Computer application data processing method based on electric digital data processing

Abstract

The invention discloses a computer application data processing method based on electrical digital data processing, relates to the technical field of electrical digital data processing, and adopts a computer security vulnerability recognition model based on incremental learning for training; carrying out manual marking on the acquired electric digital data; carrying out dimensionality division on the acquired electric digital data; a method based on numerical approximation is adopted to improve the ability of a fractional order neural network model to capture complex nonlinear computer electrical digital data features; an auto-encoder algorithm based on a high-dimensional topological boundary effect is combined with feature sparsification of a nonlinear phase change point for improvement. According to the method, the high-dimensional complex electrical digital data is processed by adopting an auto-encoder algorithm based on the high-dimensional topological boundary effect, efficient reservation of information in the dimensionality reduction process is realized by reserving boundary effect characteristics and a sparsification strategy, and the problem of information loss during dimensionality reduction of the high-dimensional computer electrical digital data is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic digital data processing, and in particular to a computer application data processing method based on electronic digital data processing. Background Art

[0002] Computer security vulnerability identification is an important field to ensure network security and stable operation of information systems. With the development of information technology, the collection and utilization of digital data plays a key role in vulnerability identification. However, in the task of computer security vulnerability identification, digital data usually exhibits high-dimensional, nonlinear and complex characteristics. When the existing technology processes these digital data, the model needs to be retrained to adapt to the newly added digital data, resulting in high computational cost and easy forgetting of historical knowledge. In addition, key information of high-dimensional digital data is easily lost during the dimensionality reduction process, which affects the recognition accuracy of the model. Moreover, a large amount of unlabeled digital data has not been effectively utilized, which limits the generalization ability of the model in the task of unknown vulnerability detection. Therefore, how to efficiently process and utilize digital data to achieve incremental learning, improve the processing ability of high-dimensional complex digital data, and effectively utilize unlabeled digital data are technical issues that need to be addressed in the field of computer security vulnerability identification.

[0003] The Chinese invention patent with publication number CN118535866A proposes a method for processing digital data of power supply output, including data collection, feature extraction, signal processing, model selection, model parameter estimation, model verification and evaluation, and model application and optimization; the present invention uses an established nonlinear dynamic system model to predict power supply output, fault detection or control optimization, etc., studies its dynamic behavior and stability, regards the digital data of power supply output as the input signal of the nonlinear dynamic system, and studies its dynamic behavior and stability by establishing a mathematical model and a system identification method, so as to further optimize the performance and control strategy of the power supply, extract useful information or perform signal processing, and establishes a nonlinear autoregressive moving average model to predict the future trend of power supply output, thereby helping to optimize the power supply control strategy and improve the stability of power supply. The Chinese invention patent with publication number CN117609750A proposes a method for calculating target recognition rate intervals based on electronic digital data processing technology, wherein the method is applied to calculate target recognition rate intervals of three types of heterogeneous sensor systems: communication signals, SAR radars and optoelectronics. The method includes: calculating the maximum target recognition rate intervals of the three types of heterogeneous sensor systems, calculating the target recognition rate intervals of two types of heterogeneous sensors: communication signal sensors and SAR radar sensors supported by target association fusion recognition algorithms, calculating the target recognition rate intervals of two types of heterogeneous sensors: communication signal sensors and SAR radar sensors supported by target association fusion recognition algorithms, and calculating the target recognition rate intervals of two types of heterogeneous sensors: optoelectronic sensors and SAR radar sensors supported by target association fusion recognition algorithms. The Chinese invention patent with the publication number CN112307093A proposes a method for processing and analyzing electrical digital data, including: S1: collecting electrical digital parameters through an external serial port, obtaining the initial information of the current electrical digital parameters, comparing the electrical digital parameters at the same time with the initial information, obtaining difference data, and marking the data check values ​​under the same data table in the electrical digital parameters and the initial information, generating multi-dimensional detailed data of real-time electrical digital numbers, and calculating the statistical results of at least one sub-dimensional data of the current electrical digital number, and then establishing the causal relationship of the current electrical digital number. The present invention improves the traditional analysis method so that it can monitor the electrical digital parameter data online, and when the actual index exceeds the preset value, it can respond in the first time, and at the same time, it can divide different types of electrical digital parameters, which is more convenient for staff to observe in practice.

[0004] The above technical solution has the following problems that still need to be further resolved: 1. In the task of identifying computer security vulnerabilities, most of the existing dimensionality reduction methods find it difficult to retain important boundary information in the process of dimensionality reduction of high-dimensional electrical digital data, resulting in the loss of key information of the electrical digital data, which in turn affects the accuracy of computer security vulnerability identification; 2. In the task of identifying computer security vulnerabilities, the classification model based on traditional neural networks cannot efficiently capture multi-scale complex features when processing high-dimensional nonlinear computer electrical digital data, and is prone to overfitting problems, thereby reducing the stability of the computer security vulnerability identification model; 3. In the task of identifying computer security vulnerabilities, the utilization rate of unlabeled computer electrical digital data is low in existing methods, and the scale of the computer electrical digital data training set cannot be effectively expanded, resulting in insufficient recognition of unknown potential vulnerabilities and limited generalization performance of the model. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the disadvantages of the above-mentioned prior art and provide a computer application data processing method based on electronic digital data processing.

[0006] The technical solution adopted to solve the above technical problems is: a computer application data processing method based on electronic digital data processing, comprising the following steps:

[0007] S1, adopts computer security vulnerability identification model training based on incremental learning, the training data used to train the computer security vulnerability identification model is computer electrical digital data, and the model training is carried out using computer electrical digital data for vulnerability identification tasks in the field of computer security;

[0008] S2, manually annotating the collected electrical digital data;

[0009] S3, dimensionality division of the collected electrical digital data: if the dimension of the collected electrical digital data is lower than or equal to a preset threshold, no dimensionality reduction operation is performed, and a fractional-order neural network classifier based on numerical approximation is directly used to identify computer security vulnerabilities; if the dimension of the collected electrical digital data is higher than the preset threshold, a dimensionality reduction operation is performed, and an autoencoder based on a high-dimensional topological boundary effect is used for dimensionality reduction, and a preset Softmax function is used to classify the low-dimensional electrical digital data feature vector after dimensionality reduction, and any classifier among random forest, support vector machine, and decision tree can also be used to classify the low-dimensional electrical digital data feature vector after dimensionality reduction;

[0010] S4, Fractional-order neural network classifier based on numerical approximation To solve the classification problem in complex computer electrical digital data scenarios, a method based on numerical approximation is used to improve the ability of the fractional-order neural network model to capture complex nonlinear computer electrical digital data features, improve the accuracy and stability of classification, and reduce the risk of overfitting;

[0011] S5 is improved by using an autoencoder algorithm based on high-dimensional topological boundary effects combined with feature sparsification of nonlinear phase transition points. By analyzing the boundary effects of high-dimensional topological structures of computer electrical digital data, it can effectively identify key local areas and retain them in the dimensionality reduction process.

[0012] Furthermore, the workflow of the incremental learning-based computer security vulnerability identification model training method in S1 includes the following steps:

[0013] S101, import a basic equipment computer security vulnerability identification model obtained after offline training of the original electrical digital data, and judge whether a new electrical digital data set is available at the next moment. The basis for judgment is: whether the accuracy of the computer security vulnerability identification model in the original electrical digital data is improved after training with the new electrical digital data. If so, the new electrical digital data set is available, and the next step is entered.

