Physical layer encryption security analysis method based on neural network
By building a security analysis model based on neural network, the problems of high complexity and low robustness of physical layer encryption security analysis are solved, and efficient and general security evaluation of physical layer encryption is achieved.
Patent Information
- Application Number
- CN202510769924.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-10
- Publication Date
- 2025-07-11
AI Technical Summary
The prior art lacks effective general methods to evaluate the security of physical layer encryption schemes, resulting in high complexity and low robustness of security analysis, making it difficult to measure its attack resistance through unified algorithm standards.
A security analysis model based on neural network is built, including the first convolutional layer, the second convolutional layer, the third convolutional layer and the fully connected layer. Through training and verification, the loss function and parameter update function are used to optimize the network structure to realize the security analysis of the encrypted ciphertext of the physical layer.
It effectively reduces the computational complexity and provides a general security analysis method that can target different encryption methods and attack methods, improving the accuracy and efficiency of security assessment.
Smart Images
Figure CN120301706A_ABST
Abstract
Description
Technical Field
[0001] This specification relates to the field of secure communication technologies, and particularly to a method for analyzing the security of physical layer encryption based on neural networks. Background Art
[0002] With the emergence of new concepts such as the Internet of Things, big data, and the metaverse entering people's lives, Internet services have expanded rapidly, and the data transmission speed and total interaction volume have also increased explosively. Against this background, data security issues have attracted more and more attention. Researchers around the world have conducted research on current security solutions from multiple perspectives and proposed various security solutions suitable for the present. However, considering multiple aspects such as usage cost, computational complexity, and hardware load, the physical layer is at the bottom of the entire network in the OSI seven-layer network model and is the foundation of the entire network system. Encryption at the physical layer is the most potential and practical method. Currently, most technical solutions for enhancing physical layer security are implemented based on the CO-OFDM system, and its core is to perturb data in the electrical domain and digital domain to achieve the purpose of encryption. Usually, three indicators are used to measure the quality of an algorithm, namely the security of the secure system, the speed of implementing encryption, and the cost of implementing encryption, among which the security of the secure system is the most important point. Compared with cryptography that studies how to design physical layer encryption algorithms, the security evaluation of physical layer encryption schemes often can only be judged by the size of the key space, lacking an effective and general security analysis method, which restricts the further development and application of physical layer encryption. Conducting security analysis and iteratively optimizing the attack performance for physical layer encryption schemes from the perspective of cryptanalysis is equally important for ensuring information security.
[0003] The most effective security analysis solution for physical layer chaotic encryption schemes is to study their computational security, that is, to analyze the ability of physical layer encryption schemes to resist various attacks. The attack methods include: chosen-plaintext attack, chosen-ciphertext attack, ciphertext-only attack, known-plaintext attack. Each type of attack method contains a large number of attack algorithms respectively. When a researcher proposes a new physical layer security scheme, it is difficult to measure the security of the scheme through a unified algorithm standard. Summary of the Invention
[0004] Aiming at the above deficiencies in the prior art, a method for analyzing the security of physical layer encryption based on neural networks provided by the present invention solves the problems of high complexity and low robustness of physical layer encryption security.
[0005] In order to achieve the above invention purpose, the technical solution adopted by the present invention is: a method for analyzing the security of physical layer encryption based on neural networks, including: S1: Using a first convolutional layer, a second convolutional layer, a third convolutional layer, and a fully connected layer to construct a security analysis neural network model; S2: Analyze the ciphertext image samples based on the security analysis neural network model, train the security analysis neural network model using the loss function and parameter update function, and verify the training results to obtain a trained security analysis neural network model; S3: Analyze the physical layer encrypted ciphertext using the trained security analysis neural network model to obtain a security analysis result, and complete the security analysis of the physical layer encryption.
[0006] The beneficial effects of the present invention are: A method for security analysis of physical layer encryption based on neural network. In this way, based on the specific characteristics of physical layer chaotic encryption, the network structure and various parameters are designed specifically, solving the problems of poor effect and overfitting, effectively evaluating the security while greatly reducing the computational complexity. At the same time, it is no longer limited to a certain type of encryption method and attack method, providing a general and effective method for the security proof and vulnerability analysis of modern optical physical layer chaotic encryption from the perspective of chosen-plaintext attack.
