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

Abstract

The invention relates to a physical layer authentication scheme based on multi-scale convolution and meta learning. In order to solve the problem of data imbalance caused by difficult collection of illegal data, the scheme proposes an SMOTE oversampling algorithm based on a subcarrier weight factor to oversample minority samples, introduces weights of different subcarriers to guide the oversampling process, and ensures that newly generated minority samples are closer to actual sample distribution. In order to extract deep features of CSI data more effectively, a CNN network MDA-Net based on multi-scale convolution and an attention mechanism is designed, and a deformable convolution module is used to improve the feature extraction capability of the model. And finally, an environment difference index EDI is proposed to measure the difference between the existing environment and the basic environment, the learning rate of meta learning is adaptively adjusted based on the EDI, and the adaptive capacity of the model in a new scene and a changing scene is improved. Experiments show that the F1 score of the model in a new scene reaches 0.966, and the false positive example rate is reduced by 2.42% compared with an existing representative algorithm.

Description

Technical Field

[0001] The present invention relates to the field of communication security, and in particular to physical layer authentication, and specifically to a physical layer authentication scheme based on multi-scale convolution and meta-learning Background Art

[0002] With the continuous development and widespread application of communication technology, more and more private data is transmitted through the network, and it is particularly important to ensure the security of communication. With the continuous increase of network access devices, the problems of traditional authentication schemes based on encryption technology, such as difficult management, high resource usage and high latency, have gradually emerged. Physical layer security authentication has become a key technology in the current wireless communication security field due to its low latency and high security. Physical layer authentication verifies the identities of both parties in communication by analyzing the unique characteristics of the wireless channel, providing an efficient and lightweight authentication scheme for networks in various emerging scenarios. The application of deep learning technology can effectively improve the accuracy of physical layer authentication, but the existing deep learning models are mainly targeted at specific scenarios and are often difficult to adapt to changes in channel characteristics caused by environmental changes. How to make the deep learning-based authentication model effectively adapt to the changing environment has become a hot issue in physical layer authentication research.

[0003] For physical layer security issues in changing environments, existing solutions can be mainly divided into two categories: data-driven strategies and model-driven strategies. Data-driven adaptive strategies emphasize using the characteristics of the data itself to enhance the adaptability of the model, such as through transfer learning, online learning, and incremental learning techniques. In order to solve 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 average-based data enhancement algorithm to improve authentication accuracy, and used ResNet to extract CSI data features of different transmitters. Chen et al. proposed a physical layer authentication mechanism based on transfer learning and triple pooling network (TP-Net), and generated training data sample sets through data enhancement methods, which improved the performance and generalization ability of the model. However, data-driven adaptive strategies mostly focus on training highly robust models to adapt to environmental changes, which requires a large amount of training data and computing resources. Model-driven adaptive strategies respond to environmental changes by adjusting the structure or learning mechanism of the model, such as adaptive model algorithms and multi-task learning. Han et al. studied the impact of dynamic environments on authentication performance, extracted deep feature information of CSI data through Bahdanau Attention Autoencoder (BAAE) to reduce inter-carrier interference, and proposed an authentication strategy based on model-driven learning algorithm to solve the problem of authentication performance degradation in dynamic scenarios. However, the model-driven learning algorithm still relies on the initial learning ability of the original training data. When encountering a new environment that is significantly different from the training data, it is difficult for the model to make timely adjustments through a small amount of data to accurately identify new channel characteristics. Summary of the invention

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

[0005] The object of the present invention is achieved in this way:

[0006] In order to solve the problem of CSI data imbalance, this solution first proposes an oversampling algorithm CSI-SMOTE based on subcarrier weight factors, and introduces subcarrier weights to guide the oversampling process, ensuring that the generated new samples can more accurately reflect the true signal feature distribution. Subsequently, the present invention designs a CNN network MDA-Net based on multi-scale convolution and attention mechanism, and uses it as the base learner of the meta-learning model, while introducing a deformable convolution module to enhance the feature extraction capability of the model. Finally, the present invention proposes an environmental difference index EDI to measure the difference between the existing environment and the initial environment, and further proposes an adaptive learning rate meta-learning method ALRM based on EDI for model training, which improves the adaptability of the authentication model in a changing environment.

