convolution operations on the complex-valued radio frequency signal and its conjugate form respectively through two convolutional filters with identical parameters, and adds the processing results of the two convolutional filters as the output of the wide linear complex-valued convolutional layer; and obtaining the identification result of the radio frequency transmitter to be identified based on the output result of the output layer.

[0008] Preferably, the complex-valued convolutional block includes a wide linear complex-valued convolutional layer, a ReLU activation function, a normalization layer, a maximum pooling layer, and an SE attention mechanism module connected along the forward propagation direction.

[0009] Preferably, the output expression of the wide linear complex-valued convolutional layer is: ; in, To enhance the input complex-valued radio frequency signal of a complex convolutional neural network, This is the conjugate form of the complex-valued radio frequency signal. and These are the lengths of the wide linear complex-valued convolutional layers. The first convolution filter and the second convolution filter, This is the conjugate transpose. This is the output of a wide linear complex-valued convolutional layer.

[0010] Preferably, the kernel size of the wide linear complex-valued convolutional layer is 3, and the number of output channels is 64.

[0011] Preferably, the processing steps of the SE attention mechanism module include: Global average pooling is used to perform global feature aggregation on each channel of the input feature map of the SE attention mechanism module to obtain a compressed channel vector. The compressed channel vector is then reduced in dimension by the first fully connected layer and subjected to a nonlinear transformation operation. After being increased in dimension by the second fully connected layer, it is then normalized to output the channel weight vector after activation. The channel weight vector is then multiplied with the input feature map of the SE attention mechanism module channel by channel to obtain the output feature map of the SE attention mechanism module.

[0012] Preferably, the input feature map of the SE attention mechanism module No. Channel vector after compression The expression is: ; in, and These are the input feature maps. Height and width, For the number of channels, Input feature map In spatial location No. The eigenvalues ​​of the channel; Compressed channel vectors The channel weight vector after activation The expression is: ; in, and These are the weight matrices for the first and second fully connected layers, respectively. It is the ReLU activation function. It is the Sigmoid activation function. ; The expression for multiplying the channel weight vector by the input feature map of the SE attention mechanism module channel by channel is as follows: ; in, For the first Channel attention weights Output feature map of the SE attention mechanism module In spatial location No. The characteristic values ​​of the channel.

[0013] Preferably, the enhanced complex convolutional neural network includes nine complex-valued convolutional blocks.

[0014] Preferably, the training process of the enhanced complex convolutional neural network includes: Communication signals transmitted by multiple identical RF transmitters were collected and stored in I / Q format to obtain a sample dataset. The sample dataset was divided into a training set, a validation set, and a test set, which were used for training, parameter tuning, and performance evaluation of the enhanced complex convolutional neural network model, respectively. In each training round, the training set samples were forward propagated to obtain the model output. The classification error between the model output and the true labels of the training set samples was calculated using the cross-entropy loss function. The gradient was calculated using the backpropagation algorithm, and the parameters of the enhanced complex convolutional neural network model after training were saved.

[0015] The present invention also provides a radio frequency fingerprint recognition device based on an enhanced complex convolutional neural network, comprising: The signal acquisition module is used to acquire the I / Q signals sent by the radio frequency transmitter to be identified and convert them into complex radio frequency signals; A network processing module is used to input complex-valued radio frequency signals and their conjugate forms into an enhanced complex convolutional neural network. The enhanced complex convolutional neural network includes multiple complex-valued convolutional blocks, fully connected layers, and an output layer. The convolutional layers of the complex-valued convolutional blocks are wide linear complex-valued convolutional layers. The wide linear complex-valued convolutional layer performs convolution operations on the complex-valued radio frequency signals and their conjugate forms through two convolutional filters with identical parameters, and adds the processing results of the two convolutional filters as the output of the wide linear complex-valued convolutional layer. The result output module is used to obtain the identification result of the radio frequency transmitter to be identified based on the output result of the output layer.

[0016] The present invention also provides a radio frequency fingerprint recognition device based on an enhanced complex convolutional neural network, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described radio frequency fingerprint recognition method based on an enhanced complex convolutional neural network.