[0014] If there is no improvement, the new electrical digital data set is unavailable and the training is stopped;

[0015] S102, selecting m sample electrical digital data of each category in the new electrical digital data set upd Sample digital data is saved, m upd The number of electrical digital data samples is preset manually. The samples of the same category of the new and old electrical digital data are clustered to obtain the feature center. The new electrical digital data sample points falling inside the two spheres are selected as the sample points, with the feature center distance as the sphere radius. The selected sample points are then added to the same category of the original electrical digital data set, and the saved fused electrical digital data set is sent to the new task, and the next step is entered at the same time.

[0016] S103, using the device computer security vulnerability identification model to start training the fused electrical digital data set after adding the new electrical digital data sample, generate a computer security vulnerability identification model corresponding to the next task, and determine the availability of the electrical digital data set corresponding to the next task. If available, return to S102, otherwise proceed to the next step;

[0017] S104, terminate the training, update the parameters of the computer security vulnerability identification model, the training data used to train the computer security vulnerability identification model is computer electrical digital data, for the vulnerability identification task in the field of computer security, use computer electrical digital data to train the model, and output the current version of the computer security vulnerability identification model.

[0018] Furthermore, the sources of the computer digital data in S1 include public vulnerability databases, runtime monitoring logs, and artificially simulated attack digital data.

[0019] Furthermore, the manually labeled categories in S2 include: label 0 represents known vulnerabilities: the digital data is labeled as the specific type of known computer vulnerabilities; label 1 represents unknown potential vulnerabilities: potential computer vulnerabilities that are labeled as abnormal but not clearly classified; label 2 represents normal behavior: behavior that is labeled as normal operation and has no computer vulnerabilities.

[0020] Furthermore, the attributes of the digital data in S2 include a severity score R of the vulnerability. a , the timestamp D when the vulnerability is triggered a , the hardware or software environment type T that triggers the vulnerability a , the input parameter S when the vulnerability is triggered a , the trigger position P of the program exception a , function call stack information F a , the key content of the recorded log L a , the component category affected by the vulnerability is C a , the interrupt flag I executed by the system a , the output exception type O generated by the vulnerability a .

[0021] Furthermore, the training process of the fractional-order neural network based on numerical approximation in S4 is as follows:

[0022] S401, initializing the parameters of the fractional-order neural network. In order to reduce the gradient disappearance or explosion in the early stage of training, the He initialization method is adopted to ensure that the fractional-order neural network has a moderate response ability in the early stage of training, and has a good foundation for numerical stability and generalization performance, which is expressed as:

[0023]

[0024] In the formula, is the weight of the lth layer of the fractional-order neural network, ~ indicates that it obeys a specific distribution, N() is a normal distribution, is the number of nodes in the l-1th layer of the fractional-order neural network, is the number of nodes in the lth layer of the fractional-order neural network, is the bias of the lth layer of the fractional-order neural network;

[0025] S402, forward propagation is performed, and the computer electrical digital data is nonlinearly activated through each layer of the fractional-order neural network, and the calculation of the fractional-order derivative is combined to capture deeper information, and multi-scale processing is performed on complex computer electrical digital data, which has better nonlinear fitting ability, expressed as:

[0026]

[0027] In the formula, is the linear transformation output of the lth layer of the fractional-order neural network, is the weight of the lth layer of the fractional-order neural network, is the output of the l-1th layer of the fractional-order neural network, is the bias of the lth layer of the fractional-order neural network;

[0028] The fractional order derivative operation is performed on the output of the activation function, so that the fractional order neural network has the effect of capturing the multi-scale memory characteristics of computer digital data, which can be expressed as:

[0029]

[0030] In the formula, is the output of the lth layer of the fractional-order neural network, is the fractional derivative operator function, α u is the derivative order, Sig() is the Sigmoid activation function, ⊙ is the element-wise product, and Fs() is the activation function of the gated layer;

[0031] The activation function of the gating layer enables the network to adjust the activation strength of each layer according to the specific characteristics of the current input. The output of the gating layer represents the open and closed state of each computer digital data feature channel, which is used to modulate the amount of information transmitted to the next layer and enhance the model's responsiveness to different computer digital data patterns. The calculation method is expressed as:

[0032]

[0033] Where Re() is the ReLU activation function, is the weight of the gating layer, is the bias of the gating layer;

[0034] S403. In order to make the fractional-order neural network have the memory effect of processing complex computer electrical digital data, the numerical approximation method is used to implement the fractional-order derivative operation in each layer, thereby enhancing the analysis ability of complex computer electrical digital data, which is expressed as:

[0035]

[0036] In the formula, Δt is the preset iteration interval step size, Δt is set to 3, N max is the maximum number of iterations of fractional-order neural network training, k is the index, representing the kth iteration, is the generalized binomial coefficient, is the linear transformation output of the lth layer of the fractional-order neural network at the tth iteration;

[0037] S404, using an adaptive feature decomposition enhancement mechanism to dynamically adjust the parameters of the computer digital data feature decomposition, optimize the digital data feature representation capability of the fractional-order neural network for the computer security vulnerability identification task, extract and enhance the computer digital data features that are helpful for classification, expressed as:

[0038]

[0039]

[0040] In the formula, is the computer digital data feature matrix enhanced by singular value decomposition in the lth layer of the fractional-order neural network, SVD() is the singular value decomposition operation function, which aims to extract the most important computer digital data feature subspace, and Re() is the ReLU activation function;

[0041] In order to adapt to the characteristics of complex computer digital data, the dynamic adjustment coefficient of the lth layer of the fractional-order neural network is calculated and expressed as:

[0042]

[0043] In the formula, is the dynamic adjustment coefficient of the lth layer of the fractional-order neural network, is the mean value of the computer electronic digital data feature matrix, which is used to evaluate the effectiveness of the current computer electronic digital data features;

[0044] The update operation of the computer digital data feature matrix enhanced by singular value decomposition at each iteration is expressed as:

[0045]

[0046] Where ← is the parameter update operation;

[0047] S405. Calculate the loss function of the fractional-order neural network and perform back propagation. The loss function consists of two parts: prediction error and fractional-order neural network regularization. When calculating the gradient through back propagation, the influence of the fractional-order derivative on the gradient is considered, so as to take into account both the control of model complexity and the capture of long-term dependent information. The calculation method is expressed as:

[0048]

[0049] Where, L u is the loss function of the fractional-order neural network, which enables the fractional-order neural network to fully understand the characteristics of computer digital data, so that the fractional-order neural network model can better learn the characteristic distribution of data, m u is the number of training computer digital data samples, i is the index, representing the i-th training sample, is the target output of the i-th computer digital data sample, is the weight of the last layer of the fractional-order neural network, is the output of the Last-1 layer of the fractional-order neural network, is the i-th input of the fractional-order neural network, is the bias of the last layer of the fractional-order neural network, λ u is the regularization coefficient of the fractional-order neural network, λ u Set to 0.4, ∥∥ F is the Frobenius norm, Last is the number of fractional-order neural network layers, l is the index, representing the lth layer of the fractional-order neural network, ρ u is the gating layer weight regularization coefficient, ρ u Set to 0.2, is the dynamic adjustment coefficient of the lth layer of the fractional-order neural network, ∥∥ is the L2 norm, μ u is the dynamic adjustment coefficient weight, μ u Set to 0.1;

[0050] S406, performing gradient correction and parameter update of the fractional-order neural network, taking into account the influence of the fractional-order derivative, adopting a modified gradient descent strategy to enable the fractional-order neural network to maintain convergence efficiency when facing non-standard and nonlinear computer digital data, expressed as:

[0051]

[0052] Where ← is the parameter update operation, is the symbol of partial derivative, η u is the learning rate of the fractional-order neural network, η u Set to 0.01;

[0053] S407, repeat the above steps until a preset stop iteration condition is met. The preset stop iteration condition is reaching a preset maximum number of iterations, which means that the model training is completed.