[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 original CSI data through the Linux 802.11n CSI Tool, calculate the amplitude and phase information of each subcarrier and splice them. Then perform noise reduction and normalization processing, and classify the CSI data according to the data collection scenario.

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

[0011] Step 3: In the meta-learning training phase based on MDA-Net, multi-scale convolution and self-attention mechanisms are used to extract deep features of 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 by training on data from different scenarios.

[0012] Step 4: In the EDI calculation phase, the EDI is calculated to measure the difference between the new scene and the baseline scene, and the initial learning rate is calculated through the EDI.

[0013] Step 5: In the new scene adaptation phase, the parameters of the meta-learner are updated and fine-tuned through a small amount of data in 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 the present invention are:

[0015] Aiming at the problem that the existing models in the field of physical layer authentication have poor adaptability in changing scenarios, the present invention proposes a physical layer authentication scheme MSC-PLA based on multi-scale convolution and meta-learning; Aiming at the problem of data imbalance caused by the difficulty in collecting illegal data, the scheme first proposes a SMOTE oversampling algorithm based on subcarrier weight factors to oversample minority class samples, and guides the oversampling process by introducing weights of different subcarriers to ensure that the newly generated minority class samples are closer to the actual sample distribution. In order to more effectively extract the deep features of CSI data, the present invention designs a CNN network MDA-Net based on multi-scale convolution and attention mechanism, and uses a deformable convolution module to improve the feature extraction ability of the model. Finally, the present invention proposes an environmental difference index EDI to measure the difference between the existing environment and the basic environment, and adaptively adjusts the learning rate of the meta-learning process based on EDI to improve the adaptability of the model in new and changing scenarios. Experiments show that the F1 score of the authentication model proposed by the present invention reaches 0.966 in the new scenario, and the false positive rate is reduced by 2.42% compared with the existing representative algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0018] Figure 3 It is an experimental architecture for comparing different parameters of multi-scale convolution modules.

[0019] Figure 4 It is a meta-learning process.

[0020] Figure 5 It is a performance comparison of different physical layer authentication schemes. DETAILED DESCRIPTION

[0021] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

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

[0023] Three ThinkPad X201 laptops with Intel 5300 network cards were used for data collection. The Ubuntu operating system with system version 14.06LTS and kernel version 4.4 was used as the information collection platform. The Linux802.11n CSI Tool toolkit was used as the CSI data collection tool. The Savitzky-Golay (SG) filter was used to denoise the CSI data. The SG filter smoothes the data by fitting a local polynomial to the data points within a sliding window, and replaces the original data points with the center value of the polynomial window. The SG filter can effectively remove random noise from the CSI data while retaining important features such as the peak, valley, and waveform trend of the signal. Min-Max Normalization is a commonly used data preprocessing method, which scales the data on each feature dimension so that all feature values ​​are limited to a fixed range. For each input data T, the data T of the lth dimension l , the process of scaling it to [0,1] through maximum and minimum normalization can be expressed as:

[0024]

[0025] Step 2: Data oversampling: Use the CSI-SMOTE algorithm proposed in the present invention to perform data oversampling 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, the present invention proposes a SMOTE oversampling algorithm CSI-SMOTE based on subcarrier weight factors, which can improve the performance of the physical layer authentication system when facing data imbalance problems. The algorithm combines the weight information of each subcarrier to measure the contribution of different subcarriers in the CSI data to the authentication performance. The oversampling process is guided according to the weights of different subcarriers in the CSI data to ensure that the generated new samples can more accurately reflect the true signal feature distribution. This not only enhances the representativeness of minority samples, but also improves the model's ability to detect abnormal behaviors in complex environments, and has 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 the weight coefficient. MI can measure the interdependence of two variables. The importance of different features can be obtained by calculating the MI value of each feature and the target variable. The weight of each subcarrier w j The calculation formula is as follows:

