Method and system for processing anomaly in ciphertext execution of privacy preserving computing algorithm

Abstract

The present invention relates to the technical field of information, and disclosed are a method and system for processing an anomaly in ciphertext execution of a privacy preserving computing algorithm. Differential privacy query parameters of multiple tenants are collected to quantify the data query sensitivity of each tenant, an ARIMA model is used to perform time series analysis on sensitivity quantification results to predict future trends thereof, and residuals between actual sensitivity and predicted sensitivity are calculated to identify anomalous tenants; a sliding window method is used to extract statistical features of the anomalous tenants to construct a feature set, and cluster analysis is performed on the feature set to identify anomaly patterns and anomaly fingerprint feature clusters corresponding to the anomalous tenants; and a model is trained on the basis of normal and anomaly fingerprint feature clusters of the anomalous tenants, and the model is used to quantify the influence of each anomaly pattern on the privacy protection level of the corresponding anomalous tenant, so as to adjust a privacy budget allocation ratio of each anomalous tenant on the basis of obtained quantitative analysis results and perform execution, thereby effectively processing anomalies in ciphertext execution of a privacy preserving computing algorithm.

Description

A method and system for handling ciphertext execution exceptions in a privacy-preserving computation algorithm. Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method and system for handling anomalies in the ciphertext execution of a privacy computing algorithm. Background Technology

[0002] With the continuous development of privacy algorithms, differential privacy algorithms, a type of privacy computing algorithm, are frequently applied in multi-tenant cloud environments. A multi-tenant cloud environment is a cloud computing architecture where multiple different users (called tenants) share the same set of physical or virtualized underlying computing resources (such as servers, storage, networks, and applications), while each tenant's data and configuration maintain logical isolation and security. Currently, tenant-submitted differential privacy queries are typically executed in encrypted form to protect privacy. The global sensitivity of the query function is a key parameter for measuring the level of privacy protection. However, because the query process is executed in encrypted form, the sensitivity calculation process cannot be directly observed. Therefore, if the sensitivity is underestimated, insufficient noise is added, resulting in a decrease in privacy protection; if the sensitivity is overestimated, although privacy is protected, excessive noise is introduced, reducing the usability of the query results. Therefore, it is necessary to dynamically identify and handle anomalies in sensitivity calculation during the execution of tenant differential privacy queries.

[0003] Therefore, how to detect and handle sensitivity anomalies of tenants during differential privacy query execution in a multi-tenant cloud environment has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention provides a method and system for handling ciphertext execution anomalies in privacy computing algorithms, addressing the problem of how to detect and handle sensitivity anomalies of tenants during differential privacy query execution in a multi-tenant cloud environment.

[0005] To address the aforementioned technical problems, this invention provides a method for handling ciphertext execution anomalies in privacy computation algorithms, comprising:

[0006] Obtain differential privacy query parameters for multiple tenants, and quantify the data query sensitivity of each tenant based on the differential privacy query parameters to obtain sensitivity calculation feature data;

[0007] The sensitivity calculation feature data is analyzed by time series analysis using the ARIMA model to obtain sensitivity prediction data. The residual between the sensitivity calculation feature data and the sensitivity prediction data is quantified to assess whether each tenant has sensitivity shift anomalies and to identify anomalous tenants.

[0008] A sliding window is used to extract statistical features from the sensitivity calculation feature data of each abnormal tenant to construct a feature set, and cluster analysis is performed on the feature set to obtain the abnormal pattern and abnormal fingerprint feature cluster corresponding to each abnormal tenant.

[0009] Obtain the normal fingerprint feature clusters of each of the abnormal tenants, and construct an abnormal pattern influence mechanism model by combining the abnormal fingerprint feature clusters. Then, quantify the impact of each abnormal pattern on the privacy protection level of each of the abnormal tenants through the abnormal pattern influence mechanism model to obtain quantitative analysis results.

[0010] Based on the quantitative analysis results, the privacy budget allocation ratio of each abnormal tenant is adjusted to obtain a privacy budget allocation scheme, and the privacy budget allocation scheme is executed.