[0007] Further, the security analysis neural network model includes: The first convolutional layer, which is used to perform the first convolutional calculation on the ciphertext image samples to obtain the first convolutional result; The second convolutional layer, which is used to perform the second convolutional calculation on the first convolutional result to obtain the second convolutional result; The third convolutional layer, which is used to perform the third convolutional calculation on the second convolutional result, remap it back to one-dimensional space, and obtain the third convolutional result; The fully connected layer, which is used to analyze the third convolutional result to obtain the security analysis result output by the security analysis neural network model.
[0008] Only use max pooling in the first convolutional module, cancel pooling in subsequent modules and increase the number of convolutional kernels. By retaining the spatial dimension, the feature resolution is improved, avoiding the loss of plaintext details caused by pooling, and the plaintext can be restored with high fidelity and more spatial information can be retained. By fixedly using 3*3 convolutional kernels, the odd-size retains the central anchor point, the stride is 1 and there is no padding, ensuring the complete extraction of ciphertext features. At the same time, by directly recovering the plaintext from the ciphertext through end-to-end training, replacing the key retrieval in traditional cryptanalysis, and using the chaotic characteristics of the neural network to fit the equivalent key of the encryption process, a security analysis scheme that does not rely on the mathematical characteristics of the key space but learns the mapping relationship through data pairs is realized, providing an unconventional methodology for the security analysis of physical layer encryption.
[0009] Further, the first convolutional layer includes: The convolutional module, which is used to extract features from the ciphertext image samples to obtain ciphertext image features; A normalization module for normalizing the ciphertext image features to obtain normalization information; A first activation module for performing hidden layer activation on the normalization information using an activation function to obtain an activation result; A pooling module for performing max pooling on the activation result to obtain a pooling result; A second activation module for activating the pooling result to obtain a first convolution result.
[0010] Using an unconventional combination of activation functions to improve the hierarchical application of Sigmoid and adaptive parameters. The hierarchical use of activation functions enhances the efficiency of non-linear fitting and gradient propagation. Inside the convolution module, Sigmoid is used as the activation function to fit the complex relationship between ciphertext and plaintext through its smooth non-linear mapping. For the activation function between layers, the α parameter is introduced to replace the function operation with a fixed slope, thereby optimizing the gradient propagation through the adaptive parameter α, avoiding gradient vanishing and explosion, enhancing the fitting ability for chaotic encryption noise features, adapting to the non-convex optimization problem of ciphertext analysis, and being more flexible than traditional fixed activation functions. Optimize the conventional linear connection or single activation function between convolution modules to enhance the non-linearity of hierarchical connections, reduce the computational complexity while retaining ciphertext features through non-linear transformation, and perform customized design for the characteristics of chaotic encryption noise.
[0011] Further, the expressions for the mean and variance of the normalization information are: ; ; where, represents the mean of the normalization information, represents the number of ciphertext image samples, represents the th ciphertext image sample, represents the standard deviation of the normalization information; The expression for the activation function is: ; ; where, represents the activation function, represents the input data, represents the derivative of the activation function; The expression for the first convolution result is: ; where, represents the first convolution result, represents the learnable parameter.
[0012] Furthermore, the expressions of the loss function and the parameter update function are as follows: ; ; ; ; where, represents the loss function, represents the number of prediction results, represents the label value, represents the prediction result, represents the parameter state at the beginning of the (t + 1)-th step, represents the parameter state at the beginning of the t-th step, represents the learning rate, represents the weighted average of the squared gradients at the t-th step, represents a small constant value, represents the hyperparameter controlling the momentum, represents the momentum estimate with bias correction, represents the hyperparameter controlling the momentum at the t-th step, represents the current gradient, represents the momentum of the gradient at the t-th step, represents the momentum of the gradient at the (t - 1)-th step, represents the parameter controlling the decay rate of the second moment, represents the weighted average of the squared gradients at the (t - 1)-th step.
[0013] The innovative combination of the optimizer adopts an unconventional optimization strategy combined with an adaptive learning rate. Based on the prospective gradient estimation of the gradient mean and variance and Nesterov momentum, an optimization method suitable for the security analysis task is formed. A correction term is introduced in the gradient update, and the learning rate is dynamically adjusted through historical gradient information to accelerate convergence and reduce oscillations, making the network applicable to the high-dimensional non-convex optimization problem of ciphertext-plaintext mapping. At the same time, targeted settings of parameter configuration are carried out, taking into account both momentum stability and learning rate adaptability, which is different from the conventional default parameters and specifically optimizes the complex mapping of the chaotic encryption feature space.
