Trusted computing method and device based on model vector

Abstract

The invention discloses a credible computing method and device based on a model vector, and the method comprises the steps: carrying out the preprocessing operation of to-be-processed data, obtaining a data credible influence factor set and a data time dimension set, the data time dimension set evaluates time factors based on a time decay function; assigning a credible score to each credible influence factor in the data credible influence factor set, and assigning a basic score to the corresponding credible influence factor as an element score based on the influence degree for each time dimension in the data time dimension set; and performing variation coefficient calculation, variation proportion calculation and weighted proportion calculation on each element in the data credible influence element set and the data time dimension set to obtain a final credible score. According to the method, the time attenuation form is perfected into model calculation in a more practical and more reasonable form, so that the business rule is consistent with the actual condition.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to a method and device for calculating the credibility of data calculated by a model engine. Background Art

[0002] With the development of big data technology, data diversification, data complexity, and other complex scenarios have emerged. The original trusted computing method of data from a single data source can no longer meet current business needs. The trusted model algorithm can handle data diversity and data with diverse influencing factors.

[0003] At present, when building a trust model, people tend to calculate the trust score of a single factor. This trust calculation can integrate multiple influencing factors, amplify the discrimination between each factor, and calculate it in a weighted form to unify the standard, so as to calculate the trust score in the form of comprehensive multi-factors.

[0004] As data situations become more and more complex and logic continues to change, the difficulty of distinguishing data and calculating the degree of data credibility increases. This is especially true in the trusted calculation methods that use the time dimension, which are more often calculated in the form of piecewise functions, making it difficult to effectively distinguish small time differences.

[0005] To this end, an invention patent with publication number CN117763496A discloses a method and device for judging attributes and association relationships of multi-source data fusion, which includes a data acquisition module, which is used to fuse data sets generated by multiple data sources to form an entity object data set; the entity object data set includes multiple entity objects, entity object attributes corresponding to each entity object, and association relationships between entity objects; a weight feature determination module, which is used to determine weight features for evaluating the credibility of observation values, wherein the observation values ​​include entity attributes and / or association relationships between entity objects; the weight features include: weight features based on the credibility of data sources, weight features based on the frequency of occurrence of observation values, and weight features based on the time of observation value acquisition; a credibility determination module, which is used to determine the credibility of the observation values ​​according to the weight features. The credibility determination module includes a first weight feature determination unit, which is used to determine the credibility of the data source corresponding to the observation value based on the machine learning algorithm of the Bayesian model; determine the weight feature of the observation value based on the credibility of the data source according to the credibility of the data source; a second weight feature determination unit, which is used to count the occurrence frequency of the observation value in the entity object data set, and divide the weight interval based on the quartile method of multi-data source grouping; determine the weight feature based on the occurrence frequency of the observation value according to the weight interval to which the occurrence frequency belongs; a third weight feature determination unit, which is used to obtain the weight feature based on the observation value acquisition time through a time decay function.

[0006] For another example, an invention patent with publication number CN111125071A discloses a real-time data credibility evaluation method, device, storage medium and process, which outputs a credibility score by preprocessing data and inputting an evaluation model.

[0007] For another example, the article published by Li Gang, Li Tianqi, Cheng Xiaorong et al.: Big Data Credibility Measurement Method ([J]. Computer Engineering and Design, 2017.) mentioned that a hierarchical dynamic big data trust analysis network model is constructed by adding weight parameters such as time factor and penalty factor to calculate the global credibility of the data.

[0008] For another example, the article published by Feng Yanling et al.: Research on Data Credibility Assessment Method Based on Multi-source Heterogeneous Information Fusion ([D]. University of Electronic Science and Technology of China, 2022.) discloses a data credibility assessment method based on multi-source heterogeneous information fusion, which evaluates the credibility of data through standardization and correlation analysis.

[0009] In the above scheme, the evaluation of data credibility is highly dependent on the quality of initial data. If the data quality after preprocessing is not high, the subsequent credibility evaluation results may be greatly affected. The current data credibility evaluation tests are all based on high data quality under ideal conditions. In actual application, the evaluation results of existing technical schemes vary greatly under different measurement conditions, which brings great troubles to the application and promotion of current data credibility evaluation schemes. Summary of the invention

[0010] A brief summary of embodiments of the present invention is provided below to provide a basic understanding of certain aspects of the present invention. It should be understood that the following summary is not an exhaustive summary of the present invention. It is not intended to identify key or important parts of the present invention, nor is it intended to limit the scope of the present invention. Its purpose is merely to provide certain concepts in a simplified form as a prelude to a more detailed description discussed later.

[0011] In order to solve the above technical problems, the present invention provides a new trusted computing method and device. First, by improving the preprocessing scheme, the quality of the initial data is greatly guaranteed; secondly, the time decay form is improved in a more practical and reasonable form to the model calculation, and at the same time, complex and difficult to distinguish data can be separated from the overall data, allowing domain experts to participate in the definition of trusted scores and trusted dimensions, so as to achieve consistency between business rules and actual situations.