[0054] Furthermore, the training process of the autoencoder algorithm based on high-dimensional topological boundary effect in S5 is as follows:

[0055] S501. In the initial stage of training, the original computer digital data is first standardized so that different features have the same scale with a mean of 0 and a variance of 1, thereby effectively reducing the impact of feature dimension differences in subsequent training. In this process, the covariance matrix is ​​used to retain the correlation between features to ensure that the subsequent dimensionality reduction analysis is more accurate. The standardized calculation method is expressed as:

[0056]

[0057] In the formula, X′ r is the standardized computer digital data, X r is the original collected computer digital data, μ r is the mean vector of the features, used to characterize the overall correlation between features, σ r Set to a non-zero constant to ensure computational stability;

[0058] The calculation method of the covariance matrix of computer digital data is expressed as:

[0059]

[0060] In the formula, C r is the covariance matrix of the features, N r is the number of samples, X (r,i) is the feature vector of the i-th sample input to the autoencoder, μ r is the mean vector of the features, which is used to characterize the overall correlation between features.

[0061] () T represents transpose;

[0062] S502, after the standardized computer electronic digital data is initialized, a topological analysis is performed on the high-dimensional computer electronic digital data to identify key boundary effects in the feature space. By calculating the similarity between the sample points of the computer electronic digital data and combining the local density distribution, the boundary areas that are extremely important for the classification or regression task are identified. These areas will be retained in the subsequent dimensionality reduction process to avoid losing key structural information. In order to characterize the density of the sample points of the computer electronic digital data in the feature space, the calculation method is expressed as:

[0063]

[0064] In the formula, ρ r () is the local density function of the computer digital data sample point in the feature space, x represents the current computer digital data sample point, N r is the number of computer digital data samples, ‖xx i ‖ r is the Euclidean distance between the current computer electronic digital data sample point and its adjacent ith computer electronic digital data sample point, x i is the ith computer digital data sample adjacent to the current computer digital data sample point, σ r is the Gaussian kernel width, σ r Set to 0.1, Δρ r (x) is the local density difference;

[0065] In order to improve the recognition accuracy of high-dimensional boundary areas, the local density difference is used to adjust the traditional density calculation method through weighted similarity. The calculation method is expressed as:

[0066]

[0067] In the formula, Δρ r (x) is the local density difference;

[0068] S503, after completing the high-dimensional topological analysis and identifying the key boundaries, use the autoencoder to perform encoding and decoding to achieve compression representation and reconstruction of the high-dimensional computer electronic digital data, and guide the initial optimization of the model through the reconstruction error. The autoencoder includes an encoder and a decoder, wherein the encoder maps the input computer electronic digital data to a low-dimensional representation, and the decoder restores the low-dimensional representation to an approximate original computer electronic digital data to retain the core structure of the computer electronic digital data and reduce information loss. The calculation method of the low-dimensional vector output by the encoder is expressed as:

[0069]

[0070] In the formula, z r is the low-dimensional representation obtained by encoding, f r () is the activation function of the encoder, f r () can use the stacked activation structure of multi-layer neural networks to capture complex features, X r It is the original collected computer digital data. is the weight matrix of the encoder, is the bias term of the encoder;

[0071] The calculation method of the decoder's reconstructed output is expressed as:

[0072]

[0073] In the formula, The decoded computer digital data is approximately reconstructed. In order to make the reconstruction more accurate, the activation function of the decoder mirrors the structure of the encoder. r () is the activation function of the decoder, is the weight matrix of the decoder, is the bias term of the decoder;

[0074] In order to measure the loss of the autoencoder in the reconstruction process, the mean square error is used as the reconstruction error loss function of the autoencoder, and the calculation method is expressed as:

[0075]

[0076] Where, L r is the reconstruction error loss function, Nr is the number of samples of computer digital data input to the autoencoder in the current batch, X (r,i) is the feature vector of the i-th sample input to the autoencoder, is the reconstructed computer digital data of the i-th sample, ‖‖ r Calculate the Euclidean distance.

[0077] In order to achieve adaptive weight update and more stable convergence, an adaptive update rule combining first-order and second-order gradient information is adopted. The weight update calculation method of the encoder is expressed as:

[0078]

[0079] In the formula, is the encoder weight at the t+1th iteration, is the encoder weight of the tth iteration, η r is the learning rate of the encoder, η r Set to 0.01, is the gradient of the reconstruction error loss function with respect to the encoder weights, δ r is the adaptive parameter of the encoder, δ r Set to 0.4, is the second-order derivative of the loss function with respect to the weight;

[0080] S504, after the autoencoder completes the initial dimensionality reduction, the features with significant changes in each dimension are identified by analyzing the nonlinear phase change points, and these key features are retained to improve the expression ability of the computer digital data after dimensionality reduction. Specifically, the second-order derivative is used to identify the phase change points, and the calculation method is expressed as:

[0081]

[0082] In the formula, Δ r (f r (x)) is the second-order derivative of the activation function of the encoder at the current computer digital data sample point;

[0083] In order to suppress insignificant features and increase the sparsity of the model, L1 regularization is used to constrain the weights of the encoder. The calculation method is expressed as:

[0084]

[0085] In the formula, L1 regularization refers to the implementation method of L1 norm. is the L1 regularization term of the encoder weights, λ r is the regularization coefficient of the encoder weight, λ r Set to 0.3, is the weight matrix of the encoder, ‖‖1 is the L1 norm;

[0086] In order to more accurately locate the nonlinear phase change position of computer electrical digital data in the feature space, the entropy-based detection method is used to calculate the information entropy of each dimensional feature, and then the entropy value change is analyzed. The calculation method of information entropy is expressed as:

[0087]

[0088] In the formula, H r (x) is the information entropy of the current computer digital data sample point, xc i is the i-th feature of the current computer digital data sample point, p r (xc i ) is the probability distribution of the i-th eigenvalue, p r (xc i ) can be obtained based on histogram or kernel density estimation;

[0089] The calculation method of entropy difference is expressed as:

[0090] ΔH r (x) = H r (x)-H r (x-δ r )

[0091] In the formula, ΔH r (x) is the entropy value difference of the current computer digital data sample point, δ r is a small constant used to measure the difference before and after the characteristic perturbation, δ r Set to 0.001;

[0092] S505. Based on the reconstruction error and feature sparsification of the above-mentioned autoencoder, in order to take into account the reconstruction quality, sparsity and phase change point recognition effect, the comprehensive loss function is minimized as a constraint during the training process. The calculation method of the comprehensive loss function is expressed as:

[0093]

[0094] Where, L total The loss function is used to constrain the training process of the autoencoder. The training is performed iteratively. Each iteration needs to calculate the loss function. The training process needs to make the loss function smaller and smaller, so that the autoencoder model has higher and higher performance in reducing the feature dimension of computer digital data. eare is the weighting coefficient of entropy value difference, γ eare Set to 0.2;

[0095] S506, repeat the above steps until a preset stop iteration condition is met. The preset stop iteration condition is reaching a preset maximum number of iterations, which means that the model training is completed.

[0096] The beneficial effects of the present invention are as follows: (1) In the task of identifying computer security vulnerabilities, the present invention adopts an autoencoder algorithm based on high-dimensional topological boundary effects to process high-dimensional complex electrical digital data. By retaining boundary effect features and sparsification strategies, efficient information retention is achieved during dimensionality reduction, solving the problem of information loss during dimensionality reduction of high-dimensional computer electrical digital data.

[0097] (2) In the task of identifying computer security vulnerabilities, the present invention constructs a fractional-order neural network and uses fractional-order derivatives to capture complex nonlinear features and multi-scale memory characteristics. Through dynamic weight adjustment and feature decomposition enhancement mechanism, the classification accuracy problem of high-dimensional nonlinear electrical digital data by the computer security vulnerability identification model is solved.