[0014] Furthermore, the verification process of the trained security analysis neural network model includes: Input the ciphertext image sample into the trained security analysis neural network model, compare the security analysis result output by the trained security analysis neural network model with the original plaintext information to obtain a similarity result, and complete the verification of the trained security analysis neural network model; where, the expression of the similarity result is: ; ; Among them, represents the similarity result, represents the security analysis result output by the trained security analysis neural network model, represents the original plaintext information, represents the mean value of represents the mean value of and represents a constant, represents and the covariance of represents the standard deviation of represents the standard deviation of and represent small positive numbers, represents a fixed parameter with a value of 255. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] This specification will be further described in the form of exemplary embodiments, and these exemplary embodiments will be described in detail through the drawings. These embodiments are not restrictive. In these embodiments, the same numbers represent the same structures, where: Figure 1 is an exemplary flowchart of a physical layer encryption security analysis method based on a neural network according to some embodiments of this specification. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0016] The following describes the specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.
[0017] Embodiment Figure 1 is an exemplary flowchart of a physical layer encryption security analysis method based on a neural network according to some embodiments of this specification. As Figure 1 shown, the process includes the following steps. In some embodiments, the process can be executed by a processor.
[0018] S1: Use the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fully connected layer to construct a security analysis neural network model.
[0019] The security analysis neural network model is used to perform security analysis on the encrypted data at the physical layer to obtain the security analysis result. The type of the security analysis neural network model can be various. For example, the type of the security analysis neural network model can include a convolutional neural network model.
[0020] In some embodiments, the input of the security analysis neural network model can be a ciphertext image sample, and the output of the security analysis neural network model can be the security analysis result.
[0021] In some embodiments, the structure of the security analysis neural network model is as follows: The security analysis neural network model includes a first convolutional layer, a second convolutional layer, a third convolutional layer, and a fully connected layer. The output of the first convolutional layer serves as the input of the second convolutional layer, the output of the second convolutional layer serves as the input of the third convolutional layer, the output of the third convolutional layer serves as the input of the fully connected layer, and the output of the fully connected layer serves as the final output of the security analysis neural network model.
[0022] The first convolutional layer is used to perform the first convolution calculation on the ciphertext image sample to obtain the first convolution result. The input of the first convolutional layer can include the ciphertext image sample, and the output can include the first convolution result.
[0023] The ciphertext image sample is a physical layer encrypted ciphertext sample. For example, the ciphertext image sample can include a physical layer encrypted ciphertext sample image with a size of 32*32.
[0024] In some embodiments, the first convolutional layer can include a convolution module, a normalization module, a first activation module, a pooling module, and a second activation module. Among them, the convolution module is used to extract features from the ciphertext image sample to obtain ciphertext image features; the normalization module is used to perform normalization processing on the ciphertext image features to obtain normalization information; the first activation module is used to use an activation function to perform hidden layer activation on the normalization information to obtain an activation result; the pooling module is used to perform max pooling processing on the activation result to obtain a pooling result; the second activation module is used to activate the pooling result to obtain the first convolution result.
[0025] The ciphertext image features are the feature information of the ciphertext image sample.
[0026] In some embodiments, the processor can use 24 convolution modules with a size of 3*3 to perform better feature extraction in the ciphertext domain for the image, output 24*30*30 pieces of information, preprocess the image size to 32*32, and finally the vertical and horizontal sliding strides of the convolution kernel are both 1 to obtain the ciphertext image features.
[0027] The normalized information is the feature of the encrypted text image after normalization. For example, the normalized information may include the corresponding mean and variance.
[0028] In some embodiments, the expressions for the mean and variance of the normalized information can be: ; ; where, represents the mean of the normalized information, represents the number of encrypted text image samples, represents the th encrypted text image sample, represents the standard deviation of the normalized information.
[0029] The activation result is the result reflecting the complex relationship between the encrypted text and the plaintext. For example, the activation result may include data of 24*30*30.
[0030] In some embodiments, the expression of the activation function can be: ; ; where, represents the activation function, represents the input data, represents the derivative of the activation function The pooling result is the activation result that reduces the matrix dimension and the number of parameters.
[0031] In some embodiments, the processor can perform max pooling on the activation result, reducing the dimension of the feature data within the spatial range to reflect more extensive feature information of the encrypted text domain, while reducing the input size of the next layer, effectively reducing the matrix dimension and the number of parameters, and finally obtaining the feature data of 24*15*15 dimensions as the pooling result.