[0012] According to one aspect of the present invention, a trusted computing method based on a model vector is provided, comprising:

[0013] Step 1: Perform preprocessing operations on the data to be processed to obtain a feature vector set with time series labels; the preprocessing operations include abnormal data cleaning, outlier removal, light deduplication aggregation, classification and time series division of the data;

[0014] Step 2: Obtain a data credibility influencing factor set and a data time dimension set of the feature vector set outputted in step 1, wherein the data credibility influencing factor set includes multi-dimensional credibility influencing factors that have an impact on the data credibility; the data time dimension set includes a data time dimension corresponding to each credibility influencing factor in the data credibility influencing factor set, and the data time dimension is used to correspond to the degree of influence of the credibility influencing factor;

[0015] The data credibility influencing factor set includes the number of data discoveries and the frequency of data discoveries;

[0016] The data time dimension set evaluates the time factor based on the time decay function, and the evaluation method of the time factor includes: defining the period of data half-decay and the degree of decay as the time decay coefficient;

[0017] Among them, the period of data half-decay is m = 1 / 2 (t / T) , m is the final time decay coefficient, t is the time difference, T is the data half-life period, with the increase of time difference, the decay degree of data will gradually increase, when the time difference t is 0, it means that the data has not decayed; when the time difference t is equal to the data half-life period T, it means that the value or importance of the data is reduced by half, the value range of time decay coefficient m is 0-1, 0 is historical data with no reference significance, 1 is data not affected by time; the algorithm replaces the conventional time segmentation function coefficient with the time decay coefficient, so that the time dimension is more distinguishable, and the small time change can be reflected in the final score, which greatly avoids the situation where the data cannot be distinguished due to the close time or in the same time score segment.

[0018] The data credibility influencing factor set is a combination of credibility influencing factors, and the data time dimension set is a set of time dimensions. The credibility influencing factors are a set of individual factors that have an impact on credibility, and the time dimension acts on each type of individual factor.

[0019] Step 3: Assign a trustworthy score to each trustworthy influencing factor in the data trustworthy influencing factor set, and the trustworthy score is not less than 0 and not greater than 1;

[0020] For each time dimension in the data time dimension set, a basic score is assigned to the corresponding credible influencing factor based on the degree of influence as the factor score, and the decay score is calculated as the time decay dimension. The decay score is calculated by the following formula: decay score = time decay coefficient × factor score. The decay score is the final score of each factor affecting credibility after decay, and is used as the input parameter for the final discrete calculation. Due to the existence of the time decay coefficient, even under the same circumstances, the final time decay score of each type of resource will be different as the discovery time is different, that is, the newer the data, the higher the score.

[0021] Step 4: Calculate the coefficient of variation for each element in the data credibility influencing factor set and the data time dimension set. The coefficient of variation calculation is to calculate the degree of dispersion of each element in the overall data. Coefficient of variation = standard deviation / mean value. In order to amplify the impact of each type of credible influencing factor on the credibility of the data, for each type of credible influencing factor, calculate the degree of dispersion of each single element in the overall data, that is, the coefficient of variation (also known as the coefficient of dispersion). The higher the degree of dispersion, the larger the coefficient, the higher the data credibility distinction, and the greater the impact on data accuracy. According to the same algorithm, calculate the coefficient of variation of each type of credible influencing factor.

[0022] Step 5: Calculation of variation ratio: Calculate the credible ratio of each factor score by the following formula: Credible ratio of factor score = coefficient of variation of each factor / sum of coefficients of variation of each factor; Calculate the coefficient of variation of each type of credible influencing factor;

[0023] Step 6: Calculate weighted proportion: Based on the proportion of the trustworthy scores of each factor calculated in step 5, calculate the proportion of each type of factor in the final trustworthy score (single factor score = trustworthy proportion of factor score × single factor assigned score). The final result is the sum of the weighted scores of each factor result, which is the final trustworthy score. The final trustworthy score is intuitively displayed in the form of 0-1 score, which can clearly reflect the accuracy of the data. Final trustworthy score = the sum of the scores of each single factor.

[0024] Although the prior art also performs preprocessing operations on the data, the current method may not fully consider key steps such as data cleaning, missing value processing and outlier detection, which may affect the accuracy of the final result. And because the trusted computing method is very dependent on the quality of the initial data, the requirements for data preprocessing are relatively high. For this reason, the present invention specially designs a neural network model suitable for credibility calculation, uses the neural network model for preprocessing, and outputs high-quality data to infinitely approach the theoretical credibility evaluation results. At the same time, although the prior art also adds the time dimension element, the prior art has the following defects: only adding the time decay function cannot fully reflect the complexity of data changes over time, especially in terms of data timeliness, and fails to fully consider the impact of data update frequency and timeliness on credibility.

