Physical layer authentication scheme based on sparse autoencoders and knowledge distillation

Abstract

This invention relates to a physical layer authentication scheme based on sparse autoencoders and knowledge distillation. To enhance the model's learning ability, the scheme first proposes SensSpec-GAN, a generative adversarial network data augmentation model based on Sensitivity-Specificity Loss. This model enhances data diversity and the model's ability to identify negative samples by introducing sensitivity and specificity constraints into the generative adversarial network. Subsequently, addressing the computational burden caused by the high dimensionality of CSI data, this invention proposes SABES, a sparse autoencoder model based on the vulture optimization algorithm, which can enhance the model's feature representation ability by automatically adjusting the model's hyperparameters. Finally, this invention proposes AMTDF, a multi-teacher knowledge distillation model with adaptive weights. This model improves the model's performance on the target task by adjusting the contributions of different teacher models to the student model through an adaptive weight mechanism. Experimental results show that the authentication model proposed in this scheme achieves an accuracy of 96.32%, and the training speed is 2.46 times faster than the benchmark algorithm.

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 sparse autoencoders and knowledge distillation. Background Technology

[0002] With the continuous development of wireless communication technology, the security challenges faced by wireless communication systems are also increasing. As the underlying layer responsible for signal transmission in a communication system, the physical layer is constantly being targeted by evolving attack methods, and traditional key-based authentication methods are gradually showing their limitations. Physical layer authentication utilizes the inherent properties of wireless channels for identity verification, offering advantages such as higher security and lower computational complexity compared to key-based authentication technologies, and has become a key technology in the current authentication field. Deep learning can significantly improve the accuracy of physical layer authentication, and deep learning-based physical layer authentication technologies have attracted increasing attention from researchers. However, with the rapid development of the Internet of Things (IoT), more and more devices need to perform complex authentication tasks in resource-constrained environments. Existing high-performance authentication schemes have large parameter scales, resulting in high authentication latency and making them difficult to deploy directly in resource-constrained environments such as the IoT. Therefore, how to make the authentication model more lightweight and efficient while ensuring secure authentication performance has become a hot issue in the field of physical layer security authentication.

[0003] For physical layer authentication (PLE) problems in resource-constrained scenarios, solutions can be broadly categorized into two types: optimization schemes based on algorithm complexity and data compression, and distributed authentication schemes. Optimization schemes based on algorithm complexity and data compression reduce the computational and storage requirements of the model through algorithm-level optimization, primarily employing lightweight algorithms, simplified model structures, and data compression techniques to reduce data dimensionality. Chen et al. proposed a PLE scheme based on a convolutional denoising autoencoder (CDAE), which uses CDAE and a weighted neural network (-NN) algorithm for feature extraction and classification, effectively eliminating noise and extracting key features from CSI data. However, when the extracted features lack sufficient discriminative power, the authentication scheme combined with the -NN algorithm suffers a significant drop in accuracy. Senigagliesi L et al. reduced the number of authentication features through data compression techniques. Experiments showed that principal component analysis, random sampling, and t-distributed random neighborhood embedding all significantly reduced the computational burden of authentication. However, when the attacker and legitimate communication participants are close together, the information loss caused by compression leads to a decrease in authentication accuracy. Distributed authentication schemes improve the efficiency and reliability of the authentication process by sharing authentication tasks among multiple nodes and leveraging network collaboration capabilities. Zhang et al. proposed a cooperative physical layer authentication scheme, COPLA, which utilizes multiple supervisory nodes for task sharing, providing a lightweight deployment and high authentication performance solution in complex wireless environments. However, compared to other physical layer authentication schemes, COPLA is not suitable for physical layer devices in isolated environments, limiting its widespread application in real-world scenarios. Zhao et al. proposed a physical layer authentication strategy for resource-constrained underwater acoustic networks (UANs). This strategy utilizes the Support Vector Machine (SVM) algorithm for distributed authentication on a small training set, which can improve node authentication accuracy while reducing resource consumption. However, this scheme is only applicable to underwater wireless acoustic networks and has poor generalization ability in other physical layer authentication scenarios. Summary of the Invention

[0004] The purpose of this invention is to propose a physical layer authentication scheme based on sparse autoencoders and knowledge distillation to solve the problem of limited resources in wireless devices.

