Cross-network deployment method of radio frequency fingerprint identification model

Abstract

The invention discloses a cross-network deployment method of a radio frequency fingerprint identification model, which comprises the following steps: connecting each edge node with a plurality of wireless devices to form a sub-identification network, and constructing a wireless authentication network based on the plurality of sub-identification networks; segmenting a radio frequency fingerprint identification model in the wireless authentication network into a feature extractor and a selector; wherein the feature extractor is used for extracting radio frequency fingerprint features of wireless equipment, and the selector is used for performing individual identification based on the radio frequency fingerprint features; training a radio frequency fingerprint identification model based on pre-training of a feature extractor and federated training of a selector; and after the radio frequency fingerprint identification model is trained, obtaining trained model parameters and deploying the trained model parameters to each edge node. By designing an extensible feature extractor and an efficient edge deployment strategy, the generalization, safety and practicability of the radio frequency fingerprint identification system are improved while the problems of quantity difference of different network devices, open set device identification, edge node computing power difference and the like are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless security research, and particularly to a method for cross-network deployment of a radio frequency fingerprint recognition model. Background Art

[0002] Radio frequency fingerprint technology identifies and verifies device identities by analyzing the tolerance characteristics of radio frequency circuits or devices of wireless devices. However, traditional radio frequency fingerprint recognition methods based on expert features have various defects. These methods rely on manually designed feature extraction algorithms, which are difficult to comprehensively capture the complex characteristics in radio frequency signals, limiting the recognition accuracy. At the same time, expert features lack generality and are difficult to adapt to changes in different environments and devices, reducing the robustness of the system. When dealing with large-scale data, such methods have low computational efficiency and are difficult to meet the requirements of real-time recognition, and there is an urgent need to introduce more advanced technical means for improvement.

[0003] To improve the accuracy and adaptability of radio frequency fingerprint recognition, radio frequency fingerprint recognition technology based on deep learning has emerged. This method constructs a deep neural network to automatically extract high-level features from radio frequency signals to achieve more accurate device recognition. Especially in the face of complex environments and diverse devices, deep learning models have shown superior performance. Further, to enhance data privacy and security, researchers have proposed a federated learning training strategy based on blockchain. It includes steps such as distributed ledger recording, encrypted communication and data protection, decentralized model aggregation, and consensus mechanism to ensure consistency.

[0004] Although the advantages of high security and low computational overhead shown by radio frequency fingerprints have been initially demonstrated, the deployment research of existing radio frequency fingerprint recognition networks based on deep learning is not sufficient and it is difficult to be widely applied in practice.

[0005] In recent years, machine learning technologies, especially deep learning, have been widely applied to the research of radio frequency fingerprints, greatly improving the recognition accuracy and efficiency, and effectively promoting the practical process of radio frequency fingerprint technology. However, existing research mostly focuses on radio frequency fingerprint recognition in a single device or homogeneous network environment, ignoring the complexity of heterogeneous network environments in practical application scenarios. For example, the number of different network devices varies significantly, resulting in different model structures trained, which are difficult to be effectively integrated, limiting the generalization ability of the models. At the same time, existing research mostly trains based on known device data, resulting in the models being difficult to recognize out-of-set devices and unable to meet the application requirements in open scenarios; in addition, the hardware computing power and transmission bandwidth vary among different edge nodes. If the same neural network is deployed, it will inevitably cause waste of computing power and decline in performance, and it is difficult to achieve efficient edge deployment. Therefore, how to solve the challenges of radio frequency fingerprint recognition in heterogeneous network environments is the key problem that needs to be solved urgently in current research. Summary of the Invention

[0006] The present invention aims to solve at least one of the technical problems in the related art to some extent.

[0007] The present invention provides a cross-network deployment method for a radio frequency fingerprint recognition model, which realizes a training strategy of decentralized federated learning based on blockchain technology, and at the same time realizes the efficient edge deployment of the radio frequency fingerprint recognition neural network by using model pruning and model quantization technologies.