[0011] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following:

[0012] (1) By quantifying the data query sensitivity of each tenant in the differential privacy algorithm and predicting the future data query sensitivity trend, it is possible to detect and handle sensitivity shift anomalies in a timely manner, thereby ensuring that the privacy protection of tenants is not affected. In this process, the ARIMA model is used for time series analysis and prediction, and the sliding window method is used to extract statistical features, which improves the accuracy and efficiency of anomaly detection.

[0013] (2) By constructing an abnormal mode influence mechanism model, the impact of each abnormal mode on the level of privacy protection is quantified, providing a scientific basis for the reasonable allocation of privacy budget, effectively detecting sensitivity anomalies in multi-tenant environments, dynamically adjusting privacy protection strategies, improving the overall level of privacy protection, effectively handling anomalies in the execution of algorithm ciphertext, and thus improving the robustness and practicality of privacy computing algorithms. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 is a flowchart of a method for handling ciphertext execution exceptions of a privacy computing algorithm according to a certain embodiment of the present invention;

[0016] Figure 2 is a structural diagram of a ciphertext execution exception handling system for a privacy computing algorithm provided in a certain embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings and examples. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0019] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. The terms "vertical," "horizontal," "left," "right," "upper," "lower," and similar expressions used herein are for illustrative purposes only and do not indicate or imply that the system or component referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0020] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0021] In one embodiment, as shown in FIG1, the first aspect of the present invention provides a method for handling ciphertext execution exceptions of a privacy computation algorithm, comprising:

[0022] S1. Obtain differential privacy query parameters for multiple tenants, and quantify the data query sensitivity of each tenant based on the differential privacy query parameters to obtain sensitivity calculation feature data; wherein, the differential privacy query parameters include data difference values ​​and weight coefficients;

[0023] In one embodiment, step S1 includes:

[0024] The data difference values ​​and weight coefficients of multiple tenants are obtained, and the first tenant isolation boundary value is obtained by performing a dot product operation and normalization on the data difference values ​​and the weight coefficients to construct a tenant difference parameter matrix; each row of the tenant difference parameter matrix contains the tenant identifier and its corresponding data query sensitivity;

[0025] A data query sensitivity threshold is set according to the tenant differential parameter matrix. Each tenant is grouped and clustered according to the data query sensitivity threshold to obtain multiple tenant groups with different data query sensitivity levels. The differential privacy query parameters of each tenant group are quantized and encoded according to a preset desensitization rule to obtain a first feature space.

[0026] Gradient boosting trees are used to associate and map the parameter range in the first feature space with the data query sensitivity level, outputting a parameter mapping vector. Random forest features are then extracted from the parameter mapping vector to obtain the second feature space.

[0027] Obtain the intermediate index for global sensitivity calculation in the second feature space, process the intermediate index for global sensitivity calculation using an adaptive clustering algorithm to obtain a set of cluster center points, and construct a feature vector dimension matrix based on the set of cluster center points;

[0028] The feature vector dimension matrix is ​​orthogonally transformed to obtain a standardized feature vector. The standardized feature vector is then weighted and summed with the first tenant isolation boundary value to generate a differential privacy query parameter set as the sensitivity calculation feature data.

[0029] Specifically, in the differential privacy algorithm, this embodiment of the invention utilizes a tenant access control library to store real-time differential privacy query requests and related data of different tenants, such as data difference values ​​and weight coefficients. The data difference value reflects the dependence of the query result on a single record in the dataset, while the weight coefficient is used to adjust the privacy protection level of different queries. When there are 300 tenants, the data difference value for each tenant is distributed between 0 and 1, and the weight coefficient is determined between 0.1 and 0.9 based on the tenant's data sensitivity. This embodiment of the invention then accesses the control library through a differential privacy query request to obtain differential privacy query parameters. Next, a dot product operation is performed on the data difference value and the weight coefficient, and the result is processed using a maximum-minimum normalization method to obtain initial tenant isolation boundary values ​​to construct a 300×2-dimensional tenant difference parameter matrix. Each row in the tenant difference parameter matrix contains the tenant identifier and its corresponding data query sensitivity. The data query sensitivity can be understood as the maximum change in the query result when the same query is executed on adjacent datasets (i.e., datasets differing by only one record).

