AI application risk identification method and system based on big data

Abstract

The invention discloses an AI application risk identification method and system based on big data. The method comprises the steps of multi-source data acquisition, multi-field application feature selection, AI application risk identification model establishment and application risk intelligent identification. The invention relates to the technical field of data processing, in particular to an AI application risk identification method and system based on big data, according to the scheme, group recommendation and personalized optimization are innovatively combined, and integrated compatibility of single-field and multi-field application risk identification is achieved; an improved clustering algorithm based on two-way neighbor equilibrium density calculation, a two-factor collaborative optimization strategy and a two-way neighbor dynamic weight distribution strategy is innovatively provided, and the accuracy of an application risk identification output result is improved. And an optimization algorithm for obtaining an optimal parameter combination of the recognition model is improved by adopting a habitat position selection strategy and a food attraction strategy, so that the accuracy of a model output result is remarkably improved, and precise recognition of application risks is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data processing, and specifically relates to an AI application risk identification method and system based on big data. BACKGROUND

[0002] The AI application risk identification method and system is based on big data technology and artificial intelligence algorithms, collects, analyzes and processes multi-source data of AI applications, monitors and evaluates risks that may exist in actual use scenarios of AI applications in real time, comprehensively monitors risks of AI applications by using big data analysis means, identifies potential risk factors, and takes preventive measures in time after identification to ensure the stability, security and reliability of AI applications.

[0003] However, the traditional AI application risk identification method cannot simultaneously meet the risk feature screening requirements of single-field AI applications and cross-field comprehensive AI applications in feature selection, leading to a technical problem that the screened risk features do not match the actual risk patterns, thereby directly affecting the accuracy of the AI application risk identification result; the existing clustering method suitable for application risk feature selection has the technical problems of density imbalance, bridge noise sensitivity and low distribution fault tolerance, leading to insufficient accuracy and robustness of the clustering result, thereby affecting the accuracy of the AI application risk identification output result; the existing model suitable for AI application risk identification has the technical problem that improper built-in parameter setting leads to inaccurate identification model output result. SUMMARY

[0004] In view of the above, in order to overcome the defects of the prior art, the present application provides an AI application risk identification method and system based on big data, which is aimed at the technical problem that the traditional AI application risk identification method cannot simultaneously meet the risk feature screening requirements of single-field AI application and cross-field comprehensive AI application in feature selection, leading to mismatch between the screened risk features and the actual risk patterns, thereby directly affecting the accuracy of the AI application risk identification result. The present scheme innovatively selects risk features through a clustering algorithm, dynamically integrates the optimal application risk feature set according to the clustering output results combined with the field attributes, without the need to separately design a feature selection system for various AI application risk identification scenarios, significantly enhances the adaptability of various AI application risk feature selection, improves the accuracy of the AI application risk identification result, realizes the integration and compatibility of single-field and multi-field AI application risk identification, and effectively improves the intelligent level of various AI application risk identification. In view of the technical problems of density imbalance, bridge noise sensitivity and low fault tolerance in the existing clustering method suitable for application risk feature selection, which leads to insufficient accuracy and robustness of the clustering results, thereby affecting the accuracy of the AI application risk identification output result, the present scheme innovatively proposes an improved clustering algorithm based on a balanced density calculation of a bidirectional neighbor, a double-factor collaborative optimization strategy and a bidirectional neighbor dynamic weight distribution strategy, which solves the problem of density imbalance, optimizes the distinguishability of the cluster center, significantly enhances the stability and fault tolerance of the algorithm in complex data, improves the accuracy of the clustering results, and thereby improves the accuracy of the AI application risk identification output result. In view of the technical problem that the existing model suitable for AI application risk identification has improper built-in parameter settings, leading to inaccurate identification model output results, the present scheme uses a habitat location selection strategy and a food attraction strategy to improve the optimization algorithm for obtaining the optimal parameter combination of the identification model, significantly improves the hyperparameter optimization efficiency of the identification model, reduces the computational cost of model training, and improves the accuracy of the identification model output results, thereby realizing intelligent and accurate identification of AI application risks.

[0005] The technical solutions adopted by the present application are as follows: The present application provides an AI application risk identification method based on big data, which comprises the following steps:

[0006] Step S1: Multi-source data acquisition;

[0007] Step S2: Multi-field application feature selection;

[0008] Step S3: Establishing an AI application risk identification model;

[0009] Step S4: Application risk intelligent identification;

[0010] Further, in step S1, the multi-source data acquisition is used to acquire data related to AI application risk identification, specifically by collecting data information from the risk identification platform to obtain application risk identification original data, and performing data optimization processing on the application risk identification original data to obtain application risk identification optimized data; the application risk identification original data includes reference application risk identification data and target application risk identification data; the data optimization processing specifically includes data cleaning and data standardization of the original data to obtain application risk identification optimized data.

[0011] Further, in step S2, the multi-field application feature selection specifically includes the following steps:

[0012] Step S21: constructing a feature selection clustering algorithm, specifically including the following steps:

[0013] Step S211: constructing a bidirectional near neighbor matrix, specifically based on the risk feature set, calculating the distance between any two risk features through the Euclidean distance formula, and for each risk feature , filtering the k nearest features through the KD-Tree algorithm to form the k-nearest neighbor set of , constructing a risk feature k-nearest neighbor matrix according to the k-nearest neighbor set of all risk features, then traversing all bidirectional near neighbor pairs in the risk feature k-nearest neighbor matrix, if , then and are bidirectional near neighbor pairs, based on all filtered bidirectional near neighbor pairs between risk features, constructing a risk feature bidirectional near neighbor matrix, wherein, represents the i-th risk feature, represents the j-th risk feature, represents the k-nearest neighbor set of ;

[0014] Step S212: calculating the balanced density, specifically based on the risk feature k-nearest neighbor matrix and the risk feature bidirectional near neighbor matrix, calculating the balanced density value of each risk feature, and performing normalization processing to obtain the normalized balanced density value ; the formula is as follows:

[0015] ;

[0016] In the formula, represents the balanced density value of the i-th risk feature, represents the Euclidean distance between and its k-nearest neighbor feature , represents the bidirectional near neighbor number of , represents The bidirectional nearest neighbor set, where k represents the number of nearest neighbors;

[0017] Step S213: Calculate the minimum equilibrium density of the path, specifically by first calculating the manifold distance between features. And for each pair of risk characteristics All connection paths are sorted by manifold distance. Sort the paths from smallest to largest, and select the path with the shortest distance as the optimal connection path. The normalized equilibrium density values ​​of all risk characteristics along the optimal connection path are calculated, and the minimum normalized equilibrium density value is selected as the minimum equilibrium density of the optimal connection path; the formula used is as follows:

[0018] ;

[0019] In the formula, Represents the i-th risk feature and the j-th risk feature The manifold distance between them express and All connecting paths between them, that is, all combinations of paths indirectly related through other risk characteristics. Indicates the connection path. Indicates the length of the connection path. Indicates the adjacent risk characteristics on path p and The Euclidean distance between them represents the manifold distance adjustment factor, and t represents the index variable of the features on the path;

[0020] Step S214: Relative distance comprehensive calculation, specifically, the relative distance is calculated using a two-factor collaborative optimization strategy of interaction similarity factor and path contribution compensation factor; including the following steps:

[0021] Step S2141: Calculate the interaction similarity factor, specifically by calculating the interaction similarity factor value based on the number of shared nearest neighbor features between features; the formula used is as follows:

[0022] ;