[0008] To achieve the above object, on the one hand, the present invention provides a cross-network deployment method for a radio frequency fingerprint recognition model, including:

[0009] Connect each edge node to multiple wireless devices to form a sub-recognition network, and construct a wireless authentication network based on multiple sub-recognition networks;

[0010] Divide the radio frequency fingerprint recognition model in the wireless authentication network into a feature extractor and a selector; wherein, the feature extractor is used to extract the radio frequency fingerprint features of the wireless device, and the selector is used to perform individual recognition based on the radio frequency fingerprint features;

[0011] Train the radio frequency fingerprint recognition model based on feature extractor pre-training and selector federated training; wherein, the feature extractor pre-training includes training to obtain the parameter information of the feature extractor part by using the pre-collected signal information, and sending it to each edge node; the selector federated training includes that the edge nodes in each sub-recognition network use blockchain for a decentralized federated training strategy and perform pruning according to the node computing power level;

[0012] After the radio frequency fingerprint recognition model is trained, obtain the trained model parameters from the blockchain and deploy them to each edge node.

[0013] The cross-network deployment method for the radio frequency fingerprint recognition model according to the embodiment of the present invention may further have the following additional technical features:

[0014] In an embodiment of the present invention, during the pre-training sample collection process of the feature extractor pre-training, each edge node j obtains signals from the M j wireless devices it serves, processes them, including synchronization, preamble extraction, frequency offset compensation and power normalization, and stores them in the training data set, expressed as:

[0015]

[0016] where, T j represents the training data set collected at the edge node j, r j represents the received IQ sample, s j represents the corresponding wireless device label, K j refers to the data set T jThe number of signals in

[0017] In one embodiment of the present invention, after the pre-training samples are collected, pre-training is performed to generate the feature extractor parameters and store them in the blockchain of the network; the input of the feature extractor is the IQ samples corresponding to the signals of different wireless devices, and the output is the extracted radio frequency fingerprint.

[0018] In one embodiment of the present invention, in the selector federated training, all participating edge nodes adopt a radio frequency fingerprint recognition model with the same structure, wherein the feature extractor part uses the pre-trained model parameters; the number of classification heads in the selector part is set to the sum of the wireless device categories in all networks.

[0019] In one embodiment of the present invention, during the training process, each edge node uses the wireless device signals for multiple rounds of training, then performs pruning on the models of each network at a preset ratio, and finally performs the federated averaging operation to update the parameters of the selector part of each network model to the arithmetic mean of all network model parameters; after the training is completed, the federated averaging operation is performed again to ensure that the model parameters of all networks are consistent.

[0020] In one embodiment of the present invention, training the radio frequency fingerprint recognition model includes:

[0021] S31, the edge node j processes the collected signal x j and converts the signal into IQ samples for training and constructs the dataset T j ;

[0022] S32, pre-train to obtain the feature extractor φ = Model(x; θ), where x is the input and θ is the weight of the model;

[0023] S33, each edge node uses the feature extractor φ to obtain the radio frequency fingerprint corresponding to the original signal and constructs the radio frequency fingerprint library H of its own network j = φ(T j );

[0024] S34, distribute the initial radio frequency fingerprint recognition model to the edge node j where the feature extractor is φ, and the parameters of the feature extractor part are frozen, and the number of classification heads in the selector part is set to the sum of the wireless device categories in all networks,

[0025] S35, start the federated training, and the edge node j uses the radio frequency fingerprint library H j to train the model for n rounds to obtain the intermediate model

[0026] S36. Perform a pruning operation on the weight parameter of the selector part, including introducing two additional parameters, weight_orig and weight_mask, for the selector. Here, weight = weight_orig × weight_mask, where weight_orig represents the original weight matrix for backpropagation, weight_mask is the weight mask, and weight represents the pruned weight value.

[0027] S37. Share the training results, including performing a federated averaging operation to calculate the arithmetic mean of the weight_orig parameter. Where is the model The i-th parameter of the selector part, and Are all updated to θ i ;

[0028] S38. Repeat steps S36 and S37 until all N rounds are completed; finally, share the training results so that

[0029] S39. Perform weight quantization on the trained f final Model.

[0030] In an embodiment of the present invention, the pre-training of the feature extractor includes quantifying the feature extraction effect based on the cross-entropy loss and the loss function based on metric learning; the cross-entropy loss is used to measure the difference between the probability distribution predicted by the model and the true label, and the mathematical expression is:

[0031]

[0032] Where N is the number of samples, C is the number of categories, y i,c Is the true label of sample i in category c, and p i,c Is the probability predicted by the model;

[0033] The center loss minimizes the Euclidean distance between the sample feature vector and the center of the category to which it belongs, and the mathematical expression is:

[0034]

[0035] Where, h i Is the feature vector of sample i, and c yi Is the center vector of category y i ;

[0036] During the training process, the model optimizes both loss functions simultaneously and balances them through the weight coefficient λ. The total loss function is:

[0037]

[0038] In one embodiment of the present invention, the selector federated training includes using federated averaging or other training result sharing strategies. Every n training cycles, the partial parameters of each edge node model are averaged. For the weight matrix W of the fully connected layer, the mathematical expression of federated averaging is:

[0039]

[0040] where J is the number of models participating in the averaging, and W j is the weight of the j-th model in this layer.