[0025] Specifically, the neural network model includes an improved decision tree for offline use and a first convolutional neural network for real-time detection, wherein the improved decision tree is a decision tree integrated with a second convolutional neural network;

[0026] Among them, the improved decision tree includes a root node and a leaf node, each node includes a second convolutional neural network; the root node is used to roughly classify the data to be processed, and then pass it to the leaf node of the next layer, further identify the roughly classified data to be processed, and remove redundant or irrelevant features of the data to be processed, and finally output the data set to be detected after abnormal data cleaning, outlier removal, and light de-duplication aggregation of the data at the leaf node; the second convolutional neural network is an existing convolutional neural network structure, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The pooling layer adopts the maximum pooling method, and the others are implemented using the general existing convolutional neural network structure. Because the neural network model has the defect of slow detection speed when detecting a large amount of data, the improved decision tree can be used offline. The large amount of data to be detected added every day is regularly processed by the improved decision tree, a large amount of abnormal data is filtered, and normal data is retained (suspected abnormal data can be retained as appropriate), thereby greatly speeding up the data processing speed of the entire model. At the same time, in order to avoid a large amount of data to be tested causing a major change in the structure of the decision tree, the depth of the decision tree is set to 3 or 4, and the pruning parameter range is set to [0.009-0.011] to control the growth of the tree.

[0027] The first convolutional neural network is a two-dimensional convolutional neural network CNN, which includes a first convolutional layer, a first batch normalization layer, a first maximum pooling layer, a second convolutional layer, a second batch normalization layer, a second maximum pooling layer, a third convolutional layer, a third batch normalization layer and a random dropout layer in sequence. The first convolutional layer, the second convolutional layer and the third convolutional layer are all 3×3 convolutional layers, and the first maximum pooling layer and the second maximum pooling layer are 2×2 maximum pooling layers. After the data to be processed is processed by the improved decision tree, a data set to be detected with a label (judgment result) is output, which passes through the first convolutional layer, the first maximum pooling layer, the second convolutional layer, the second maximum pooling layer, the third convolutional layer and the flattening layer in sequence, and finally a feature vector set with a time series label is output.

[0028] The process of training the convolutional neural network CNN is as follows: the data set is input into the first convolution layer, the first convolution layer generates 32 feature maps, and then is sent to the first maximum pooling layer after being processed by the first batch normalization layer, and then is input into the second convolution layer to generate 64 feature maps, and then is processed by the second batch normalization layer and the second maximum pooling layer, and then is input into the third convolution layer to generate 128 feature maps, and the feature maps are flattened into one-dimensional vectors through the third batch normalization layer, random inactivation layer and flattening layer to generate a feature vector set after classification and time series division (which can be designed as different labels). This feature vector set has a time series label, which takes into account the update frequency, timeliness and historical trends of the data to more accurately reflect the influence of time factors; it fully considers the influence of data update frequency and timeliness on credibility, and the labels of this feature vector set include time dimension labels to capture time series dependencies, so as to fully reflect the complexity of data changes over time.

[0029] The first convolutional neural network has been improved by adding a batch normalization layer after each convolution layer and a random dropout layer, Dropout, to reduce overfitting; two maximum pooling layers have been introduced to help the model capture richer features and improve generalization capabilities. The random dropout layer is a regularization technique used to prevent overfitting during the training process of the neural network. Batch normalization can speed up the training process, reduce dependence on initialization, and also help reduce overfitting.

[0030] This neural network model combines the powerful feature extraction capabilities of CNN and the decision logic of the decision tree. When in use, the newly added data to be tested is periodically passed through the improved decision tree of the neural network model to obtain the data set to be tested, and then the data set to be tested is input into the first convolutional neural network to obtain a feature vector set for accurate classification.

[0031] The trusted computing solutions of the prior art are generally only applicable to specific fields. The present invention can be widely used in data prediction in various fields through data preprocessing of the above-mentioned neural network model. At the same time, the coefficient of variation calculation in the prior art cannot fully capture the discreteness and volatility of the data, especially in high-dimensional data or complex data structures, which may lead to deviations in the evaluation results. The present invention overcomes this problem through the preprocessing of the above-mentioned neural network model. In addition, the prior art lacks transparency and explanatory power of the model decision-making process. The present invention greatly facilitates users to understand the source and rationality of the evaluation results through the decision tree in the above-mentioned neural network model.

[0032] According to another aspect of the present invention, there is also provided a trusted computing device based on a model vector, the device being used to execute the above-mentioned trusted computing method, comprising:

[0033] Preprocessing module: used to perform preprocessing operations on the data to be processed and obtain a feature vector set with time series labels; preprocessing operations include abnormal data cleaning, outlier removal, light deduplication aggregation, classification and time series division of data;

[0034] Data processing module: obtaining a data credibility influencing factor set and a data time dimension set of the feature vector set output by the preprocessing module, wherein the data credibility influencing factor set includes multi-dimensional credibility influencing factors that have an impact on the data credibility; the data time dimension set includes a data time dimension corresponding to each credibility influencing factor in the data credibility influencing factor set, and the data time dimension is used to correspond to the degree of influence of the credibility influencing factor;

[0035] Among them, the data half-decay period is m = 1 / 2^(t / T), m is the final time decay coefficient, t is the time difference, and T is the data half-decay period; the value range of the time decay coefficient m is 0-1, 0 is historical data with no reference significance, and 1 is data that is not affected by time;