[0028]

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

[0030]

[0031] After obtaining the weight of each subcarrier, randomly select a sample from the minority class samples and use the weighted Euclidean distance calculation to obtain its nearest k samples of the same class. p and sc q , which is based on the weighted Euclidean distance D of the weight factor w (sc p ,sc q ) 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 by interpolation. new During the process, each subcarrier sc newj The value of is calculated according to the following weighted formula:

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

[0035] Where sc pj and sc qj They represent the subcarrier values ​​in the original sample respectively, and δ is a random number in [0,1].

[0036] When processing unbalanced data sets, the oversampling algorithm based on CSI-SMOTE calculates the mutual information of each subcarrier as a weight factor, making important features more influential when generating new samples and improving the quality of synthetic samples. The weighted Euclidean distance is used to calculate the distance between samples, taking into account the weight of each dimensional feature, which has a more practical physical meaning. At the same time, the weight factor is considered in the interpolation process, which reflects the complex relationship and signal characteristics between subcarriers in the original data, and can better simulate real data.

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

[0038] Step 3: In the meta-learning training phase based on MDA-Net, multi-scale convolution and self-attention mechanisms are used to extract deep features of 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 by training on data from different scenarios.

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

[0040] Based on the characteristics of the original CSI data and the matrix size, the present invention uses three different sizes of convolution kernels, 1×1, 3×3 and 5×5, to extract data features. The 1×1 convolution kernel can capture local detail features and enhance the model's sensitivity to small spatial changes; the 3×3 convolution kernel can balance local details and large-scale spatial features; and the 5×5 convolution kernel enables the model to capture spatial information in a larger range. Through the multi-scale convolution strategy, the model can extract and fuse spatial features of different scales of data.

[0041] After using multi-scale convolution for feature extraction, the features of different dimensions are fused using the feature concatenation method, that is, the outputs of different parts are merged along the channel dimension to obtain a richer feature representation. In order to reduce the number of parameters and the amount of calculation for subsequent model training, 2×2 maximum pooling is used to compress the feature space.

[0042] In order to make the model better extract information between channels, the present invention uses the self-attention mechanism to further process the fusion features extracted by multi-scale convolution. The present invention uses the fusion feature matrix T processed by the multi-scale convolution layer as the input sequence of the attention mechanism. For any given sequence Where D x Represents the feature dimension, N represents the length of the sequence, and the self-attention score can be calculated by the following formula:

[0043]

[0044] In order to better utilize the ability of deformable convolution to process complex and variable samples, the present invention replaces all the remaining convolution blocks of the CNN network except the first multi-scale convolution module with deformable convolution blocks.

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

[0046] Figure 3 A comparative experiment of different parameters of the multi-scale convolution module is demonstrated.

[0047] Step 4: In the EDI calculation phase, the EDI is calculated to measure the difference between the new scene and the baseline scene, and the initial learning rate is calculated through the EDI.

[0048] The present invention proposes an environmental difference index (EDI) to measure the difference between the new environment and the basic environment, and uses EDI to guide the setting of the initial learning rate in the meta-training phase. When facing different tasks, if the current environment and the basic environment are quite different, a relatively high initial learning rate is used to speed up the convergence of the model; when the environmental difference is small, a smaller initial learning rate is provided to help the model make more detailed adjustments based on the existing parameters.

[0049] Based on the weighted combination of the KL divergence, mean and variance between the distributions of the CSI training set data, the final calculation result of EDI can be obtained, and the formula is as follows:

[0050]

[0051] Where α, β and γ are weight coefficients, which are used to balance the contribution of KL divergence, variance and mean to the learning rate adjustment. The weight coefficients can be optimized according to the specific usage scenario. The hyperbolic tangent function tanh can be used to map the value to the [-1,1] interval to ensure that EDI adjusts the learning rate within a stable and reasonable range.