[0023] In the formula, express and Interaction similarity factor between express and The number of shared nearest neighbor features;

[0024] Step S2142: Calculate the path contribution compensation factor, specifically by calculating the path contribution compensation factor value based on the relationship between the minimum equilibrium density of the optimal connection path and the feature densities at both ends; the formula used is as follows:

[0025] ;

[0026] In the formula, express and Path contribution compensation factor between express and The minimum equilibrium density value of the optimal connection path between them;

[0027] Step S2143: Calculate the optimized manifold distance, specifically by multiplying the manifold distance by the interaction similarity factor and the path contribution compensation factor to obtain the optimized manifold distance value;

[0028] Step S2144: Calculate the relative distance value of risk features. Specifically, for each risk feature, first determine whether it is the risk feature with the highest normalized equilibrium density value among all risk features. Then, calculate the relative distance value based on the determination result. The formula used is as follows:

[0029] ;

[0030] In the formula, express The relative distance value, This represents the normalized equilibrium density value of the v-th risk feature. Index variables representing risk characteristics, express and Optimized manifold distance between them;

[0031] Step S215: Cluster center selection, specifically, selecting risk features that simultaneously meet the contribution threshold condition and the difference threshold condition as cluster centers to obtain a set of feature cluster centers;

[0032] Step S216: Core risk feature allocation, specifically, first calculating the density threshold. Next, the cluster centers are placed into the initialization queue Q, and a unique cluster label is assigned to each cluster center. Then, the label propagation process is performed. When the queue Q is not empty, the feature of the head of the queue is retrieved. The risk features are traversed through the bidirectional nearest neighbor matrix of the risk features. ,like Unassigned labels and their normalized equilibrium density values , then bestowing Same cluster label, and Add to queue Q, and repeat the above process until the queue is empty, thus completing the cluster assignment of the core risk features; the formula used is as follows:

[0033] ;

[0034] In the formula, This represents the mean of the normalized equilibrium density values ​​of all risk characteristics. The standard deviation of the normalized equilibrium density values ​​representing all risk characteristics;

[0035] Step S217: Residual risk feature allocation, specifically, dynamically allocating residual risk features based on a bidirectional nearest neighbor dynamic weight allocation strategy; including the following steps:

[0036] Step S2171: Calculate the bidirectional nearest neighbor ratio weights using the following formula:

[0037] ;

[0038] In the formula, express and Bidirectional nearest neighbor ratio weights between them This represents the m-th risk characteristic. express and The manifold distance value, This represents the l-th risk characteristic. express and The manifold distance value, express bidirectional nearest neighbor set, express and Bidirectional nearest neighbor ratio weights between them;

[0039] Step S2172: Fuzzy membership calculation, specifically, the risk features that have been assigned to the c-th cluster among the k nearest neighbors of the risk feature are multiplied by their normalized proportional weights and bidirectional nearest neighbor proportional weights and then summed.

[0040] The normalized proportional weight specifically refers to the relative weight proportion of the bidirectional nearest neighbor proportional weight in all k nearest neighbor sets of the risk feature; the formula used is as follows:

[0041] ;

[0042] In the formula, Indicates risk characteristics The fuzzy membership value belonging to the c-th cluster. express and Normalized proportional weights between them express It belongs to the c-th cluster;

[0043] Step S2173: Residual feature allocation, specifically for each residual feature Calculate its fuzzy membership degree among all clusters, and then... Assign to the cluster corresponding to the maximum fuzzy membership degree, then update. The membership degree of all risk features in the k nearest neighbor set is calculated, and this process is repeated until all remaining features are assigned. If there are still unassigned features, they are sorted in descending order of normalized equilibrium density value, and each feature is assigned to the cluster of the risk feature with the nearest Euclidean distance.

[0044] Step S218: Clustering results are generated, specifically after all risk features have been assigned, multiple independent clusters are formed, each cluster corresponding to a set of risk features for a domain AI application;

[0045] Step S22: Real-time clustering result identification, specifically, inputting the reference application risk identification data and the target application risk identification data from the application risk identification optimization data into the feature selection clustering algorithm to obtain the real-time risk feature selection clustering result;

[0046] Step S23: Obtain the optimal application risk feature set. Specifically, for each cluster, calculate the Euclidean distance between all risk features within the cluster and the cluster center, sort them in ascending order according to the Euclidean distance, and select the top-ranked features within each cluster. The characteristics are used as candidate application risk features, and the optimal application risk feature set is determined according to the domain attributes of the AI ​​application. If the AI ​​application is a single domain application, the candidate application risk features selected from the clusters corresponding to that domain are used as the optimal application risk feature set. If the AI ​​application is a multi-domain application, the candidate application risk features selected from each cluster are merged as the optimal application risk feature set.

[0047] Furthermore, in step S3, establishing the AI ​​application risk identification model specifically includes the following steps:

[0048] Step S31: Construct the AI ​​application risk identification model architecture, specifically by establishing an AI application risk identification model based on a deep neural network (DNN).

[0049] Step S32: Preliminary training of the identification model, specifically, using the optimal application risk feature set as the training data, and using the preset benchmark model hyperparameters to perform preliminary model training on the AI ​​application risk identification model, to obtain the preliminary trained AI application risk identification model;

[0050] Step S33: Identify model hyperparameter optimization, specifically by obtaining the optimal hyperparameter combination of the identification model through an improved optimization algorithm; including the following steps:

[0051] Step S331: Initial population generation, specifically, encoding the hyperparameters of the AI ​​application risk identification model into individual location vectors, and randomly initializing and generating populations based on the upper and lower boundaries of the hyperparameters. The initial population is obtained by analyzing the position vectors of each individual.

[0052] Step S332: Determine the global optimal position. Specifically, calculate the fitness value of individuals in the population, use the performance of the AI ​​application risk identification model after preliminary training based on the individual position as the individual fitness value, sort the individuals according to their fitness values ​​from best to worst, and determine the position vector of the individual ranked first as the initial global optimal position.

[0053] Step S333: Habitat exploration behavior stage, specifically when an individual's exploration factors... In this case, the individual's location is updated during the habitat exploration phase using a habitat location selection strategy; the formula used is as follows:

[0054] ;

[0055] ;

[0056] In the formula, Indicates the globally optimal position. This represents the optimal position of the current individual across all iterations. Indicates the reference location of the habitat. Indicates the overall quality score of the habitat. Indicates that the i-th individual is in the i-th position. The position of the j-th dimension in the next iteration of the population. Indicates that the i-th individual is in the i-th position. The position in the j-th dimension of the population in the next iteration, where e represents the current search iteration number. Indicates the maximum number of search iterations. , , , , and They all said Randomness parameters within a certain range;

[0057] Step S334: Individual foraging behavior development stage, specifically when the individual explores factors At this time, the individual's location is updated during the foraging development stage through food attraction strategies; the formula used is as follows:

[0058] ;

[0059] ;

[0060] ;

[0061] ;

[0062] In the formula, Indicates the location of the food. and Let these represent the upper and lower bounds of the j-th dimension, respectively. This represents the position of the globally optimal individual in the j-th dimension. Indicates that the i-th individual is in the i-th position. The food attraction factor in the next iteration Indicates that the i-th individual is in the i-th position. Position in the next iteration of the population This indicates the probability of the individual being searched for food. Indicates the size of the food. Indicates the number of dimensions. and express Randomness parameters within a range and They all said Randomness parameters within a certain range;

[0063] Step S335: The iterative search terminates. Specifically, for all updated individuals, the individual fitness values ​​are recalculated, and the fitness values ​​of all individuals are compared to update the global optimal position. When the global optimal position is higher than the fitness threshold and the number of search iterations reaches the maximum number of search iterations, the search is terminated and the global optimal position is obtained. The global optimal position specifically refers to the optimal combination of hyperparameters of the recognition model.