[0017] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: The radio frequency fingerprinting method based on enhanced complex convolutional neural networks described in this invention employs complex convolutional blocks with wide linear complex convolutional layers. This allows for better utilization of the amplitude and phase information in complex radio frequency signals, enhancing the model's ability to extract and represent features of complex signals, thereby improving the network model's feature recognition performance. The wide linear complex convolutional layers perform convolution operations on the complex radio frequency signal and its conjugate form using two convolutional filters with identical parameters. This fully utilizes the non-circular characteristics of non-circular complex data, better models the nonlinear relationships in non-circular complex data, and mines deeper features. It also better preserves and utilizes the phase information of complex signals, extracting features of complex signals more completely, and improving the network model's ability to process complex data, thus enhancing the accuracy and efficiency of radio frequency fingerprinting.

[0018] Since wide linear complex-valued convolutional layers require convolution operations on the input signal and its conjugate signal respectively, this introduces twice the number of convolution kernel parameters. In order to balance the performance of the model and the computational efficiency, this invention also introduces an SE attention mechanism module in the complex-valued convolutional block as a lightweight attention mechanism, which can effectively enhance the feature representation ability of the network model, making its feature recognition performance in complex radio frequency environments more excellent. This improves the accuracy of the radio frequency fingerprint recognition method while ensuring recognition efficiency. Attached Figure Description

[0019] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart illustrating the steps of an RFID fingerprint recognition method based on an enhanced complex convolutional neural network provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the enhanced complex convolutional neural network provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the SE attention mechanism module in a complex convolutional block provided in an embodiment of the present invention; Figure 4 This is a visual representation of the semantic features of three network models on the ADS-B dataset with a labeling ratio of 10% and 20%. Figure 4 (a) in the figure shows the semantic feature visualization results of Att-RV-Net under the ADS-B dataset with a label ratio of 10%; Figure 4 (b) in the figure shows the semantic feature visualization results of Att-SL-Net under the ADS-B dataset with a label ratio of 10%; Figure 4 (c) in the figure shows the semantic feature visualization results of Att-WL-Net under the ADS-B dataset with a label ratio of 10%; Figure 4 In the diagram, (d) shows the semantic feature visualization results of Att-RV-Net under the ADS-B dataset with a label ratio of 20%; Figure 4 In the diagram, (e) represents the semantic feature visualization results of Att-SL-Net under the ADS-B dataset with a label ratio of 20%. Figure 4 In the figure, (f) is the semantic feature visualization result of Att-WL-Net under the ADS-B dataset with a label ratio of 20%; Figure 5 This is a schematic diagram of the semantic feature visualization results of three network models under the Wi-Fi (62ft) dataset with a labeling ratio of 10% and 20%; Figure 5(a) shows the semantic feature visualization results of Att-RV-Net under the Wi-Fi (62ft) dataset with a label ratio of 10%; Figure 5 (b) in the figure shows the semantic feature visualization results of Att-SL-Net under the Wi-Fi (62ft) dataset with a label ratio of 10%; Figure 5 (c) in the figure shows the semantic feature visualization results of Att-WL-Net under the Wi-Fi (62ft) dataset with a label ratio of 10%; Figure 5 In the diagram, (d) shows the semantic feature visualization results of Att-RV-Net under the Wi-Fi (62ft) dataset with a label ratio of 20%; Figure 5 (e) in the figure represents the semantic feature visualization results of Att-SL-Net under the Wi-Fi (62ft) dataset with a label ratio of 20%; Figure 5 (f) in the figure is the semantic feature visualization result of Att-WL-Net under the Wi-Fi (62ft) dataset with a label ratio of 20%. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0021] Reference Figure 1 As shown in the figure, the radio frequency fingerprint recognition method based on an enhanced complex convolutional neural network provided by this invention includes the following steps: S1: Construct an enhanced complex convolutional neural network, which includes multiple complex-valued convolutional blocks, fully connected layers, and an output layer; the convolutional layers of the complex-valued convolutional blocks are wide linear complex-valued convolutional layers.