[0030] Next, a data query sensitivity threshold, preferably 0.8, is set based on the distribution characteristics of the data in the tenant differential parameter matrix. Tenants are then grouped and clustered based on the data query sensitivity threshold. Euclidean distance can be used to measure the similarity between tenants, and hierarchical clustering is used to divide them into multiple tenant groups with different data query sensitivity levels. For example, 300 tenants can be divided into 25 groups, each group corresponding to a data query sensitivity level, with the number of tenants in each group between 8 and 15. Then, the differential privacy query parameters of each tenant group are quantized and encoded according to preset desensitization rules such as fuzzing, replacement, encryption, and masking to map character parameters into numerical features, and then the first feature space is constructed using these features.

[0031] In the first feature space, the quantized query parameters and their ranges are layered according to parameter type and preset range to construct a gradient boosting tree with a depth of 5 layers. Each layer of this model contains 20 decision nodes, and the Gini coefficient is used as the node splitting criterion. Then, the parameter ranges are associated with the data sensitivity level through the gradient boosting tree model to obtain a 300×10-dimensional parameter mapping vector. The parameter mapping vector is then used to train a random forest model containing 50 decision trees. The maximum depth of each tree in this model is 8, and the feature sampling ratio is 0.7, generating the second feature space.

[0032] Multiple intermediate indicators for global sensitivity calculation are extracted from the second feature space, such as the change in the query function, the scale parameter of noise, the change in data distribution, and the sensitivity boundary. A density-based adaptive clustering algorithm is used to process these indicator data, with the neighborhood radius set to 0.3 and the minimum number of samples set to 5. This yields a set of cluster centers, which are then organized into an 8×10 dimensional feature vector matrix.

[0033] The feature vector dimension matrix is ​​decomposed by singular value decomposition and orthogonal transformation. Feature vectors with a variance explanation rate of 95% are retained as standardized feature vectors. The standardized feature vectors are then weighted and summed with the tenant isolation boundary values. The preferred weights for the two are 0.25 and 0.75, respectively. Other relevant information (such as query type, query timestamp, query frequency, data access range, operation type, etc.) are also included in the feature data to generate a differential privacy query parameter set for 300 tenants. The differential privacy query parameter set is used as the sensitivity calculation feature data. The parameter values ​​are distributed between 0.2 and 0.9. Larger parameter values ​​correspond to tenants with higher data query sensitivity, reflecting the adaptive characteristics of differential privacy protection.

[0034] In practical applications, a user information query system contains sensitive data such as power plant output power, substation load, user electricity consumption patterns, and historical energy consumption records. The differential privacy query parameters calculated using the above method can automatically adjust the amount of noise added according to the sensitivity of different types of data. This protects user privacy while ensuring the availability of query results. For example, for credit score queries, the system adds Laplace noise with a mean of 0 and a standard deviation of 0.8 to the high-sensitivity tenant group, while adding noise with a standard deviation of 0.3 to the low-sensitivity tenant group, thus fully realizing the dynamic adjustment of the differential privacy protection strength.

[0035] This invention, by comprehensively considering the data difference values ​​and weight coefficients in the differential privacy query parameters, can more accurately assess the data query sensitivity of tenants. It further improves the accuracy of data query sensitivity assessment by utilizing machine learning algorithms such as gradient boosting trees and random forests for feature extraction and mapping. Grouping and clustering tenants according to data query sensitivity thresholds helps implement different privacy protection measures for tenants with different sensitivity levels. Quantitative encoding of differential privacy query parameters through preset de-identification rules enhances data privacy protection capabilities. Combining multiple algorithms and technologies makes the data processing workflow more efficient and flexible, and allows for adjustment of parameters and algorithms according to actual needs to adapt to different data processing scenarios and requirements.

[0036] S2. Perform time series analysis on the sensitivity calculation feature data using the ARIMA model to obtain sensitivity prediction data, and quantify the residual between the sensitivity calculation feature data and the sensitivity prediction data to assess whether each tenant has sensitivity shift anomalies and to identify anomalous tenants.