[0064] Step S34: Final training of the identification model. Specifically, based on the optimal hyperparameter combination of the identification model, the hyperparameters of the AI ​​application risk identification model after initial training are adjusted, and the reference optimal application risk feature set is used as training data to retrain the AI ​​application risk identification model after initial training to obtain the AI ​​application risk identification model with the best performance.

[0065] Furthermore, in step S4, the intelligent identification of application risks specifically involves inputting the target optimal application risk feature set into the best-performing AI application risk identification model to obtain the real-time application risk level result. Based on the real-time application risk level result, the corresponding security response strategy is triggered to achieve real-time monitoring and intelligent identification of AI application risks.

[0066] The technical solution adopted by the present invention is as follows: The present invention provides an AI application risk identification system based on big data, including a multi-source data acquisition module, a multi-domain application feature selection module, an AI application risk identification model establishment module, and an application risk intelligent identification module;

[0067] The multi-source data acquisition module specifically obtains application risk identification optimization data through data information collection and data optimization processing, and sends the data to the multi-domain application feature selection module.

[0068] The multi-domain application feature selection module receives data sent by the multi-source data acquisition module, improves the clustering algorithm by using a balanced density calculation based on bidirectional nearest neighbors, a two-factor collaborative optimization strategy, and a bidirectional nearest neighbor dynamic weight allocation strategy, thereby completing the construction of the feature selection clustering algorithm. Then, the data is input into the feature selection clustering algorithm, and the real-time risk feature selection clustering result is output. Based on the clustering result, candidate application risk features are selected. Finally, the features are integrated according to the attributes of the AI ​​application domain to obtain the optimal application risk feature set. The data is then sent to the AI ​​application risk identification model building module and the application risk intelligent identification module.

[0069] The AI ​​application risk identification model module receives data sent by the multi-domain application feature selection module, establishes an AI application risk identification model based on a deep neural network, and sequentially performs preliminary training of the identification model, hyperparameter optimization of the identification model using an improved optimization algorithm, and final training of the identification model to obtain the AI ​​application risk identification model with optimal performance, and sends the data to the application risk intelligent identification module.

[0070] The application risk intelligent identification module receives data from the multi-domain application feature selection module and the AI ​​application risk identification model establishment module. Specifically, it inputs the target optimal application risk feature set into the optimal model and outputs the risk level result in real time, thereby realizing real-time monitoring of AI application risks.

[0071] The beneficial effects achieved by the present invention using the above solution are as follows:

[0072] (1) In view of the technical problem that traditional AI application risk identification methods cannot simultaneously meet the risk feature selection requirements of single-domain AI applications and cross-domain comprehensive AI applications, resulting in the selected risk features not matching the actual risk patterns, thus directly affecting the accuracy of AI application risk identification results, this solution innovatively selects risk features through clustering algorithms. Based on the clustering output results and domain attributes, it dynamically integrates the optimal application risk feature set, eliminating the need to design a separate feature selection system for various AI application risk identification scenarios. It also significantly enhances the adaptability of various AI application risk feature selections, improves the accuracy of AI application risk identification results, realizes the integrated compatibility of single-domain and multi-domain AI application risk identification, and effectively improves the intelligence level of various AI application risk identification.

[0073] (2) In view of the technical problems of density imbalance, bridge noise sensitivity and low allocation fault tolerance in the existing clustering methods applicable to the selection of application risk features, which lead to insufficient accuracy and robustness of clustering results and thus affect the accuracy of AI application risk identification output results, this solution innovatively proposes an improved clustering algorithm based on bidirectional nearest neighbor balanced density calculation, two-factor collaborative optimization strategy and bidirectional nearest neighbor dynamic weight allocation strategy. This solves the problem of density imbalance, optimizes the distinguishability of cluster centers, significantly enhances the stability and fault tolerance of the algorithm in complex data, improves the accuracy of clustering results, and thus improves the accuracy of AI application risk identification output results.

[0074] (3) In view of the technical problem that the existing models applicable to AI application risk identification have improper built-in parameter settings, resulting in inaccurate identification model output results, this solution adopts habitat location selection strategy and food attraction strategy to improve the optimization algorithm for obtaining the optimal parameter combination of the identification model. The optimal parameter combination of the model is obtained, which significantly improves the hyperparameter optimization efficiency of the identification model, reduces the computational cost of model training, and improves the accuracy of the identification model output results, thereby realizing the intelligent and accurate identification of AI application risks. Attached Figure Description

[0075] Figure 1 A flowchart illustrating a big data-based AI application risk identification method provided by this invention;

[0076] Figure 2 A schematic diagram of a big data-based AI application risk identification system provided by the present invention;

[0077] Figure 3 A flowchart illustrating the multi-domain application feature selection process in step S2;

[0078] Figure 4A flowchart illustrating the process of establishing an AI application risk identification model for step S3;

[0079] Figure 5 A flowchart illustrating the process of constructing the feature selection clustering algorithm for step S21;

[0080] Figure 6 A flowchart illustrating the process of identifying and optimizing model hyperparameters in step S33;

[0081] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0082] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0083] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0084] Example 1, see Figure 1 This invention provides a method for identifying risks in AI applications based on big data, which includes the following steps:

[0085] Step S1: Multi-source data acquisition, specifically, obtaining application risk identification optimization data through data information collection and data optimization processing;

[0086] Step S2: Multi-domain application feature selection. Specifically, the clustering algorithm is improved by using balanced density calculation based on bidirectional nearest neighbors, a two-factor collaborative optimization strategy, and a bidirectional nearest neighbor dynamic weight allocation strategy to complete the construction of the feature selection clustering algorithm. Then, the data is input into the feature selection clustering algorithm, and the real-time risk feature selection clustering results are output. Candidate application risk features are selected based on the clustering results. Finally, the features are integrated according to the attributes of the AI ​​application domain to obtain the optimal application risk feature set.

[0087] Step S3: Establish an AI application risk identification model. Specifically, this involves establishing an AI application risk identification model based on a deep neural network, and then sequentially performing initial training of the identification model, hyperparameter optimization of the identification model using an improved optimization algorithm, and final training of the identification model to obtain the AI ​​application risk identification model with optimal performance.

[0088] Step S4: Apply intelligent risk identification, specifically by inputting the target optimal application risk feature set into the optimal model and outputting the risk level result in real time, thereby realizing real-time monitoring of AI application risks.

[0089] By performing the above operations, this solution addresses the technical problem that traditional AI application risk identification methods cannot simultaneously meet the risk feature selection needs of both single-domain and cross-domain integrated AI applications, leading to a mismatch between the selected risk features and the actual risk patterns, thus directly affecting the accuracy of AI application risk identification results. This solution innovatively uses a clustering algorithm for risk feature selection. Based on the clustering output and domain attributes, it dynamically integrates the optimal application risk feature set. This eliminates the need to design separate feature selection systems for various AI application risk identification scenarios, significantly enhances the adaptability of risk feature selection for various AI applications, improves the accuracy of AI application risk identification results, achieves integrated compatibility for single-domain and multi-domain AI application risk identification, and effectively improves the intelligence level of risk identification for various AI applications.