[0022] like Figure 2 As shown, the enhanced complex convolutional neural network provided in this embodiment of the invention includes multiple complex-valued convolutional blocks, fully connected layers, and an output layer (of size K), where K is the number of transmitters in the sample dataset. Each complex-valued convolutional block includes a wide linear complex-valued convolutional layer connected along the forward propagation direction, a ReLU activation function, a normalization layer, a 2-kernel max-pooling layer, and an SE attention mechanism module. The wide linear complex-valued convolutional layer has a kernel size of 3 and 64 output channels.

[0023] The wide linear complex-valued convolutional layer performs convolution operations on the complex-valued radio frequency signal and its conjugate form using two convolution filters with identical parameters. The results of the two convolution filters are added together to obtain the output of the wide linear complex-valued convolutional layer. The output expression of the wide linear complex-valued convolutional layer is as follows: ; in, To enhance the input complex-valued radio frequency signal of a complex convolutional neural network, This is the conjugate form of the complex-valued radio frequency signal. and These are the lengths of the wide linear complex-valued convolutional layers. The first and second convolutional filters act on the input complex signal and its conjugate form, respectively; This is the conjugate transpose. This is the output of a wide linear complex-valued convolutional layer.

[0024] In additive white Gaussian noise environments, wide linear models can achieve up to twice the performance gain compared to strictly linear models. This indicates that traditional linear convolutions cannot fully utilize the complete feature information of complex signals, thus limiting the network's feature recognition capabilities. This invention applies the wide linear model to radio frequency fingerprinting tasks, employing wide linear complex-valued convolutional layers within the network's convolutional layers. By fully utilizing the amplitude and phase characteristics of complex signals, it improves the network's feature extraction capabilities for complex-valued radio frequency signals.

[0025] Since wide linear convolution requires convolution operations on the input signal and its conjugate signal separately, this introduces twice the number of convolution kernel parameters. To balance the model's recognition accuracy and efficiency, this embodiment of the invention introduces a squeeze-and-excitation (SE) attention mechanism module in the complex-valued convolution block. The structure of the SE attention mechanism module is as follows: Figure 3 As shown, the SE attention mechanism module, as a lightweight attention mechanism, is used to model the dependencies between features in the channel dimension, thereby adaptively adjusting the channel weights and improving the feature representation ability of the network.

[0026] The core idea of ​​the SE attention mechanism module is to achieve feature reweighting through three steps: (1) Compression (squeeze): Global average pooling is used to aggregate global features for each channel, extracting the global description at the channel level, thereby capturing the degree of contribution of each channel to the overall features, and outputting a 1-dimensional vector. , dimension , Input feature map No. The channel is compressed The channel vector after The expression is: ; in, and These are the input feature maps. Height and width, For the number of channels, Input feature map In spatial location No. The characteristic values ​​of the channel.

[0027] (2) Excitation: Based on the extracted global features, a set of fully connected layers is used to generate weight vectors for each channel. This step, through nonlinear transformation and normalization operations, enables the network to learn the relative importance of different channels to the current task. Compressed channel vectors After being motivated The channel weight vector after The expression is: ; in, and These are the weight matrices for the first and second fully connected layers, respectively. It is the ReLU activation function. For the Sigmoid activation function, the channel weight vector Dimensions .

[0028] (3) Scaling: The generated weights adaptively adjust the amplitude of the original feature channels to enhance the focus on task-related features while suppressing redundant or irrelevant information, thereby improving the model's feature representation capability. Input feature map Scaled The expression after that is: ; in, For the first Channel attention weights Output feature map of the SE attention mechanism module In spatial location No. The characteristic values ​​of the channel.

