Access Processing Method and System for Integrating Structured and Unstructured Data

Abstract

An embodiment of the present application provides a method and system for accessing and processing integrated structured and unstructured data. By loading each template access session data into multiple deep learning units respectively, a confidence sequence of sensitive operation labels obtained for each template access session data under multiple deep learning units is generated. Based on the confidence sequence of sensitive operation labels of each template access session data and the sensitive operation annotation data corresponding to each template access session data, a sensitive recognition error value corresponding to each template access session data is determined. After selecting target template access session data from multiple template access session data, weight parameters of multiple deep learning units are optimized, and combined with multiple deep learning units to jointly perform sensitive recognition learning of access operations. Compared with using a single deep learning unit for knowledge learning, the deep learning effect can be significantly improved, and the accuracy of subsequent sensitive operation recognition can be improved.

Description

Technical Field

[0001] This application relates to the technical field of access processing. Specifically, it relates to a method and system for access processing that integrates structured and unstructured data. Background Art

[0002] In the current information age, the identification and processing of access behavior data are particularly important. However, inevitably, there are some users who need to perform various access behaviors frequently, thus generating various structured access behavior data and unstructured access behavior data. To ensure data reliability, it is necessary to identify sensitive operations on these access session behavior data. Usually, in related technologies, corresponding sample training is carried out in combination with a deep learning network and then sensitive operation identification is performed. However, related technologies usually use a single deep learning unit for knowledge learning. This method has poor deep learning effect and is difficult to ensure the accuracy of subsequent sensitive operation identification. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a method and system for access processing that integrates structured and unstructured data.

[0004] According to the first aspect of this application, there is provided a method for access processing that integrates structured and unstructured data, which is applied to a cloud service data system. The method includes:

[0005] Obtain access operation sensitive identification learning data, where the access operation sensitive identification learning data includes multiple template access session data and sensitive operation annotation data corresponding to each template access session data. The template access session data integrates structured access behavior data and unstructured access behavior data;

[0006] Load each template access session data into multiple deep learning units respectively, and generate a sequence of sensitive operation label confidence levels obtained by each template access session data under the multiple deep learning units;

[0007] Based on the sequence of sensitive operation label confidence levels of each template access session data and the sensitive operation annotation data corresponding to each template access session data, determine the sensitive identification error value corresponding to each template access session data;

[0008] Based on the numerical distribution of the sensitive identification error values corresponding to each template access session data, select target template access session data from the multiple template access session data, and optimize the weight parameters of the multiple deep learning units according to the target template access session data to generate optimized multiple deep learning units;

[0009] Return to execute loading the multiple template access session data into the optimized multiple deep learning units respectively, generating a sensitive operation label confidence sequence obtained by each template access session data under the optimized multiple deep learning units and target sample access session data after scrambling the corresponding sensitive operations, and performing loop access operation sensitive recognition learning until it is detected that the weight parameters of the multiple deep learning units no longer change, generating the optimized multiple deep learning units that end the optimization;

[0010] Based on the optimized multiple deep learning units, perform sensitive operation recognition on the access session data to be analyzed, and generate sensitive operation recognition data.

[0011] In a possible implementation manner of the first aspect, the determining the sensitive recognition error value corresponding to each template access session data based on the sensitive operation label confidence sequence of each template access session data and the sensitive operation annotation data corresponding to each template access session data includes:

[0012] Calculate the first mean square error loss function value between each sensitive operation label confidence in the sensitive operation label confidence sequence corresponding to each template access session data and the corresponding sensitive operation annotation data:

[0013] Add up the determined first mean square error loss function values to generate the first target sensitive recognition error value corresponding to each template access session data;

[0014] Determine the sensitive recognition error value corresponding to each template access session data based on the first target sensitive recognition error value corresponding to each template access session data.

[0015] In a possible implementation manner of the first aspect, the method further includes:

[0016] Calculate the information divergence value between every N sensitive operation labels in the sensitive operation label confidence sequence corresponding to each template access session data:

[0017] Add up the information divergence values to generate the second target sensitive recognition error value corresponding to each template access session data;

[0018] The determining the sensitive recognition error value corresponding to each template access session data based on the first target sensitive recognition error value corresponding to each template access session data includes:

[0019] Perform weight fusion on the first target sensitive recognition error value and the second target sensitive recognition error value corresponding to each template access session data to generate the sensitive recognition error value corresponding to each template access session data.

[0020] In a possible implementation manner of the first aspect, the method further includes:

[0021] Obtain the confidence distribution of the sensitive operation annotation data in the template access session data, and generate corresponding training label features according to the confidence distribution;

[0022] Calculate the value of the second mean square error loss function between the training label features and the confidence sequence of the sensitive operation labels corresponding to each template access session data:

[0023] Add up the determined values of the second mean square error loss function to generate the third target sensitive recognition error value corresponding to each template access session data;

[0024] The weight fusion of the first target sensitive recognition error value and the second target sensitive recognition error value corresponding to each template access session data to generate the sensitive recognition error value corresponding to each template access session data includes:

[0025] Perform weight fusion on the first target sensitive recognition error value, the second target sensitive recognition error value, and the third target sensitive recognition error value corresponding to each template access session data to generate the sensitive recognition error value corresponding to each template access session data.