[0041] In one embodiment of the present invention, performing a pruning operation on the weight parameter of the selector part includes:

[0042] At the beginning of training, create a binary mask matrix mask with the same size as the weight matrix of each layer, and initialize the mask to all 1s;

[0043] For the weight matrix W of each layer, calculate the absolute value of each weight:

[0044] |W ij | for i = 1, 2, …, N and j = 1, 2, …, M

[0045] where W ij is the weight of the i-th row and j-th column, and N and M are the number of rows and columns of the weight matrix respectively;

[0046] Based on the absolute value of the weight, calculate the threshold θ, and prune the weights whose absolute value is less than the threshold θ. The threshold θ is determined by a preset pruning ratio; including at the end of each training cycle, select the smallest p% of the weights for pruning:

[0047] θ = quantile(|W|, p)

[0048] where quantile(|W|, p) means that after sorting by absolute value, select the latter p% of the weights as the pruning target;

[0049] According to the pruning threshold, update the mask matrix mask, set the positions corresponding to the pruned weights to 0, and keep the positions of the remaining weights as 1;

[0050] In each forward propagation, apply the pruned weight matrix W effective = W·mask to the calculation of the network;

[0051] After the training process is completed, fix the weight matrix W as the final W effective matrix, and remove the mask matrix.

[0052] The cross-network deployment method of the radio frequency fingerprint recognition model according to the embodiment of the present invention, by designing an extensible feature extractor and an efficient edge deployment strategy, while solving problems such as differences in the number of different network devices, open-set device recognition, and edge node computing power differences, improves the generalization, security, and practicality of the radio frequency fingerprint recognition system.

[0053] Additional aspects and advantages of the present invention will be given in part in the following description, become apparent in part from the following description, or be understood through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The above and / or additional aspects and advantages of the present invention will become apparent and be readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0055] Figure 1 is a flowchart of the cross-network deployment method of the radio frequency fingerprint recognition model according to the embodiment of the present invention;

[0056] Figure 2 is a structural diagram of the radio frequency fingerprint recognition network according to the embodiment of the present invention;

[0057] Figure 3 is a structural diagram of the radio frequency fingerprint recognition model according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0059] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0060] The cross-network deployment method of the radio frequency fingerprint recognition model according to the embodiment of the present invention will be described below with reference to the accompanying drawings.

[0061] Figure 1 is a flowchart of the cross-network deployment method of the radio frequency fingerprint recognition model according to the embodiment of the present invention. As Figure 1 shown, the method includes:

[0062] S1, connecting each edge node to multiple wireless devices to form a sub-recognition network, and constructing a wireless authentication network based on the multiple sub-recognition networks;

[0063] S2. Split the radio frequency fingerprint recognition model in the wireless authentication network into a feature extractor and a selector. Among them, the feature extractor is used to extract the radio frequency fingerprint features of wireless devices, and the selector is used to perform individual recognition based on the radio frequency fingerprint features.

[0064] S3. Train the radio frequency fingerprint recognition model based on the pre-training of the feature extractor and the federated training of the selector. Among them, the pre-training of the feature extractor includes training the parameter information of the feature extractor part using the pre-collected signal information and sending it to each edge node. The federated training of the selector includes that the edge nodes in each sub-recognition network use the blockchain for a decentralized federated training strategy and perform pruning according to the node computing power level.

[0065] S4. After the radio frequency fingerprint recognition model is trained, obtain the trained model parameters from the blockchain and deploy them to each edge node.

[0066] It can be understood that in response to the challenges faced by the radio frequency fingerprint recognition network in actual deployment, the present invention proposes a complete solution. This solution combines cutting-edge technologies such as deep learning, federated learning, and blockchain to construct a secure, efficient, and scalable radio frequency fingerprint recognition network. Specifically, this solution designs a new type of feature extractor. By introducing the metric learning method, it effectively extracts the unique radio frequency fingerprint features contained in the wireless signal and maps them to a high-dimensional feature space, thereby improving the accuracy and robustness of recognition. At the same time, this solution innovatively proposes a decentralized federated learning training strategy based on the blockchain. On the premise of ensuring data security and privacy, it realizes the collaborative training and model update of multiple edge nodes. In addition, in response to the problem of limited resources of edge devices, this solution also studies the efficient edge deployment technology of neural network models. Through means such as federated model pruning and quantization, it minimizes resource consumption and realizes the efficient operation of the radio frequency fingerprint recognition network in a heterogeneous edge environment. The deployed recognition network architecture is as Figure 2 shown.