[0029] Preferably, in this embodiment of the invention, the enhanced complex convolutional neural network includes nine complex-valued convolutional blocks. Multi-level feature extraction: Different complex-valued convolutional blocks can extract features from the input complex data at different scales and levels. Nine complex-valued convolutional blocks provide sufficient depth, enabling the network to analyze complex data from multiple perspectives, thus representing the data features more comprehensively. Multiple complex-valued convolutional blocks increase the network's nonlinear transformation capability. As the number of complex-valued convolutional blocks increases, the network can learn more complex function mapping relationships, thereby better fitting the distribution patterns in the training data and improving the performance of feature recognition signals in complex radio frequency environments. When the network has multiple complex-valued convolutional blocks, it avoids the network's over-reliance on certain local features, reducing the risk of overfitting. Nine complex-valued convolutional blocks provide a reasonable balance between computational cost and model performance. Compared to a small number of complex-valued convolutional blocks, it provides better performance; while compared to an excessive number of complex-valued convolutional blocks, it does not lead to an exponential increase in computational cost. When processing complex signals on resource-constrained computing devices, this setup can control the consumption of computing resources while ensuring a certain level of accuracy.

[0030] S2: Collect communication signals transmitted by multiple radio frequency transmitters of the same model, and store the collected communication signals in I / Q format to obtain a sample dataset.

[0031] Communication signals transmitted by multiple radio frequency transmitters of the same model are collected. The collected signals are stored in in-phase / quadrature (I / Q) format, which can completely preserve their amplitude and phase information, thereby constructing a sample dataset containing signals from multiple transmitters.

[0032] S3: Divide the sample dataset into training, validation, and test sets, which are used to enhance the training, parameter tuning, and performance evaluation of the complex convolutional neural network model, respectively.

[0033] S4: Supervised learning of the constructed enhanced complex convolutional neural network is performed using the training set, with the cross-entropy loss function and the network parameters updated through backpropagation.

[0034] The expression for the cross-entropy loss function is: Where K represents the total number of transmitters, This represents the actual tag corresponding to the i-th transmitter. Let be the predicted probability of the model for the i-th transmitter. In each training round, the training set samples are forward-propagated to obtain the model output. The classification error between the model output and the true labels of the training set samples is calculated using the cross-entropy loss function. The gradient is calculated using the backpropagation algorithm, and the parameters of the trained augmented complex convolutional neural network model are saved.

[0035] S5: During the training of the enhanced complex convolutional neural network, the model performance is monitored and optimized in real time using the validation set.

[0036] S6: After the enhanced complex convolutional neural network is trained, the recognition performance of the enhanced complex convolutional neural network is finally evaluated using the test set.

[0037] The test set samples are forward-propagated to calculate the difference between the classification results and the true labels, and the recognition accuracy is statistically analyzed, thereby achieving accurate differentiation and recognition of the transmitter.

[0038] S7: Acquire the I / Q signal sent by the radio frequency transmitter to be identified and convert it into a complex-valued radio frequency signal. Input the complex-valued radio frequency signal and its conjugate form into the trained enhanced complex convolutional neural network and output the identification result of the radio frequency transmitter to be identified.

[0039] To demonstrate the performance of the enhanced complex convolutional neural network model provided in this embodiment, two publicly available datasets were selected for model performance evaluation: the ADS-B dataset collected by automatic dependent surveillance-broadcast (ADS-B) technology and the Wi-Fi dataset collected from the USRPX310 radio.

[0040] The ADS-B dataset contains ADS-B signals collected in a large-scale real-world airspace environment. The data generation process is as follows: the aircraft's position and other information are used to generate a signal through the ADS-B system's signal generation module, which then transmits the signal. The signal is received by a ground signal tower and processed by the control terminal at the console. The high-performance software-defined radio device SM200B was used during signal acquisition. This device can flexibly and accurately capture complex radio signals in the airspace. The acquired signals undergo digital processing to optimize quality and are ultimately stored in I / Q signal format, thus fully preserving the amplitude and phase information of the signal, providing a reliable data foundation for subsequent analysis and model training. The ADS-B dataset covers ADS-B signals collected in a real, vast airspace environment, including 3080 training samples and 1000 test samples. Each sample is a complex signal in 4800×2 format, originating from 10 different transmitters of the same model.