[0026] In a possible implementation manner of the first aspect, the selection of the target template access session data from the multiple template access session data based on the numerical distribution of the sensitive recognition error values corresponding to each template access session data includes:

[0027] Obtain the number of learning rounds for sensitive recognition learning of the cyclic access operation on the multiple deep learning units;

[0028] Calculate the learning scale of the target template access session data based on the number of learning rounds of the cyclic access operation sensitive recognition learning;

[0029] Select the template access session data corresponding to the learning scale according to the ascending order of the sensitive recognition error values to generate the target template access session data.

[0030] In a possible implementation manner of the first aspect, the calculation of the learning scale of the target template access session data based on the number of learning rounds of the cyclic access operation sensitive recognition learning includes:

[0031] Obtain the set training optimization parameters, and the set training optimization parameters are configured to optimize the weight parameters of the multiple template access session data;

[0032] Calculate the proportion of the target template access session data in the multiple template access session data based on the set training optimization parameters and the number of learning rounds for sensitive recognition learning of the loop access operation;

[0033] Determine the learning scale of the target template access session data based on the proportion and the number of the multiple template access session data.

[0034] In a possible implementation manner of the first aspect, the sensitive operation recognition of the to-be-analyzed access session data based on the multiple deep learning units after the end of optimization to generate sensitive operation recognition data includes:

[0035] Load the to-be-analyzed access session data into the multiple deep learning units after the end of optimization for sensitive operation recognition to generate multiple sensitive operation label confidences;

[0036] Calculate the mean confidence of the multiple sensitive operation label confidences to generate the sensitive operation prediction label corresponding to the to-be-analyzed access session data.

[0037] In a possible implementation manner of the first aspect, the sensitive operation recognition of the to-be-analyzed access session data based on the multiple deep learning units after the end of optimization to generate sensitive operation recognition data includes:

[0038] Obtain the sensitive operation recognition performance metrics of the multiple deep learning units after the end of optimization to generate multiple sensitive operation recognition performance metrics;

[0039] Sort the multiple sensitive operation recognition performance metrics in descending order, and determine the deep learning unit with the largest sensitive operation recognition performance metric as the target deep learning unit;

[0040] Load the to-be-analyzed access session data into the target deep learning unit for sensitive operation recognition to generate the sensitive operation prediction label corresponding to the to-be-analyzed access session data.

[0041] According to the second aspect of the present application, there is provided a cloud service data system, which includes a processor and a readable storage medium. The readable storage medium stores a program, and when the program is executed by the processor, the foregoing access processing method for fusing structured and unstructured data is implemented.

[0042] According to the third aspect of the present application, there is provided a computer-readable storage medium, in which computer-executable instructions are stored, and when it is monitored that the computer-executable instructions are executed, the foregoing access processing method for fusing structured and unstructured data is implemented.

[0043] According to any of the above aspects, in the present application, by obtaining access operation sensitive recognition learning data, each template access session data is respectively loaded into a plurality of deep learning units to generate a sensitive operation label confidence sequence obtained by each template access session data under the plurality of deep learning units. Based on the sensitive operation label confidence sequence of each template access session data and the sensitive operation annotation data corresponding to each template access session data, the sensitive recognition error value corresponding to each template access session data is determined, and after selecting the target template access session data from the plurality of template access session data, the weight parameters of the plurality of deep learning units are optimized. Then, after performing cyclic access operation sensitive recognition learning, the sensitive operation recognition of the access session data to be analyzed is performed to generate sensitive operation recognition data. Thus, by jointly performing access operation sensitive recognition learning with a plurality of deep learning units, compared with using a single deep learning unit for knowledge learning, the deep learning effect can be significantly improved, and the accuracy of subsequent sensitive operation recognition can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0045] Figure 1 The flowchart of the access processing method for integrating structured and unstructured data provided by the embodiments of the present application;

[0046] Figure 2 The component structure diagram of the cloud service data system for implementing the above access processing method for integrating structured and unstructured data provided by the embodiments of the present application is shown. EMBODIMENTS

[0047] In order to enable those skilled in the art to better understand the solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0048] It should be noted that the terms "first", "second", etc. in the description, claims and the above-mentioned drawings of this application are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application described here can be implemented in an order other than those illustrated or described here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0049] Figure 1 The flowchart of the access processing method for integrating structured and unstructured data provided by the embodiment of the present application is shown. It should be understood that in other embodiments, the order of some steps of the access processing method for integrating structured and unstructured data in this embodiment can be interchanged according to actual needs, or some of the steps can also be omitted or deleted. The detailed steps of the access processing method for integrating structured and unstructured data are introduced as follows.

[0050] Step S101, obtain sensitive recognition learning data for access operations.

[0051] In an alternative embodiment, the template access session data may include multiple template access session data and the sensitive operation annotation data corresponding to each template access session data. Among them, the sensitive operation annotation data corresponding to the template access session data is the confidence level corresponding to each sensitive operation label. Among them, the sensitive operation annotation data is not necessarily very accurate, so that the sensitive operation annotation data of the template access session data contains some interference features, that is, the reliability of the confidence level corresponding to some sensitive operation labels is not high.