[0036] Credibility score assignment module: assigns a credibility score to each credibility factor in the data credibility factor set, and the credibility score is not less than 0 and not greater than 1;

[0037] Attenuation score calculation module: For each time dimension in the data time dimension set, a basic score is assigned to the corresponding credible influencing factor based on the degree of influence as the factor score, and the attenuation score is calculated as the time attenuation dimension. The attenuation score is calculated by the following formula: Attenuation score = time attenuation coefficient × factor score;

[0038] Coefficient of variation calculation module: Calculate the coefficient of variation for each element in the data credibility influencing element set and the data time dimension set. The coefficient of variation calculation is to calculate the degree of dispersion of each element in the overall data. Coefficient of variation = standard deviation / mean value;

[0039] Variation weight calculation module: used to calculate the credible weight of each factor score, credible weight of factor score = coefficient of variation of each factor / sum of coefficients of variation of each factor;

[0040] Weighted weight calculation module: Based on the trustworthy weight of each factor calculated by the variation weight calculation module, calculate the proportion of each type of factor under the final trustworthy score. Single factor score = trustworthy weight of factor score × single factor assigned score. The final result is the weighted sum of the results of each factor, which is the final trustworthy score.

[0041] The model vector-based trusted computing device is used to implement the above-mentioned model vector-based trusted computing method.

[0042] The trust calculation method of the present invention utilizes data statistics method, time half-decay method, discrete coefficient method and weighted average method. Based on the above methods, the data measurement comparison dimension can be unified, and the influence of time development on the trustworthiness of the data itself can be amplified. At the same time, it can produce substantial score value distinction for small time changes.

[0043] At the same time, the present invention provides an overall model building framework, which can freely and flexibly adjust the data impact dimension, time decay period and credibility score impact ratio, so as to facilitate the user's accumulation and adjustment during the development and use process.

[0044] In addition, in the present invention, there is no limit on the number of factors affecting the trust score, and they can be freely increased or decreased. Perfect differentiation of dimensions is conducive to the accuracy of the final trust calculation results. The dimensions can include all dimensions that affect the accuracy of the factors, such as time, environment, object characteristics, historical calculation factors, etc., which are related to the final results.

[0045] When in use, as long as the basic scores and influencing conditions of the data credibility influencing factors are assigned to the model, the calculation logic of the final credibility score can be realized, and the final result is intuitively displayed in the form of 0-1 scores, which clearly reflects the accuracy of the data. Therefore, the present invention has good practicality.

[0046] Compared with the prior art, the present invention obtains a higher quality initial data set by using an improved preprocessing scheme of the neural network model, which greatly enhances the evaluation results of the trusted computing method in the actual application process; at the same time, it also has the advantages of being widely used in data prediction in various fields, enhancing the transparency and explanation ability of the model decision-making process, etc. Secondly, the time decay form is improved in a more practical and reasonable form in the model calculation, and at the same time, it can separate complex and difficult to distinguish data from the overall data, allowing field experts to participate in the definition of trusted scores and trusted dimensions, so that business rules are consistent with the actual situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The present invention may be better understood by referring to the following description in conjunction with the accompanying drawings, wherein the same or similar reference numerals are used throughout the drawings to represent the same or similar components. The accompanying drawings, together with the following detailed description, are incorporated in and form a part of this specification and are used to further illustrate the preferred embodiments of the present invention and to explain the principles and advantages of the present invention.

[0048] In the figure:

[0049] Figure 1 This is a model architecture diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0050] Embodiments of the present invention will be described below with reference to the accompanying drawings. Elements and features described in one drawing or one embodiment of the present invention may be combined with elements and features shown in one or more other drawings or embodiments. It should be noted that for the purpose of clarity, the representation and description of components and processes that are not related to the present invention and are known to those of ordinary skill in the art are omitted in the drawings and descriptions.

[0051] In the description of the present invention, it should be understood that the terms “first”, “second” and “third” are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0052] In order to solve the above-mentioned problems of the prior art, an embodiment of the present invention provides a trusted computing method based on a model vector, which can effectively distinguish data and trustworthy score evaluation and scoring, and effectively reflect the impact of the time dimension on the trustworthy score.

[0053] The trusted computing method based on model vectors can be applied to distinguish complex data, including: distinguishing similar data, distinguishing data affected by multiple dimensions and multiple factors. The trusted computing method based on model vectors uses data statistics methods, time half-decay method, discrete coefficient method and weighted average method to unify the data measurement comparison dimensions, and amplify the influence of time development on the credibility of the data itself, while being able to produce substantial score value distinctions for small time changes.

[0054] The embodiment of the present invention also provides a framework for building an overall model of trusted computing, which can freely and flexibly adjust the data impact dimension, time decay period and trust score impact ratio, so as to facilitate the user's accumulation and adjustment during the development and use process.

[0055] The trusted computing method of the present invention can be applied to various fields and various data forms. By performing offline training on the data set of the corresponding field, a useful data set can be output. Then, as long as the data trust influencing factors and data time dimensions are extracted from the useful data set and their basic scores and influencing conditions are assigned to the model, the calculation logic of the final trusted score can be realized. The final result is intuitively displayed in the form of a 0-1 score, which clearly reflects the accuracy of the data.