[0041] The Wi-Fi dataset consists of raw I / Q data samples collected by 16 advanced X310 USRP software-defined radio (SDR) devices, with the B210 model radio device used uniformly at the receiving end. At the transmitting end, wireless data frames conforming to the IEEE 802.11a standard are generated using the MATLAB WLAN toolbox. These data frames contain randomly generated payload information, but the address field remains unchanged. Real-time wireless signal transmission is then achieved using a specified SDR. At the receiving end, the Wi-Fi signal with a center frequency of 2.45 GHz is demodulated and sampled at a sampling rate of 5 MS / s using the SDR. The experiments in this section use a dataset with a transmitter-receiver distance of 62 feet, referred to as the Wi-Fi (62ft) dataset. This dataset contains 3080 training samples and 16004 test samples, each sample being a complex signal in 6000×2 format, from 16 different transmitters of the same model.

[0042] This embodiment evaluates the performance of the proposed RF fingerprint recognition method based on enhanced CV-CNN using the ADS-B dataset and the Wi-Fi (62ft) dataset.

[0043] This embodiment selects three network models to assess their ability to recognize radio frequency fingerprint signals under specific electromagnetic environments: RV-CNN, CV-CNN, and the enhanced CV-CNN proposed in this invention. All three networks incorporate an SE attention mechanism module and are designated Att-RV-Net, Att-SL-Net, and Att-WL-Net, respectively. The overall architecture of the three network models is largely consistent, differing only in that their convolutional layers employ real-valued convolutional layers, traditional complex-valued convolutional layers, and complex-valued convolutional layers based on wide linear convolutions, respectively. During training, all network models utilize the Adam optimizer, undergoing 300 iterations of training with a uniform learning rate of 0.001 and a batch size of 32. 30% of the labeled samples in the training data are used as the validation set. The experiments are implemented using PyTorch (v1.13.0, Python 3.8.2) and computed using an NVIDIA GeForce RTX 1060Ti GPU. In this embodiment, all network models use two fully connected layers of size 512 and 128 respectively in the ADS-B dataset, and one fully connected layer of size 1024 in the Wi-Fi (62ft) dataset.

[0044] In the field of radio frequency fingerprint recognition, the proportion of labeled samples plays a crucial role in model performance, and its variation directly affects the effectiveness of the network model in different application scenarios. This embodiment studies the performance differences of three network models—Att-RV-Net, Att-SL-Net, and Att-WL-Net—in radio frequency fingerprint recognition under different proportions of labeled samples, and demonstrates the generalization ability of the network models in diverse environments.

[0045] This embodiment uses recognition accuracy as a performance metric. Accuracy is a commonly used metric for classification performance, calculated as the ratio of correctly identified samples to the total number of test samples. It is used to demonstrate the overall recognition capability of the network model in classification tasks. In multi-class classification tasks, accuracy is the proportion of correctly predicted N-class samples out of all samples, calculated using the following formula: ; in, This represents the total number of correctly classified samples across all categories. Indicates the true category is And the predicted category is also The number of samples, i.e., the number of correctly classified samples; This indicates the total number of samples in the dataset.

[0046] To achieve intuitive semantic representation, this embodiment employs the t-distributed stochastic neighbor embedding (t-SNE) algorithm. This algorithm reduces the dimensionality of complex high-dimensional data to two dimensions for visualization in low-dimensional space. The t-SNE algorithm can capture complex nonlinear structures, facilitating intuitive visualization analysis. A silhouette coefficient is introduced as a metric to evaluate the t-SNE algorithm. The silhouette coefficient is a quantitative indicator of clustering effectiveness; it can be understood as a metric describing the clarity of the silhouettes of each cluster, used to evaluate the quality of clustering. For a given point i in the cluster, the silhouette coefficient... The calculation formula is: ; Among them, parameters This represents the average distance from a given point i to other samples in the same cluster, reflecting intra-class compactness; parameters The silhouette coefficient represents the average distance from a given point i to the nearest heterogeneous sample, reflecting the inter-class separation. The overall silhouette coefficient is obtained by averaging the silhouette coefficients of all points. The silhouette coefficient ranges from -1 to 1; the closer it is to 1, the higher the cohesion and separation, and the better the clustering quality; the closer it is to -1, the lower the cohesion and separation, and the worse the clustering quality.