[0067] Specifically, in a complex wireless authentication network, each edge node is connected to multiple wireless devices to form an identification network. A complex wireless authentication network usually contains multiple identification subnets. The radio frequency fingerprint recognition model deployed in the network is divided into a feature extractor and a selector. The former is used to extract the radio frequency fingerprint features of wireless devices, and the latter performs individual identification based on the fingerprint features. The training of the radio frequency fingerprint recognition model can be divided into two stages: pre-training of the feature extractor and federated training of the selector. The pre-training is carried out by the deployer of the entire system using the pre-collected signal information to train the parameter information of the feature extractor part and distribute it to each edge node. The edge nodes in each identification subnet implement a decentralized federated training strategy with the help of the blockchain and perform pruning according to the node computing power level. Finally, the trained selector model is deployed to the edge nodes.

[0068] Specifically, during the pre-training sample collection process, each edge node j obtains signals from the M j wireless devices it serves and processes them, including synchronization, preamble extraction, frequency offset compensation, and power normalization, and then stores them in the training dataset, which is expressed as follows:

[0069]

[0070] where T j represents the training dataset collected at edge node j, r j represents the received IQ samples, s j represents the corresponding wireless device label. K j refers to the number of signals in the dataset T j .

[0071] The number of wireless devices connected to each edge node can be different, that is, M 1 ≠M 2 ≠…M J . This is a more practical setting because each edge node can serve different numbers of wireless devices in the network. After the sample collection is completed, the identification network deployer performs pre-training to generate the feature extractor parameters and saves them in the blockchain of the network. The input of the feature extractor is the IQ samples corresponding to the signals of different wireless devices, and the output is the extracted radio frequency fingerprint. By training the feature extractor instead of the entire classification neural network, the model architecture can be independent of the number of connected wireless devices, thus solving the model heterogeneity problem in the federated radio frequency fingerprint recognition model training.

[0072] During the federated learning training process, all participating edge nodes adopt a radio frequency fingerprint recognition model with the same structure. Among them, the pre-trained model parameters are used in the feature extractor part. To adapt to different numbers of wireless device categories, the number of classification heads in the selector part is set to the sum of the wireless device categories in all networks. During the training process, each edge node uses its specific wireless device signal for multiple rounds (n epochs) of training, then performs pruning on the models of each network according to a preset ratio, and finally performs the Federated Averaging operation to update the parameters of the selector part of each network model to the arithmetic mean of the parameters of all network models. After the training is completed, the Federated Averaging operation is performed again to ensure that the model parameters of all networks are consistent, thus achieving the consistency of the global model. Throughout the process, the learning behaviors of all participating parties (edge nodes) are recorded and verified through blockchain technology to ensure data privacy and the transparency of the model training process. Each edge node participating in the federated learning acts as a blockchain verification node during this process, responsible for verifying and recording transactions to ensure the normal operation and security of the blockchain network.

[0073] Based on the above scheme objectives, the main architecture of the proposed radio frequency fingerprint recognition model can be based on CNN or other deep neural networks. For example, 3 convolutional layers can be used to construct the feature extractor, and 2 fully connected layers can be used to construct the selector. The specific structure is as Figure 3 shown:

[0074] The specific training process of the radio frequency fingerprint recognition model proposed in the embodiment of the present invention is as follows:

[0075] S31, Edge node j first processes the collected signal x j and converts it into IQ samples that can be used for training and constructs a dataset T j .

[0076] S32, The recognition network deployer pre-trains to obtain a feature extractor φ = Model(x; θ), where x is the input and θ is the weight of the model.

[0077] S33, Each edge node uses the feature extractor φ to obtain the radio frequency fingerprint corresponding to the original signal and constructs a radio frequency fingerprint library H j = φ(T j ).