[0098] (3) In the task of identifying computer security vulnerabilities, the present invention adopts a pseudo-label generation and fusion mechanism, and uses the model prediction results to generate pseudo-labels in unlabeled data. By combining with labeled data and using incremental learning, the scale of the training set is expanded, thereby improving the computer security vulnerability identification model's ability to identify unknown potential vulnerabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 It is a workflow diagram of the computer security vulnerability identification model training method based on incremental learning. DETAILED DESCRIPTION

[0100] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0101] The computer application data processing method based on electronic digital data processing of this embodiment comprises the following steps:

[0102] S1, adopts computer security vulnerability identification model training based on incremental learning, the training data used to train the computer security vulnerability identification model is computer electrical digital data, and the model training is carried out using computer electrical digital data for vulnerability identification tasks in the field of computer security.

[0103] Incremental learning can acquire knowledge from old computer security vulnerability identification model reasoning tasks, enabling the computer security vulnerability identification model to learn to solve new reasoning tasks while retaining the knowledge learned from previous reasoning tasks, avoiding retraining of computer security vulnerability identification model parameters when new electronic digital data is obtained.

[0104] The computer security vulnerability identification model training framework based on incremental learning inputs the real-time collected electrical digital data into the computer security vulnerability identification model, so that the computer security vulnerability identification model can learn the characteristics of the real-time collected electrical digital data, and learn new knowledge and capabilities from the real-time electrical digital data while maintaining the original computer security vulnerability identification model capabilities.

[0105] Incremental learning is a continuous learning process. During the training process, it is assumed that the computer security vulnerability identification model has learned the first m tasks. When facing a new task (the m+1th task) T m+1 and its corresponding electrical digital data D m+1 At the same time, the computer security vulnerability identification model trained with historical electronic digital data can use the prior knowledge learned from the old reasoning tasks to help learn new reasoning tasks, and then use the learned knowledge to update the computer security vulnerability identification model.

[0106] During the training process, for the new unlabeled electrical digital data collected, the pseudo-label method is used to use the unlabeled electrical digital data to assist the original labeled electrical digital data for training, and the computer security vulnerability identification model is used to identify the new task T. i The unlabeled electrical digital data is predicted and the prediction result R i As pseudo labels, they are added to the original training set to form a new training set D i * , and train the network again, so that new categories can be gradually learned and the performance of the computer security vulnerability identification model can be improved.

[0107] like Figure 1 As shown, the workflow of the incremental learning-based computer security vulnerability identification model training method includes the following steps:

[0108] S101, import a basic equipment computer security vulnerability identification model obtained after offline training with the original electrical digital data, and judge whether a new electrical digital data set is available at the next moment. The basis for judgment is: whether the accuracy of the computer security vulnerability identification model in the original electrical digital data is improved after training with the new electrical digital data. If so, the new electrical digital data set is available and enters the next step. If not, the new electrical digital data set is not available and training is stopped.

[0109] S102, selecting m sample electrical digital data of each category in the new electrical digital data set upd Sample digital data is saved, m updThe number of electrical digital data samples is preset manually. The samples of the same category of the new and old electrical digital data are clustered to obtain the feature center. The new electrical digital data sample points falling inside the two spheres with the feature center distance as the sphere radius are selected as the sample points. The selected sample points are then added to the same category of the original electrical digital data set, and the saved fused electrical digital data set is sent to the new task and enters the next step at the same time.

[0110] S103, use the device computer security vulnerability identification model to start training the fused electrical digital data set after adding the new electrical digital data sample, generate a computer security vulnerability identification model corresponding to the next task, and determine the availability of the electrical digital data set corresponding to the next task. If available, return to S102, if not available, proceed to the next step.

[0111] S104, terminate the training, update the parameters of the computer security vulnerability identification model, the training data used to train the computer security vulnerability identification model is computer electrical digital data, for the vulnerability identification task in the field of computer security, use computer electrical digital data to train the model, and output the current version of the computer security vulnerability identification model.

[0112] Sources of computer digital data include public vulnerability databases: such as the CVE (Common Vulnerabilities and Exposures) database, which obtains known vulnerabilities and their related descriptions, especially as the initial digital training data source in incremental learning; runtime monitoring logs: real-time recording of abnormal behaviors in program execution through the security module of the computer system; artificial simulation of attack digital data: by simulating various known and unknown attack scenarios, digital test data containing vulnerability characteristics is generated.

[0113] The electrical digital data collection methods include directly downloading the organized electrical digital data files (such as JSON or XML format) and real-time capture through a specific computer interface. All collected electrical digital data are stored as structured files in binary compression format for efficient reading, writing and processing.

[0114] The attributes of the digital data include the severity score R of the vulnerability a , the timestamp D when the vulnerability is triggered a , the hardware or software environment type T that triggers the vulnerability a , the input parameter S when the vulnerability is triggered a , the trigger position P of the program exception a , function call stack information F a , the key content of the log recorded is L a , the component category affected by the vulnerability is C a , the interrupt flag I executed by the systema , the output exception type O generated by the vulnerability a .

[0115] In this embodiment, five examples of electrical digital data are listed in the following table:

[0116]

[0117] S2, manually label the collected electrical digital data.

[0118] The manually annotated categories include:

[0119] Label 0 indicates a known vulnerability: the digital data is labeled with the specific type of known computer vulnerability;

[0120] Label 1 indicates unknown potential vulnerability: a potential computer vulnerability that is marked as anomaly but not clearly classified;

[0121] Label 2 represents normal behavior: behavior that is labeled as normal operation and without computer vulnerabilities.

[0122] The attributes of the digital data include the severity score R of the vulnerability a , the timestamp D when the vulnerability is triggered a , the hardware or software environment type T that triggers the vulnerability a , the input parameter S when the vulnerability is triggered a , the trigger position P of the program exception a , function call stack information F a , the key content of the recorded log L a , the component category affected by the vulnerability is C a , the interrupt flag I executed by the system a , the output exception type O generated by the vulnerability a .

[0123] S3, dimensionality division of the collected digital data: if the dimension of the collected digital data is lower than or equal to a preset threshold (such as 50 dimensions, i.e., corresponding to 50 attributes of the digital data), no dimensionality reduction operation is performed, and a fractional-order neural network classifier based on numerical approximation is directly used to identify computer security vulnerabilities;

[0124] If the dimension of the collected electrical digital data is higher than a preset threshold (such as 50 dimensions, corresponding to 50 attributes of the electrical digital data), a dimensionality reduction operation is performed, and an autoencoder based on a high-dimensional topological boundary effect is used for dimensionality reduction, and a preset Softmax function is used to classify the low-dimensional electrical digital data feature vectors after dimensionality reduction. Any classifier among random forests, support vector machines, and decision trees can also be used to classify the low-dimensional electrical digital data feature vectors after dimensionality reduction.

[0125] The dimensionality of the collected electrical digital data is too high, which can easily lead to a decrease in the accuracy of the computer security vulnerability identification model. Especially when faced with large-scale, high-dimensional, and nonlinear electrical digital data, the computer security vulnerability identification model based on the neural network principle is difficult to efficiently process the electrical digital data.

[0126] S4, fractional-order neural network classifier based on numerical approximation. In order to solve the classification problem in complex computer electrical digital data scenarios, a method based on numerical approximation is used to enhance the ability of the fractional-order neural network model to capture complex nonlinear computer electrical digital data features, improve the accuracy and stability of classification, and reduce the risk of overfitting.