[0047] In this embodiment, five different scenarios were designed. By adjusting the proportion of labeled samples in the total training samples to 5%, 10%, 20%, 50%, and 100%, respectively, the performance of different network models in the radio frequency fingerprint signal recognition task was evaluated. Table 1 shows the recognition accuracy of three different network models on the ADS-B dataset; Table 2 shows the recognition accuracy of three different network models on the Wi-Fi (62ft) dataset.

[0048] Table 1. Recognition accuracy of three different network models on the ADS-B dataset.

[0049] Table 2. Recognition accuracy of three different network models on the Wi-Fi (62ft) dataset.

[0050] As shown in Tables 1 and 2, with a labeled sample ratio of only 5%, Att-WL-Net achieves recognition accuracies of 64.30% and 34.79% on the two datasets, respectively, both higher than Att-RV-Net and Att-SL-Net. With an increasing labeled sample ratio, Att-WL-Net's overall recognition performance remains superior to Att-RV-Net and Att-SL-Net, especially demonstrating a significant advantage in large-scale datasets and scenarios with low labeled sample ratios. These results indicate that the enhanced CV-CNN proposed in this embodiment exhibits superior performance in radio frequency fingerprint recognition.

[0051] Then, t-SNE is used to reduce the dimensionality of the semantic features extracted by the model to a two-dimensional space. Figure 4 The results show the semantic feature visualization based on the ADS-B dataset at 10% and 20% labeled sample ratios. Figure 4 (a) in the figure shows the semantic feature visualization results of Att-RV-Net under the ADS-B dataset with a label ratio of 10%; Figure 4 (b) in the figure shows the semantic feature visualization results of Att-SL-Net under the ADS-B dataset with a label ratio of 10%; Figure 4 (c) in the figure shows the semantic feature visualization results of Att-WL-Net under the ADS-B dataset with a label ratio of 10%; Figure 4 In the diagram, (d) shows the semantic feature visualization results of Att-RV-Net under the ADS-B dataset with a label ratio of 20%; Figure 4 In the diagram, (e) represents the semantic feature visualization results of Att-SL-Net under the ADS-B dataset with a label ratio of 20%. Figure 4(f) shows the semantic feature visualization results of Att-WL-Net on the ADS-B dataset with a labeling ratio of 20%. When the labeled samples are 10%, the silhouette coefficients of the three network models Att-RV-Net, Att-SL-Net, and Att-WL-Net are 0.1733, 0.1390, and 0.2255, respectively; when the labeled samples are 20%, the silhouette coefficients of the three network models are 0.3251, 0.3731, and 0.3248, respectively.

[0052] Figure 5 The results show the semantic feature visualizations based on the Wi-Fi (62ft) dataset at 10% and 20% labeled sample ratios. Figure 5 (a) shows the semantic feature visualization results of Att-RV-Net under the Wi-Fi (62ft) dataset with a label ratio of 10%; Figure 5 (b) in the figure shows the semantic feature visualization results of Att-SL-Net under the Wi-Fi (62ft) dataset with a label ratio of 10%; Figure 5 (c) shows the semantic feature visualization results of Att-WL-Net under the Wi-Fi (62ft) dataset with a label ratio of 10%; Figure 5 (d) in the figure represents the semantic feature visualization results of Att-RV-Net under the Wi-Fi (62ft) dataset with a label ratio of 20%; Figure 5 (e) in the figure represents the semantic feature visualization results of Att-SL-Net under the Wi-Fi (62ft) dataset with a label ratio of 20%; Figure 5 (f) shows the semantic feature visualization results of Att-WL-Net on the Wi-Fi (62ft) dataset with a labeling ratio of 20%. When the labeled samples are 10%, the silhouette coefficients of the three network models Att-RV-Net, Att-SL-Net, and Att-WL-Net are 0.0100, 0.0252, and 0.0507, respectively; when the labeled samples are 20%, the silhouette coefficients of the three network models are 0.0832, 0.0986, and 0.2134, respectively.

