Physical layer authentication scheme based on multi-scale convolution and meta-learning

Abstract

This invention relates to a physical layer authentication scheme based on multi-scale convolution and meta-learning. Addressing the data imbalance problem caused by the difficulty in collecting illegitimate data, this scheme proposes a SMOTE oversampling algorithm based on subcarrier weight factors to oversample minority class samples. Different subcarrier weights are introduced to guide the oversampling process, ensuring that the newly generated minority class samples more closely resemble the actual sample distribution. To more effectively extract deep features from CSI data, this invention designs a CNN network MDA-Net based on multi-scale convolution and attention mechanisms, using deformable convolution modules to enhance the model's feature extraction capabilities. Finally, this invention proposes an Environmental Dissimilarity Index (EDI) to measure the difference between the existing environment and the baseline environment, and adaptively adjusts the learning rate of meta-learning based on the EDI to improve the model's adaptability in new and changing scenarios. Experiments show that the model proposed in this invention achieves an F1 score of 0.966 in new scenarios, and the false positive rate is reduced by 2.42% compared to existing representative algorithms.

Description

Technical Field

[0001] This invention relates to the field of communication security, and more particularly to physical layer authentication, specifically a physical layer authentication scheme based on multi-scale convolution and meta-learning. Background Technology

[0002] With the continuous development and widespread application of communication technologies, more and more private data is being transmitted through networks, making communication security paramount. As the number of network access devices increases, traditional encryption-based authentication schemes are increasingly revealing problems such as management difficulties, high resource consumption, and excessive latency. Physical layer authentication (PLA), due to its low latency and high security, has become a key technology in the field of wireless communication security. PLA verifies the identities of communicating parties by analyzing the unique characteristics of wireless channels, providing an efficient and lightweight authentication solution for networks in various emerging scenarios. The application of deep learning technology can effectively improve the accuracy of PLA; however, existing deep learning models are mainly designed for specific scenarios and often struggle to adapt to changes in channel characteristics caused by environmental variations. How to enable deep learning-based authentication models to effectively adapt to constantly changing environments has become a hot research topic in PLA.

[0003] To address physical layer security issues in changing environments, existing solutions can be broadly categorized into two types: data-driven strategies and model-driven strategies. Data-driven adaptive strategies emphasize leveraging the inherent characteristics of data to enhance the model's adaptability, such as through transfer learning, online learning, and incremental learning techniques. To address the impact of channel diversity on authentication accuracy in mobile environments, Jing et al. proposed a parameter-based transfer learning authentication strategy, combined with an exponential averaging-based data augmentation algorithm to improve authentication accuracy, and used ResNet to extract CSI data features from different transmitters. Chen et al. proposed a physical layer authentication mechanism based on transfer learning and a triple pooling network (TP-Net), and improved model performance and generalization ability by generating training data sample sets through data augmentation methods. However, data-driven adaptive strategies mostly focus on training highly robust models to adapt to environmental changes, requiring substantial training data and computational resources. Model-driven adaptive strategies, on the other hand, adapt to environmental changes by adjusting the model's structure or learning mechanism, such as adaptive model algorithms and multi-task learning. Han et al. investigated the impact of dynamic environments on authentication performance. They used the Bahdanau attention autoencoder (BAAE) to extract deep feature information from CSI data to mitigate inter-carrier interference and proposed an authentication strategy based on a model-driven learning algorithm to address the performance degradation in dynamic scenarios. However, model-driven learning algorithms still rely on the initial learning ability of the original training data. When encountering a new environment significantly different from the training data, the model struggles to accurately identify new channel characteristics by making timely adjustments with limited data. Summary of the Invention

[0004] The purpose of this invention is to propose a physical layer authentication scheme based on multi-scale convolution and meta-learning to solve the problem of complex and ever-changing wireless environments.