[0078] S34, Distribute the initial radio frequency fingerprint recognition model to edge node j where the feature extractor is φ, and the parameters of the feature extractor part are frozen to ensure that they do not change during subsequent training. The number of classification heads in the selector part is set to the sum of the wireless device categories in all networks. At this time

[0079] S35. Start the federated training. Edge node j uses the RF fingerprint database H j Train the model for n epochs to obtain an intermediate model Since the RF fingerprint databases of each network are different, at this time

[0080] S36. Perform a pruning operation on the weight parameters of the selector part. One feasible method is to introduce two additional parameters, weight_orig and weight_mask, for the selector. Here, weight = weight_orig × weight_mask. weight_orig represents the original weight matrix for backpropagation, weight_mask is the weight mask used to indicate which weights need to be pruned and which weights are retained, and its value is {0, 1}. weight represents the pruned weight value for forward propagation

[0081] S37. Share the training results. For example, perform a federated averaging operation to calculate the arithmetic mean of the weight_orig parameters Where is the i-th parameter of the model selector part, and are all updated to θ i .

[0082] S38. Repeat steps (S36) and (S37) until all N epochs are completed. Finally, share the training results again to ensure

[0083] S39. Perform a weight quantization operation on the trained f final model. The quantization precision can be selected as 8-bit integer (int8) or 16-bit integer (int16).

[0084] After the RF fingerprint recognition model training is completed, each client obtains the trained model parameters from the blockchain and deploys them locally to achieve efficient edge deployment of the neural network model

[0085] It can be known that the present invention divides the RF fingerprint recognition model into a feature extractor and a selector. This design improves the interpretability and recognition scalability, simplifies the neural network representation, enhances the ability to distinguish signals of different devices, and supports open-set recognition. In addition, a metric learning method with center loss is introduced in the feature extractor training to optimize the distribution of signals in the feature space, improve the recognition accuracy and robustness, and break through the limitations of traditional open-set recognition

[0086] Blockchain technology is used to achieve decentralization of federated learning, ensuring transparency and traceability of the training process. Edge nodes serve as blockchain verification nodes to verify and record learning behaviors and improve data security. Through periodic sharing and averaging of model parameters, a unified recognition model across network devices is achieved. In view of the differences in hardware computing power and bandwidth of different edge nodes, model tailoring is innovatively integrated with the federated learning process to achieve dynamic optimization of the model. Model quantization is used to further improve inference speed and reduce resource consumption, ensuring efficient operation in diverse edge environments while maintaining superior recognition performance.

[0087] Furthermore, based on the proposed network architecture, according to the goal of the above scheme, a reasonable loss function is designed to achieve the desired purpose. In the pre-training process of the feature extractor, the cross-entropy loss (Cross-EntropyLoss) and the loss function based on metric learning are combined to quantify the feature extraction effect. The cross-entropy loss is used to measure the difference between the probability distribution predicted by the model and the true label. Its mathematical expression is:

[0088]

[0089] Where N is the number of samples, C is the number of categories, and y i,c is the true label of sample i in category c, p i,c is the probability predicted by the model. This loss function encourages the model to maximize the probability of the correct category, thereby improving the classification accuracy.

[0090] Center loss is a common loss function in metric learning. It further enhances the distinguishability of features by minimizing the Euclidean distance between the sample feature vector and the center of its category. Its mathematical expression is:

[0091]

[0092] Among them, h i is the feature vector of sample i, c yi For category y i By using the training strategy of metric learning, the model optimizes classification accuracy while ensuring that features of the same category are clustered together in the feature space, features of different categories are as dispersed as possible, and features are guaranteed to be redundant in the feature space, thereby improving the recognition ability of the model on open set devices. This strategy can effectively improve the interpretability and robustness of features.

[0093] During the training process, the model optimizes two loss functions at the same time, balancing the importance of the two through the weight coefficient λ. Specifically, the total loss function is:

[0094]

[0095] Set λ = 1.0 in the code, which can be adjusted according to specific requirements in actual applications to achieve the best performance. The optimizer uses the Adam algorithm or other model optimization algorithms to independently update the model parameters and the center loss parameters respectively, ensuring the stability and efficiency of the training process. After N training epochs, the model parameters of the feature extractor are saved for subsequent use.

[0096] The training of the selector adopts Federated Averaging or other training result sharing strategies. Every n training epochs, the partial parameters of each edge node model are averaged to achieve the synchronous update of the global model. Specifically, for each layer to be averaged, such as the weight matrix W of the fully connected layer, the mathematical expression of Federated Averaging is:

[0097]

[0098] where J is the number of models participating in the averaging, and W j is the weight of the j-th model in this layer. This process realizes the aggregation of the model parameters of multiple edge nodes, ensuring that the global model can reflect the learning results of each node.