[0052] Among them, the template access session data integrates structured access behavior data and unstructured access behavior data.

[0053] Step S102, load each template access session data into multiple deep learning units respectively, and generate a sequence of confidence levels of sensitive operation labels obtained by each template access session data under the multiple deep learning units.

[0054] In an alternative embodiment, joint knowledge learning is performed based on multiple deep learning units. The multiple deep learning units can be deep learning units that have not yet performed knowledge learning related to sensitive recognition, or can also be deep learning units that have undergone certain knowledge learning related to sensitive recognition.

[0055] Multiple template access session data included in the template access session data are loaded one by one into multiple deep learning units for sensitive operation recognition. Herein, the sensitive operation recognition is to predict the sensitive operation labels of the template access session data, so as to obtain the corresponding sensitive operation label confidence levels. Herein, the sensitive operation label confidence level is the possibility that the template access session data matches the corresponding sensitive operation label. For any target template access session data, when it is loaded into multiple deep learning units, the sensitive operation label confidence levels obtained by each deep learning unit can be obtained, and multiple sensitive operation label confidence levels corresponding to the target template access session data are generated. The multiple sensitive operation label confidence levels constitute the sensitive operation label confidence level sequence corresponding to the target template access session data. Similarly, for other template access session data, when it is loaded into multiple deep learning units, multiple sensitive operation label confidence level sequences obtained by the multiple deep learning units can also be generated, thereby generating the sensitive operation label confidence level sequence corresponding to each template access session data.

[0056] Step S103, determining the sensitive recognition error value corresponding to each template access session data based on the sensitive operation label confidence level sequence of each template access session data and the sensitive operation annotation data corresponding to each template access session data.

[0057] Wherein, after determining the sensitive operation label confidence level sequence of each template access session data, the sensitive recognition error value corresponding to each template access session data is calculated based on the sensitive operation label confidence level sequence and the sensitive operation annotation data corresponding to each template access session data. Herein, the sensitive recognition error value is an index for evaluating the difference between the sensitive operation label confidence level of the template access session data and the sensitive operation prediction confidence level of the model. The purpose of joint knowledge learning is to minimize the sensitive recognition error value corresponding to the template access session data, that is, to make the sensitive operation label confidence level of the template access session data and the sensitive operation prediction confidence level of the model infinitely close.

[0058] In an alternative implementation manner, determining the sensitive recognition error value corresponding to each template access session data based on the sensitive operation label confidence level sequence of each template access session data and the sensitive operation annotation data corresponding to each template access session data includes:

[0059] 1. Calculating the value of the first mean square error loss function between each sensitive operation label confidence level in the sensitive operation label confidence level sequence corresponding to each template access session data and the corresponding sensitive operation annotation data:

[0060] 2. Adding the determined values of the first mean square error loss function to generate the first target sensitive recognition error value corresponding to each template access session data;

[0061] 3. Determine the sensitive recognition error value corresponding to each template access session data based on the first target sensitive recognition error value corresponding to the template access session data.

[0062] In an alternative embodiment, the sensitive recognition error value corresponding to the template access session data can be determined based on the mean square error loss function value between the confidence sequence of sensitive operation labels formed by the sensitive operation labels in the sensitive operation label confidence sequence corresponding to each template access session data and the confidence sequence of sensitive operation labels formed by the sensitive operation labels of the template access session data. Among them, the confidence sequence of sensitive operation labels formed by the sensitive operation labels of the template access session data is a set composed of the sensitive operation label confidences of multiple template access session data.

[0063] After determining the mean square error loss function value between the confidence sequence of sensitive operation labels corresponding to any target template access session data and the corresponding sensitive operation annotation data, generate multiple mean square error loss function values corresponding to the target template access session data. Then add up the multiple mean square error loss function values corresponding to the target template access session data to generate the first target sensitive recognition error value corresponding to the target template access session data, and determine this first target sensitive recognition error value as the sensitive recognition error value of the target template access session data. Then, the sensitive recognition error value corresponding to each template access session data can be determined in the same way as in the above embodiments.

[0064] In an alternative embodiment, the above embodiments may further include:

[0065] A. Calculate the information divergence value between every N sensitive operation label confidences in the sensitive operation label confidence sequence corresponding to each template access session data:

[0066] B. Add up the information divergence values to generate the second target sensitive recognition error value corresponding to each template access session data;

[0067] C. Determining the sensitive recognition error value corresponding to each template access session data based on the first target sensitive recognition error value corresponding to each template access session data includes:

[0068] Perform weight fusion on the first target sensitive recognition error value and the second target sensitive recognition error value corresponding to each template access session data to generate the sensitive recognition error value corresponding to each template access session data.