[0056] Figure 1 Schematic diagram of the structure of the trusted computing method based on model vector provided by the embodiment of the present invention. Figure 1 The trusted computing method based on the model vector of the embodiment of the present invention includes initialization data, data preprocessing, confirmation of influencing factors, definition of time dimension, scoring of influencing factors, time decay, calculation of coefficient of variation, calculation of variation proportion, and calculation of weighted proportion. The initialization data is the data that needs to be trusted for computing. With the enrichment of data resources and the amount of data, the overall trusted computing of data is improved. There is no limit on the number of factors affecting the trusted score, and they can be freely increased or decreased. The perfect distinguishing dimension is conducive to the accuracy of the final trusted computing result.

[0057] The technical solution of the embodiment of the present invention can be based on Figure 1 The model framework of the architecture shown in the example is used for specific implementation. The following is a detailed description of each step of the framework:

[0058] Basic data preparation: Perform basic data processing on data that needs to be trusted: Since excessive abnormal data will affect the trusted computing results to a certain extent, the basic data must be processed, including but not limited to abnormal data cleaning and outlier removal. Perform light deduplication and aggregation on the data to reduce duplicate data and resource waste in the computing process. Then classify the deduplication and aggregation data and divide the time series; this step is achieved through a specially designed neural network model.

[0059] Factor confirmation: including confirmation of factors affecting data credibility and confirmation of data time dimension. Data influencing factors: confirm the impact of each dimension on data credibility, including but not limited to data source, number of data discoveries, frequency of data discoveries, etc. Time dimension confirmation: confirm the degree of influence of time factors on the current data credibility influencing factors, and define the data half-life cycle (m = 1 / 2 (t / T)), m is the final attenuation coefficient, t is the time difference, T is the data half-life period) and the degree of attenuation are used as the time attenuation coefficient. The attenuation coefficient ranges from 0 to 1, where 0 is historical data with no reference significance and 1 is data that is not affected by time. This algorithm replaces the conventional time segmentation function coefficient with the time attenuation coefficient, making the time dimension more distinguishable. Small time changes can be reflected in the final score, which greatly avoids the situation where data cannot be distinguished due to time proximity or being in the same time segment.

[0060] Scoring of influencing factors: For the defined credible influencing factors, a credible score of no less than 0 and no more than 1 is assigned to each single factor based on data conditions, impact level, and professional evaluation.

[0061] Time decay: Based on the basic score given by the influencing factor, add the time decay dimension, decay score = decay coefficient * factor score, and calculate the decay score. Use the same method to calculate the time decay score of each data and each type of single factor. Due to the existence of the time decay coefficient, even under the same circumstances, the total weight of each type of resource will have different time decay scores as the discovery time is different, that is, the newer the data, the higher the score.

[0062] Calculation of dispersion coefficient: In order to magnify the impact of each type of factor on the credibility of the data, for each type of influencing factor, the dispersion degree of each single factor in the overall data is calculated, that is, the coefficient of variation (also known as dispersion coefficient) (coefficient of variation = standard deviation / mean value). The higher the dispersion degree, the larger the coefficient, the higher the credibility of the data, and the greater the impact on the accuracy of the data. According to the same algorithm, the coefficient of variation of each type of credible influencing factor is calculated.

[0063] Calculation of variation ratio: Due to the coefficient of variation, the unit standard is normalized, eliminating the impact of different standards on different dimensions, and now all data are at the same level. Therefore, the credible score ratio of each type of data is calculated (the credible ratio of factor score = the coefficient of variation of a single factor / the sum of the coefficients of variation of all factors). Calculate the credible ratio of the factor score of each type of factor

[0064] Weighted proportion calculation: Based on the calculated proportion of the credibility score of each factor, calculate the proportion of each type of factor in the final credibility score (single factor score = credibility proportion of factor score × single factor assigned score). The final result is the sum of the weighted scores of the results of each factor, that is, the final credibility score (final credibility score = sum of the scores of each single factor).

[0065] Multi-dimensional evaluation: By constructing a set of factors influencing data credibility and a set of data time dimensions, the solution can comprehensively evaluate the credibility of data from multiple perspectives, including factors such as the source of the data, the collection process, and timeliness, which helps to fully reflect the credibility of the data.

[0066] Time factor consideration: Introducing a time decay function to evaluate the time factor can dynamically reflect the credibility of data changes over time, which is very useful for application scenarios that require long-term tracking of data changes.

[0067] Intuitive display of scores: The final credibility score is displayed in the form of 0-1 points, which is intuitive and easy to understand, allowing users to quickly judge the accuracy of the data.

[0068] Coefficient of variation and weighted calculation: Through the calculation of coefficient of variation, variation proportion and weighted proportion, the impact of each factor on data credibility can be evaluated more scientifically, thus improving the accuracy of the evaluation results.

