GMM clustering and neural network mixed model structure inversion method
Through the GMM clustering and neural network hybrid method, combined with multimodal data and dynamic threshold adjustment, the problems of high computational cost and poor accuracy in reverse engineering of neural network models are solved, and efficient and accurate model reverse analysis is achieved.
Patent Information
- Application Number
- CN202410941178.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-15
- Publication Date
- 2024-11-08
AI Technical Summary
In the prior art, the neural network model reverse engineering method has high computational cost, complexity of sample feature extraction and signal processing, and accuracy problems, especially in poor performance on complex networks.
Using the model structure inverse method of mixing GMM clustering and neural network, the power data and multimodal physical signal data of neural network models under different input and weight conditions were collected, and statistical analysis and principal component analysis were performed, combined with the Expectation-Maximization algorithm Cluster, dynamically adjust the threshold and use neural networks to make final judgments, reducing computational costs and improving accuracy.
A more accurate neural network model inversion is achieved, which significantly reduces computational costs, improves classification accuracy and robustness, and ensures efficient and accurate classification performance in different environments.
Smart Images

Figure CN118916756A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of neural network model reverse engineering, and in particular to a model structure reverse method based on side channel attack GMM clustering and neural network mixing. Background Art
[0002] Neural network model reverse engineering refers to a series of methods and techniques that attempt to understand the internal working principles of a neural network model and infer the logic and decision-making process behind the model. This reverse engineering aims to explore how a neural network converts input data into output predictions and attempts to obtain information from the network's structure, parameters or outputs. Obtaining information about the internal model structure of the chip will help us further study the chip device.
[0003] Generally speaking, there are currently two methods for inverting the structural information of neural network models, one of which is based on meta-learning, and the other is based on hardware side channels;
[0004] The core idea of the meta-learning-based method is to use the training outputs of the meta-learning model and multiple white-box models to obtain a meta-model through training, which can predict the structural properties of the target black-box model. The method includes: in the meta-training stage, white-box models of various structural types are collected as meta-training sets, and the training data of these models are used to train them to make them close to the target model. In the meta-model construction stage, a "meta-model" is obtained based on the training on the meta-training set, and the relationship between the output of the white-box model and the structural properties is established. In the reasoning stage, the meta-model is used to predict the structural properties of the target black-box model given its output;
[0005] This method requires the construction of a large number of white-box models and their output data, so the computational cost is high. Although it works well in reverse deduction of simple neural network structures, its performance on complex networks still needs further verification.
[0006] The other is a method based on hardware side channels. This method recovers the model architecture details by analyzing the trace information left by the underlying hardware (such as cache side channels, bus monitoring, etc.). From a hardware perspective, when the model performs inference tasks, architecture-dependent traces will be left on the hardware. These trace information are analyzed to recover the model architecture details. Hardware side channel methods include inference time analysis, resource usage monitoring (CPU usage, memory occupancy, etc.), power consumption, electromagnetic radiation, and current and voltage monitoring. These methods also have problems with sample feature extraction, signal processing analysis and complexity, as well as accuracy.
[0007] Based on this, a model structure inversion method that mixes GMM clustering and neural network is needed. Summary of the invention
[0008] To achieve the above object, the present invention provides the following scheme: a GMM clustering and neural network hybrid model structure inversion method, comprising the following steps:
[0009] S1. Collect power data and other multimodal physical signal data of the neural network model under different input and weight conditions on the embedded device;
[0010] S2. Perform statistical analysis on the collected data and calculate the mean, median, standard deviation, quartile, quartile difference, and coefficient of variation statistics;
[0011] S3. Using principal component analysis (PCA) technology, high-dimensional statistical features are projected into low-dimensional space to enhance the separability between features and reduce computational complexity.
[0012] S4, input the reduced feature data into the Gaussian mixture model (GMM), perform cluster analysis through the expectation-maximization (EM) algorithm, and output the probability of each model category;
[0013] S5. In the GMM judgment process, a dynamic threshold adjustment module is added to analyze the data distribution and classification results in real time, and the threshold is adjusted dynamically. When the GMM output probability of all classifications is less than the dynamic threshold, the neural network is enabled for further judgment;
[0014] S6. Input the feature data into the pre-trained neural network, use the neural network to make the final model category determination, and output the neural network classification result.