[0005] The objective of this invention is achieved as follows:

[0006] To address the issue of imbalanced CSI data, this invention first proposes an oversampling algorithm, CSI-SMOTE, based on subcarrier weight factors. By introducing subcarrier weights to guide the oversampling process, it ensures that the generated new samples more accurately reflect the true signal feature distribution. Subsequently, this invention designs a CNN network, MDA-Net, based on multi-scale convolution and attention mechanisms, and uses it as the base learner for the meta-learning model. A deformable convolution module is also introduced to enhance the model's feature extraction capabilities. Finally, this invention proposes an Environmental Dissimilarity Index (EDI) to measure the difference between the current environment and the initial environment. Based on EDI, an adaptive learning rate meta-learning method, ALRM, is further proposed for model training, improving the certification model's adaptability to changing environments.

[0007] The specific method is as follows:

[0008] A physical layer authentication scheme based on multi-scale convolution and meta-learning includes the following steps:

[0009] Step 1: Perform data preprocessing. After obtaining the raw CSI data through the Linux 802.11n CSI Tool, calculate the amplitude and phase information of each subcarrier and stitch them together. Then, perform noise reduction and normalization processing, and classify the CSI data according to the data acquisition scenario.

[0010] Step 2: Data oversampling. The CSI-SMOTE algorithm proposed in this invention is used to oversample the data to increase the number of negative samples and enhance the model's learning ability for negative samples.

[0011] Step 3: The meta-learning training stage based on MDA-Net uses multi-scale convolution and self-attention mechanisms to extract deep features from CSI data. Then, the MAML strategy in meta-learning is used to train CSI data in different scenarios to improve the model's adaptability to new tasks. The pre-trained meta-learner parameters are obtained through training on data from different scenarios.

[0012] Step 4: Environmental Difference Index (EDI) Calculation Stage. The Environmental Difference Index (EDI) is calculated to measure the difference between the new scenario and the baseline scenario, and the initial learning rate is calculated through the EDI.

[0013] Step 5, the new scene adaptation stage, involves updating and fine-tuning the parameters of the meta-learner using a small amount of data from the new scene to improve the model's adaptability to the new scene. Finally, the fine-tuned model is used for physical layer authentication in the new scene.

[0014] The positive effects of this invention are:

[0015] This invention addresses the poor adaptability of existing physical layer authentication (PSE) models in changing scenarios by proposing the MSC-PLA PSE authentication method based on multi-scale convolution and meta-learning. To address the data imbalance problem caused by the difficulty in collecting illegitimate data, this method first proposes the SMOTE oversampling algorithm based on subcarrier weight factors to oversample minority class samples. By introducing different subcarrier weights to guide the oversampling process, it ensures that the newly generated minority class samples more closely resemble the actual sample distribution. To more effectively extract deep features from CSI data, this invention designs a CNN network MDA-Net based on multi-scale convolution and attention mechanisms, while using deformable convolution modules to enhance the model's feature extraction capabilities. Finally, this invention proposes an Environmental Dissimilarity Index (EDI) to measure the difference between the existing environment and the baseline environment, and adaptively adjusts the learning rate of the meta-learning process based on the EDI to improve the model's adaptability in new and changing scenarios. Experiments show that the authentication model proposed in this invention achieves an F1 score of 0.966 in new scenarios, and the false positive rate is reduced by 2.42% compared to existing representative algorithms. Attached Figure Description

[0016] Figure 1 This is a comparison of the performance of different oversampling techniques.

[0017] Figure 2 It is the MDA-Net network.

[0018] Figure 3 This is an experimental architecture for comparing different parameters of a multi-scale convolution module.

[0019] Figure 4 This is the meta-learning process.

[0020] Figure 5 This is a performance comparison of different physical layer authentication schemes. Detailed Implementation

[0021] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0022] Step 1: Perform data preprocessing. After obtaining the raw CSI data through the Linux 802.11n CSI Tool, calculate the amplitude and phase information of each subcarrier and stitch them together. Then, perform noise reduction and normalization processing, and classify the CSI data according to the data acquisition scenario.