[0069] In the above preprocessing, the embodiment of the present invention specially designs a neural network model suitable for credibility calculation, uses the neural network model for preprocessing, and outputs high-quality data to infinitely approach the theoretical credibility evaluation result. At the same time, although the prior art also adds the time dimension element, the prior art has the following defects: only adding the time decay function cannot fully reflect the complexity of data changes over time, especially in terms of data timeliness, and fails to fully consider the impact of data update frequency and timeliness on credibility.

[0070] Specifically, the neural network model includes an improved decision tree for offline use and a first convolutional neural network (CNN) for real-time detection. The improved decision tree is a decision tree (DT) integrated with a second convolutional neural network, which can be recorded as CNN-DT. The neural network model can be recorded as a convolutional neural network-decision tree-convolutional neural network model (CNN-DT-CNN).

[0071] Among them, the improved decision tree includes a root node and a leaf node, and each node (root node and leaf node) includes a second convolutional neural network; the root node is used to roughly classify the data to be processed (normal, abnormal, suspected abnormal), and then pass it to the leaf node of the next layer, further identify the roughly classified data to be detected, and remove redundant or irrelevant features of the data to be detected, and finally output the abnormal data cleaning, outlier removal, and slightly de-aggregated data to be detected data set at the leaf node; the second convolutional neural network is an existing convolutional neural network structure, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The pooling layer adopts the maximum pooling method, and the others are implemented using the general existing convolutional neural network structure. Because the neural network model has the defect of slow detection speed when detecting a large amount of data, the improved decision tree can be used offline. The large amount of data to be detected added every day is regularly processed by the improved decision tree, a large amount of abnormal data is filtered, and normal data is retained (suspected abnormal data can be retained as appropriate), thereby greatly speeding up the data processing speed of the entire model. At the same time, in order to avoid a large amount of data to be tested causing a major change in the structure of the decision tree, the depth of the decision tree is set to 3 or 4, and the pruning parameter range is set to [0.009-0.011] to control the growth of the tree.

[0072] The first convolutional neural network is a two-dimensional convolutional neural network CNN, which includes a first convolutional layer, a first batch normalization layer, a first maximum pooling layer, a second convolutional layer, a second batch normalization layer, a second maximum pooling layer, a third convolutional layer, a third batch normalization layer and a random dropout layer in sequence, the first convolutional layer, the second convolutional layer and the third convolutional layer are all 3×3 convolutional layers, and the first maximum pooling layer and the second maximum pooling layer are 2×2 maximum pooling layers; the data to be processed is processed by the improved decision tree to output a data set to be detected with a label (judgment result), which passes through the first convolutional layer, the first maximum pooling layer, the second convolutional layer, the second maximum pooling layer, the third convolutional layer and the flattening layer in sequence, and finally outputs a feature vector set with a time series label;

[0073] CNN is a two-dimensional convolutional neural network, including three 3×3 convolutional layers (Conv2D), three batch normalization layers, two 2×2 maximum pooling layers (MixedPooling2D), one random inactivation layer, and one flattening layer (Flatten). The specific structure is shown in Table 1.

[0074] Table 1 The first convolutional neural network structure diagram

[0075]

[0076]

[0077] The process of training the convolutional neural network CNN is as follows: the data set is input into the first convolution layer, the first convolution layer generates 32 feature maps, and then is sent to the first maximum pooling layer after being processed by the first batch normalization layer, and then is input into the second convolution layer to generate 64 feature maps, and then is processed by the second batch normalization layer and the second maximum pooling layer, and then is input into the third convolution layer to generate 128 feature maps, and the feature maps are flattened into one-dimensional vectors through the third batch normalization layer, random inactivation layer and flattening layer to generate a feature vector set after classification and time series division (which can be designed as different labels). This feature vector set has a time series label, which takes into account the update frequency, timeliness and historical trends of the data to more accurately reflect the influence of time factors; it fully considers the influence of data update frequency and timeliness on credibility, and the labels of this feature vector set include time dimension labels to capture time series dependencies, so as to fully reflect the complexity of data changes over time.

[0078] The first convolutional neural network has been improved by adding a batch normalization layer after each convolution layer and a random dropout layer, Dropout, to reduce overfitting; two maximum pooling layers have been introduced to help the model capture richer features and improve generalization capabilities. The random dropout layer is a regularization technique used to prevent overfitting during the training process of the neural network. Batch normalization can speed up the training process, reduce dependence on initialization, and also help reduce overfitting.

[0079] This neural network model combines the powerful feature extraction capabilities of CNN and the decision logic of the decision tree. When in use, the newly added data to be tested is periodically passed through the improved decision tree of the neural network model to obtain the data set to be tested, and then the data set to be tested is input into the first convolutional neural network to obtain a feature vector set for accurate classification.

[0080] Although existing technologies also perform data preprocessing operations, current methods may not fully consider key steps such as data cleaning, missing value processing, and outlier detection, which may affect the accuracy of the final result. And because trusted computing methods are highly dependent on the quality of initial data, they have high requirements for data preprocessing.