[0015] Furthermore, the data collection step includes sampling power data at a high frequency, specifically at a frequency of 400 Hz, to ensure that the collected power data has sufficient time resolution so as to more accurately capture subtle changes in the neural network model under different input and weight conditions, thereby improving the accuracy of statistical calculations and the accuracy of the inverse model.
[0016] Furthermore, the characteristic statistics calculation step further includes calculating the statistics of multimodal physical signals, and the signals include but are not limited to electromagnetic radiation, temperature changes, current and voltage monitoring. By introducing multimodal data, the ability to comprehensively capture the operating characteristics of the neural network model is improved, and the accuracy and robustness of the reverse analysis are increased.
[0017] Furthermore, the principal component analysis (PCA) technique is used in the high-dimensional feature space projection and dimensionality reduction steps to project multidimensional statistical features into a low-dimensional space to reduce computational complexity and improve the separability between features, thereby effectively performing cluster analysis and subsequent neural network classification.
[0018] Furthermore, the GMM cluster analysis step includes iteratively optimizing parameters by an Expectation-Maximization (EM) algorithm until the parameter change is less than a preset threshold or a maximum number of iterations is reached, and updating weights, means and covariance matrices in each iteration.
[0019] Furthermore, the dynamic adaptive threshold adjustment mechanism adjusts the judgment thresholds of GMM and neural network in real time by analyzing historical data and current environmental variables, thereby improving the flexibility and accuracy of the model, and ensuring efficient and accurate classification performance under different data distributions and environments through dynamic adjustment.
[0020] Furthermore, the neural network in the neural network classification step is a multilayer perceptron (MLP), which includes at least three hidden layers, each layer includes at least 128 neurons, and uses a ReLU (Rectified Linear Unit) activation function to improve the nonlinear expression ability and classification performance of the neural network.
[0021] Furthermore, the training data of the neural network includes power statistics of neural network models of various structural types, and the training is performed through a supervised learning method so that the neural network can effectively identify and classify different neural network model structures.
[0022] Furthermore, the real-time online learning and updating step includes continuously collecting new data in actual applications and updating GMM and neural network model parameters in real time to dynamically adapt to new side channel attack patterns and environmental changes, thereby ensuring the stability and efficiency of the model during long-term operation.
[0023] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0024] 1. The present invention collects a large amount of power consumption data of a single neural network model under different input and weight conditions through side channel attacks, and calculates the statistics of these data. Different neural network models generate different power statistics during operation. These characteristics can be used to more accurately distinguish different neural network models, thereby achieving more accurate model inversion. In addition, the introduction of high-dimensional statistical features helps to more clearly distinguish different models in high-dimensional space and improve classification accuracy.
[0025] 2. The present invention obtains power consumption data through side channel attacks, and performs reverse model structure through a hybrid method of GMM and neural network. There is no need to build a large number of white box models, which significantly reduces the computing cost. GMM is more efficient in dealing with simpler classification problems, while neural networks are more advantageous in dealing with complex problems. The combination of the two can achieve a balance between accuracy and efficiency.
[0026] 3. The present invention introduces more statistics for feature extraction, and each statistic is a dimension. These rich statistical features help to more comprehensively understand the distribution pattern, central tendency and degree of dispersion of the data, thereby improving the accuracy of classification. The introduction of multimodal data further improves the ability to comprehensively capture the operating characteristics of the neural network model and increases the accuracy and robustness of reverse analysis.