[0069] In an alternative implementation, the information divergence value between the confidence levels of sensitive operation labels obtained from the same template access session data under different models can be further calculated. When the number of deep learning units for joint knowledge learning is 2, the information divergence value corresponding to the template access session data is one; when the number of deep learning units for joint knowledge learning is 3, the information divergence value corresponding to the template access session data is three; when the number of deep learning units for joint knowledge learning is k, the information divergence value corresponding to the template access session data is k*(k - 1) / 2. After determining all the information divergence values corresponding to the template access session data, these information divergence values are added together to generate the second target sensitive recognition error value corresponding to the template access session data. Moreover, the above first target sensitive recognition error value and the second target sensitive recognition error value are weighted and fused to generate the sensitive recognition error value corresponding to the template access session data, thereby determining the sensitive recognition error value corresponding to each template access session data. Loading the information divergence value of the prediction confidence levels of the same template access session data in different deep learning units into the sensitive recognition error value of the template access session data can make the prediction confidence levels of different deep learning units approach infinitely when performing joint knowledge learning, thereby improving the reliability of joint knowledge learning.

[0070] In an alternative implementation, the above method embodiments may further include:

[0071] a. Obtain the confidence distribution of the sensitive operation annotation data in the template access session data, and generate the corresponding training label features based on the confidence distribution;

[0072] b. Calculate the second mean square error loss function value between the training label features and the confidence sequence of the sensitive operation labels corresponding to each template access session data:

[0073] c. Add up the determined second mean square error loss function values to generate the third target sensitive recognition error value corresponding to each template access session data;

[0074] d. Perform weighted fusion on the first target sensitive recognition error value and the second target sensitive recognition error value corresponding to each template access session data to generate the sensitive recognition error value corresponding to each template access session data, including:

[0075] Perform weighted fusion on the first target sensitive recognition error value, the second target sensitive recognition error value, and the third target sensitive recognition error value corresponding to each template access session data to generate the sensitive recognition error value corresponding to each template access session data.

[0076] In an alternative implementation, it is possible to first determine the sensitive operation annotation data for accessing session data of multiple templates, and then obtain the confidence distribution of the sensitive operation annotation data in the template access session data based on the sensitive operation annotation data for accessing session data of the multiple templates. Then, it is possible to generate its corresponding training label features based on this confidence distribution for calculating the mean squared error loss function value. On this basis, it is possible to calculate the mean squared error loss function value between this confidence distribution and the confidence sequence of the sensitive operation labels corresponding to each template access session data, and then add up the obtained mean squared error loss function values to generate the third target sensitive recognition error value corresponding to each template access session data. Thus, it is possible to perform weighted fusion on the above first target sensitive recognition error value, second target sensitive recognition error value, and third target sensitive recognition error value to generate the sensitive recognition error value corresponding to each template access session data.

[0077] Step S104: Select target template access session data from multiple template access session data based on the numerical distribution of the sensitive recognition error values corresponding to each template access session data, and optimize the weight parameters of multiple deep learning units based on the target template access session data to generate optimized multiple deep learning units.

[0078] Among them, after determining the sensitive recognition error value corresponding to each template access session data, a set number of target template access session data with smaller sensitive recognition error values are determined from the template access session data based on the numerical distribution of the sensitive recognition error values corresponding to the template access session data. Then, the weight parameters of multiple deep learning units are optimized based on the set number of target template access session data, and the weight parameters of the initial multiple deep learning units are optimized based on the optimized multiple deep learning units to generate optimized multiple deep learning units. Among them, the smaller the sensitive recognition error value of the template access session data, the closer the obtained prediction confidence is to the sensitive operation annotation data of the template access session data. And the larger the sensitive recognition error value of the template access session data, the lower the accuracy of the obtained prediction confidence. Therefore, some template access session data with larger sensitive recognition error values can be removed from the template access session data, so that the accuracy of the sensitive operation annotation data of the remaining template access session data is higher, in order to improve the accuracy of subsequent joint knowledge learning.

[0079] Among them, in an alternative implementation, selecting target template access session data from multiple template access session data based on the numerical distribution of the sensitive recognition error values corresponding to each template access session data includes:

[0080] 1. Obtain the number of learning rounds for sensitive recognition learning of the cyclic access operation of multiple deep learning units;

[0081] 2. Calculate the learning scale of the target template access session data based on the number of learning rounds of the learning sensitive to cyclic access operations;

[0082] 3. Select the template access session data corresponding to the learning scale in ascending order of the sensitive recognition error value to generate the target template access session data.

[0083] Among them, a set number of target template access session data is determined from multiple template access session data, and after optimizing the weight parameters of multiple deep learning units based on the target template access session data, the sensitive operation recognition is performed again on each template access session data based on the optimized multiple deep learning units to generate a sensitive operation label confidence sequence corresponding to each template access session data; then, based on the new sensitive operation label confidence sequence and the prior annotation value of the sensitive operation of each template access session data, calculate the new sensitive recognition error value of each template access session data, and re-determine the target template access session data based on the new sensitive recognition error value and perform sensitive operation recognition and optimization on the optimized multiple deep learning units based on the new target template access session data, so as to perform multiple rounds of learning sensitive to cyclic access operations on multiple deep learning units.