[0081] The trusted computing solutions of the prior art are generally only applicable to specific fields. The present invention can be widely used in data prediction in various fields through data preprocessing of the above-mentioned neural network model. At the same time, the coefficient of variation calculation in the prior art cannot fully capture the discreteness and volatility of the data, especially in high-dimensional data or complex data structures, which may lead to deviations in the evaluation results. The present invention overcomes this problem through the preprocessing of the above-mentioned neural network model. In addition, the prior art lacks transparency and explanatory power of the model decision-making process. The present invention greatly facilitates users to understand the source and rationality of the evaluation results through the decision tree in the above-mentioned neural network model.

[0082] The embodiment of the present invention further provides a trusted computing device based on a model vector, the device is used to execute the above-mentioned trusted computing method, and includes:

[0083] Preprocessing module: used to perform preprocessing operations on the data to be processed and obtain a feature vector set with time series labels; preprocessing operations include abnormal data cleaning, outlier removal, light deduplication aggregation, classification and time series division of data;

[0084] Data processing module: obtains a data credibility influencing factor set and a data time dimension set of the feature vector set output by the preprocessing module. The data credibility influencing factor set includes multi-dimensional credibility influencing factors that have an impact on the credibility of the data. The data time dimension is used to confirm the degree of influence of the time factor on the credibility influencing factor;

[0085] Among them, the data half-decay period is m = 1 / 2^(t / T), m is the final time decay coefficient, t is the time difference, and T is the data half-decay period; the value range of the time decay coefficient m is 0-1, 0 is historical data with no reference significance, and 1 is data that is not affected by time;

[0086] Credibility score assignment module: assigns a credibility score to each credibility factor in the data credibility factor set, and the credibility score is not less than 0 and not greater than 1;

[0087] Attenuation score calculation module: For each time dimension in the data time dimension set, a basic score is assigned to the corresponding credible influencing factor based on the degree of influence as the factor score, and the attenuation score is calculated as the time attenuation dimension. The attenuation score is calculated by the following formula: Attenuation score = time attenuation coefficient × factor score;

[0088] Coefficient of variation calculation module: Calculate the coefficient of variation for each element in the data credibility influencing element set and the data time dimension set. The coefficient of variation calculation is to calculate the degree of dispersion of each element in the overall data. Coefficient of variation = standard deviation / mean value;

[0089] Variation weight calculation module: used to calculate the credible weight of each factor score, credible weight of factor score = coefficient of variation of each factor / sum of coefficients of variation of each factor;

[0090] Weighted proportion calculation module: Based on the credible score proportion of each factor calculated by the variation proportion calculation module, calculate the proportion of each type of factor in the final credible score. The score of a single factor = the credible proportion of the factor score × the score assigned to a single factor. The final result is the sum of the weighted scores of the results of each factor, which is the final credible score.

[0091] The present invention obtains a higher quality initial data set by using an improved preprocessing scheme of the neural network model, which greatly enhances the evaluation results of the trusted computing method in the actual application process; at the same time, it also has the advantages of being widely used in data prediction in various fields, enhancing the transparency and explanation ability of the model decision-making process, etc. Secondly, the time decay form is improved in a more practical and reasonable form in the model calculation, and at the same time, it can separate complex and difficult to distinguish data from the overall data, allowing field experts to participate in the definition of trusted scores and trusted dimensions, so that business rules are consistent with actual situations.

[0092] It should be emphasized that the method of the present invention is not limited to being executed in the time sequence described in the specification, and can also be executed in other time sequences, in parallel or independently. Therefore, the execution order of the method described in this specification does not limit the technical scope of the present invention.

[0093] Although the present invention has been disclosed above by describing specific embodiments of the present invention, it should be understood that all the above embodiments and examples are exemplary rather than restrictive. Those skilled in the art may design various modifications, improvements or equivalents of the present invention within the spirit and scope of the appended claims. These modifications, improvements or equivalents should also be considered to be included in the protection scope of the present invention.

Claims

1. A trusted computing method based on a model vector, characterized in that: include: Step 1: Preprocess the data to be processed to obtain a feature vector set with time series labels; Preprocessing operations include abnormal data cleaning, outlier removal, light deduplication aggregation, classification, and time series segmentation of data; Step 2: Obtain a data credibility influencing factor set and a data time dimension set of the feature vector set outputted in step 1, wherein the data credibility influencing factor set includes multi-dimensional credibility influencing factors that have an impact on the data credibility; the data time dimension set includes a data time dimension corresponding to each credibility influencing factor in the data credibility influencing factor set, and the data time dimension is used to correspond to the degree of influence of the credibility influencing factor; The data time dimension set evaluates the time factor based on the time decay function, and the evaluation method of the time factor includes: defining the period of data half-decay and the degree of decay as the time decay coefficient; Among them, the period of data half-decay is m = 1 / 2 (t / T) , m is the final time decay coefficient, t is the time difference, and T is the data half-life period; the value range of the time decay coefficient m is 0-1, 0 is historical data with no reference significance, and 1 is data that is not affected by time; Step 3: Assign a trustworthy score to each trustworthy influencing factor in the data trustworthy influencing factor set, and the trustworthy score is not less than 0 and not greater than 1; For each data time dimension in the data time dimension set, a basic score is assigned to the corresponding credible influencing factor based on the degree of influence as the factor score, and a decay score is calculated as the time decay dimension. The decay score is calculated by the following formula: decay score = time decay coefficient × factor score; Step 4: Calculate the coefficient of variation for each element in the data credibility influencing element set and the data time dimension set. The coefficient of variation calculation is to calculate the degree of dispersion of each element in the overall data. The coefficient of variation = standard deviation / mean value; Step 5: Calculate the credible weight of each factor score: credible weight of factor score = coefficient of variation of each factor / sum of coefficients of variation of each factor; Step 6: Calculate weighted proportion: Based on the credible score proportion of each factor calculated in step 5, calculate the proportion of each type of credible influencing factor in the final credible score. The score of a single factor = the credible proportion of the factor score × the score assigned to the single factor. The final result is the weighted sum of the results of each factor.