[0032] In some embodiments, the expression of the first convolution result can be: ; where, represents the first convolution result, represents the learnable parameter.
[0033] The second convolutional layer is used to perform a second convolution calculation on the first convolution result to obtain the second convolution result. The input of the second convolutional layer may include the first convolution result, and the output may include the second convolution result.
[0034] The third convolutional layer is used to perform a third convolution calculation on the result of the second convolution, remap it back to a one-dimensional space, and obtain the result of the third convolution. The input of the third convolutional layer can include the result of the second convolution, and the output can include the result of the third convolution.
[0035] The fully connected layer is used to analyze the result of the third convolution to obtain the security analysis result output by the security analysis neural network model. The input of the fully connected layer can include the result of the third convolution, and the output can include the security analysis result.
[0036] The security analysis result is an equivalent key, which is used to reflect the security situation of the ciphertext image sample.
[0037] In some embodiments, the processor can utilize the security analysis result to recover the plaintext information without obtaining the decryption key by means of the constructed equivalent key, reflecting the security situation of the ciphertext image sample.
[0038] In some embodiments, the convolutional kernel sizes of the first convolutional layer, the second convolutional layer, and the third convolutional layer are selected as 3*3, and 32 convolutional kernels and 64 convolutional kernels are respectively used for convolution operations. Pooling calculations are not performed on the second convolutional layer and the third convolutional layer; the information dimension finally output by the first convolutional module is 24*15*15, and after passing through the second convolutional layer and the third convolutional layer in sequence, data of 64*11*11 is obtained. It is remapped back to a one-dimensional space by using the Flatten function, and the data size becomes 7744. It is passed through a fully connected layer with 1024 dimensions, and finally 1024 data are obtained as the security analysis result.
[0039] S2: Analyze the ciphertext image sample based on the security analysis neural network model, train the security analysis neural network model by using a loss function and a parameter update function, and verify the training result to obtain a trained security analysis neural network model.
[0040] In some embodiments, the security analysis neural network model can be trained by multiple labeled training samples. For example, multiple labeled training samples can be input into the initial security analysis neural network model, a loss function can be constructed based on the labels and the results of the initial security analysis neural network model, and the parameters of the initial security analysis neural network model can be iteratively updated based on the loss function by gradient descent or other methods. When the preset conditions are met, the model training is completed, and a trained security analysis neural network model is obtained. Among them, the preset conditions can be that the loss function converges, the number of iterations reaches a threshold, etc.
[0041] In some embodiments, the training samples can at least include ciphertext image samples. The label can be the plaintext information corresponding to the ciphertext image sample. The label can be determined from historical data.
[0042] In some embodiments, the expressions of the loss function and the parameter update function can be: ; ; ; ; where, represents the loss function, represents the number of prediction results, represents the label value, represents the prediction result, represents the parameter state at the start of step t + 1, represents the parameter state at the start of step t, represents the learning rate, represents the weighted average of the squared gradients at step t, represents a small constant, represents the hyperparameter controlling the momentum, represents the momentum estimate with bias correction, represents the hyperparameter controlling the momentum at step t, represents the current gradient, represents the momentum of the gradient at step t, represents the momentum of the gradient at step t - 1, represents the parameter controlling the decay rate of the second moment, represents the weighted average of the squared gradients at step t - 1.
[0043] In some embodiments, the verification process of the trained security analysis neural network model includes: inputting the ciphertext image sample into the trained security analysis neural network model, comparing the security analysis result output by the trained security analysis neural network model with the original plaintext information to obtain a similarity result, and completing the verification of the trained security analysis neural network model.
[0044] In some embodiments, the expression of the similarity result can be: ; ; where, represents the similarity result, represents the security analysis result output by the trained security analysis neural network model, represents the original plaintext information, represents the mean of, represents The mean value of and represents a constant, represents and the covariance of represents the standard deviation of represents the standard deviation of and represents a small positive number, represents a fixed parameter with a value of 255.
[0045] S3: Use the trained security analysis neural network model to analyze the physical layer encrypted ciphertext, obtain the security analysis result, and complete the security analysis of the physical layer encryption.
[0046] The physical layer encrypted ciphertext is the result of encrypting the binary data at the sending end before serial-to-parallel conversion and high-order quadrature amplitude modulation mapping.