[0053] When the labeled sample ratio drops to 10%, the silhouette coefficients of the ADS-B and Wi-Fi (62ft) datasets are low across all three network models. This indicates that under conditions of scarce samples, the various network models struggle to effectively distinguish the boundaries between different data categories, resulting in suboptimal intra-class sample density and inter-class sample separation. However, Att-WL-Net, based on enhanced CV-CNN, demonstrates stronger intra-class density and inter-class separation in feature visualization compared to the other two models, suggesting that this model can still more effectively capture the deep semantic features of the data.

[0054] In summary, this invention, for the first time, proposes using wide linear complex-valued convolutional layers to perform convolution operations on the input complex-valued signal and its conjugate form. By fully utilizing the amplitude and phase characteristics of complex signals, it improves the network's feature extraction capability for complex-valued radio frequency signals. Furthermore, the introduced SE module, as a lightweight attention mechanism, effectively enhances the feature representation capability of the network model, making its feature recognition performance in complex radio frequency environments even better.

[0055] This invention provides a radio frequency fingerprint recognition device based on an enhanced complex convolutional neural network; the specific device may include: The signal acquisition module is used to acquire the I / Q signals sent by the radio frequency transmitter to be identified and convert them into complex radio frequency signals; A network processing module is used to input complex-valued radio frequency signals and their conjugate forms into an enhanced complex convolutional neural network. The enhanced complex convolutional neural network includes multiple complex-valued convolutional blocks, fully connected layers, and an output layer. The convolutional layers of the complex-valued convolutional blocks are wide linear complex-valued convolutional layers. The wide linear complex-valued convolutional layer performs convolution operations on the complex-valued radio frequency signals and their conjugate forms through two convolutional filters with identical parameters, and adds the processing results of the two convolutional filters as the output of the wide linear complex-valued convolutional layer. The result output module is used to obtain the identification result of the radio frequency transmitter to be identified based on the output result of the output layer.

[0056] The radio frequency fingerprint recognition device based on the enhanced complex convolutional neural network in this embodiment is used to implement the aforementioned radio frequency fingerprint recognition method based on the enhanced complex convolutional neural network. Therefore, the specific implementation of the radio frequency fingerprint recognition device based on the enhanced complex convolutional neural network can be found in the embodiment section of the radio frequency fingerprint recognition method based on the enhanced complex convolutional neural network mentioned above. So, its specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0057] A specific embodiment of the present invention also provides a radio frequency fingerprint recognition device based on an enhanced complex convolutional neural network, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described radio frequency fingerprint recognition method based on an enhanced complex convolutional neural network.

[0058] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0059] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0060] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0061] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0062] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A radio frequency fingerprint recognition method based on an enhanced complex convolutional neural network, characterized in that, include: Acquire the I / Q signals transmitted by the radio frequency transmitter to be identified and convert them into complex radio frequency signals; A complex-valued radio frequency signal and its conjugate form are input into an enhanced complex convolutional neural network, which includes multiple complex-valued convolutional blocks, fully connected layers, and an output layer. The convolutional layer of the complex-valued convolutional blocks is a wide linear complex-valued convolutional layer. The wide linear complex-valued convolutional layer performs convolution operations on the complex-valued radio frequency signal and its conjugate form through two convolutional filters with identical parameters, and adds the processing results of the two convolutional filters as the output of the wide linear complex-valued convolutional layer. The identification result of the radio frequency transmitter to be identified is obtained based on the output result of the output layer.

2. The radio frequency fingerprint recognition method based on enhanced complex convolutional neural networks according to claim 1, characterized in that, The complex-valued convolutional block includes a wide linear complex-valued convolutional layer connected along the forward propagation direction, a ReLU activation function, a normalization layer, a maximum pooling layer, and an SE attention mechanism module.