[0005] The specific method is as follows:

[0006] The physical layer authentication scheme based on sparse autoencoders and knowledge distillation includes the following steps:

[0007] Step 1: Data Acquisition and Preprocessing. CSI data from real-world communication scenarios is acquired using the Linux 802.11n CSI Tool. The amplitude and phase of each subcarrier are then calculated based on the subcarrier complex matrix. Subsequently, a Savitzky-Golay filter is used to denoise the CSI data. Finally, the data is normalized for easier subsequent analysis and processing.

[0008] Step 2: Perform data augmentation. Use a generative adversarial network based on Sensitivity-Specificity Loss to generate high-quality experimental data, improving the model's ability to identify negative samples.

[0009] Step 3: Perform data dimensionality reduction. Reconstruct the input data using a sparse autoencoder based on the Bald Eagle optimization algorithm. The Bald Eagle optimization algorithm can automatically find the hyperparameters of the sparse autoencoder, improving its performance.

[0010] Step 4: Train the model using a knowledge distillation strategy. First, train multiple teacher models using a large amount of labeled data to perform complex security authentication tasks. Then, to verify the effectiveness of the teacher models in real-world applications, evaluate the performance of each teacher model on independent test sets and perform model selection to ensure they can provide high-accuracy security authentication in diverse scenarios. Finally, use the Adaptive Weighted Multi-Teacher Knowledge Distillation Model (AMTDF) to distill the large teacher models into lightweight student models, allowing the student models to better simulate the output distribution of the teacher models and promoting the effective transfer of complex knowledge.

[0011] The positive effects of this invention are:

[0012] This scheme uses knowledge distillation to distill multiple large teacher models into a lightweight student model, effectively reducing model complexity while maintaining authentication accuracy. Specifically, this invention first proposes SensSpec-GAN, a generative adversarial network data augmentation model based on Sensitivity-Specificity Loss. This model enhances data diversity and the model's ability to identify negative samples by introducing sensitivity and specificity constraints into the generative adversarial network. Subsequently, addressing the computational burden caused by the high dimensionality of CSI data, this invention proposes SABES, a sparse autoencoder model based on the Vulture Optimization algorithm, to reduce the dimensionality of CSI data. SABES automatically adjusts the model's hyperparameters using the Vulture Optimization algorithm, enhancing the model's feature representation ability. Finally, this invention proposes AMTDF, a multi-teacher knowledge distillation model with adaptive weights. This model adjusts the contribution of different teacher models to the student model through an adaptive weight mechanism, improving the model's performance on the target task. Experimental results show that the accuracy of this scheme reaches 96.32%, and the training speed is 2.46 times faster than the benchmark algorithm. Attached Figure Description

[0013] Figure 1 This is the algorithm structure of SensSpec-GAN.

[0014] Figure 2 These are performance metrics for different algorithms at a 30% data augmentation ratio.

[0015] Figure 3 This refers to the impact of different dimensionality reduction techniques on performance metrics.

[0016] Figure 4 This is the overall architecture diagram of AMTDF.

[0017] Figure 5 It is the relationship between the number of teacher models and the accuracy of student models. Detailed Implementation

[0018] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0019] Step 1: Data Acquisition and Preprocessing. CSI data from real-world communication scenarios is acquired using the Linux 802.11n CSI Tool. The amplitude and phase of each subcarrier are then calculated based on the subcarrier complex matrix. Subsequently, a Savitzky-Golay filter is used to denoise the CSI data. Finally, the data is normalized for easier subsequent analysis and processing.

[0020] CSI data was collected using three ThinkPad X201 laptops equipped with Intel 5300 network cards. These three laptops represented Alice (sender), Bob (receiver), and Eve (attacker) in the communication system. An Ubuntu operating system (version 14.06LTS, kernel version 4.4) was used as the data acquisition platform, and the Linux 802.11n CSITool package was used as the CSI data acquisition tool. The received data was a complex matrix of dimension (T, R, N), where T and R represent the number of antennas at the transmitter and receiver, respectively, and N represents the number of subcarriers between each pair of antennas in the CSI data.