[0023] Data was collected using three ThinkPad X201 laptops equipped with Intel 5300 network cards. The data acquisition platform was an Ubuntu operating system running version 14.06 LTS with kernel version 4.4. The Linux 802.11n CSI Toolkit was used for CSI data acquisition, and a Savitzky-Golay (SG) filter was employed.

[73] To denoise CSI data, the SG filter smooths the data by performing local polynomial fitting on the data points within a sliding window, and replaces the original data points with the center value of the polynomial window. The SG filter effectively removes random noise from CSI data while preserving important features such as signal peaks, valleys, and waveform trends. Min-MaxNormalization is a common data preprocessing technique that scales the data in each feature dimension, limiting all feature values ​​to a fixed range. For each input data T, the data in the l-th dimension... l The process of scaling it to the range [0,1] using max-min normalization can be represented as:

[0024]

[0025] Step 2: Data oversampling. The CSI-SMOTE algorithm proposed in this invention is used to oversample the data to increase the number of negative samples and enhance the model's learning ability for negative samples.

[0026] Based on the characteristics of CSI data, this invention proposes a SMOTE oversampling algorithm, CSI-SMOTE, based on subcarrier weighting factors, which can improve the performance of physical layer authentication systems when facing data imbalance problems. This algorithm combines the weight information of each subcarrier to measure the contribution of different subcarriers in the CSI data to authentication performance. The oversampling process is guided by the weights of different subcarriers in the CSI data, ensuring that the generated new samples more accurately reflect the true signal feature distribution. This not only enhances the representativeness of minority class samples but also improves the model's ability to detect abnormal behavior in complex environments, resulting in higher accuracy in cross-scenario authentication tasks.

[0027] CSI-SMOTE uses the mutual information (MI) of different subcarriers and class labels in the CSI information as weighting coefficients. MI measures the interdependence between two variables, and the importance of different features can be obtained by calculating the MI values ​​of each feature and the target variable. The weight w for each subcarrier... j The calculation formula is as follows:

[0028]

[0029] In the formula, m is the total number of subcarriers, and I(sc) j ;L) for each subcarrier sc j Mutual information with category label L. Assuming X represents the subcarrier variable, Y represents the label variable, p(x,y) represents the joint probability distribution of X and Y, and p(x) and p(y) represent the marginal probability distributions of X and Y, then the formula for calculating the mutual information (MI) between X and Y can be expressed as follows:

[0030]

[0031] After obtaining the weight of each subcarrier, samples are randomly selected from the minority class samples, and their k nearest neighbors (similar to each other) are calculated using weighted Euclidean distance. For any two samples sc p and sc q Its weighted Euclidean distance D based on weight factors w (sc p ,sc q It can be defined as follows:

[0032]

[0033] For each minority class sample, a sample is randomly selected from its k nearest neighbors as an auxiliary sample, and a new sample is synthesized through interpolation. This process is repeated when generating the new sample sc. new During the process, each subcarrier sc newj The value is calculated using the following weighted formula:

[0034] sc new j = sc pj +w j ·(sc qj -sc pj )·δ

[0035] In the formula sc pj and sc qj These represent the values ​​of the subcarriers in the original sample, and δ is a random number in [0,1].

[0036] The CSI-SMOTE-based oversampling algorithm, when processing imbalanced datasets, calculates the mutual information of each subcarrier as a weighting factor, giving important features a greater influence when generating new samples and improving the quality of synthesized samples. Calculating the distance between samples using weighted Euclidean distance takes into account the weight of each dimension of the feature, providing more practical physical meaning. Furthermore, considering the weighting factor during interpolation reflects the complex relationships and signal characteristics between subcarriers in the original data, enabling a better simulation of real-world data.

[0037] Figure 1The performance comparison of different oversampling techniques is shown.

[0038] Step 3: The meta-learning training stage based on MDA-Net uses multi-scale convolution and self-attention mechanisms to extract deep features from CSI data. Then, the MAML strategy in meta-learning is used to train CSI data in different scenarios to improve the model's adaptability to new tasks. The pre-trained meta-learner parameters are obtained through training on data from different scenarios.