[0099] In an embodiment of the present invention, in the training loop, each client uses local data to train the model respectively and performs Federated Averaging every n epochs. The specific training steps are as follows:

[0100] Local training: Each edge node independently trains its local model, and uses the Cross-Entropy Loss function as the optimization objective. The mathematical expression of the Cross-Entropy Loss function is the same as (1).

[0101] Federated Averaging: The parameters of the specified layer are averaged every n epochs. This process realizes the synchronous update of the global model parameters, making the edge node models consistent on the key layers and promoting the integration and sharing of global knowledge.

[0102] Performance evaluation and best model saving: After each Federated Averaging, evaluate the average loss of the current model on the test set. If the current average loss is better than the previous best loss, update the best model state. This strategy ensures that the model parameters with the best performance are always retained during the training process, improving the robustness and generalization ability of the model.

[0103] In one embodiment of the present invention, when there are problems such as limited computing resources in edge nodes, a gradual pruning process can be introduced during model training to reduce the number of model parameters, thereby reducing the storage space of the model and improving the inference efficiency. The gradual pruning reduces the computational overhead by gradually setting the weights with smaller absolute values in the network to zero, without significantly affecting the performance of the model. The specific pruning steps are as follows:

[0104] Initialize the mask: At the beginning of training, create a binary mask matrix mask of the same size as the weight matrix for each layer. The mask is initialized to all 1s, indicating that all connections will be retained.

[0105] Calculate the absolute value of the weights: For the weight matrix W of each layer, calculate the absolute value of each weight

[0106] |W ij | for i = 1, 2, …, N and j = 1, 2, …, M

[0107] where W ij is the weight at the i-th row and j-th column, and N and M are the number of rows and columns of the weight matrix respectively.

[0108] Determine the pruning threshold: Based on the absolute value of the weights, calculate a threshold θ, and regard the weights with absolute values less than this threshold as redundant connections for pruning. This threshold is determined by presetting the pruning ratio. For example, at the end of each training epoch, select the smallest p% of the weights for pruning:

[0109] θ = quantile(|W|, p)

[0110] where quantile(|W|, p) means that after sorting by absolute value, select the last p% of the weights as the pruning target.

[0111] Update the mask: According to the pruning threshold, update the mask matrix mask, set the positions corresponding to the pruned weights to 0, and keep the positions of the retained weights as 1.

[0112] Apply the mask: In each forward propagation, use the mask to apply the pruned weight matrix W effective = W · mask to the calculation of the network.

[0113] Perform pruning: After the training process is completed, fix the weight matrix W as the final W effective matrix, and remove the mask matrix mask.

[0114] Furthermore, after completing the training and pruning processes, the model can also be quantized to convert the floating-point weights and activation values of the model into a low-precision representation (usually integers) to reduce storage requirements and improve computational efficiency. This process aims to reduce the consumption of hardware resources during model inference, especially in edge computing devices or embedded systems, where it can significantly improve the computational speed and reduce memory occupancy. The specific quantization steps are as follows:

[0115] Quantization Precision Selection and Quantization Criteria: To perform the quantization operation, the quantization precision (i.e., the bit width of the data) needs to be selected. Common quantization precisions include 8-bit integers (int8) and 16-bit integers (int16). When selecting the quantization precision, factors such as computational resources, storage capacity, and model accuracy need to be considered comprehensively.

[0116] The quantization operation also requires setting a scaling factor and a zero point. The scaling factor is used to determine the size of the numerical range, and the zero point is the offset of the quantized value. The mapping relationship of the quantized data can be described by the following formula:

[0117]

[0118] where x is the original floating-point value, s is the scaling factor, z is the zero point, Q(x) is the quantized integer value, and round(·) represents the rounding operation.

[0119] Weight Quantization: Weight quantization quantizes the trained model parameters in floating-point format into integer format. The weight quantization operation usually uses the following formula:

[0120]

[0121] where, W {float} is the trained floating-point weight, and S w is the scaling factor calculated for this weight matrix. Through the rounding operation, the floating-point weight is mapped to an integer value, and at the same time, the scaling factor is used for dequantization to restore the approximate numerical range.

[0122] Dequantization and Inference: During the actual inference process, the quantized weights and activation values need to be dequantized to restore their floating-point representation for subsequent calculations. The dequantization operation is usually performed through the following formula:

[0123] W dequantized = W quantized · S w

[0124] where, W dequantized is the dequantized floating-point weight, W quantized is the quantized integer weight, and Sw is the scaling factor calculated during the quantization process.