[0090] Example 2, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S1, the multi-source data acquisition is used to acquire data related to AI application risk identification. Specifically, it involves collecting data information from the risk identification platform to obtain raw application risk identification data, and performing data optimization processing on the raw application risk identification data to obtain optimized application risk identification data.

[0091] The raw data for application risk identification includes reference application risk identification data and target application risk identification data;

[0092] The data optimization process is used to construct a high-quality, structured input dataset; specifically, it involves data cleaning and data standardization of the original data to obtain optimized data for application risk identification.

[0093] Both the reference application risk identification data and the target application risk identification data include application user behavior data, application log data, application attribute data, and application user feedback data.

[0094] The application user behavior data includes the number of sensitive words in the questions, the topic classification of the questions, the length of the questions, the feedback data on the answers, and the frequency of the keywords in the questions.

[0095] The application log data includes application error logs and application runtime logs;

[0096] The application's own attribute data includes application domain, application response time, application throughput, application latency, and application update frequency;

[0097] The user problem description data includes the number of sensitive words, problem topic category, problem length, and keyword frequency;

[0098] The application areas include finance, healthcare, IT, e-commerce, entertainment, agriculture, and comprehensive applications.

[0099] The fields of finance, healthcare, IT, e-commerce, entertainment, and agriculture are defined as single application areas, and their names are used as cluster labels in the clustering.

[0100] The reference application risk identification data also includes historical application risk levels;

[0101] The application risk levels include high risk, medium risk, low risk, and no risk;

[0102] The data cleaning process is used to remove noise, erroneous records, and invalid information from the original data. Specifically, it involves imputing missing values ​​and removing outliers from the original structured data. The imputation of missing values ​​is specifically done by filling in missing values ​​using the mean imputation method. The removal of outliers is specifically done by detecting and removing extreme values ​​and logical outliers in the original data using the Z-Score algorithm.

[0103] The data standardization process is used to unify the dimensions of structured numerical data. Specifically, it maps all continuous variables to the [0,1] interval using the min-max normalization method and encodes the category fields in the original data using the one-hot encoding method, converting discrete text and label variables into sparse numerical vectors.

[0104] Example 3, see Figure 1 , Figure 2 , Figure 3 and Figure 5 This embodiment is based on the above embodiment. In step S2, the multi-domain application feature selection is used to select the features that best identify application risks from the original data of AI application risk identification in different domains. Specifically, it involves constructing a feature selection clustering algorithm, identifying real-time clustering results, and obtaining a risk feature set to obtain the optimal application risk feature set. This includes the following steps:

[0105] Step S21: Construct a feature selection clustering algorithm to cluster data from different AI applications, grouping similar risk features from different fields into different clusters. This includes the following steps:

[0106] Step S211: Construct a bidirectional nearest neighbor matrix to quantify the risk identification correlation between risk features. The nearest neighbor matrix focuses on locally similar groups of features, providing a core basis for balanced density calculation and label propagation. Specifically, based on the risk feature set, the distance between any two risk features is calculated using the Euclidean distance formula, and for each risk feature... The KD-Tree algorithm is used to select the k nearest features to form a group. k-nearest neighbor set Construct a risk feature k-nearest neighbor matrix based on the k-nearest neighbor set of all risk features. Then, iterate through all bidirectional nearest neighbor pairs in the risk feature k-nearest neighbor matrix. ,but and As bidirectional nearest neighbor pairs, a bidirectional nearest neighbor matrix for risk features is constructed based on all selected bidirectional nearest neighbor pairs, where, This represents the i-th risk characteristic. This represents the j-th risk characteristic. express The set of k nearest neighbors;

[0107] Specifically, the risk feature set is obtained by taking each field in the reference application risk identification data and the target application risk identification data from the application risk identification optimization data as a risk feature. ,in, This indicates the first risk characteristic. This indicates the second risk characteristic. This represents the nth risk feature, where n is the total number of features;

[0108] The formula used is as follows:

[0109] ;

[0110] In the formula, Represents the i-th risk feature and the j-th risk feature The Euclidean distance between them, where m represents the number of data entries in the identification data. This represents the specific value of the i-th risk feature in the q-th identification data. This represents the specific value of the j-th risk feature in the q-th identification data;

[0111] Step S212: Calculate the equilibrium density to identify the density differences of different feature clusters in risk identification for multi-domain AI applications, providing an equilibrium density benchmark for feature cluster center identification; specifically, based on the risk feature k-nearest neighbor matrix and the risk feature bidirectional nearest neighbor matrix, calculate the equilibrium density value of each risk feature and normalize it to obtain the normalized equilibrium density value. ;

[0112] The normalization process uses the min-max normalization method to map the equilibrium density value to the [0,1] interval;

[0113] The formula used is as follows:

[0114] ;

[0115] In the formula, This represents the equilibrium density value of the i-th risk characteristic. express Features of its k nearest neighbors The Euclidean distance between them express The number of bidirectional nearest neighbors, express The bidirectional nearest neighbor set, where k represents the number of nearest neighbors;

[0116] Step S213: Calculate the minimum equilibrium density of the path to identify key contribution information in the feature connection path, enhance the identification of different risk feature clusters, and filter out irrelevant feature interference; specifically, first calculate the manifold distance between features. And for each pair of risk characteristics All connection paths are sorted by manifold distance. Sort the paths from smallest to largest, and select the path with the shortest distance as the optimal connection path. The normalized equilibrium density values ​​of all risk characteristics along the optimal connection path are calculated, and the minimum normalized equilibrium density value is selected as the minimum equilibrium density of the optimal connection path; the formula used is as follows:

[0117] ;

[0118] In the formula, Represents the i-th risk feature and the j-th risk feature The manifold distance between them express and All connecting paths between them, that is, all combinations of paths indirectly related through other risk characteristics. Indicates the connection path. Indicates the length of the connection path. Indicates the adjacent risk characteristics on path p and The Euclidean distance between them This represents the manifold distance adjustment factor, set to 90, which amplifies the distance differences between clusters with different risk characteristics, preserves the compactness of features within the same cluster, and enhances inter-cluster distinguishability. The exponential function is used to highlight the influence of long-distance feature edges between clusters and weaken the weight of short-distance feature edges within clusters. t represents the index variable of the feature on the path.

[0119] Step S214: Relative distance comprehensive calculation, used to amplify the distance differences between cluster centers with different risk characteristics, accurately distinguish core feature connections within clusters from inter-cluster interference feature connections, and enhance the identifiability of cluster center features. Specifically, it calculates the relative distance through a two-factor collaborative optimization strategy using interaction similarity factor and path contribution compensation factor; including the following steps:

[0120] Step S2141: Calculate the interaction similarity factor, which measures the local correlation between feature pairs. Specifically, the interaction similarity factor value is calculated based on the number of shared nearest neighbor features between features. The formula used is as follows:

[0121] ;

[0122] In the formula, express and Interaction similarity factor between express and The number of shared nearest neighbor features;

[0123] Step S2142: Calculate the path contribution compensation factor to distinguish between core paths within a cluster and inter-cluster interference paths, and filter out spurious association paths with low contribution features. Specifically, the path contribution compensation factor value is calculated based on the relationship between the minimum equilibrium density of the optimal connection path and the feature densities at both ends; the formula used is as follows:

[0124] ;

[0125] In the formula, express and Path contribution compensation factor between express and The minimum equilibrium density value of the optimal connection path between them;