[0021] To further improve CSI data quality, this invention uses a Savitzky-Golay (SG) filter for denoising. 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 can effectively remove random noise from CSI data while preserving important features such as signal peaks, valleys, and waveform trends. The SG filtering process can be expressed by the following formula:

[0022]

[0023] In the formula, It is the filtered value, y i+j It is the value of the (i+j)th point in the original data sequence, c j is the pre-computed convolution coefficient, and m is the window radius. By adjusting the window size and the polynomial order, the smoothness of the filtering and the accuracy of the fitting can be controlled.

[0024] Min-Max Normalization is a commonly used data preprocessing technique that scales the data along each feature dimension to limit all feature values ​​to a fixed range. For each input data T, the data in the l-th dimension T... l The process of scaling it to the range [0,1] using max-min normalization can be represented as:

[0025]

[0026] In the formula T l Represents the original data value, min(T) l ) and max(T l () represents the minimum and maximum values ​​in the l-th dimension of data. This represents the result after normalization.

[0027] Step 2: Perform data augmentation. Use a generative adversarial network based on Sensitivity-Specificity Loss to generate high-quality experimental data, improving the model's ability to identify negative samples.

[0028] SensSpec-GAN uses a fully connected network with two hidden layers as the generator, and to fully utilize the sequence features in CSI data, it uses a deep neural network based on a combination of Sensitivity-Specificity Loss and cross-entropy loss as the discriminator. Its algorithm structure is as follows: Figure 1 As shown.

[0029] The definition of Sensitivity-Specificity Loss is as follows:

[0030] L ss =w·sensitivity+(1-w)·specificity

[0031] In the formula

[0032]

[0033]

[0034] In SensSpec-GAN, the total loss function L of the discriminator D It can be defined by combining Sensitivity-Specificity Loss and Cross-Entropy Loss, and can be expressed in the following form:

[0035] L D =α·L SSD +β·L CED

[0036] In the formula, α and β are weighting coefficients used to balance the contributions of Sensitivity-Specificity Loss and Cross-Entropy Loss to the total discriminator loss. SSD and L CED These represent the specific values ​​of Sensitivity-Specificity Loss and Cross-Entropy Loss, respectively.

[0037] SensSpec-GAN uses a generator for final data augmentation, but in the discriminator D, SensSpec-GAN introduces a Sensitivity-Specificity loss function, which is combined with the cross-entropy loss function to generate samples, so that the discriminator achieves a balance between sensitivity and specificity. In practical use, the model's emphasis on sensitivity and specificity can be adjusted through parameters.

[0038] Performance metrics of different algorithms at a 30% data augmentation ratio, such as Figure 2 As shown.

[0039] Step 3: Perform data dimensionality reduction. Reconstruct the input data using a sparse autoencoder based on the Bald Eagle optimization algorithm. The Bald Eagle optimization algorithm can automatically find the hyperparameters of the sparse autoencoder, improving its performance.

[0040] Compressing CSI data using a sparse autoencoder before inputting it into a neural network can effectively improve the efficiency of physical layer authentication systems and significantly reduce latency during the authentication process. Furthermore, sparse autoencoders can learn the inherent patterns in the data, thereby efficiently extracting key features from the CSI data and enabling the model to adapt to different wireless propagation environments and devices. However, sparse autoencoders have a series of hyperparameters, such as sparsity penalty terms and L2 regularization terms. Improper hyperparameter settings can lead to poor model performance, requiring extensive experimentation for hyperparameter optimization and adjustment, which limits its widespread practical application.

[0041] This invention proposes a sparse autoencoder based on the Bald Eagle Search Algorithm (SABES). The Bald Eagle Search Algorithm is an efficient global search algorithm that mimics the hunting strategy of a bald eagle, using iterative search to obtain the globally optimal solution for the model's hyperparameters. Before training the sparse autoencoder, this method automatically adjusts the sparsity penalty factor and other key hyperparameters to optimize the sparse autoencoder's performance. Applying the Bald Eagle Search Algorithm to the hyperparameter optimization of the sparse autoencoder overcomes the difficulties in hyperparameter optimization, thereby achieving more efficient and accurate physical layer authentication in resource-constrained environments.