[0027] 4. The present invention adopts a dynamic adaptive threshold adjustment mechanism to adjust the judgment thresholds of GMM and neural network in real time by analyzing historical data and current environmental variables, ensuring efficient and accurate classification performance under different data distributions and environmental conditions. This dynamic adjustment mechanism makes the model more flexible and robust in practical applications. When the GMM output probability of all classifications is less than the set threshold, the neural network is enabled for further judgment to ensure the accuracy of the results. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0029] Figure 1 It is the overall process step diagram of the present invention;
[0030] Figure 2 The flowchart of the algorithm of the present invention as a whole;
[0031] Figure 3 It is a flow chart of the GMM clustering algorithm of the present invention;
[0032] Figure 4 This is a neural network flow chart for the present invention. DETAILED DESCRIPTION
[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0034] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0035] Example
[0036] like Figure 1-4 As shown, a GMM clustering and neural network hybrid model structure inversion method in this embodiment includes the following steps:
[0037] S1. Collect power data and other multimodal physical signal data of the neural network model under different input and weight conditions on the embedded device;
[0038] In the specific implementation process, data collection is achieved through side channel attacks. An external data acquisition card is used to measure energy consumption when the model is running. For example, it samples at a frequency of 400Hz to capture energy consumption data. For a single neural network model structure, a number of data (such as 10 pictures) are used as a set of inputs to calculate the statistics of all collected power data in the set of data (mean, median, standard deviation, quartile, quartile difference, coefficient of variation). These statistics constitute a multidimensional data. Under the same weight condition, different data groups are input to collect the statistics of each group of power data, and then the model weight is changed and data is collected again. This is repeated to collect power data statistics of a single model structure under different weights and different input conditions. Finally, for each model, a set of power data statistics under different weights and different input conditions can be obtained.
[0039] S2. Perform statistical analysis on the collected data and calculate the mean, median, standard deviation, quartile, quartile difference, and coefficient of variation statistics;
[0040] The data obtained by the side channel attack mainly includes the power consumption data of the embedded device under different input and weight conditions. The original form of the data is time series data, that is, the power value that changes over time. In order to effectively analyze and use these data, it is first necessary to perform statistical analysis on these time series data and extract meaningful statistical features. These statistical features include but are not limited to the mean, median, standard deviation, quartile, quartile difference, and coefficient of variation.
[0041] The mean is the average value of all values in the data set. By calculating the mean of the time series data, we can get the average power consumption of the neural network model when it is running on the embedded device under specific input and weight conditions, which reflects the overall power consumption level of the device under specific conditions.
[0042] The median is the middle value in a data set, that is, the value in the middle after the data is arranged in order of size. Compared with the mean, the median is not sensitive to extreme values and can better reflect the central trend of the data. In power consumption data, calculating the median can help identify typical power consumption levels and avoid statistical deviations caused by extreme values.
[0043] The standard deviation is the average difference between each value in the data set and the mean, which is used to measure the degree of data dispersion. By calculating the standard deviation of power consumption data, we can understand the fluctuation of power consumption of the neural network model under different input and weight conditions. A larger standard deviation means that the power consumption varies greatly between different time points, reflecting the existence of high and low peak power consumption during the operation of the model.
[0044] Quartiles are three values that divide a data set into four equal parts in order of size, namely the first quartile (Q1), the second quartile (i.e. the median, Q2) and the third quartile (Q3). Quartiles can provide detailed information on data distribution and help identify the central tendency and distribution range of data;
[0045] The quartile difference is the difference between the third quartile and the first quartile (i.e., Q3-Q1), which is used to measure the dispersion of the data. Compared with the standard deviation, the quartile difference is more robust to extreme values and better reflects the overall distribution of the data. In power consumption data, the quartile difference can help identify the range of variation in model operating power consumption.
[0046] The coefficient of variation is the ratio of the standard deviation to the mean, which is used to measure the relative dispersion of data. By calculating the coefficient of variation of power consumption data, we can understand the magnitude of power consumption fluctuation relative to the average power consumption. Under different input and weight conditions, the coefficient of variation can help identify situations with poor power consumption stability.
[0047] The original time series data are grouped according to different input and weight conditions, and then the above statistical features are calculated for each group of data. Through the calculation of statistical features, the original high-dimensional time series data are converted into low-dimensional statistical features. These features will serve as the basis for subsequent GMM clustering and neural network classification.
[0048] S3. Using principal component analysis (PCA) technology, high-dimensional statistical features are projected into low-dimensional space to enhance the separability between features and reduce computational complexity.