[0127] The training process of the fractional-order neural network based on numerical approximation is as follows:

[0128] S401, initializing the parameters of the fractional-order neural network. In order to reduce the gradient disappearance or explosion in the early stage of training, the He initialization method is adopted to ensure that the fractional-order neural network has a moderate response ability in the early stage of training, and has a good foundation for numerical stability and generalization performance, which is expressed as:

[0129]

[0130] In the formula, is the weight of the lth layer of the fractional-order neural network, ~ indicates that it obeys a specific distribution, N() is a normal distribution, is the number of nodes in the l-1th layer of the fractional-order neural network, is the number of nodes in the lth layer of the fractional-order neural network, is the bias of the lth layer of the fractional-order neural network.

[0131] For the first layer of the fractional-order neural network, the input is the collected computer digital data, and the input attributes include: a Score the severity of the vulnerability, D a is the timestamp when the vulnerability is triggered, T a is the type of hardware or software environment that triggers the vulnerability, S a is the input parameter when the vulnerability is triggered, P a is the trigger position of the program exception, F a is the function call stack information, L a C is the key content of the log. a is the component category affected by the vulnerability, I a It is the interrupt flag executed by the system. a is the output exception type caused by the vulnerability, then the number of nodes in the first layer of the fractional-order neural network is 10.

[0132] S402, forward propagation is performed, and the computer electrical digital data is nonlinearly activated through each layer of the fractional-order neural network, and the calculation of the fractional-order derivative is combined to capture deeper information, and multi-scale processing is performed on complex computer electrical digital data, which has better nonlinear fitting ability, expressed as:

[0133]

[0134] In the formula, is the linear transformation output of the lth layer of the fractional-order neural network, is the weight of the lth layer of the fractional-order neural network, is the output of the l-1th layer of the fractional-order neural network, is the bias of the lth layer of the fractional-order neural network.

[0135] The fractional order derivative operation is performed on the output of the activation function, so that the fractional order neural network has the effect of capturing the multi-scale memory characteristics of computer digital data, which can be expressed as:

[0136]

[0137] In the formula, is the output of the lth layer of the fractional-order neural network, is the fractional derivative operator function, α u is the derivative order, such as Sig() is the Sigmoid activation function, ⊙ is the element-wise product, and Fs() is the activation function of the gated layer.

[0138] The activation function of the gating layer enables the network to adjust the activation strength of each layer according to the specific characteristics of the current input. The output of the gating layer represents the open and closed state of each computer digital data feature channel, which is used to modulate the amount of information transmitted to the next layer and enhance the model's responsiveness to different computer digital data patterns. The calculation method is expressed as:

[0139]

[0140] Where Re() is the ReLU activation function, is the weight of the gating layer, is the bias of the gating layer.

[0141] S403. In order to make the fractional-order neural network have the memory effect of processing complex computer electrical digital data, the numerical approximation method is used to implement the fractional-order derivative operation in each layer, thereby enhancing the analysis ability of complex computer electrical digital data, which is expressed as:

[0142]

[0143] In the formula, Δt is the preset iteration interval step size, Δt is set to 3, N max is the maximum number of iterations of fractional-order neural network training, k is the index, representing the kth iteration, is the generalized binomial coefficient, is the linear transformation output of the lth layer of the fractional-order neural network at the tth iteration.

[0144] S404, using an adaptive feature decomposition enhancement mechanism to dynamically adjust the parameters of the computer digital data feature decomposition, optimize the digital data feature representation capability of the fractional-order neural network for the computer security vulnerability identification task, extract and enhance the computer digital data features that are helpful for classification, expressed as:

[0145]

[0146]

[0147] In the formula, is the computer digital data feature matrix enhanced by singular value decomposition in the lth layer of the fractional-order neural network. SVD() is the singular value decomposition operation function, which aims to extract the most important computer digital data feature subspace. Re() is the ReLU activation function.

[0148] In order to adapt to the characteristics of complex computer digital data, the dynamic adjustment coefficient of the lth layer of the fractional-order neural network is calculated and expressed as:

[0149]

[0150] In the formula, is the dynamic adjustment coefficient of the lth layer of the fractional-order neural network, It is the mean value of the computer electronic digital data feature matrix, which is used to evaluate the effectiveness of the current computer electronic digital data features.

[0151] The update operation of the computer digital data feature matrix enhanced by singular value decomposition at each iteration is expressed as:

[0152]

[0153] Where ← is the parameter update operation.

[0154] S405. Calculate the loss function of the fractional-order neural network and perform back propagation. The loss function consists of two parts: prediction error and fractional-order neural network regularization. When calculating the gradient through back propagation, the influence of the fractional-order derivative on the gradient is considered, so as to take into account both the control of model complexity and the capture of long-term dependent information. The calculation method is expressed as:

[0155]

[0156] Where, L u is the loss function of the fractional-order neural network, which enables the fractional-order neural network to fully understand the characteristics of computer digital data, so that the fractional-order neural network model can better learn the characteristic distribution of data, m u is the number of training computer digital data samples, i is the index, representing the i-th training sample, is the target output of the i-th computer digital data sample, label 0 indicates a known vulnerability: the data is labeled with the specific type of a known computer vulnerability; label 1 indicates an unknown potential vulnerability: a potential computer vulnerability that is labeled as abnormal but not clearly classified; label 1 indicates normal behavior: behavior that is labeled as normal operation and has no computer vulnerabilities.

[0157] is the weight of the last layer of the fractional-order neural network, is the output of the Last-1 layer of the fractional-order neural network, is the i-th input of the fractional-order neural network, is the bias of the last layer of the fractional-order neural network, λ u is the regularization coefficient of the fractional-order neural network, λ u Set to 0.4, ∥∥ F is the Frobenius norm, Last is the number of fractional-order neural network layers, l is the index, representing the lth layer of the fractional-order neural network, ρ u is the gating layer weight regularization coefficient, ρ u Set to 0.2, is the dynamic adjustment coefficient of the lth layer of the fractional-order neural network, ∥∥ is the L2 norm, μ u is the dynamic adjustment coefficient weight, μ u Set to 0.1.

[0158] S406, performing gradient correction and parameter update of the fractional-order neural network, taking into account the influence of the fractional-order derivative, adopting a modified gradient descent strategy to enable the fractional-order neural network to maintain convergence efficiency when facing non-standard and nonlinear computer digital data, expressed as:

[0159]

[0160]

[0161] Where ← is the parameter update operation, is the symbol of partial derivative, η u is the learning rate of the fractional-order neural network, η u Set to 0.01.

[0162] S407, repeat the above steps until a preset stop iteration condition is met. The preset stop iteration condition is reaching a preset maximum number of iterations. The preset maximum number of iterations is set to 1000 times, which means that the model training is completed.

[0163] S5, Autoencoder based on high-dimensional topological boundary effect In order to solve the information loss and sparsity problems that may occur in the process of feature dimensionality reduction in high-dimensional computer electrical digital data space, an autoencoder algorithm based on high-dimensional topological boundary effect is improved by combining the feature sparsification of nonlinear phase transition points. By analyzing the boundary effect of the high-dimensional topological structure of computer electrical digital data, the key local areas can be effectively identified and retained in the process of dimensionality reduction. At the same time, the sparsification mechanism of nonlinear phase transition points is used to screen out the most important features, avoiding the influence of redundancy and noise, thereby improving the representation effect of computer electrical digital data in low-dimensional space.

[0164] The training process of the autoencoder algorithm based on high-dimensional topological boundary effect is as follows:

[0165] S501. In the initial stage of training, the original computer digital data is first standardized so that different features have the same scale with a mean of 0 and a variance of 1, thereby effectively reducing the impact of feature dimension differences in subsequent training. In this process, the covariance matrix is ​​used to retain the correlation between features to ensure that the subsequent dimensionality reduction analysis is more accurate. The standardized calculation method is expressed as:

[0166]

[0167] In the formula, X′ r is the standardized computer digital data, X r is the original collected computer digital data, μ r is the mean vector of the features, used to characterize the overall correlation between features, σ r Set to a non-zero constant to ensure computational stability.