[0126] Step S2143: Calculate the optimized manifold distance to enhance intra-cluster compactness and inter-cluster discriminability. Specifically, the optimized manifold distance is obtained by multiplying the manifold distance by the interaction similarity factor and the path contribution compensation factor; the formula used is as follows:

[0127] ;

[0128] In the formula, express and Optimized manifold distance between them;

[0129] Step S2144: Calculate the relative distance value of risk features. Specifically, for each risk feature, first determine whether it is the risk feature with the highest normalized equilibrium density value among all risk features. Then, calculate the relative distance value based on the determination result. The formula used is as follows:

[0130] ;

[0131] In the formula, express The relative distance value, This represents the normalized equilibrium density value of the v-th risk feature. Index variables representing risk characteristics;

[0132] Step S215: Cluster center selection, used to screen the core features of each risk feature cluster, specifically, the risk features that simultaneously meet the contribution threshold condition and the difference threshold condition are taken as cluster centers to obtain the feature cluster center set;

[0133] The specific condition for the contribution threshold is the normalized equilibrium density value of the risk characteristic. This is used to ensure that cluster centers are high-contribution features that significantly contribute to risk identification;

[0134] The specific condition for the difference threshold is the relative distance value of the risk characteristics. This ensures strong differentiation between cluster centers and other core features, avoiding cluster overlap. This represents the mean of the relative distance values ​​of all risk characteristics. The standard deviation of the relative distance values ​​of all risk characteristics;

[0135] Step S216: Core risk feature allocation, used to screen core risk features in multi-domain AI applications and exclude weakly correlated interference features. Specifically, this involves first calculating the density threshold. Next, the cluster centers are placed into the initialization queue Q, and a unique cluster label is assigned to each cluster center. Then, the label propagation process is performed. When the queue Q is not empty, the feature of the head of the queue is retrieved. The risk features are traversed through the bidirectional nearest neighbor matrix of the risk features. ,like Unassigned labels and their normalized equilibrium density values , then bestowing Same cluster label, and Add to queue Q, repeat the above process until the queue is empty, and complete the cluster affiliation assignment of the core risk features;

[0136] The core risk feature specifically refers to the bidirectional nearest neighbor risk feature of the cluster center, and its normalized equilibrium density value. ;

[0137] The formula used is as follows:

[0138] ;

[0139] In the formula, This represents the mean of the normalized equilibrium density values ​​of all risk characteristics. The standard deviation of the normalized equilibrium density values ​​representing all risk characteristics;

[0140] Step S217: Residual risk feature allocation, used to achieve accurate allocation of residual risk features, ensuring that all risk features belong to the correct cluster. Specifically, it dynamically allocates residual risk features based on a bidirectional nearest neighbor dynamic weight allocation strategy; including the following steps:

[0141] Step S2171: Calculate the bidirectional nearest neighbor ratio weights to quantify the correlation strength between the remaining risk features and the assigned core feature clusters. The formula used is as follows:

[0142] ;

[0143] In the formula, express and Bidirectional nearest neighbor ratio weights between them This represents the m-th risk characteristic. express and The manifold distance value, This represents the l-th risk characteristic. express and The manifold distance value, express bidirectional nearest neighbor set, express and Bidirectional nearest neighbor ratio weights between them;

[0144] Step S2172: Fuzzy membership calculation, used to quantify the probability that the remaining features belong to the c-th risk cluster; specifically, the risk features that have been assigned to the c-th cluster among the k nearest neighbors of the risk feature are multiplied by their normalized proportional weights and bidirectional nearest neighbor proportional weights and then summed.

[0145] The normalized proportional weight specifically refers to the relative weight ratio of the bidirectional nearest neighbor proportional weight in all k nearest neighbor sets of the risk feature.

[0146] The formula used is as follows:

[0147] ;

[0148] ;

[0149] In the formula, Indicates risk characteristics The fuzzy membership value belonging to the c-th cluster. express and Normalized proportional weights between them express It belongs to the c-th cluster;

[0150] Step S2173: Residual feature allocation, specifically for each residual feature Calculate its fuzzy membership degree among all clusters, and then... Assign to the cluster corresponding to the maximum fuzzy membership degree, then update. The membership degree of all risk features in the k nearest neighbor set is calculated, and this process is repeated until all remaining features are assigned. If there are still unassigned features, they are sorted in descending order of normalized equilibrium density value, and each feature is assigned to the cluster of the risk feature with the nearest Euclidean distance.

[0151] The aforementioned unassigned features specifically refer to features for which no effective membership reference can be provided;

[0152] Step S218: Clustering results are generated, specifically after all risk features have been assigned, multiple independent clusters are formed, each cluster corresponding to a set of risk features for a domain AI application;

[0153] Step S22: Real-time clustering result identification, specifically, inputting the reference application risk identification data and the target application risk identification data from the application risk identification optimization data into the feature selection clustering algorithm to obtain the real-time risk feature selection clustering result;

[0154] Step S23: Obtain the optimal application risk feature set, used to select the features that best represent the core risk patterns of the domain from each cluster. Specifically, for each cluster, calculate the Euclidean distance between all risk features within the cluster and the cluster center, sort them in ascending order according to the Euclidean distance, and select the top features within each cluster. The characteristics are used as candidate application risk features, and the optimal application risk feature set is determined according to the domain attributes of the AI ​​application. If the AI ​​application is a single domain application, the candidate application risk features selected from the clusters corresponding to that domain are used as the optimal application risk feature set. If the AI ​​application is a multi-domain application, the candidate application risk features selected from each cluster are merged as the optimal application risk feature set.

[0155] By performing the above operations, this solution addresses the technical problems of density imbalance, bridge noise sensitivity, and low assignment tolerance in existing clustering methods applicable to risk feature selection, which lead to insufficient accuracy and robustness of clustering results and thus affect the accuracy of AI application risk identification output. This solution innovatively proposes an improved clustering algorithm based on bidirectional nearest neighbor balanced density calculation, a two-factor collaborative optimization strategy, and a bidirectional nearest neighbor dynamic weight allocation strategy. This solves the density imbalance problem, optimizes the distinguishability of cluster centers, significantly enhances the algorithm's stability and fault tolerance in complex data, and improves the accuracy of clustering results, thereby enhancing the accuracy of AI application risk identification output.

[0156] Example 4, see Figure 1 , Figure 2 , Figure 4 and Figure 6 This embodiment is based on the above embodiment. In step S3, the establishment of the AI ​​application risk identification model is used to construct a risk identification model with high adaptability and high accuracy, so as to quickly determine the risks and classify the risk levels of AI applications in a single or multiple fields; specifically, it includes the following steps:

[0157] Step S31: Construct an AI application risk identification model architecture, specifically by establishing an AI application risk identification model based on a deep neural network (DNN); the AI ​​application risk identification model includes an input layer, a hidden layer, and an output layer;

[0158] Specifically, the input layer uses a fully connected layer to receive the optimal application risk feature vector from the optimal application risk feature set;

[0159] The hidden layer is specifically set as an Ny fully connected hidden layer, with ReLU as the activation function;

[0160] Specifically, the output layer sets the number of neurons to 4, selects Softmax as the activation function, and uses the category with the highest probability as the application risk level result output by the AI ​​application risk identification model.

[0161] Step S32: Preliminary training of the identification model, which is used to perform preliminary training of the identification model under the benchmark hyperparameter conditions. Specifically, the optimal application risk feature set is used as the training data, and the AI ​​application risk identification model is trained in advance using the preset benchmark model hyperparameters to obtain the AI ​​application risk identification model after preliminary training.