2. The trusted computing method based on model vector according to claim 1, characterized in that: In the step 1, preprocessing operation is performed on the data to be processed, which is specifically implemented through a neural network model; The neural network model includes an improved decision tree for offline use and a first convolutional neural network for real-time detection; Among them, the improved decision tree includes a root node and a leaf node, each node includes a second convolutional neural network; the root node is used to classify the data to be processed, and then pass it to the leaf node of the next layer to further identify the classified data to be processed, and at the same time remove redundant or irrelevant features of the data to be detected, and finally output the data set to be detected at the leaf node; the second convolutional neural network is an existing convolutional neural network structure, which includes an input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer, and the pooling layer adopts the maximum pooling method.

3. The trusted computing method based on model vector according to claim 2, characterized in that: The specific structure of the first convolutional neural network is shown in Table 1: Table 1 The first convolutional neural network structure diagram The first convolutional neural network includes a first convolutional layer, a first batch normalization layer, a first maximum pooling layer, a second convolutional layer, a second batch normalization layer, a second maximum pooling layer, a third convolutional layer, a third batch normalization layer and a random inactivation layer; the first convolutional layer, the second convolutional layer and the third convolutional layer are all 3×3 convolutional layers, and the first maximum pooling layer and the second maximum pooling layer are 2×2 maximum pooling layers; the data to be processed is processed by the improved decision tree to output a labeled data set to be detected, which passes through the first convolutional layer, the first maximum pooling layer, the second convolutional layer, the second maximum pooling layer, the third convolutional layer and the flattening layer in sequence, and finally outputs a feature vector set with time series labels.

4. The trusted computing method based on model vector according to claim 2, characterized in that: The depth of the improved decision tree is 3 or 4, and the pruning parameter range is set to [0.009-0.011] to control the growth of the tree.

5. The trusted computing method based on model vector according to claim 1, characterized in that: In step 1, the data credibility influencing factor set includes the number of data discoveries and the frequency of data discoveries.

6. A trusted computing device based on a model vector, characterized in that: include: Preprocessing module: used to perform preprocessing operations on the data to be processed and obtain a feature vector set with time series labels; Preprocessing operations include abnormal data cleaning, outlier removal, light deduplication aggregation, classification, and time series segmentation of data; Data processing module: used to obtain a data credibility influencing factor set and a data time dimension set of the feature vector set output by the preprocessing module, wherein the data credibility influencing factor set includes multi-dimensional credibility influencing factors that have an impact on the data credibility; the data time dimension set includes a data time dimension corresponding to each credibility influencing factor in the data credibility influencing factor set, and the data time dimension is used to correspond to the degree of influence of the credibility influencing factor; Among them, the period of data half-decay is m = m = 1 / 2 (t / T) , m is the final time decay coefficient, t is the time difference, and T is the data half-life period; the value range of the time decay coefficient m is 0-1, 0 is historical data with no reference significance, and 1 is data that is not affected by time; Credibility score assignment module: assigns a credibility score to each credibility factor in the data credibility factor set, and the credibility score is not less than 0 and not greater than 1; Attenuation score calculation module: For each time dimension in the data time dimension set, a basic score is assigned to the corresponding credible influencing factor based on the degree of influence as the factor score, and the attenuation score is calculated as the time attenuation dimension. The attenuation score is calculated by the following formula: Attenuation score = time attenuation coefficient × factor score; Coefficient of variation calculation module: Calculate the coefficient of variation for each element in the data credibility influencing element set and the data time dimension set. The coefficient of variation calculation is to calculate the degree of dispersion of each element in the overall data. Coefficient of variation = standard deviation / mean value; Variation weight calculation module: used to calculate the credible weight of each factor score, credible weight of factor score = coefficient of variation of each factor / sum of coefficients of variation of each factor; Weighted weight calculation module: Based on the trustworthy score ratio of each factor calculated by the variation weight calculation module, calculate the proportion of each type of factor in the final trustworthy score, the score of a single factor = the trustworthy ratio of the factor score × the score assigned to a single factor, and the final result is the sum of the weighted scores of the results of each factor, which is the final trustworthy score; the trusted computing device based on the model vector is used to implement the trusted computing method based on the model vector as described in any one of claims 1-5.

Images