[0049] In the specific implementation process, the PCA technology can be used to project multidimensional statistical features into a low-dimensional space so as to more effectively perform cluster analysis and subsequent neural network classification. The process includes calculating the covariance matrix of the data, performing eigenvalue decomposition, and selecting principal components to reduce the dimensionality;
[0050] Calculate the covariance matrix of high-dimensional statistical feature data. The covariance matrix is used to describe the linear relationship between each feature. Each element of the covariance matrix represents the covariance value between two features. The formula for calculating the covariance matrix is as follows:
[0051]
[0052] Among them, X is the standardized data matrix, μX is the mean vector of each feature, and n is the number of samples. After calculating the covariance matrix, eigenvalue decomposition is performed. The purpose of eigenvalue decomposition is to find the eigenvalues and eigenvectors of the covariance matrix. The eigenvalues and eigenvectors are used to determine the main components of the data, that is, the principal components. The eigenvalues of the covariance matrix represent the variance of the data in the direction of the corresponding eigenvector, and the eigenvector represents the unit vector in that direction. The formula for eigenvalue decomposition is as follows:
[0053]
[0054] Where Λ is the diagonal eigenvalue matrix and P is the eigenvector matrix;
[0055] After eigenvalue decomposition, the principal components are selected according to the size of the eigenvalues. Usually, the eigenvectors with larger corresponding eigenvalues are selected as principal components. These principal components can explain the most variance in the data. The standard for selecting principal components can be determined based on the cumulative explained variance ratio. A common method is to select the number of principal components whose cumulative explained variance ratio reaches 95% or 99%.
[0056] After selecting the principal components, data dimensionality reduction is achieved by projecting the high-dimensional statistical feature data onto the directions of these principal components. The reduced data has a lower dimension but retains most of the original data information. The formula for the dimensionality reduction process is as follows:
[0057] Y=XP k
[0058] Among them, Y is the data matrix after dimension reduction, P k It is a matrix composed of the first k principal component vectors selected. The reduced-dimensional data will be used as the input for subsequent GMM clustering and neural network classification. The PCA technology is used to project high-dimensional statistical features into low-dimensional space, which not only reduces the computational complexity but also enhances the separability between features. This is because PCA can maximize the variance of the original data and concentrate it on a few principal components, making the separation of different categories of data in low-dimensional space more obvious.
[0059] Example:
[0060] Assume that 100 statistical features are collected and calculated in step S2, including mean, median, standard deviation, quartile, quartile difference, and coefficient of variation. In order to perform effective cluster analysis and classification, these 100 high-dimensional features need to be reduced to a lower dimension (e.g., 10 dimensions);
[0061] First, the 100 high-dimensional feature data are standardized to make their mean zero and standard deviation one. Then, the covariance matrix of the standardized feature data is calculated. The size of the covariance matrix is 100x100, which represents the covariance value between every two features.
[0062] Then, eigenvalue decomposition is performed to obtain 100 eigenvalues and corresponding eigenvectors. According to the size of the eigenvalues, the first 10 eigenvectors are selected as principal components. These 10 principal components can explain more than 95% of the variance in the data.
[0063] Finally, the high-dimensional feature data is projected onto the 10 principal component directions to obtain the 10-dimensional data after dimensionality reduction. The reduced-dimensional data will be used as input for further GMM clustering analysis and neural network classification.
[0064] S4, input the reduced feature data into the Gaussian mixture model (GMM), perform cluster analysis through the expectation-maximization (EM) algorithm, and output the probability of each model category;
[0065] GMM clustering is a clustering method based on probability density model. The main steps include initializing parameters (number of clusters, mean, covariance matrix and weight). Usually, random initialization or data-based method is used to determine the initial value, and then iterative optimization is performed through EM algorithm.
[0066] The first is the E step (Expectation Step), which uses the current parameter value to calculate the probability (posterior probability) that each data point belongs to each Gaussian distribution, and labels each data point according to the probability. The calculation method is as follows:
[0067]
[0068] Next is the M step (Maximization Step), which updates the mean μ, covariance matrix σ, and weight α of each Gaussian distribution according to the label of the data point. After each iteration, check whether the stopping condition is met. Common stopping conditions include parameter change less than a threshold, log-likelihood change less than a threshold, or reaching the maximum number of iterations.