[0037] In one embodiment, the step of performing time series analysis on the sensitivity calculation feature data using an ARIMA model to obtain sensitivity prediction data, and quantifying the residual between the sensitivity calculation feature data and the sensitivity prediction data, includes:

[0038] The sensitivity calculation feature data is sampled at fixed intervals by a sliding time window to obtain sampled data, and the time series feature matrix is ​​constructed by quantifying the autocorrelation function and partial autocorrelation function of the sampled data.

[0039] The time series feature matrix is ​​subjected to stationarity processing to obtain a stationary time series dataset for constructing an ARIMA model. The autoregressive term coefficients and moving average term coefficients of the ARIMA model are estimated by the least squares method, and the goodness-of-fit index is obtained by performing a significance test on the parameters of the ARIMA model.

[0040] The optimal parameter combination is selected by the goodness-of-fit index, and the ARIMA model with the optimal parameter combination is subjected to multi-step iterative prediction on the validation dataset to calculate the confidence interval of the prediction results and obtain the sensitivity prediction sequence.

[0041] A Kalman filter is constructed based on the sensitivity prediction sequence, a state transition matrix and an observation matrix are set, sensitivity prediction data is obtained through prediction update and measurement update, and the residual is calculated based on the sensitivity prediction data and the sampled data.

[0042] Specifically, in the tenant sensitivity assessment, this embodiment of the invention selects a fixed-size time window and slides the window across the sensitivity calculation feature data, moving it by a fixed time interval each time to obtain a series of sampled data. This smooths the data and reduces noise while preserving the trend and periodicity of the time series. Next, the autocorrelation function (ACF) and partial autocorrelation function (PACF) of the sampled data are calculated (to reveal the autocorrelation and patterns in the time series data), and the values ​​of the ACF and PACF are quantified to construct a time series feature matrix. Taking a cross-tenant data access scenario of an organization as an example, its sensitivity calculation feature data includes dimensions such as query frequency, data access scope, and operation type. A daily sensitivity sequence of 144 sampling points is established for each tenant, with a sampling interval of 10 minutes. The sampled data shows that the sensitivity calculation feature data reaches significant correlation at a lag of 12, with an autocorrelation coefficient of 0.85 and a partial autocorrelation coefficient of 0.72, indicating that the data has obvious periodic characteristics.

[0043] Next, the sampling difference operation is used to perform stationarity processing on the time series feature matrix. The variance of the sequence after first-order difference is reduced from 0.46 to 0.08. The test statistic obtained by the augmented Dickey-Fowler test is -3.85, which is less than the critical value of -3.43 at the 1% significance level, confirming that the differenced sequence is stationary. The obtained stationary time series dataset can be used to construct an autoregressive difference moving average model. Based on the Akaike information criterion, the optimal combination of model parameters is selected: autoregressive order p = 2, difference order d = 1, moving average order q = 1. The coefficients of the model are estimated by least squares: AR(1) coefficient is 0.65, AR(2) coefficient is 0.28, MA(1) coefficient is -0.42, and residual standard deviation is 0.06. Then, the parameters of the ARIMA model are tested for significance, such as t-test or F-test, to evaluate the statistical significance of the parameters and obtain the goodness-of-fit index.

[0044] The optimal ARIMA model parameter combination was selected based on the goodness-of-fit index, and the ARIMA model with the optimal parameter combination was used to perform rolling predictions on 24 points on the validation dataset. The upper and lower limits of the 95% confidence interval of the prediction results were ±0.12 of the predicted value, respectively, to obtain the sensitivity prediction sequence. Then, a Kalman filter was constructed using the sensitivity prediction sequence. The state transition matrix of the Kalman filter adopted a 2×2 identity matrix, the observation matrix was [1,0], the diagonal elements of the state noise covariance matrix were set to 0.01, and the observation noise variance was set to 0.04. These matrices were used to describe the changes in the system state and the generation process of the observation data. Then, a prediction-update loop was used to enable the constructed Kalman filter to effectively track the changing trend of the sensitivity sequence. Finally, the prediction information and the observation information were fused to obtain more accurate sensitivity prediction data for residual calculation with the sampled data.