[0168] The calculation method of the covariance matrix of computer digital data is expressed as:

[0169]

[0170] In the formula, C r is the covariance matrix of the features, N r is the number of samples, X (r,i) is the feature vector of the i-th sample input to the autoencoder, μ r is the mean vector of the feature, μ r Set to the global mean obtained by statistics on all samples to characterize the overall correlation between features, () T Indicates transpose.

[0171] S502, after the standardized computer electronic digital data is initialized, a topological analysis is performed on the high-dimensional computer electronic digital data to identify key boundary effects in the feature space. By calculating the similarity between the sample points of the computer electronic digital data and combining the local density distribution, the boundary areas that are extremely important for the classification or regression task are identified. These areas will be retained in the subsequent dimensionality reduction process to avoid losing key structural information. In order to characterize the density of the sample points of the computer electronic digital data in the feature space, the calculation method is expressed as:

[0172]

[0173] In the formula, ρ r () is the local density function of the computer digital data sample point in the feature space, x represents the current computer digital data sample point, N r is the number of computer digital data samples, ‖xx i ‖ r is the Euclidean distance between the current computer electronic digital data sample point and its adjacent ith computer electronic digital data sample point, x i is the ith computer digital data sample adjacent to the current computer digital data sample point, σ r is the Gaussian kernel width, σ r Set to 0.1, Δρ r (x) is the local density difference.

[0174] In order to improve the recognition accuracy of high-dimensional boundary areas, the local density difference is used to adjust the traditional density calculation method through weighted similarity. The calculation method is expressed as:

[0175]

[0176] In the formula, Δρ r (x) is the local density difference.

[0177] S503, after completing the high-dimensional topological analysis and identifying the key boundaries, use the autoencoder to perform encoding and decoding to achieve compression representation and reconstruction of the high-dimensional computer electronic digital data, and guide the initial optimization of the model through the reconstruction error. The autoencoder includes an encoder and a decoder, wherein the encoder maps the input computer electronic digital data to a low-dimensional representation, and the decoder restores the low-dimensional representation to an approximate original computer electronic digital data to retain the core structure of the computer electronic digital data and reduce information loss. The calculation method of the low-dimensional vector output by the encoder is expressed as:

[0178]

[0179] In the formula, zr is the low-dimensional representation obtained by encoding, f r () is the activation function of the encoder, f r () can use the stacked activation structure of multi-layer neural networks to capture complex features, X r It is the original collected computer digital data. is the weight matrix of the encoder, is the encoder bias term.

[0180] The calculation method of the decoder's reconstructed output is expressed as:

[0181]

[0182] In the formula, The decoded computer digital data is approximately reconstructed. In order to make the reconstruction more accurate, the activation function of the decoder mirrors the structure of the encoder. r () is the activation function of the decoder, is the weight matrix of the decoder, is the bias term of the decoder.

[0183] In order to measure the loss of the autoencoder in the reconstruction process, the mean square error is used as the reconstruction error loss function of the autoencoder, and the calculation method is expressed as:

[0184]

[0185] Where, L r is the reconstruction error loss function, N r is the number of samples of computer digital data input to the autoencoder in the current batch, X (r,i) is the feature vector of the i-th sample input to the autoencoder, is the reconstructed computer digital data of the i-th sample, ‖‖ r The Euclidean distance calculation paradigm.

[0186] In order to achieve adaptive weight update and more stable convergence, an adaptive update rule combining first-order and second-order gradient information is adopted. The weight update calculation method of the encoder is expressed as:

[0187]

[0188] In the formula, is the encoder weight at the t+1th iteration, is the encoder weight of the tth iteration, η r is the learning rate of the encoder, η r Set to 0.01, is the gradient of the reconstruction error loss function with respect to the encoder weights, δ ris the adaptive parameter of the encoder, δ r Set to 0.4, is the second-order derivative of the loss function with respect to the weight.

[0189] S504, after the autoencoder completes the initial dimensionality reduction, the features with significant changes in each dimension are identified by analyzing the nonlinear phase change points, and these key features are retained to improve the expression ability of the computer digital data after dimensionality reduction. Specifically, the second-order derivative is used to identify the phase change points, and the calculation method is expressed as:

[0190]

[0191] In the formula, Δ r (f r (x)) is the second-order derivative of the encoder's activation function at the current computer digital data sample point.

[0192] In order to suppress insignificant features and increase the sparsity of the model, L1 regularization is used to constrain the weights of the encoder. The calculation method is expressed as:

[0193]

[0194] In the formula, L1 regularization refers to the implementation method of L1 norm. is the L1 regularization term of the encoder weights, λ r is the regularization coefficient of the encoder weight, λ r Set to 0.3, is the weight matrix of the encoder, and ‖‖1 is the L1 norm.

[0195] In order to more accurately locate the nonlinear phase change position of computer electrical digital data in the feature space, the entropy-based detection method is used to calculate the information entropy of each dimensional feature, and then the entropy value change is analyzed. The calculation method of information entropy is expressed as:

[0196]

[0197] In the formula, H r (x) is the information entropy of the current computer digital data sample point, xc i is the i-th feature of the current computer digital data sample point, p r (xc i ) is the probability distribution of the i-th eigenvalue, p r (xc i ) can be obtained based on the histogram or kernel density estimation.

[0198] The calculation method of entropy difference is expressed as:

[0199] ΔH r (x) = Hr (x)-H r (x-δ r )

[0200] In the formula, ΔH r (x) is the entropy value difference of the current computer digital data sample point, δ r is a small constant used to measure the difference before and after the characteristic perturbation, δ r Set to 0.001.

[0201] S505. Based on the reconstruction error and feature sparsification of the above-mentioned autoencoder, in order to take into account the reconstruction quality, sparsity and phase change point recognition effect, the comprehensive loss function is minimized as a constraint during the training process. The calculation method of the comprehensive loss function is expressed as:

[0202]

[0203] Where, L total The loss function is used to constrain the training process of the autoencoder. The training is performed iteratively. Each iteration needs to calculate the loss function. The training process needs to make the loss function smaller and smaller, so that the autoencoder model has higher and higher performance in reducing the feature dimension of computer digital data. eare is the weighting coefficient of entropy value difference, γ eare Set to 0.2.

[0204] S506, repeat the above steps until a preset stop iteration condition is met. The preset stop iteration condition is reaching a preset maximum number of iterations, which means that the model training is completed.

[0205] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention.

Claims

1. A computer application data processing method based on electronic digital data processing, characterized in that: The following steps are involved: S1, adopts computer security vulnerability identification model training based on incremental learning, the training data used to train the computer security vulnerability identification model is computer electrical digital data, and the model training is carried out using computer electrical digital data for vulnerability identification tasks in the field of computer security; S2, manually annotating the collected electrical digital data; S3, dimensionality division of the collected electrical digital data: if the dimension of the collected electrical digital data is lower than or equal to a preset threshold, no dimensionality reduction operation is performed, and a fractional-order neural network classifier based on numerical approximation is directly used to identify computer security vulnerabilities; If the dimension of the collected electrical digital data is higher than a preset threshold, a dimensionality reduction operation is performed, and an autoencoder based on a high-dimensional topological boundary effect is used for dimensionality reduction, and a preset Softmax function is used to classify the low-dimensional electrical digital data feature vector after dimensionality reduction. Any classifier among random forest, support vector machine, and decision tree can also be used to classify the low-dimensional electrical digital data feature vector after dimensionality reduction; S4, Fractional-order neural network classifier based on numerical approximation To solve the classification problem in complex computer electrical digital data scenarios, a method based on numerical approximation is used to improve the ability of the fractional-order neural network model to capture complex nonlinear computer electrical digital data features, improve the accuracy and stability of classification, and reduce the risk of overfitting; S5 is improved by using an autoencoder algorithm based on high-dimensional topological boundary effects combined with feature sparsification of nonlinear phase transition points. By analyzing the boundary effects of high-dimensional topological structures of computer electrical digital data, it can effectively identify key local areas and retain them in the dimensionality reduction process.