[0047] In some embodiments, the processor can perturb the bit data through chaotic mapping combined with coding rules, perform serial-to-parallel conversion and 16QAM symbol mapping on the perturbed scrambled data, and finally obtain an OFDM signal through inverse fast Fourier transform (IFFT), convert it into an optical signal through electro-optic modulation, and transmit it to the receiving end through an optical fiber. The receiving end performs coherent detection on the optical signal and reconverts it into an electrical signal to obtain the physical layer encrypted ciphertext.
[0048] In some embodiments of this specification, a method for security analysis of physical layer encryption based on a neural network is proposed. In this way, based on the specific characteristics of physical layer chaotic encryption, the network structure and various parameters are designed specifically, solving the problems of poor effect and overfitting, effectively evaluating the security while greatly reducing the computational complexity, and at the same time not being limited to a certain type of encryption method and attack method. From the perspective of chosen-plaintext attack, it provides a general and effective method for the security proof and vulnerability analysis of modern optical physical layer chaotic encryption.
Claims
1. A method for analyzing the security of physical layer encryption based on neural network, characterized in that, Including: S1: Construct a security analysis neural network model using the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fully connected layer; S2: Analyze the ciphertext image samples based on the security analysis neural network model, train the security analysis neural network model using the loss function and the parameter update function, and verify the training results to obtain a trained security analysis neural network model; S3: Analyze the physical layer encrypted ciphertext using the trained security analysis neural network model to obtain a security analysis result, and complete the security analysis of the physical layer encryption.
2. The method for analyzing the physical layer encryption security based on a neural network according to claim 1, wherein The security analysis neural network model includes: The first convolutional layer, which is used to perform the first convolutional calculation on the ciphertext image samples to obtain the first convolutional result; The second convolutional layer, which is used to perform the second convolutional calculation on the first convolutional result to obtain the second convolutional result; The third convolutional layer, which is used to perform the third convolutional calculation on the second convolutional result and remap it back to one-dimensional space to obtain the third convolutional result; The fully connected layer, which is used to analyze the third convolutional result to obtain the security analysis result output by the security analysis neural network model.
3. The method for analyzing the physical layer encryption security based on a neural network according to claim 2, wherein The first convolutional layer includes: The convolutional module, which is used to extract features from the ciphertext image samples to obtain ciphertext image features; The normalization module, which is used to perform normalization processing on the ciphertext image features to obtain normalized information; The first activation module, which is used to perform hidden layer activation on the normalized information using an activation function to obtain an activation result; The pooling module, which is used to perform max pooling processing on the activation result to obtain a pooling result; The second activation module, which is used to activate the pooling result to obtain the first convolutional result.
4. The method for analyzing the physical layer encryption security based on a neural network according to claim 3, wherein The expressions for the mean and variance of the normalized information are: ； ； Among them, represents the mean of the standardized information, represents the number of ciphertext image samples, represents the th ciphertext image sample, represents the standard deviation of the standardized information; The expression for the activation function is: ； ； Among them, represents the activation function, represents the input data, represents the derivative of the activation function; The expression for the first convolutional result is: ； Among them, represents the result of the first convolution, represents the learnable parameters.
5. The physical layer encryption security analysis method based on neural network according to claim 1, wherein The expressions for the loss function and the parameter update function are: ； ； ； ； Among them, represents the loss function, represents the number of prediction results, represents the label value, represents the prediction result, represents the parameter state at the start of the (t + 1)-th step, represents the parameter state at the start of the t-th step, represents the learning rate, represents the weighted average of the squared gradients at the t-th step, represents a small value constant, represents a hyperparameter that controls momentum, represents the momentum estimate with bias correction, represents a hyperparameter that controls the momentum at the t-th step, represents the current gradient, represents the momentum of the gradient at the t-th step, represents the momentum of the gradient at the (t - 1)-th step, represents a parameter that controls the decay rate of the second moment, represents the weighted average of the squared gradients at the (t - 1)-th step.
6. The physical layer encryption security analysis method based on neural network according to claim 1, characterized in that, The verification process of the trained security analysis neural network model includes: Input the ciphertext image samples into the trained security analysis neural network model, compare the security analysis result output by the trained security analysis neural network model with the original plaintext information to obtain a similarity result, and complete the verification of the trained security analysis neural network model; where the expression for the similarity result is: ； ； Among them, represents the similarity result, represents the security analysis result output by the trained security analysis neural network model, represents the original plaintext information, represents the mean value of, represents the mean value of, and represents a constant, represents and the covariance of, represents the standard deviation of, represents the standard deviation of, and represents a small positive number, represents a fixed parameter with a value of 255.