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

[0053] η=η base (1+EDI)

[0054] Step 5: In the new scene adaptation phase, the parameters of the meta-learner are updated and fine-tuned through a small amount of data in 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] The present invention first divides the data set into different tasks according to different scenarios of collecting CSI data. Each task contains corresponding legal device and illegal device data. In each task, the data set is divided into a query set and a support set, and the parameters θ of the model are randomly initialized.

[0056] Then, we enter the meta-learning training phase. The model training process can be divided into two optimization phases: inner loop and outer loop. In the inner loop optimization, for each task T i , use T i The query set in the model is updated by gradient descent to obtain the updated parameters θ of the task i ', and then use task T i The support set in is used to evaluate the model performance and calculate the loss function about the parameter θ. After calculating the loss function of a meta-batch, the outer loop phase is entered to calculate each task T using the query set. i About the adjusted model parameters θ i ′, and use the optimization algorithm to update the model parameter θ. Repeat the above steps until the predetermined number of iterations is met, and output the final optimal parameter θ * .

[0057] In the meta-test phase, a new task T is used new Test parameter θ * First, use T new The support set is used to update the model parameters of the base learner to effectively adapt to the new task and output the latest parameters θ new as the parameters of the final model. Then use T new The query set is used to evaluate the performance of the model and the effect 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 is shown, and the performance comparison of different physical layer authentication schemes is shown. Figure 5 shown.

Claims

1. A physical layer authentication scheme based on multi-scale convolution and meta-learning, characterized in that: The method comprises the following steps: Step 1: Perform data preprocessing. After obtaining the original CSI data through the Linux 802.11n CSI Tool, calculate the amplitude and phase information of each subcarrier and splice them. Then perform noise reduction and normalization processing, and classify the CSI data according to the data collection scenario. Step 2: Data oversampling: using the CSI-SMOTE algorithm proposed in the present invention to perform data oversampling to increase the number of negative samples and enhance the model's learning ability for negative samples; Step 3: In the meta-learning training phase based on MDA-Net, the deep features of CSI data are extracted by using multi-scale convolution and self-attention mechanism. 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 by training on data from different scenarios. Step 4: Environmental Difference Index (EDI) calculation phase: calculate the environmental difference index (EDI) to measure the difference between the new scene and the benchmark scene, and calculate the initial learning rate through EDI; Step 5: In the new scenario adaptation phase, the parameters of the meta-learner are updated and fine-tuned through a small amount of data in the new scenario to improve the model's adaptability to the new scenario. Finally, the fine-tuned model is used for physical layer authentication in the new scenario.

2. The process of using the CSI-SMOTE algorithm to perform data oversampling in step 2 is characterized in that: The mutual information (MI) of different subcarriers and category labels in the CSI information is used as the weight coefficient. MI can measure the interdependence of two variables. The importance of different features can be obtained by calculating the MI value of each feature and the target variable. The weight w of each subcarrier j The calculation formula is as follows: Where m is the total number of subcarriers, I(sc j ; L) for each subcarrier sc j The calculation formula of the mutual information between X and Y and the category label L can be expressed as follows: Where 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 distribution of X and Y.

3. The network structure of MDA-Net described in step 3 is characterized in that: Three convolution kernels of different sizes, 1×1, 3×3, and 5×5, are used to extract data features. The 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.

4. The EDI calculation process in step 4 is characterized by: The EDI is calculated using KL divergence, variance and mean, as follows: Where α, β and γ are weight coefficients, which are used to balance the contribution of KL divergence, variance and mean to the learning rate adjustment; Based on EDI, the formula for initial calculation rate can be expressed as follows: the=the base ·(1+EDI) The weight coefficient can be optimized according to the specific usage scenario. The hyperbolic tangent function tanh can be used to map the value to the [-1,1] interval to ensure that EDI adjusts the learning rate within a stable and reasonable range.

5. The meta-learning process of adaptive initial learning rate in step 5 is characterized by: Applies an initial learning rate to the MAML training process.

Images