[0084] Next, the learning scale of the target template access session data to be retained can be calculated based on the number of learning rounds. Finally, select the template access session data corresponding to the learning scale in ascending order of the sensitive recognition error value of each template access session data to generate the target template access session data. That is, determine that the template access session data with a smaller sensitive recognition error value corresponding to the learning scale among multiple template access session data is the target template access session data.

[0085] In an alternative implementation, calculating the learning scale of the target template access session data based on the number of learning rounds of the learning sensitive to cyclic access operations includes:

[0086] 2.1. Obtain the set training optimization parameters, and the set training optimization parameters are configured to optimize the weight parameter data of multiple template access session data;

[0087] 2.2. Calculate the proportion of the target template access session data in multiple template access session data based on the set training optimization parameters and the number of learning rounds of the learning sensitive to cyclic access operations;

[0088] 2.3. Determine the learning scale of the target template access session data based on the proportion and the number of multiple template access session data.

[0089] Step S105: Return and execute the operation of separately loading multiple template access session data into the optimized multiple deep learning units, generating the confidence sequences of sensitive operation labels obtained by each template access session data under the optimized multiple deep learning units, and the target sample access session data after scrambling the corresponding sensitive operations, and performing cyclic access operation sensitive recognition learning until it is detected that the weight parameters of the multiple deep learning units no longer change, generating the multiple deep learning units with optimized termination.

[0090] Among them, the above steps S102 to S104 are a training stage in the cyclic access operation sensitive recognition learning of the deep learning units. That is, based on the multiple deep learning units, sensitive operation recognition is performed on the multiple template access session data, and the confidence sequence of sensitive operation labels corresponding to each template access session data is output. According to the confidence sequence of sensitive operation labels of each template access session data and the prior annotation value of the sensitive operation of each template access session data, the sensitive recognition error value corresponding to each template access session data is calculated. Then, based on the sensitive recognition error value of each template access session data, the target template access session data is determined. Further, based on the target template access session data, the weight parameters of the multiple deep learning units are optimized and updated. This is a training stage in the cyclic access operation sensitive recognition learning of the multiple deep learning units.

[0091] After obtaining the optimized multiple deep learning units, it is also necessary to substitute the optimized multiple deep learning units into step S102 for processing in the next training stage. That is, the multiple template access session data are separately loaded into the optimized multiple deep learning units, generating the confidence sequence of sensitive operation labels obtained by each template access session data under the optimized multiple deep learning units. Then, based on this confidence sequence of sensitive operation labels and the prior annotation value of the sensitive operation of each template access session data, the new sensitive recognition error value corresponding to each template access session data is calculated again. Further, based on the sensitive recognition error value of each template access session data and the discussion of cyclic access operation sensitive recognition learning, the new target template access session data is determined, and based on the new target template access session data, the weight parameters of the optimized multiple deep learning units are optimized and updated again. In this way, cyclic access operation sensitive recognition learning is performed on the multiple deep learning units until it is detected that the weight parameters of the multiple deep learning units no longer change, generating the multiple deep learning units with optimized termination.

[0092] Step S106: Based on the multiple deep learning units with optimized termination, perform sensitive operation recognition on the access session data to be analyzed, generating sensitive operation recognition data.

[0093] Among them, after optimizing the weight parameters of multiple deep learning units and generating multiple deep learning units with optimized ends, sensitive operation recognition is performed on the access session data to be analyzed based on the multiple deep learning units with optimized ends, and sensitive operation recognition data is generated.

[0094] Among them, in an alternative implementation, sensitive operation recognition is performed on the access session data to be analyzed based on the multiple deep learning units with optimized ends, and sensitive operation recognition data is generated, including:

[0095] 1. Load the access session data to be analyzed into the multiple deep learning units with optimized ends for sensitive operation recognition, and generate multiple sensitive operation label confidence levels;

[0096] 2. Calculate the mean confidence level of the multiple sensitive operation label confidence levels to generate a sensitive operation prediction label corresponding to the access session data to be analyzed.

[0097] In an alternative implementation, after performing sensitive recognition learning on the loop access operation of multiple deep learning units to obtain multiple deep learning units with optimized ends, the access session data to be analyzed is loaded into the multiple deep learning units with optimized ends for sensitive operation recognition, and the sensitive operation label confidence levels obtained by each deep learning unit with optimized ends for sensitive operation recognition of the access session data to be analyzed are generated, that is, multiple sensitive operation label confidence levels are obtained. Then, an average calculation is performed on the multiple sensitive operation label confidence levels to generate a final sensitive operation prediction label, and this sensitive operation prediction label is the sensitive operation recognition result obtained by the multiple deep learning units with optimized ends for sensitive operation recognition of the access session data to be analyzed.

[0098] Among them, in an alternative implementation, sensitive operation recognition is performed on the access session data to be analyzed based on the multiple deep learning units with optimized ends, and sensitive operation recognition data is generated, including:

[0099] A. Obtain the sensitive operation recognition performance indicators of the multiple deep learning units with optimized ends to generate multiple sensitive operation recognition performance indicators;

[0100] B. Sort the multiple sensitive operation recognition performance indicators in descending order, and determine the deep learning unit with the largest sensitive operation recognition performance indicator as the target deep learning unit;

[0101] C. Load the access session data to be analyzed into the target deep learning unit for sensitive operation recognition, and generate a sensitive operation prediction label corresponding to the access session data to be analyzed.