[0039] This invention designs a convolutional neural network MDA-Net based on multi-scale convolution and self-attention mechanism as a base learner for meta-learning. At the same time, it introduces deformable convolutional networks (DCN) into the subsequent CNN network to further improve the model's adaptability to unknown changes.

[0040] Based on the characteristics and matrix size of the original CSI data, this invention uses three convolutional kernels of different sizes—1×1, 3×3, and 5×5—to extract data features. The 1×1 kernel captures local details, enhancing the model's sensitivity to subtle spatial changes; the 3×3 kernel balances local details with large-scale spatial features; and the 5×5 kernel enables the model to capture spatial information over a wider range. Through this multi-scale convolution strategy, the model can extract and fuse spatial features from different scales of the data.

[0041] After feature extraction using multi-scale convolution, the obtained features from different dimensions are fused using feature concatenation, which merges the outputs of different parts along the channel dimension to obtain a richer feature representation. To reduce the number of parameters and the computational cost of subsequent model training, 2×2 max pooling is used to compress the feature space.

[0042] To improve the model's ability to extract information between channels, this invention employs a self-attention mechanism to further process the fused features extracted by multi-scale convolution. This invention uses the fused feature matrix T, processed by the multi-scale convolutional layers, as the input sequence for the attention mechanism. For any given sequence... Where D x The self-attention score can be calculated using the following formula: (where N represents the feature dimension and N represents the sequence length)

[0043]

[0044] To better utilize the ability of deformable convolutions to process complex and variable samples, this invention replaces all remaining convolutional blocks in the CNN network with deformable convolutional blocks, except for the first multi-scale convolutional module.

[0045] MDA-Net network architecture such as Figure 2 As shown.

[0046] Figure 3 The experiment demonstrates a comparison of different parameters for a multi-scale convolution module.

[0047] Step 4: Environmental Difference Index (EDI) Calculation Stage. The Environmental Difference Index (EDI) is calculated to measure the difference between the new scenario and the baseline scenario, and the initial learning rate is calculated through the EDI.

[0048] This invention proposes an Environmental Difference Index (EDI) to measure the difference between the new environment and the baseline environment, and uses EDI to guide the setting of the initial learning rate during the meta-training phase. When facing different tasks, if the difference between the current environment and the baseline environment is large, a relatively high initial learning rate is used to accelerate the convergence speed of the model; when the environmental difference is small, a smaller initial learning rate is provided to help the model make more refined adjustments based on the existing parameters.

[0049] The final EDI result can be obtained by weighting the KL divergence, mean, and variance of the distributions in the CSI training set data. The formula is as follows:

[0050]

[0051] In the formula, α, β, and γ are weighting coefficients used to balance the contributions of KL divergence, variance, and mean to the learning rate adjustment. The weighting coefficients can be optimized according to the specific use case. Using the hyperbolic tangent function tanh can map the value to the interval [-1, 1], ensuring that the EDI adjusts the learning rate within a stable and reasonable range.

[0052] Based on EDI, the formula for the initial calculation rate can be expressed as follows:

[0053] η = η base ·(1+EDI)

[0054] Step 5, the new scene adaptation stage, involves updating and fine-tuning the parameters of the meta-learner using a small amount of data from the new scene to improve the model's adaptability to the new scene. Finally, the fine-tuned model is used for physical layer authentication in the new scene.

[0055] This invention first divides the dataset into different tasks according to different scenarios of CSI data collection. Each task contains corresponding legitimate and illegitimate device data. In each task, the dataset is divided into a query set and a support set, and the model parameters θ are randomly initialized.

[0056] The model then enters the meta-learning training phase, which can be divided into two levels of optimization: an inner loop and an outer loop. In the inner loop optimization, for each task T... i Using T i The query set is used to update the model using gradient descent, resulting in the updated parameters θ' for the task. i Then use task T i The model performance is evaluated using the support set in the query set, and the loss function with respect to parameter θ is computed. After computed the loss function for a meta-batch of meta-learning tasks, the outer loop phase is entered, where the loss function for each task T is computed using the query set. i Regarding the adjusted model parameter θ' i The loss is calculated, and the model parameters θ are updated using an optimization algorithm. This process is repeated until a predetermined number of iterations is reached, and the final optimal parameters θ are output. * .