[0042] The specific steps for searching model hyperparameters in SABES are as follows:

[0043] (1) Initialization

[0044] Generate an initial bald eagle colony, with each bald eagle representing a specific set of decision variable values ​​(l i ,s i ), where l i s represents the L2 regularization parameter. i This represents the sparse regularization parameter. For each parameter P... i ∈(l i ,s i These variables can be randomly initialized to form an initial population in the following manner:

[0045] Pi =P min +rand()*(P max -P min )

[0046] (2) Selection phase

[0047] During the selection phase, the vulture will fly to the vicinity of the current best individual, and then randomly select a search area, choosing the area with the most prey to obtain the best position. The following formula is used to select the search area:

[0048]

[0049] In the formula, α is a constant in the range [1.5, 2] used to control the range of position changes, and r is a random number in the range [0, 1]. This represents the average position of the group.

[0050] (3) Search phase

[0051] During the search phase, the vulture will update its position around its current search space using an Archimedean spiral, and accelerate the search by moving in different directions, as specifically implemented as follows:

[0052] P i,new =P i +y(i)×(P i -P i+1 )+x(i)×(P i -P mean )

[0053]

[0054] xr(i)=r(i)×sin(θ(i)), yr(i)=r(i)×cos(θ(i))

[0055] θ(i)=α′×π×S, r(i)=θ(i)+R×S

[0056] In the formula, α' and R are constant parameters that control the flight trajectory, where α'∈[5,10] and R∈[0.5,2]. x(i) and y(i) represent the current position of the bald eagle in the coordinate system, S is a random number in the range of [0,1], and θ(i) and r(i) represent the polar angle and polar radius of the spiral equation.

[0057] (4) Dive phase

[0058] Once a vulture has locked onto prey in its search area during the search phase, it will spiral towards the prey from its current position, updating its position using the following methods:

[0059] Pi,new =S×P best +x1(i)×(P i -c1×P mean )+y1(i)×(P i -c2×P best )

[0060]

[0061] xr(i)=r(i)×sinh(θ(i)), yr(i)=r(i)×cosh(θ(i))

[0062] θ(i)=α×π×S, r(i)=θ(i)

[0063] In the formula, x(i) and y(i) represent the current position of the vulture in the coordinate system, c1 represents the flight speed to the optimal position, which ranges from c1∈(-1,1), and c2 represents the flight speed to the average position, which ranges from c1∈(1,2). These two parameters can increase the randomness of the vulture's movement intensity.

[0064] The impact of different dimensionality reduction techniques on performance metrics, such as Figure 3 As shown.

[0065] Step 4: Train the model using a knowledge distillation strategy. First, train multiple teacher models using a large amount of labeled data to perform complex security authentication tasks. Then, to verify the effectiveness of the teacher models in real-world applications, evaluate the performance of each teacher model on independent test sets and perform model selection to ensure they can provide high-accuracy security authentication in diverse scenarios. Finally, use the Adaptive Weighted Multi-Teacher Knowledge Distillation Model (AMTDF) to distill the large teacher models into lightweight student models, allowing the student models to better simulate the output distribution of the teacher models and promoting the effective transfer of complex knowledge.

[0066] This scheme consists of two parts: teacher model selection and adaptive weight allocation. In the model selection part, the F-value of the teacher model on the prediction set is first calculated. 0.5 The score is used as an evaluation metric for the degree of positive knowledge the teacher model can impart to students. Then, the F-score is calculated for all teacher models. 0.5 The score is processed to obtain the model selection threshold. The selection process is as follows:

[0067] First, calculate the model F for all teachers using the formula. 0.5 The mean and standard deviation of the score, where N represents the total number of teacher models, F 0.5,i F represents the i-th teacher model. 0.5 Score:

[0068]

[0069]

[0070] To improve the generalization ability of student models, this scheme uses the mean minus 0.5 times the standard deviation as the screening threshold for teacher models, ensuring that while retaining as many teacher models as possible, F-values ​​are removed. 0.5 A teacher model with an excessively low score. The threshold formula is as follows:

[0071]

[0072] This scheme sets the initial number of teacher models to 3. After all teacher models have been trained, the F-value for each model on the test set is calculated. 0.5 The model is evaluated using a score metric, and models with scores below a threshold are removed. The remaining models then participate in the subsequent training and computation.