2. The computer application data processing method based on electronic digital data processing according to claim 1 is characterized in that: The workflow of the incremental learning-based computer security vulnerability identification model training method in S1 includes the following steps: S101, importing a basic equipment computer security vulnerability identification model obtained after offline training with the original electrical digital data, and judging whether a new electrical digital data set is available at the next moment, the basis for judging is: whether the accuracy of the computer security vulnerability identification model in the original electrical digital data is improved after training with the new electrical digital data, if so, the new electrical digital data set is available, and the next step is entered; if not, the new electrical digital data set is not available, and the training is stopped; S102, selecting m sample electrical digital data of each category in the new electrical digital data set upd Sample digital data is saved, m upd The number of electrical digital data samples is preset manually. The samples of the same category of the new and old electrical digital data are clustered to obtain the feature center. The new electrical digital data sample points falling inside the two spheres are selected as the sample points, with the feature center distance as the sphere radius. The selected sample points are then added to the same category of the original electrical digital data set, and the saved fused electrical digital data set is sent to the new task, and the next step is entered at the same time. S103, using the device computer security vulnerability identification model to start training the fused electrical digital data set after adding the new electrical digital data sample, generate a computer security vulnerability identification model corresponding to the next task, and determine the availability of the electrical digital data set corresponding to the next task. If available, return to S102, otherwise proceed to the next step; S104, terminate the training, update the parameters of the computer security vulnerability identification model, the training data used to train the computer security vulnerability identification model is computer electrical digital data, for the vulnerability identification task in the field of computer security, use computer electrical digital data to train the model, and output the current version of the computer security vulnerability identification model.

3. The computer application data processing method based on electronic digital data processing according to claim 1, characterized in that: The sources of the computer digital data in S1 include public vulnerability databases, runtime monitoring logs, and artificially simulated attack digital data.

4. The computer application data processing method based on electronic digital data processing according to claim 1, characterized in that: The manually labeled categories in S2 include: label 0 indicates known vulnerabilities: the digital data is labeled as the specific type of known computer vulnerabilities; label 1 indicates unknown potential vulnerabilities: potential computer vulnerabilities that are labeled as abnormal but not clearly classified; label 2 indicates normal behavior: behavior that is labeled as normal operation and has no computer vulnerabilities.

5. The computer application data processing method based on electronic digital data processing according to claim 1 is characterized in that: The attributes of the S2 digital data include the severity score of the vulnerability R a , the timestamp D when the vulnerability is triggered a , the hardware or software environment type T that triggers the vulnerability a , the input parameter S when the vulnerability is triggered a , the trigger position P of the program exception a , function call stack information F a , the key content of the log recorded is L a , the component category affected by the vulnerability is C a , the interrupt flag I executed by the system a , the output exception type O generated by the vulnerability a .

6. The computer application data processing method based on electronic digital data processing according to claim 1, characterized in that: The training process of the fractional-order neural network based on numerical approximation in S4 is as follows: S401, initializing the parameters of the fractional-order neural network. In order to reduce the gradient disappearance or explosion in the early stage of training, the He initialization method is adopted to ensure that the fractional-order neural network has a moderate response ability in the early stage of training, and has a good foundation for numerical stability and generalization performance, which is expressed as: In the formula, is the weight of the lth layer of the fractional-order neural network, ~ indicates that it obeys a specific distribution, N() is a normal distribution, is the number of nodes in the l-1th layer of the fractional-order neural network, is the number of nodes in the lth layer of the fractional-order neural network, is the bias of the lth layer of the fractional-order neural network; S402, forward propagation is performed, and the computer electrical digital data is nonlinearly activated through each layer of the fractional-order neural network, and the calculation of the fractional-order derivative is combined to capture deeper information, and multi-scale processing is performed on complex computer electrical digital data, which has better nonlinear fitting ability, expressed as: In the formula, is the linear transformation output of the lth layer of the fractional-order neural network, is the weight of the lth layer of the fractional-order neural network, is the output of the l-1th layer of the fractional-order neural network, is the bias of the lth layer of the fractional-order neural network; The fractional order derivative operation is performed on the output of the activation function, so that the fractional order neural network has the effect of capturing the multi-scale memory characteristics of computer digital data, which can be expressed as: In the formula, is the output of the lth layer of the fractional-order neural network, is the fractional derivative operator function, α u is the derivative order, Sig() is the Sigmoid activation function, ⊙ is the element-wise product, and Fs() is the activation function of the gated layer; The activation function of the gating layer enables the network to adjust the activation strength of each layer according to the specific characteristics of the current input. The output of the gating layer represents the open and closed state of each computer digital data feature channel, which is used to modulate the amount of information transmitted to the next layer and enhance the model's responsiveness to different computer digital data patterns. The calculation method is expressed as: Where Re() is the ReLU activation function, is the weight of the gating layer, is the bias of the gating layer; S403. In order to make the fractional-order neural network have the memory effect of processing complex computer electrical digital data, the numerical approximation method is used to implement the fractional-order derivative operation in each layer, thereby enhancing the analysis ability of complex computer electrical digital data, which is expressed as: In the formula, Δt is the preset iteration interval step size, Δt is set to 3, N max is the maximum number of iterations for training fractional-order neural networks, k is an index representing the kth iteration, is the generalized binomial coefficient, is the linear transformation output of the lth layer of the fractional-order neural network at the tth iteration; S404, using an adaptive feature decomposition enhancement mechanism to dynamically adjust the parameters of the computer digital data feature decomposition, optimize the digital data feature representation capability of the fractional-order neural network for the computer security vulnerability identification task, extract and enhance the computer digital data features that are helpful for classification, expressed as: In the formula, is the computer digital data feature matrix enhanced by singular value decomposition in the lth layer of the fractional-order neural network, SVD() is the singular value decomposition operation function, which aims to extract the most important computer digital data feature subspace, and Re() is the ReLU activation function; In order to adapt to the characteristics of complex computer digital data, the dynamic adjustment coefficient of the lth layer of the fractional-order neural network is calculated and expressed as: In the formula, is the dynamic adjustment coefficient of the lth layer of the fractional-order neural network, is the mean value of the computer electronic digital data feature matrix, which is used to evaluate the effectiveness of the current computer electronic digital data features; The update operation of the computer digital data feature matrix enhanced by singular value decomposition at each iteration is expressed as: Where ← is the parameter update operation; S405. Calculate the loss function of the fractional-order neural network and perform back propagation. The loss function consists of two parts: prediction error and fractional-order neural network regularization. When calculating the gradient through back propagation, the influence of the fractional-order derivative on the gradient is considered, so as to take into account both the control of model complexity and the capture of long-term dependent information. The calculation method is expressed as: Where, L u is the loss function of the fractional-order neural network, which enables the fractional-order neural network to fully understand the characteristics of computer digital data, so that the fractional-order neural network model can better learn the characteristic distribution of data, m u is the number of training computer digital data samples, i is the index, representing the i-th training sample, is the target output of the i-th computer digital data sample, is the weight of the last layer of the fractional-order neural network, is the output of the Last-1 layer of the fractional-order neural network, is the i-th input of the fractional-order neural network, is the bias of the last layer of the fractional-order neural network, λ u is the regularization coefficient of the fractional-order neural network, λ u Set to 0.4, ∥∥ F is the Frobenius norm, Last is the number of fractional-order neural network layers, l is the index, representing the lth layer of the fractional-order neural network, ρ u is the gating layer weight regularization coefficient, ρ u Set to 0.2, is the dynamic adjustment coefficient of the lth layer of the fractional-order neural network, ∥∥ is the L2 norm, μ u is the dynamic adjustment coefficient weight, μ u Set to 0.1; S406, performing gradient correction and parameter update of the fractional-order neural network, taking into account the influence of the fractional-order derivative, adopting a modified gradient descent strategy to enable the fractional-order neural network to maintain convergence efficiency when facing non-standard and nonlinear computer digital data, expressed as: Where ← is the parameter update operation, is the symbol of partial derivative, η u is the learning rate of the fractional-order neural network, η u Set to 0.01; S407, repeat the above steps until a preset stop iteration condition is met. The preset stop iteration condition is reaching a preset maximum number of iterations, which means that the model training is completed.