3. The radio frequency fingerprint recognition method based on enhanced complex convolutional neural networks according to claim 1, characterized in that, The output expression of a wide linear complex-valued convolutional layer is: ； in, To enhance the input complex-valued radio frequency signal of a complex convolutional neural network, This is the conjugate form of the complex-valued radio frequency signal. and These are the lengths of wide linear complex-valued convolutional layers. The first convolution filter and the second convolution filter, It is the conjugate transpose. This is the output of a wide linear complex-valued convolutional layer.

4. The radio frequency fingerprint recognition method based on enhanced complex convolutional neural networks according to claim 1, characterized in that, The kernel size of the wide linear complex-valued convolutional layer is 3, and the number of output channels is 64.

5. The radio frequency fingerprint recognition method based on an enhanced complex convolutional neural network according to claim 2, characterized in that, The processing steps of the SE attention mechanism module include: Global average pooling is used to perform global feature aggregation on each channel of the input feature map of the SE attention mechanism module to obtain a compressed channel vector; The compressed channel vector is reduced in dimension by the first fully connected layer and then subjected to nonlinear transformation. After being increased in dimension by the second fully connected layer, it is normalized and the channel weight vector after excitation is output. The channel weight vector is multiplied channel by channel with the input feature map of the SE attention mechanism module to obtain the output feature map of the SE attention mechanism module.

6. The radio frequency fingerprint recognition method based on an enhanced complex convolutional neural network according to claim 5, characterized in that, Input feature map of SE attention mechanism module No. Channel vector after channel compression The expression is: ； in, and These are the input feature maps. Height and width, For the number of channels, Input feature map In spatial location No. The eigenvalues ​​of the channel; Compressed channel vectors The channel weight vector after activation The expression is: ； in, and These are the weight matrices for the first and second fully connected layers, respectively. It is the ReLU activation function. It is the Sigmoid activation function. ; The expression for multiplying the channel weight vector by the input feature map of the SE attention mechanism module channel by channel is as follows: ； in, For the first Channel attention weights Output feature map of the SE attention mechanism module In spatial location No. The characteristic values ​​of the channel.

7. The radio frequency fingerprint recognition method based on an enhanced complex convolutional neural network according to claim 1, characterized in that, The enhanced complex convolutional neural network includes 9 complex-valued convolutional blocks.

8. The radio frequency fingerprint recognition method based on enhanced complex convolutional neural networks according to claim 1, characterized in that: The training process of the enhanced complex convolutional neural network includes: The communication signals transmitted by multiple radio frequency transmitters of the same model are collected and stored in I / Q format to obtain a sample dataset. The sample dataset is divided into a training set, a validation set, and a test set, which are used to enhance the training, parameter tuning, and performance evaluation of the complex convolutional neural network model, respectively. In each training round, the training set samples are forward propagated to obtain the model output. The cross-entropy loss function is used to calculate the classification error between the model output and the true labels of the training set samples. The gradient is calculated through the backpropagation algorithm and the parameters of the enhanced complex convolutional neural network model are saved after training.

9. A radio frequency fingerprint recognition device based on an enhanced complex convolutional neural network, characterized in that, include: The signal acquisition module is used to acquire the I / Q signals sent by the radio frequency transmitter to be identified and convert them into complex radio frequency signals; A network processing module is used to input complex-valued radio frequency signals and their conjugate forms into an enhanced complex convolutional neural network. The enhanced complex convolutional neural network includes multiple complex-valued convolutional blocks, fully connected layers, and an output layer. The convolutional layers of the complex-valued convolutional blocks are wide linear complex-valued convolutional layers. The wide linear complex-valued convolutional layer performs convolution operations on the complex-valued radio frequency signals and their conjugate forms through two convolutional filters with identical parameters, and adds the processing results of the two convolutional filters as the output of the wide linear complex-valued convolutional layer. The result output module is used to obtain the identification result of the radio frequency transmitter to be identified based on the output result of the output layer.

10. A radio frequency fingerprint recognition device based on an enhanced complex convolutional neural network, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of a radio frequency fingerprinting method based on an enhanced complex convolutional neural network as described in any one of claims 1 to 8.

Images