[0073] To further enhance the performance of the student model, this invention uses an exponentially weighted allocation method to calculate the weights of the teacher models, enabling the student model to learn richer and more comprehensive information from multiple teacher models. The weight w of each teacher model... i The calculation formula is as follows:

[0074]

[0075] Considering the complexity of CSI data and the resource constraints and efficiency requirements in practical applications, this invention uses three deep neural networks with residual connections as the teacher model, and adaptively adjusts the weights using the algorithm of this invention. A simple, shallow, fully connected network is used as the student model to ensure efficient physical layer security authentication in resource-constrained environments.

[0076] The overall architecture diagram of the solution is as follows: Figure 4 As shown in the figure. The relationship between the number of teacher models and the accuracy of student models is as follows. Figure 5 As shown.

Claims

1. A physical layer authentication method based on sparse autoencoders and knowledge distillation, characterized in that, The method includes the following steps: Step 1: Data Acquisition and Preprocessing. Use Linux 802.11n CSITool to acquire CSI data in real communication scenarios. Then, calculate the amplitude and phase of each subcarrier based on the subcarrier complex matrix. Subsequently, use Savitzky-Golay filter to denoise the CSI data. Finally, normalize the data for easier subsequent analysis and processing. Step 2: Perform data augmentation by using a generative adversarial network based on Sensitivity-Specificity Loss to generate high-quality experimental data and improve the model's ability to identify negative samples. Step 3: Perform data dimensionality reduction and reconstruct the input data using a sparse autoencoder based on the vulture optimization algorithm. The vulture optimization algorithm can automatically find the hyperparameters of the sparse autoencoder and improve its performance. Step 4: Train the model using a knowledge distillation strategy. First, train multiple teacher models using a large amount of labeled data; then calculate the F-square of each teacher model on the independent test set. 0.5 Score metric, and based on the F-value of all teacher models 0.5 The mean and standard deviation of the score are used to set thresholds, and teacher models with scores exceeding the thresholds are selected. Using the selected teacher models, a multi-teacher knowledge distillation method based on exponential weighting is employed to distill large teacher models into lightweight student models. The weights in the exponential weighting method are based on the F-value of each teacher model. 0.5 Score calculation.

2. The physical layer authentication method based on sparse autoencoders and knowledge distillation as described in claim 1, characterized in that, The process of generating high-quality experimental data using a generative adversarial network based on Sensitivity-Specificity Loss in step 2 is characterized by using a fully connected network with two hidden layers as the generator, and using a deep neural network based on a combination of Sensitivity-Specificity Loss and cross-entropy loss as the discriminator in order to make full use of the sequence features in the CSI data. The definition of Sensitivity-Specificity Loss is as follows: L ss =w·sensitivity+(1-w)·specificity In the formula In SensSpec-GAN, the total loss function L of the discriminator D Defined by combining Sensitivity-Specificity Loss and Cross-Entropy Loss, it is expressed in the following form: L D =α·L SSD +β·L CED In the formula, α and β are weighting coefficients used to balance the contributions of Sensitivity-Specificity Loss and Cross-Entropy Loss to the total discriminator loss, L SSD and L CED These represent the specific values ​​of Sensitivity-Specificity Loss and Cross-Entropy Loss, respectively.

3. The physical layer authentication method based on sparse autoencoders and knowledge distillation as described in claim 1, characterized in that, The process of using a sparse autoencoder based on the vulture optimization algorithm for data dimensionality reduction in step 3 is characterized in that, before training the sparse autoencoder, the vulture optimization algorithm is used to automatically adjust the sparsity penalty factor and other key hyperparameters to optimize the performance of the sparse autoencoder.

4. The physical layer authentication method based on sparse autoencoders and knowledge distillation as described in claim 1, characterized in that, Step 4 describes the process of training the model using a knowledge distillation strategy, characterized by calculating the F-value of the teacher model on the prediction set. 0.5 The score is used as an evaluation metric for the degree of positive knowledge that the teacher model can impart to students. Then, the F-score is calculated for all teacher models. 0.5 The score is processed to obtain the model selection threshold. The selection process is as follows: First, calculate the model F for all teachers using the formula. 0.5 The mean and standard deviation of the score, where N represents the total number of teacher models, F 0.5,i F represents the i-th teacher model. 0.5 Score: Calculate the F-value for each model on the test set. 0.5 The model is evaluated using a score metric, and models with scores below a threshold are removed. The remaining models participate in the subsequent training and computation. The weights of the teacher models are calculated using an exponential weighted allocation method.

Images