[0125] In summary, the present invention effectively solves the key problems in radio frequency fingerprint recognition through multiple innovative technologies. First, separating the feature extraction network from the recognition network improves the interpretability and scalability of the system. The feature extractor independently optimizes the representation of radio frequency signals to ensure that the features have high distinctiveness and stability, while the recognition network focuses on classification, supports more device categories, and enables open-set recognition, ensuring a high recognition rate and robustness for signals of un-trained devices. The center loss in metric learning is used to optimize the feature space, mapping the signals of the same device closer and pulling the signals of different devices farther apart, enhancing the discriminative ability of the features and the adaptability of the system to new devices, and solving the problems of accuracy and robustness in open-set recognition.

[0126] In addition, the decentralized federated learning strategy based on blockchain significantly improves data security and training transparency. Blockchain records and verifies all learning behaviors, preventing data tampering and malicious interference. Edge nodes only upload model update information to protect data privacy. At the same time, by periodically sharing and averaging model parameters, the heterogeneity of classifier structures is solved, and a unified recognition model for cross-network devices is achieved.

[0127] Finally, the efficient edge deployment and model optimization technology ensure the stable operation of the system in diverse edge environments. The dynamic pruning model is optimized according to the hardware and bandwidth conditions of edge nodes to avoid resource waste. Combining model quantization to compress the model size reduces storage and computing overheads, improves the inference speed, while maintaining high recognition performance and reducing resource consumption.

[0128] According to the cross-network deployment method of the radio frequency fingerprint recognition model of the embodiment of the present invention, the radio frequency fingerprint recognition scheme based on decentralized federated learning, by designing an extensible feature extractor and an efficient edge deployment strategy, while solving problems such as differences in the number of different network devices, open-set device recognition, and differences in computing power of edge nodes, improves the generalization, security, and practicality of the radio frequency fingerprint recognition system.

[0129] In the description of this specification, the description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0130] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise specifically defined.

Claims

1. A cross-network deployment method of a radio frequency fingerprint recognition model, characterized in that: include: Connecting each edge node to multiple wireless devices to form a sub-identification network, and building a wireless authentication network based on the multiple sub-identification networks; The radio frequency fingerprint recognition model in the wireless authentication network is divided into a feature extractor and a selector; wherein the feature extractor is used to extract the radio frequency fingerprint features of the wireless device, and the selector is used to perform individual recognition based on the radio frequency fingerprint features; The radio frequency fingerprint recognition model is trained based on feature extractor pre-training and selector federation training; wherein the feature extractor pre-training includes using the pre-collected signal information to train the parameter information of the feature extractor part and sending it to each edge node; the selector federation training includes the edge nodes in each sub-recognition network using the blockchain to implement a decentralized federation training strategy and tailoring according to the node computing power level; After the RF fingerprint recognition model training is completed, the trained model parameters are obtained from the blockchain and deployed to each edge node.

2. The method according to claim 1, characterized in that: In the process of collecting pre-training samples for the feature extractor pre-training, each edge node j obtains the M j A wireless device acquires the signal and processes it, including synchronization, preamble extraction, frequency offset compensation, and power normalization, and stores it in the training data set, indicating: Among them, T j represents the training dataset collected at edge node j, r j represents the received IQ sample, s j Indicates the corresponding wireless device tag, K j Refers to the data set T j The number of signals in .

3. The method according to claim 2, characterized in that After the pre-training samples are collected, pre-training is performed to generate feature extractor parameters and save them in the network's blockchain; the input of the feature extractor is the IQ samples corresponding to different wireless device signals, and the output is the extracted RF fingerprint.

4. The method according to claim 3, characterized in that In selector federation training, all The edge nodes participating in the training adopt the RF fingerprint recognition model with the same structure, in which the feature extractor part uses the pre-trained model parameters; the number of classification heads in the selector part is set to the sum of the wireless device categories in all networks.

5. The method according to claim 4, characterized in that During the training process, each edge node uses wireless device signals to perform multiple rounds of training, then performs a preset proportion of pruning on the models of each network, and finally performs a federated averaging operation to update the parameters of the selector part of each network model to the arithmetic mean of all network model parameters; After training is completed, the federated averaging operation is performed again to ensure that the model parameters of all networks are consistent.