[0162] The reference optimal application risk feature set is specifically obtained by selecting reference application risk identification data from the preliminary data of watershed ecological security monitoring through multi-domain application feature selection.

[0163] The model hyperparameters include learning rate, number of hidden layers Ny, number of hidden layer neurons, batch size, and Dropout ratio, and upper and lower boundaries are set for the value of each hyperparameter.

[0164] The model training uses the category cross-entropy loss function, and iteratively updates the weight matrix and bias parameters of the recognition model through backpropagation algorithm and gradient descent optimization method. The model parameters are continuously optimized through multiple rounds of iteration. When the preset maximum number of training times is reached or the loss function converges to a set threshold, the iterative training stops.

[0165] Step S33: Optimize hyperparameters of the recognition model. This step involves tuning the hyperparameters of the recognition model based on the initial training to obtain the optimal model structure and training performance. Specifically, it involves obtaining the optimal combination of hyperparameters for the recognition model through an improved optimization algorithm. This includes the following steps:

[0166] Step S331: Initial population generation, used to initialize the hyperparameter search space of the recognition model. Specifically, the hyperparameters of the AI ​​application risk recognition model are encoded into individual position vectors, and the hyperparameters are randomly initialized and generated according to their upper and lower boundaries. The initial population is obtained by analyzing the position vectors of each individual.

[0167] Step S332: Determine the global optimal position. Specifically, calculate the fitness value of individuals in the population, use the performance of the AI ​​application risk identification model after preliminary training based on the individual position as the individual fitness value, sort the individuals according to their fitness values ​​from best to worst, and determine the position vector of the individual ranked first as the initial global optimal position.

[0168] Step S333: Habitat exploration behavior stage, used to expand the hyperparameter search range and avoid local optima, specifically when individual exploration factors... In this case, the individual's location is updated during the habitat exploration phase using a habitat location selection strategy; the formula used is as follows:

[0169] ;

[0170] ;

[0171] ;

[0172] ;

[0173] In the formula, and They all said Randomness parameters within a range and This indicates the upper and lower limits of the preset habitat quality score. Indicates the globally optimal position. This represents the optimal position of the current individual across all iterations. Indicates that the individual actively... move, This indicates behavior that represents individuals competing for habitat location, if If it means the competition has failed, then it means the competition has failed. If the result is positive, the search is expanded based on the difference between the current position and the global optimum. Indicates the reference location of the habitat. Indicates the overall quality score of the habitat. Indicates that the i-th individual is in the i-th position. The position of the j-th dimension in the next iteration of the population. Indicates that the i-th individual is in the i-th position. The position in the j-th dimension of the population in the next iteration, where e represents the current search iteration number. Indicates the maximum number of search iterations. , , , , and They all said Randomness parameters within a certain range;

[0174] Step S334: Individual foraging behavior development stage, used to focus on fine-grained search of high-quality hyperparameter regions, specifically when the individual explores factors At this time, the individual's location is updated during the foraging development stage through food attraction strategies; the formula used is as follows:

[0175] ;

[0176] ;

[0177] ;

[0178] ;

[0179] ;

[0180] In the formula, This represents the fitness value of the current individual. This represents the globally optimal fitness value. express Randomness parameters within a range Indicates the location of the food. and Let these represent the upper and lower bounds of the j-th dimension, respectively. This represents the position of the globally optimal individual in the j-th dimension. Indicates that the i-th individual is in the i-th position. The food attraction factor in the next iteration Indicates that the i-th individual is in the i-th position. Position in the next iteration of the population This indicates the probability of the individual being searched for food. Indicates the size of the food. This indicates the number of dimensions, or the number of hyperparameters. and express Randomness parameters within a range and They all said Randomness parameters within a certain range;

[0181] Step S335: The iterative search terminates. Specifically, for all updated individuals, the individual fitness values ​​are recalculated, and the fitness values ​​of all individuals are compared to update the global optimal position. When the global optimal position is higher than the fitness threshold and the number of search iterations reaches the maximum number of search iterations, the search is terminated and the global optimal position is obtained. The global optimal position specifically refers to the optimal combination of hyperparameters of the recognition model.

[0182] Step S34: Final training of the identification model. Specifically, based on the optimal hyperparameter combination of the identification model, the hyperparameters of the AI ​​application risk identification model after initial training are adjusted, and the reference optimal application risk feature set is used as training data to retrain the AI ​​application risk identification model after initial training to obtain the AI ​​application risk identification model with the best performance.

[0183] By performing the above operations, this solution addresses the technical problem of inaccurate output results caused by improper built-in parameter settings in existing models applicable to AI application risk identification. It employs habitat location selection and food attraction strategies to improve the optimization algorithm for obtaining the optimal parameter combination of the identification model. This significantly improves the hyperparameter optimization efficiency of the identification model, reduces the computational cost of model training, and enhances the accuracy of the identification model's output results, thereby achieving intelligent and precise identification of AI application risks.

[0184] Example 5, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S4, the application risk intelligent identification specifically involves inputting the target optimal application risk feature set into the performance optimal AI application risk identification model to obtain the real-time application risk level result. Based on the real-time application risk level result, the corresponding security response strategy is triggered to realize real-time monitoring and intelligent identification of AI application risks.

[0185] The security response strategy is as follows: if the result is high risk, the system automatically triggers an emergency response mechanism, immediately suspends high-risk functions of the AI ​​application, and generates a risk alarm ticket to be pushed to the operations manager; if the result is medium risk, a semi-automatic intervention process is initiated, the system displays a risk warning pop-up, restricts sensitive operation permissions of the AI ​​application, and requires manual review and approval before it can continue to run; if the result is low risk, the AI ​​application continues to run normally, the system continuously monitors the changing trends of risk characteristics, and periodically summarizes low-risk events to form a statistical report; if the result is no risk, the AI ​​application runs normally.

[0186] The target optimal application risk feature set is specifically obtained by selecting target application risk identification data from the preliminary data of watershed ecological security monitoring through multi-domain application feature selection.

[0187] Example 6, see Figure 1 and Figure 2 Based on the above embodiments, this embodiment provides an AI application risk identification system based on big data, including a multi-source data acquisition module, a multi-domain application feature selection module, an AI application risk identification model establishment module, and an application risk intelligent identification module.

[0188] The multi-source data acquisition module specifically obtains application risk identification optimization data through data information collection and data optimization processing, and sends the data to the multi-domain application feature selection module.

[0189] The multi-domain application feature selection module receives data sent by the multi-source data acquisition module, improves the clustering algorithm by using a balanced density calculation based on bidirectional nearest neighbors, a two-factor collaborative optimization strategy, and a bidirectional nearest neighbor dynamic weight allocation strategy, thereby completing the construction of the feature selection clustering algorithm. Then, the data is input into the feature selection clustering algorithm, and the real-time risk feature selection clustering result is output. Based on the clustering result, candidate application risk features are selected. Finally, the features are integrated according to the attributes of the AI ​​application domain to obtain the optimal application risk feature set. The data is then sent to the AI ​​application risk identification model building module and the application risk intelligent identification module.

[0190] The AI ​​application risk identification model module receives data sent by the multi-domain application feature selection module, establishes an AI application risk identification model based on a deep neural network, and sequentially performs preliminary training of the identification model, hyperparameter optimization of the identification model using an improved optimization algorithm, and final training of the identification model to obtain the AI ​​application risk identification model with optimal performance, and sends the data to the application risk intelligent identification module.