7. The computer application data processing method based on electronic digital data processing according to claim 1 is characterized in that: The training process of the autoencoder algorithm based on high-dimensional topological boundary effect in S5 is as follows: S501. In the initial stage of training, the original computer digital data is first standardized so that different features have the same scale with a mean of 0 and a variance of 1, thereby effectively reducing the impact of feature dimension differences in subsequent training. In this process, the covariance matrix is ​​used to retain the correlation between features to ensure that the subsequent dimensionality reduction analysis is more accurate. The standardized calculation method is expressed as: In the formula, X′ r is the standardized computer digital data, X r is the original collected computer digital data, μ r is the mean vector of the features, used to characterize the overall correlation between features, σ r Set to a non-zero constant to ensure computational stability; The calculation method of the covariance matrix of computer digital data is expressed as: In the formula, C r is the covariance matrix of the features, N r is the number of samples, X (r,i) is the feature vector of the i-th sample input to the autoencoder, μ r is the mean vector of the features, () T represents transpose; S502, after the standardized computer electronic digital data is initialized, a topological analysis is performed on the high-dimensional computer electronic digital data to identify key boundary effects in the feature space. By calculating the similarity between the sample points of the computer electronic digital data and combining the local density distribution, the boundary areas that are extremely important for the classification or regression task are identified. These areas will be retained in the subsequent dimensionality reduction process to avoid losing key structural information. In order to characterize the density of the sample points of the computer electronic digital data in the feature space, the calculation method is expressed as: In the formula, ρ r () is the local density function of the computer digital data sample point in the feature space, x represents the current computer digital data sample point, N r is the number of computer digital data samples, ‖xx i ‖ r is the Euclidean distance between the current computer electronic digital data sample point and its adjacent ith computer electronic digital data sample point, x i is the ith computer digital data sample adjacent to the current computer digital data sample point, σ r is the Gaussian kernel width, σ r Set to 0.1, Δρ r (x) is the local density difference; In order to improve the recognition accuracy of high-dimensional boundary areas, the local density difference is used to adjust the traditional density calculation method through weighted similarity. The calculation method is expressed as: In the formula, Δρ r (x) is the local density difference; S503, after completing the high-dimensional topological analysis and identifying the key boundaries, use the autoencoder to perform encoding and decoding to achieve compression representation and reconstruction of the high-dimensional computer electronic digital data, and guide the initial optimization of the model through the reconstruction error. The autoencoder includes an encoder and a decoder, wherein the encoder maps the input computer electronic digital data to a low-dimensional representation, and the decoder restores the low-dimensional representation to an approximate original computer electronic digital data to retain the core structure of the computer electronic digital data and reduce information loss. The calculation method of the low-dimensional vector output by the encoder is expressed as: In the formula, z r is the low-dimensional representation obtained by encoding, f r () is the activation function of the encoder, f r () can use the stacked activation structure of multi-layer neural networks to capture complex features, X r It is the original collected computer digital data. is the weight matrix of the encoder, is the bias term of the encoder; The calculation method of the decoder's reconstructed output is expressed as: In the formula, The decoded computer digital data is approximately reconstructed. In order to make the reconstruction more accurate, the activation function of the decoder mirrors the structure of the encoder. r () is the activation function of the decoder, is the weight matrix of the decoder, is the bias term of the decoder; In order to measure the loss of the autoencoder in the reconstruction process, the mean square error is used as the reconstruction error loss function of the autoencoder, and the calculation method is expressed as: Where, L r is the reconstruction error loss function, N r is the number of samples of computer digital data input to the autoencoder in the current batch, X (r,i) is the feature vector of the i-th sample input to the autoencoder, is the reconstructed computer digital data of the i-th sample, ‖‖ r Calculate the Euclidean distance. In order to achieve adaptive weight update and more stable convergence, an adaptive update rule combining first-order and second-order gradient information is adopted. The weight update calculation method of the encoder is expressed as: In the formula, is the encoder weight at the t+1th iteration, is the encoder weight of the tth iteration, η r is the learning rate of the encoder, η r Set to 0.01, is the gradient of the reconstruction error loss function with respect to the encoder weights, δ r is the adaptive parameter of the encoder, δ r Set to 0.4, is the second-order derivative of the loss function with respect to the weight; S504, after the autoencoder completes the initial dimensionality reduction, the features with significant changes in each dimension are identified by analyzing the nonlinear phase change points, and these key features are retained to improve the expression ability of the computer digital data after dimensionality reduction. Specifically, the second-order derivative is used to identify the phase change points, and the calculation method is expressed as: In the formula, Δ r (f r (x)) is the second-order derivative of the activation function of the encoder at the current computer digital data sample point; In order to suppress insignificant features and increase the sparsity of the model, L1 regularization is used to constrain the weights of the encoder. The calculation method is expressed as: In the formula, L1 regularization refers to the implementation method of L1 norm. is the L1 regularization term of the encoder weights, λ r is the regularization coefficient of the encoder weight, λ r Set to 0.3, is the weight matrix of the encoder, ‖‖1 is the L1 norm; In order to more accurately locate the nonlinear phase change position of computer electrical digital data in the feature space, the entropy-based detection method is used to calculate the information entropy of each dimensional feature, and then the entropy value change is analyzed. The calculation method of information entropy is expressed as: In the formula, H r (x) is the information entropy of the current computer digital data sample point, xc i is the i-th feature of the current computer digital data sample point, p r (xc i ) is the probability distribution of the i-th eigenvalue, p r (xc i ) can be obtained based on histogram or kernel density estimation; The calculation method of entropy difference is expressed as: ΔH r (x)=H r (x)-H r (x-δ r ) In the formula, ΔH r (x) is the entropy value difference of the current computer digital data sample point, δ r is a small constant used to measure the difference before and after the characteristic perturbation, δ r Set to 0.001; S505. Based on the reconstruction error and feature sparsification of the above-mentioned autoencoder, in order to take into account the reconstruction quality, sparsity and phase change point recognition effect, the comprehensive loss function is minimized as a constraint during the training process. The calculation method of the comprehensive loss function is expressed as: Where, L total The loss function is used to constrain the training process of the autoencoder. The training is performed iteratively. Each iteration needs to calculate the loss function. The training process needs to make the loss function smaller and smaller, so that the autoencoder model has higher and higher performance in reducing the feature dimension of computer digital data. eare is the weighting coefficient of entropy value difference, γ eare Set to 0.2; S506, repeat the above steps until a preset stop iteration condition is met. The preset stop iteration condition is reaching a preset maximum number of iterations, which means that the model training is completed.

Images