[0069]
[0070] After the final convergence, the mean, covariance and weight of each Gaussian distribution are determined according to the convergence results.
[0071] S5. In the GMM judgment process, a dynamic threshold adjustment module is added to analyze the data distribution and classification results in real time, and the threshold is adjusted dynamically. When the GMM output probability of all classifications is less than the dynamic threshold, the neural network is enabled for further judgment;
[0072] The dynamic adaptive threshold adjustment mechanism adjusts the judgment thresholds of GMM and neural network in real time by analyzing historical data and current environmental variables, thereby improving the flexibility and accuracy of the model. When the GMM output probability of all classifications is less than the threshold, the neural network is enabled for further judgment. The dynamic threshold adjustment module adjusts the threshold to optimize the accuracy of the model by analyzing the previous data distribution and the current classification results.
[0073] S6, input the feature data into the pre-trained neural network, use the neural network to make the final model category determination, and output the neural network classification result;
[0074] Neural networks have advantages in dealing with complex situations. The input of the neural network training set is consistent with the input of the GMM clustering algorithm, namely the mean, median, standard deviation, quartile, quartile difference, and coefficient of variation statistics. The output is the corresponding model category. The trained neural network can effectively classify the input data, especially when GMM cannot accurately classify. The high nonlinear expression ability of the neural network enables it to effectively classify complex data.
[0075] Furthermore, the data collection step includes sampling the power data at a high frequency, specifically at a frequency of 400 Hz, to ensure that the collected power data has sufficient time resolution so as to more accurately capture the subtle changes of the neural network model under different input and weight conditions, thereby improving the accuracy of statistical calculations and the accuracy of the inverse model.
[0076] In this embodiment, the feature statistics calculation step further includes calculating the statistics of multimodal physical signals, and the signals include but are not limited to electromagnetic radiation, temperature changes, and current and voltage monitoring. By introducing multimodal data, the ability to comprehensively capture the operating characteristics of the neural network model is improved, and the accuracy and robustness of the reverse analysis are increased. By introducing multimodal data, the feature information of the neural network model during operation can be more comprehensively captured. This multi-dimensional feature capture not only improves the richness of the data, but also enhances the accurate description of the model's operating status, thereby increasing the accuracy and robustness of the reverse analysis.
[0077] In this embodiment, principal component analysis (PCA) technology is used in the high-dimensional feature space projection and dimensionality reduction steps to project multidimensional statistical features into a low-dimensional space to reduce computational complexity and improve the separability between features, so as to effectively perform clustering analysis and subsequent neural network classification. By reducing the dimensionality, PCA can remove redundant information, retain the main features of the data, and ensure the efficiency and accuracy of classification and clustering.
[0078] In this embodiment, the GMM clustering analysis step includes iteratively optimizing parameters through an (EM) algorithm. The EM algorithm updates the model parameters through continuous iteration until the parameter change is less than a preset threshold or the maximum number of iterations is reached. In each iteration, the algorithm updates the weights, means, and covariance matrices respectively to improve the model's fit to the data. First, in the expectation step (E step), the posterior probability that each data point belongs to each Gaussian distribution is calculated; then, in the maximization step (M step), the model parameters are updated according to these probabilities. In this way, GMM can gradually optimize the parameters and achieve accurate clustering of the data.
[0079] In this embodiment, the dynamic adaptive threshold adjustment mechanism adjusts the judgment thresholds of GMM and neural network in real time by analyzing historical data and current environmental variables, thereby improving the flexibility and accuracy of the model, and ensuring efficient and accurate classification performance under different data distributions and environments through dynamic adjustment.
[0080] In this embodiment, the neural network in the neural network classification step is a multilayer perceptron (MLP), which includes at least three hidden layers, each layer includes at least 128 neurons, and uses a ReLU activation function to improve the nonlinear expression ability and classification performance of the neural network. By configuring multiple hidden layers and an appropriate number of neurons, the MLP has stronger classification performance and can provide more accurate and effective classification results when processing complex high-dimensional feature data.
[0081] In this embodiment, the training data of the neural network includes power statistics of neural network models of various structural types, and the training is performed through a supervised learning method so that the neural network can effectively identify and classify different neural network model structures.