[0191] The application risk intelligent identification module receives data from the multi-domain application feature selection module and the AI ​​application risk identification model establishment module. Specifically, it inputs the target optimal application risk feature set into the optimal model and outputs the risk level result in real time, thereby realizing real-time monitoring of AI application risks.

[0192] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0193] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0194] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for identifying risks in AI applications based on big data, characterized in that: The method includes the following steps: Step S1: Multi-source data acquisition, specifically, obtaining application risk identification optimization data through data information collection and data optimization processing; Step S2: Multi-domain application feature selection. Specifically, the clustering algorithm is improved by using balanced density calculation based on bidirectional nearest neighbors, a two-factor collaborative optimization strategy, and a bidirectional nearest neighbor dynamic weight allocation strategy to complete the construction of the feature selection clustering algorithm. Then, the data is input into the feature selection clustering algorithm, and the real-time risk feature selection clustering results are output. Candidate application risk features are selected based on the clustering results. Finally, the features are integrated according to the attributes of the AI ​​application domain to obtain the optimal application risk feature set. Step S3: Establish an AI application risk identification model. Specifically, establish an AI application risk identification model based on a deep neural network, and perform preliminary training of the identification model in sequence. Then, optimize the hyperparameters of the identification model by using an improved optimization algorithm that introduces habitat location selection strategy and food attraction strategy to obtain the optimal hyperparameter combination of the identification model. Finally, complete the training of the identification model based on the optimal hyperparameter combination to obtain the AI ​​application risk identification model with the best performance. Step S4: Apply intelligent risk identification, specifically by inputting the target optimal application risk feature set into the optimal model and outputting the risk level result in real time, thereby realizing real-time monitoring of AI application risks.

2. The method for identifying AI application risks based on big data according to claim 1, characterized in that: In step S2, the multi-domain application feature selection specifically includes the following steps: Step S21: Construct a feature selection clustering algorithm; Step S22: Real-time clustering result identification, specifically, inputting the reference application risk identification data and the target application risk identification data from the application risk identification optimization data into the feature selection clustering algorithm to obtain the real-time risk feature selection clustering result; Step S23: Obtain the optimal application risk feature set. Specifically, for each cluster, calculate the Euclidean distance between all risk features within the cluster and the cluster center, sort them in ascending order according to the Euclidean distance, and select the top-ranked features within each cluster. The characteristics are used as candidate application risk features, and the optimal application risk feature set is determined according to the domain attributes of the AI ​​application. If the AI ​​application is a single domain application, the candidate application risk features selected from the clusters corresponding to that domain are used as the optimal application risk feature set. If the AI ​​application is a multi-domain application, the candidate application risk features selected from each cluster are merged as the optimal application risk feature set.

3. The method for identifying AI application risks based on big data according to claim 2, characterized in that: In step S21, the construction of the feature selection clustering algorithm specifically includes the following steps: Step S211: Construct a bidirectional nearest neighbor matrix. Specifically, based on the risk feature set, calculate the distance between any two risk features using the Euclidean distance formula, and then perform a bidirectional nearest neighbor matrix for each risk feature. The KD-Tree algorithm is used to select the k nearest features to form a group. k-nearest neighbor set Construct a risk feature k-nearest neighbor matrix based on the k-nearest neighbor set of all risk features. Then, traverse all bidirectional nearest neighbor pairs in the risk feature k-nearest neighbor matrix. ,but and As bidirectional nearest neighbor pairs, a bidirectional nearest neighbor matrix for risk features is constructed based on all selected bidirectional nearest neighbor pairs, where, This represents the i-th risk characteristic. This represents the j-th risk characteristic. express The set of k nearest neighbors; Step S212: Calculate the equilibrium density. Specifically, based on the k-nearest neighbor matrix and the bidirectional nearest neighbor matrix of the risk features, calculate the equilibrium density value for each risk feature, and normalize it to obtain the normalized equilibrium density value. The formula used is as follows: ； In the formula, This represents the equilibrium density value of the i-th risk characteristic. express Features of its k nearest neighbors The Euclidean distance between them express The number of bidirectional nearest neighbors, express The bidirectional nearest neighbor set, where k represents the number of nearest neighbors; Step S213: Calculate the minimum equilibrium density of the path, specifically by first calculating the manifold distance between features. And for each pair of risk characteristics All connection paths are sorted by manifold distance. Sort the paths from smallest to largest, and select the path with the shortest distance as the optimal connection path. The normalized equilibrium density values ​​of all risk characteristics along the optimal connection path are calculated, and the minimum normalized equilibrium density value is selected as the minimum equilibrium density of the optimal connection path; the formula used is as follows: ； In the formula, Represents the i-th risk feature and the j-th risk feature The manifold distance between them express and All connecting paths between them, that is, all combinations of paths indirectly related through other risk characteristics. Indicates the connection path. Indicates the length of the connection path. Indicates the adjacent risk characteristics on path p and The Euclidean distance between them represents the manifold distance adjustment factor, and t represents the index variable of the features on the path; Step S214: Comprehensive calculation of relative distance; Step S215: Cluster center selection, specifically, selecting risk features that simultaneously meet the contribution threshold condition and the difference threshold condition as cluster centers to obtain a set of feature cluster centers; Step S216: Core risk feature allocation, specifically, first calculating the density threshold. Next, the cluster centers are placed into the initialization queue Q, and a unique cluster label is assigned to each cluster center. Then, the label propagation process is performed. When the queue Q is not empty, the feature of the head of the queue is retrieved. The risk features are traversed through the bidirectional nearest neighbor matrix. ,like Unassigned labels and their normalized equilibrium density values , then bestowing Same cluster label, and Add to queue Q, and repeat the above process until the queue is empty, thus completing the cluster assignment of the core risk features; the formula used is as follows: ； In the formula, This represents the mean of the normalized equilibrium density values ​​of all risk characteristics. The standard deviation of the normalized equilibrium density values ​​representing all risk characteristics; Step S217: Residual risk characteristic allocation; Step S218: Clustering results are generated. Specifically, after all risk features are assigned, multiple independent clusters are formed, and each cluster corresponds to a set of risk features for a domain AI application.

4. The method for identifying AI application risks based on big data according to claim 3, characterized in that: In step S214, the relative distance comprehensive calculation specifically involves calculating the relative distance using a two-factor collaborative optimization strategy of interaction similarity factor and path contribution compensation factor; this includes the following steps: Step S2141: Calculate the interaction similarity factor, specifically by calculating the interaction similarity factor value based on the number of shared nearest neighbor features between features; the formula used is as follows: ； In the formula, express and Interaction similarity factor between express and The number of shared nearest neighbor features; Step S2142: Calculate the path contribution compensation factor, specifically by calculating the path contribution compensation factor value based on the relationship between the minimum equilibrium density of the optimal connection path and the feature densities at both ends; the formula used is as follows: ； In the formula, express and Path contribution compensation factor between express and The minimum equilibrium density value of the optimal connection path between them; Step S2143: Calculate the optimized manifold distance, specifically by multiplying the manifold distance by the interaction similarity factor and the path contribution compensation factor to obtain the optimized manifold distance value; Step S2144: Calculate the relative distance value of risk features. Specifically, for each risk feature, first determine whether it is the risk feature with the highest normalized equilibrium density value among all risk features. Then, calculate the relative distance value based on the determination result. The formula used is as follows: ； In the formula, express The relative distance value, This represents the normalized equilibrium density value of the v-th risk feature. Index variables representing risk characteristics, express and The optimized value of the manifold distance between them.