6. The method according to claim 5, characterized in that Training of RF fingerprint recognition models, including: S31, edge node j collects the signal x j Processing is performed to convert the signal into IQ samples for training And construct the dataset T j ; S32, pre-training to obtain a feature extractor φ = Model (x; θ), where x is the input and θ is the weight of the model; S33, each edge node uses the feature extractor φ to obtain the RF fingerprint corresponding to the original signal And build the radio frequency fingerprint library H of each network j =φ(T j ); S34, distribute the initial RF fingerprint recognition model f to edge node j j init , where the feature extractor is φ, and the parameters of the feature extractor part are frozen, and the number of classification heads in the selector part is set to the sum of the wireless device categories in all networks, S35, start federated training, edge node j uses the radio frequency fingerprint library H j Train the model for n rounds to get the intermediate model f j 1 ; S36, performing a pruning operation on the weight parameter of the selector part, including introducing two additional parameters weight_orig and weight_mask for the selector, where weight = weight_orig × weight_mask, weight_orig represents the original weight matrix for back propagation, weight_mask is the weight mask, and weight represents the weight value after pruning; S37, sharing training results, including performing federated averaging operations and calculating the arithmetic mean of the weight_orig parameter in It is a model The i-th parameter of the selector part and are updated to θ i ; S38, repeat steps S36 and S37 until all N rounds are completed; finally, share the training results so that S39, for the trained f final The model performs weight quantization.

7. The method according to claim 1, characterized in that The feature extractor pre-training includes quantifying the feature extraction effect based on cross entropy loss and metric learning-based loss functions; cross entropy loss is used to measure the difference between the probability distribution predicted by the model and the true label, and the mathematical expression is: Where N is the number of samples, C is the number of categories, and y i,c is the true label of sample i in category c, p i,c The probability predicted by the model; The center loss minimizes the Euclidean distance between the sample feature vector and the center of the category to which it belongs. The mathematical expression is: Among them, h i is the feature vector of sample i, For category y i The center vector of During the training process, the model optimizes two loss functions at the same time, balancing them through the weight coefficient λ. The total loss function is:

8. The method according to claim 1, characterized in that The selector federated training includes adopting federated averaging or other training result sharing strategies, averaging some parameters of each edge node model every n training cycles. For the weight matrix Q of the fully connected layer, the mathematical expression of the federated average is: Where H is the number of models involved in the average, W j is the weight of the j-th model at this layer.

9. The method according to claim 6, characterized in that Perform pruning operations on the weight parameter of the selector part, including: At the beginning of training, a binary mask matrix mask of the same size is created for the weight matrix of each layer, and the mask is initialized to all 1s; For each layer's weight matrix W, calculate the absolute value of each weight: |W ij |fori=1,2,…,Nandj=1,2,…,M Where W ij is the weight of the i-th row and j-th column, N and M are the number of rows and columns of the weight matrix respectively; Based on the absolute value of the weight, a threshold θ is calculated, and weight connections whose absolute value is less than the threshold θ are pruned. The threshold θ is determined by presetting the pruning ratio; including selecting the smallest p% of weights for pruning at the end of each training cycle: θ=quantile(|W|,p) Where quantile(|W|,p) means that after sorting by absolute value, the weights of p% are selected as pruning targets; According to the pruning threshold, update the mask matrix mask, set the corresponding pruned weight position to 0, and keep the retained weight position to 1; In each forward propagation, the pruned weight matrix W is masked effective =W·mask is applied to the network calculation; After the training process is completed, the fixed weight matrix W is the final W effective Matrix and remove the mask matrix.

10. The method according to claim 6, characterized in that Perform weight quantization operations on the trained RF fingerprint recognition model, including: Set the scaling factor and zero point. The scaling factor is used to determine the size of the value range, and the zero point is the offset of the quantized value. The mapping relationship of the quantized data is described by the following formula: Where x is the original floating value, s is the scaling factor, z is the zero point, Q(x) is the quantized integer value, and round(·) indicates the rounding operation; Weight quantization is to quantize the trained model parameters in floating-point format into integer format. The weight quantization operation uses the following formula: Among them, W {float} is the trained floating point weight, S w is the scaling factor calculated for the weight matrix; In the actual reasoning process, the quantized weights and activation values ​​are dequantized to restore their floating number representation. The dequantization operation is performed using the following formula: W dequantized =W quantized ·S w Among them, W dequantized is the dequantized floating number weight, W quantized is the quantized integer weight, s w is the scaling factor calculated during quantization.

Images