[0057] During the meta-testing phase, the new task T is used. new Test parameter θ * The generalization ability. First, use T. new The support set is used to update the model parameters of the base learner, so as to effectively adapt to new tasks and output the latest parameters θ. new As parameters for the final model. Then T is used. new The query set is used to evaluate the model's performance and the effectiveness of meta-learning, and the model parameter update mechanism is adjusted and optimized based on the evaluation results.

[0058] Figure 4 The meta-learning process was demonstrated, and the performance comparisons of different physical layer authentication schemes were shown. Figure 5 As shown.

Claims

1. A physical layer authentication scheme based on multi-scale convolution and meta-learning, characterized in that, The solution includes the following steps: Step 1: Perform data preprocessing. After obtaining the raw CSI data through the Linux 802.11n CSI Tool, calculate the amplitude and phase information of each subcarrier and stitch them together. Then, perform noise reduction and normalization processing, and classify the CSI data according to the data acquisition scenario. Step 2: Data oversampling. The CSI-SMOTE algorithm is used for data oversampling. Mutual information (MI) between different subcarriers and class labels in the CSI information is used as weighting coefficients. MI measures the interdependence between two variables. By calculating the MI values ​​of each feature and the target variable, the importance of different features can be obtained, thereby increasing the number of negative samples and enhancing the model's learning ability for negative samples. The weights of each subcarrier are... In the formula, m is the total number of subcarriers, and I(sc) j ;L) for each subcarrier sc j The mutual information between category label L and category label L; the formula for calculating the mutual information (MI) between X and Y can be expressed as: Where X represents the subcarrier variable, Y represents the tag variable, p(x,y) represents the joint probability distribution of X and Y, and p(x) and p(y) represent the marginal probability distributions of X and Y; Step 3: The meta-learning training stage based on MDA-Net uses multi-scale convolution and self-attention mechanisms to extract deep features from CSI data. Then, the MAML strategy in meta-learning is used to train CSI data in different scenarios to improve the model's adaptability to new tasks. The pre-trained meta-learner parameters are obtained through training on data from different scenarios. Step 4: Environmental Difference Index (EDI) Calculation Stage. The Environmental Difference Index (EDI) is calculated to measure the difference between the new scenario and the baseline scenario, and the initial learning rate is calculated through the EDI. Step 5, New Scene Adaptation Stage: Using a small amount of data from the new scene, the parameters of the meta-learner are updated and fine-tuned to improve the model's adaptability to the new scene. Finally, the fine-tuned model is used for physical layer authentication in the new scene.

2. The physical layer authentication scheme based on multi-scale convolution and meta-learning according to claim 1, characterized in that, In step three, the MDA-Net network structure uses three convolutional kernels of different sizes, 1×1, 3×3 and 5×5, to extract data features. At the same time, a self-attention mechanism is used to further process the fused features extracted by multi-scale convolution. Finally, deformable convolution is used to extract CSI data features.

3. The physical layer authentication scheme based on multi-scale convolution and meta-learning according to claim 1, characterized in that, In step four, the EDI calculation process uses KL divergence, variance, and mean to calculate the EDI, as shown in the following formula: In the formula, α, β and γ are weighting coefficients used to balance the contributions of KL divergence, variance and mean to the learning rate adjustment; Based on EDI, the formula for the initial calculation rate can be expressed as follows: the=the base ·(1+EDI) The weight coefficients can be optimized according to the specific use case. Using the hyperbolic tangent function tanh can map the values ​​to the interval [-1,1], ensuring that the EDI adjusts the learning rate within a stable and reasonable range.

4. The physical layer authentication scheme based on multi-scale convolution and meta-learning according to claim 1, characterized in that, In step five, the meta-learning process with adaptive initial learning rate applies the initial learning rate to the MAML training process.

Images