[0102] In an alternative embodiment, after optimizing the weight parameters of multiple deep learning units to obtain multiple deep learning units with optimized ends, it is not necessary to use all the deep learning units with optimized ends to perform sensitive operation recognition on the access session data to be analyzed. Instead, the sensitive operation recognition performance metrics of each deep learning unit among the multiple deep learning units with optimized ends are obtained, and then the deep learning unit with the maximum sensitive operation recognition performance metric is determined as the target deep learning unit. Finally, based on the target deep learning unit, sensitive operation recognition is performed on the access session data to be analyzed, the confidence level of the sensitive operation label obtained by the target deep learning unit is generated, and the confidence level of the sensitive operation label obtained by the target deep learning unit is determined as the sensitive operation data of the access session data to be analyzed.

[0103] Based on the above steps, by obtaining access operation sensitive recognition learning data, the access operation sensitive recognition learning data includes multiple template access session data and sensitive operation annotation data corresponding to each template access session data; each template access session data is respectively loaded into multiple deep learning units to generate a sequence of confidence levels of sensitive operation labels obtained by each template access session data under the multiple deep learning units; based on the sequence of confidence levels of sensitive operation labels of each template access session data and the sensitive operation annotation data corresponding to each template access session data, the sensitive recognition error value corresponding to each template access session data is determined; based on the numerical distribution of the sensitive recognition error values corresponding to each template access session data, the target template access session data is selected from the multiple template access session data, and the weight parameters of the multiple deep learning units are optimized according to the target template access session data to generate multiple optimized deep learning units; return to execute loading the multiple template access session data into the multiple optimized deep learning units respectively, generating a sequence of confidence levels of sensitive operation labels obtained by each template access session data under the multiple optimized deep learning units and the target sample access session data after scrambling the corresponding sensitive operations, and performing cyclic access operation sensitive recognition learning until it is detected that the weight parameters of the multiple deep learning units no longer change, generating multiple deep learning units with optimized ends; based on the multiple deep learning units with optimized ends, sensitive operation recognition is performed on the access session data to be analyzed to generate sensitive operation recognition data. Thus, by combining multiple deep learning units to jointly perform access operation sensitive recognition learning, compared with using a single deep learning unit for knowledge learning, the deep learning effect can be significantly improved, and the accuracy of subsequent sensitive operation recognition can be improved.

[0104] Correspondingly, the embodiments of the present application will further describe in detail the image detection method provided by the present application from the perspective of a computer device, where the computer device can be a terminal or a server. Another image detection method provided by the present application includes:

[0105] Step S201: Obtain access operation sensitive recognition learning data that includes multiple template access session data and sensitive operation annotation data for each template access session data.

[0106] Step S202: Load the multiple template access session data into N deep learning units respectively for abnormal permission operation recognition, and generate the confidence distribution of N abnormal permission operations obtained by each template access session data in the N deep learning units.

[0107] In an alternative implementation, the weight parameters of the abnormal permission operation recognition network can be optimized based on the joint knowledge learning method.

[0108] After obtaining the multiple template access session data and the confidence distribution of the abnormal permission operations for each template access session data, load the multiple template access session data into N deep learning units respectively, and generate the confidence distribution of N abnormal permission operations obtained by each template access session data in the N deep learning units.

[0109] Step S203: Calculate the mean square error loss function value between the confidence distribution of N abnormal permission operations and the sensitive operation annotation data, and generate the first target sensitive recognition error value.

[0110] Among them, after determining the confidence distribution of the abnormal permission operations obtained by each template access session data under N deep learning units, calculate the mean square error loss function value corresponding to each template access session data based on the confidence distribution of the abnormal permission operations of each template access session data and the sensitive operation annotation data.

[0111] Then, add the determined N mean square error loss function values to generate the first target sensitive recognition error value.

[0112] Step S204: Calculate the information divergence value between the confidence distributions of N abnormal permission operations, and generate the second target sensitive recognition error value.

[0113] Step S205: Calculate the information divergence value between the confidence distribution of N abnormal permission operations and the template access session data label distribution, and generate the third target sensitive recognition error value.

[0114] Step S206: Calculate the sensitive recognition error value corresponding to each template access session data based on the first target sensitive recognition error value, the second target sensitive recognition error value, and the third target sensitive recognition error value.

[0115] Step S207: Determine the target template access session data based on the sensitive recognition error value of each template access session data.

[0116] Among them, after determining the sensitive recognition error value corresponding to each template access session data, it is necessary to optimize the weight parameter data of the template access session data based on the sensitive recognition error value of the template access session data, so as to eliminate the template data with larger interference features (the prior annotation value of sensitive operations is not accurate enough).

[0117] After determining the proportion of the target access session data in multiple template access session data, based on this proportion, select the template access session data with the smallest proportion of sensitive recognition error values from the multiple template access session data as the target template access session data.