[0082] In this embodiment, the real-time online learning and updating step includes continuously collecting new data in actual applications and updating the GMM and neural network model parameters in real time to dynamically adapt to new side channel attack patterns and environmental changes, thereby ensuring the stability and efficiency of the model during long-term operation.
[0083] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0084] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A model structure inversion method of hybrid GMM clustering and neural network, characterized in that: The following steps are involved: S1. Collect power data and other multimodal physical signal data of the neural network model under different input and weight conditions on the embedded device; S2. Perform statistical analysis on the collected data and calculate the mean, median, standard deviation, quartile, quartile difference, and coefficient of variation statistics; S3. Using principal component analysis (PCA) technology, high-dimensional statistical features are projected into low-dimensional space to enhance the separability between features and reduce computational complexity. S4, input the reduced feature data into the Gaussian mixture model (GMM), perform cluster analysis through the Expectation-Maximization (EM) algorithm, and output the probability of each model category; S5. In the GMM judgment process, a dynamic threshold adjustment module is added to analyze the data distribution and classification results in real time, and the threshold is adjusted dynamically. When the GMM output probability of all classifications is less than the dynamic threshold, the neural network is enabled for further judgment; S6. Input the feature data into the pre-trained neural network, use the neural network to make the final model category determination, and output the neural network classification result.
2. The GMM clustering and neural network hybrid model structure inversion method according to claim 1, characterized in that: The data collection step includes sampling power data at a high frequency, specifically at a frequency of 400 Hz, to ensure that the collected power data has sufficient time resolution to more accurately capture subtle changes in the neural network model under different input and weight conditions, thereby improving the accuracy of statistical calculations and the accuracy of the inverse model.
3. The GMM clustering and neural network hybrid model structure inversion method according to claim 1, characterized in that: The characteristic statistics calculation step further includes calculating the statistics of multimodal physical signals, and the signals include but are not limited to electromagnetic radiation, temperature changes, and current and voltage monitoring. By introducing multimodal data, the comprehensive capture capability of the operation characteristics of the neural network model is improved, and the accuracy and robustness of the reverse analysis are increased.
4. The GMM clustering and neural network hybrid model structure inversion method according to claim 1, characterized in that: The principal component analysis (PCA) technique is used in the high-dimensional feature space projection and dimension reduction steps to project multidimensional statistical features into a low-dimensional space to reduce computational complexity and improve the separability between features, so as to effectively perform cluster analysis and subsequent neural network classification.
5. The GMM clustering and neural network hybrid model structure inversion method according to claim 1, characterized in that: The GMM cluster analysis step includes iteratively optimizing parameters through the Expectation-Maximization (EM) algorithm until the parameter change is less than a preset threshold or the maximum number of iterations is reached, and updating the weight, mean and covariance matrix in each iteration.
6. The GMM clustering and neural network hybrid model structure inversion method according to claim 5, characterized in that: The dynamic adaptive threshold adjustment mechanism adjusts the judgment thresholds of GMM and neural network in real time by analyzing historical data and current environmental variables, thereby improving the flexibility and accuracy of the model and ensuring efficient and accurate classification performance under different data distributions and environments through dynamic adjustment.
7. The GMM clustering and neural network hybrid model structure inversion method according to claim 1, characterized in that: The neural network in the neural network classification step is a multilayer perceptron (MLP), which includes at least three hidden layers, each layer includes at least 128 neurons, and uses a ReLU (Rectified Linear Unit) activation function to improve the nonlinear expression ability and classification performance of the neural network.
8. The GMM clustering and neural network hybrid model structure inversion method according to claim 1, characterized in that: The training data of the neural network includes power statistics of neural network models of various structural types, and the training is performed through a supervised learning method so that the neural network can effectively identify and classify different neural network model structures.
9. The GMM clustering and neural network hybrid model structure inversion method according to claim 1, characterized in that: The real-time online learning and updating step includes continuously collecting new data in actual applications and updating GMM and neural network model parameters in real time to dynamically adapt to new side channel attack patterns and environmental changes, thereby ensuring the stability and efficiency of the model during long-term operation.