5. The method for identifying AI application risks based on big data according to claim 3, characterized in that: In step S217, the allocation of the remaining risk characteristics specifically includes the following steps: Step S2171: Calculate the bidirectional nearest neighbor ratio weights using the following formula: ； In the formula, express and Bidirectional nearest neighbor ratio weights between them This represents the m-th risk characteristic. express and The manifold distance value, This represents the l-th risk characteristic. express and The manifold distance value, express bidirectional nearest neighbor set, express and Bidirectional nearest neighbor ratio weights between them; Step S2172: Fuzzy membership calculation, specifically, the risk features that have been assigned to the c-th cluster among the k nearest neighbors of the risk feature are multiplied by their normalized proportional weights and bidirectional nearest neighbor proportional weights and then summed. The normalized proportional weight specifically refers to the relative weight proportion of the bidirectional nearest neighbor proportional weight in all k nearest neighbor sets of the risk feature; the formula used is as follows: ； In the formula, Indicates risk characteristics The fuzzy membership value belonging to the c-th cluster. express and Normalized proportional weights between them express It belongs to the c-th cluster; Step S2173: Residual feature allocation, specifically for each residual feature Calculate its fuzzy membership degree among all clusters, and then... Assign to the cluster corresponding to the maximum fuzzy membership degree, then update. The membership degree of all risk features in the k-nearest neighbor set is calculated, and this process is repeated until all remaining features are assigned. If there are still unassigned features, they are sorted in descending order of normalized equilibrium density value, and each feature is assigned to the cluster of the risk feature with the nearest Euclidean distance.

6. The method for identifying AI application risks based on big data according to claim 1, characterized in that: In step S3, establishing the AI ​​application risk identification model specifically includes the following steps: Step S31: Construct the AI ​​application risk identification model architecture, specifically by establishing an AI application risk identification model based on a deep neural network (DNN). Step S32: Preliminary training of the identification model, specifically, using the optimal application risk feature set as the training data, and using the preset benchmark model hyperparameters to perform preliminary model training on the AI ​​application risk identification model, to obtain the preliminary trained AI application risk identification model; Step S33: Identify model hyperparameter optimization; Step S34: Final training of the identification model. Specifically, based on the optimal hyperparameter combination of the identification model, the hyperparameters of the AI ​​application risk identification model after initial training are adjusted, and the reference optimal application risk feature set is used as training data to retrain the AI ​​application risk identification model after initial training to obtain the AI ​​application risk identification model with the best performance.

7. The method for identifying AI application risks based on big data according to claim 6, characterized in that: In step S33, the hyperparameter optimization of the recognition model specifically includes the following steps: Step S331: Initial population generation, specifically, encoding the hyperparameters of the AI ​​application risk identification model into individual location vectors, and randomly initializing and generating populations based on the upper and lower boundaries of the hyperparameters. The position vectors of each individual are used to obtain the initial population; Step S332: Determine the global optimal position. Specifically, calculate the fitness value of individuals in the population, use the performance of the AI ​​application risk identification model after preliminary training based on the individual position as the individual fitness value, sort the individuals according to their fitness values ​​from best to worst, and determine the position vector of the individual ranked first as the initial global optimal position. Step S333: Habitat exploration behavior stage, specifically when an individual's exploration factors... In this case, the individual's location is updated during the habitat exploration phase using a habitat location selection strategy; the formula used is as follows: ； ； In the formula, Indicates the globally optimal position. This represents the optimal position of the current individual across all iterations. Indicates the reference location of the habitat. Indicates the overall quality score of the habitat. Indicates that the i-th individual is in the i-th position. The position of the j-th dimension in the next iteration of the population. Indicates that the i-th individual is in the i-th position. The position in the j-th dimension of the population in the next iteration, where e represents the current search iteration number. Indicates the maximum number of search iterations. , , , , and They all said Randomness parameters within a certain range; Step S334: Individual foraging behavior development stage, specifically when the individual explores factors When this occurs, the individual's location is updated during the foraging development phase using food attraction strategies; the formula used is as follows: ； ； ； ； In the formula, Indicates the location of the food. and Let these represent the upper and lower bounds of the j-th dimension, respectively. This represents the position of the globally optimal individual in the j-th dimension. Indicates that the i-th individual is in the i-th position. The food attraction factor in the next iteration Indicates that the i-th individual is in the i-th position. Position in the next iteration of the population This indicates the probability of the individual being searched for food. Indicates the size of the food. Indicates the number of dimensions. and express Randomness parameters within a range and They all said Randomness parameters within a certain range; Step S335: The iterative search terminates. Specifically, for all updated individuals, the individual fitness values ​​are recalculated, and the fitness values ​​of all individuals are compared to update the global optimal position. When the global optimal position is higher than the fitness threshold and the number of search iterations reaches the maximum number of search iterations, the search is terminated and the global optimal position is obtained. The global optimal position specifically refers to the optimal combination of hyperparameters of the recognition model.

8. The method for identifying AI application risks based on big data according to claim 1, characterized in that: In step S4, the intelligent identification of application risks specifically involves inputting the target optimal application risk feature set into the best-performing AI application risk identification model to obtain the real-time application risk level result. Based on the real-time application risk level result, the corresponding security response strategy is triggered to achieve real-time monitoring and intelligent identification of AI application risks.

9. The method for identifying AI application risks based on big data according to claim 1, characterized in that: In step S1, the multi-source data acquisition is used to acquire data related to AI application risk identification. Specifically, it involves collecting data information from the risk identification platform to obtain raw application risk identification data, and then performing data optimization processing on the raw application risk identification data to obtain optimized application risk identification data. The raw application risk identification data includes reference application risk identification data and target application risk identification data; the data optimization processing specifically involves cleaning and standardizing the raw data to obtain optimized application risk identification data.

10. A big data-based AI application risk identification system, used to implement the big data-based AI application risk identification method as described in any one of claims 1-9, characterized in that: It includes a multi-source data acquisition module, a multi-domain application feature selection module, an AI application risk identification model building module, and an application risk intelligent identification module; The multi-source data acquisition module specifically obtains application risk identification optimization data through data information collection and data optimization processing, and sends the data to the multi-domain application feature selection module. The multi-domain application feature selection module receives data sent by the multi-source data acquisition module, improves the clustering algorithm by using a balanced density calculation based on bidirectional nearest neighbors, a two-factor collaborative optimization strategy, and a bidirectional nearest neighbor dynamic weight allocation strategy, thereby completing the construction of the feature selection clustering algorithm. Then, the data is input into the feature selection clustering algorithm, and the real-time risk feature selection clustering result is output. Based on the clustering result, candidate application risk features are selected. Finally, the features are integrated according to the attributes of the AI ​​application domain to obtain the optimal application risk feature set. The data is then sent to the AI ​​application risk identification model building module and the application risk intelligent identification module. The AI ​​application risk identification model module receives data sent by the multi-domain application feature selection module, establishes an AI application risk identification model based on a deep neural network, and sequentially performs preliminary training of the identification model, hyperparameter optimization of the identification model using an improved optimization algorithm, and final training of the identification model to obtain the AI ​​application risk identification model with optimal performance, and sends the data to the application risk intelligent identification module. The application risk intelligent identification module receives data from the multi-domain application feature selection module and the AI ​​application risk identification model establishment module. Specifically, it inputs the target optimal application risk feature set into the optimal model and outputs the risk level result in real time, thereby realizing real-time monitoring of AI application risks.

Images