[0118] Step S208, optimize the weight parameters of the N deep learning units based on the target template access session data, and optimize the weight parameters of the N deep learning units based on the N deep learning units with the optimization ended.

[0119] Among them, after determining the target template access session data, optimize the weight parameters of the N deep learning units based on the target template access session data and its corresponding prior annotation value of sensitive operations, so as to optimize the weight parameters of the N deep learning units and generate the optimized N deep learning units. Then, perform iterative optimization based on the optimized N deep learning units.

[0120] Step S209, determine whether the number of sensitive recognition learning rounds of the loop access operation is greater than the set number of rounds.

[0121] Among them, after each optimization of the weight parameters of the N deep learning units, it is necessary to judge the number of sensitive recognition learning rounds of the loop access operation to determine whether the preset number of sensitive recognition learning rounds of the loop access operation is reached. If it is not greater than, return to step S202, re-identify the abnormal permission operation for each template access session data based on the optimized N deep learning units, generate a new confidence distribution of the abnormal permission operation, and then further calculate the new sensitive recognition error value of each template access session data based on the new confidence distribution of the abnormal permission operation. Then, re-determine the new target template access session data, and perform advanced optimization on the optimized N deep learning units based on the new target template access session data.

[0122] Step S210, determine the updated N deep learning units as the N deep learning units with the optimization ended.

[0123] If the number of sensitive recognition learning rounds of the loop access operation is greater than the set number of rounds, determine the finally obtained N deep learning units as the finally optimized deep learning units.

[0124] Step S211: Based on the N deep learning units with optimized termination, perform abnormal permission operation recognition on the access session data to be analyzed, and generate an abnormal permission operation recognition result.

[0125] Among them, after determining the N deep learning units with optimized termination, abnormal permission operation recognition can be performed on the access session data to be analyzed based on the N deep learning units with optimized termination. Specifically, a target deep learning unit with better abnormal permission operation recognition performance can be determined from the N deep learning units with optimized termination to perform abnormal permission operation recognition on the access session data to be analyzed.

[0126] Therefore, by combining multiple deep learning units for joint access operation sensitive recognition learning, compared with using a single deep learning unit for knowledge learning, the deep learning effect can be significantly improved, and the accuracy of subsequent sensitive operation recognition can be improved.

[0127] Furthermore, Figure 2 shows a schematic hardware structure diagram of a device for implementing the method provided in the embodiments of the present application. As Figure 2 shown, the cloud service data system 100 may include one or more processors 102 (the processor 102 may include, but is not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, a transmission device 106 for communication functions, and a controller 108. Those of ordinary skill in the art can understand that Figure 2 the structure shown is only schematic and does not limit the structure of the above cloud service data system 100. For example, the cloud service data system 100 may further include more or fewer components than those Figure 2 shown, or have a different configuration from that Figure 2 shown.

[0128] The memory 104 can be used to store software programs and modules of application software, such as the program instructions corresponding to the above method embodiments in the embodiments of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, to implement the above account label prediction method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, a flash memory, or other non-volatile solid-state memories. In some instances, the memory 104 may further include a memory remotely set relative to the processor 102, and these remote memories can be connected to the cloud service data system 100 through a network. Examples of the above network include, but are not limited to, the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0129] The transmission device 106 is used to receive or send data via a network. Specific examples of the above-mentioned network may include a wireless network provided by a communication provider of the cloud service data system 100. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a Radio Frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0130] It should be noted that: the above sequence of embodiments of the present application is only for description and does not represent the superiority or inferiority of the embodiments. And the above specific embodiments of the present application have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve the desired result. Additionally, the processes depicted in the figures do not necessarily require the particular order or sequential order shown to achieve the desired result. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0131] The various embodiments in the embodiments of the present application are all described in a progressive manner. For the same or similar parts among the various embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the above different embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and reference can be made to the corresponding parts of the method embodiments for the relevant parts.

[0132] Those of ordinary skill in the art can understand that all or part of the steps to implement the above embodiments can be completed by hardware, or can be completed by a program instructing the relevant hardware. The above program can be stored in a computer-readable storage medium, and the above-mentioned storage medium can be a read-only memory, a magnetic disk or an optical disc, etc.

Claims

1. An access processing method for integrating structured and unstructured data, characterized in that The method includes: Obtaining access operation sensitive recognition learning data, where the access operation sensitive recognition learning data includes multiple template access session data and sensitive operation annotation data corresponding to each template access session data, and the template access session data integrates structured access behavior data and unstructured access behavior data; Loading each template access session data into multiple deep learning units respectively to generate a sequence of sensitive operation label confidence levels obtained by each template access session data under the multiple deep learning units; Determining a sensitive recognition error value corresponding to each template access session data based on the sequence of sensitive operation label confidence levels of each template access session data and the sensitive operation annotation data corresponding to each template access session data; Selecting target template access session data from the multiple template access session data based on the numerical distribution of the sensitive recognition error values corresponding to each template access session data, and optimizing the weight parameters of the multiple deep learning units according to the target template access session data to generate optimized multiple deep learning units; Return to execute loading the multiple template access session data into the optimized multiple deep learning units respectively to generate a sequence of sensitive operation label confidence levels obtained by each template access session data under the optimized multiple deep learning units and target sample access session data after scrambling the corresponding sensitive operations, and perform cyclic access operation sensitive recognition learning until it is detected that the weight parameters of the multiple deep learning units no longer change, generating multiple deep learning units with optimization ended; Performing sensitive operation recognition on the access session data to be analyzed based on the multiple deep learning units with optimization ended to generate sensitive operation recognition data; The determining a sensitive recognition error value corresponding to each template access session data based on the sequence of sensitive operation label confidence levels of each template access session data and the sensitive operation annotation data corresponding to each template access session data includes: Calculating the value of the first mean square error loss function between each sensitive operation label confidence level in the sequence of sensitive operation label confidence levels corresponding to each template access session data and the corresponding sensitive operation annotation data; Adding the determined values of the first mean square error loss function to generate a first target sensitive recognition error value corresponding to each template access session data; Determining a sensitive recognition error value corresponding to each template access session data based on the first target sensitive recognition error value corresponding to each template access session data; The performing sensitive operation recognition on the access session data to be analyzed based on the multiple deep learning units with optimization ended to generate sensitive operation recognition data includes: Loading the access session data to be analyzed into the multiple deep learning units with optimization ended for sensitive operation recognition to generate multiple sensitive operation label confidence levels; Calculating the mean confidence level of the multiple sensitive operation label confidence levels to generate a sensitive operation prediction label corresponding to the access session data to be analyzed.

2. The access processing method for integrating structured and unstructured data according to claim 1, wherein The method further includes: Calculating the information divergence value between every N sensitive operation label confidence levels in the sequence of sensitive operation label confidence levels corresponding to each template access session data; Add the information divergence values to generate a second target sensitive recognition error value corresponding to each template access session data; Determining the sensitive recognition error value corresponding to each template access session data based on the first target sensitive recognition error value corresponding to each template access session data includes: Performing weighted fusion on the first target sensitive recognition error value and the second target sensitive recognition error value corresponding to each template access session data to generate a sensitive recognition error value corresponding to each template access session data.

3. The access processing method for integrating structured and unstructured data according to claim 2, characterized in that The method further includes: Obtain the confidence distribution of the sensitive operation annotation data in the template access session data, and generate corresponding training label features according to the confidence distribution; Calculate the second mean square error loss function value between the training label features and the sensitive operation label confidence sequence corresponding to each template access session data: Add the determined second mean square error loss function values to generate a third target sensitive recognition error value corresponding to each template access session data; Performing weighted fusion on the first target sensitive recognition error value and the second target sensitive recognition error value corresponding to each template access session data to generate a sensitive recognition error value corresponding to each template access session data includes: Performing weighted fusion on the first target sensitive recognition error value, the second target sensitive recognition error value, and the third target sensitive recognition error value corresponding to each template access session data to generate a sensitive recognition error value corresponding to each template access session data.

4. The access processing method for integrating structured and unstructured data according to claim 1, characterized in that Selecting target template access session data from the multiple template access session data based on the numerical distribution of the sensitive recognition error values corresponding to each template access session data includes: Obtain the number of learning rounds for sensitive recognition learning of the cyclic access operation on the multiple deep learning units; Calculate the learning scale of the target template access session data based on the number of learning rounds of the cyclic access operation sensitive recognition learning; Select the template access session data corresponding to the learning scale according to the ascending order of the sensitive recognition error values to generate the target template access session data.

5. The access processing method for integrating structured and unstructured data according to claim 4, wherein Calculating the learning scale of the target template access session data based on the number of learning rounds of the cyclic access operation sensitive recognition learning includes: Obtain the set training optimization parameters, which are configured to optimize the weight parameters of the multiple template access session data; Calculate the proportion of the target template access session data in the multiple template access session data based on the set training optimization parameters and the number of learning rounds of the cyclic access operation sensitive recognition learning; Determine the learning scale of the target template access session data based on the proportion and the number of the multiple template access session data.

6. The access processing method for integrating structured and unstructured data according to claim 1, wherein Performing sensitive operation recognition on the access session data to be analyzed based on the multiple deep learning units after the end of optimization to generate sensitive operation recognition data includes: Obtain the sensitive operation recognition performance indicators of the multiple deep learning units after the end of optimization to generate multiple sensitive operation recognition performance indicators; Sort the multiple sensitive operation recognition performance metrics in descending order, and determine the deep learning unit with the largest sensitive operation recognition performance metric as the target deep learning unit; Load the access session data to be analyzed into the target deep learning unit for sensitive operation recognition, and generate sensitive operation prediction labels corresponding to the access session data to be analyzed.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program, which when executed by a processor implements the access processing method for fusing structured and unstructured data according to any one of claims 1-6.

8. A cloud service data system, characterized in that, The cloud service data system includes a processor and a readable storage medium, and the readable storage medium stores a program, which when executed by the processor implements the access processing method for fusing structured and unstructured data according to any one of claims 1-6.

Images