[0045] This invention improves prediction accuracy by extracting the inherent patterns and regularities of time series data through sliding time window sampling and time series feature construction. The ARIMA model captures the long-term dependencies and periodicity of time series data, while the Kalman filter integrates prediction information and real-time observation data, further enhancing prediction accuracy. Significance testing is used to select the optimal parameter combination, avoiding overfitting and underfitting and strengthening the model's robustness. The Kalman filter updates prediction results in real time, adapting to changes in time series data, enabling the solution to achieve real-time prediction and dynamic adjustment in practical applications, thus improving the timeliness and practicality of predictions.

[0046] In one embodiment, the step of performing time series analysis on the sensitivity calculation feature data using an ARIMA model to obtain sensitivity prediction data, and quantifying the residual between the sensitivity calculation feature data and the sensitivity prediction data, includes:

[0047] The sensitivity calculation feature data is sampled at fixed intervals using a sliding time window to obtain sampled data;

[0048] By quantifying the Pearson autocorrelation coefficient and partial autocorrelation coefficient of the sampled data, a multidimensional sensitivity initial sequence is obtained, and the stationarity of the multidimensional sensitivity initial sequence is tested to obtain a multidimensional stationary sensitivity sequence.

[0049] An ARIMA model is constructed based on the multidimensional stationary sensitivity sequence, and the log-likelihood function is optimized by the maximum likelihood estimation method to calculate the autoregressive coefficients and moving average coefficients of the ARIMA model, thereby obtaining the set of fitting parameters.

[0050] Residual analysis is performed on the multidimensional stationary sensitivity sequence based on the fitted parameter set, and the optimal model order combination is selected based on the Akaike information criterion and the Bayesian information criterion, so as to obtain the prediction parameter set based on the ARIMA model with the optimal model order combination.

[0051] A radial basis kernel support vector regression model is constructed based on the predicted parameter set, and the kernel function bandwidth parameter and penalty factor are set. The optimal hyperparameters of the radial basis kernel support vector regression model are determined through cross-validation, so as to predict the multidimensional sensitivity prediction sequence based on the radial basis kernel support vector regression model with the optimal hyperparameters.

[0052] The features of each dimension in the multidimensional sensitivity prediction sequence are standardized to obtain sensitivity prediction data, and the Euclidean distance between the sensitivity prediction data and the sampled data is calculated to obtain the residual.

[0053] Specifically, in this embodiment of the invention, the sampled data is processed to calculate the Pearson autocorrelation coefficient and partial autocorrelation coefficient for each dimension to identify patterns in the sampled data, such as periodicity and trends, and to generate a multidimensional sensitivity initial sequence. Then, the stationarity of the multidimensional sensitivity initial sequence is tested, such as the Augmented Dickey-Fuller test (ADF test), to ensure that the sequence is stationary, thereby ensuring the accuracy of the ARIMA model prediction. Multidimensional sensitivity time series analysis takes cross-departmental data access within an organization as a specific scenario. It monitors sampled data formed by three dimensions—query frequency, data volume, and computational complexity—in the sensitivity calculation feature data. Under normal conditions, a tenant's query frequency is 50 times per hour, the size of a single query result set is between 200 and 500 records, and the query execution time is between 100 and 300 milliseconds. After sampling through a sliding window, the autocorrelation coefficients for the query frequency dimension are calculated to be 0.82, the data volume dimension to be 0.75, and the complexity dimension to be 0.68, indicating that all three dimensions have significant time series correlations.

[0054] Next, based on the stationary multidimensional sensitivity sequence, an ARIMA(p,d,q) model is constructed, where p is the autoregressive order, d is the difference order (usually determined through a stationarity test), and q is the moving average order. The maximum likelihood estimation method is used to optimize the log-likelihood function, and the autoregressive coefficients and moving average coefficients are calculated to obtain the fitting parameter set. Specifically, when constructing the autoregressive difference-moving average model, the preferred log-likelihood function value obtained through maximum likelihood estimation is -856.4, the preferred diagonal elements of the autoregressive coefficient matrix are 0.72, 0.68, and 0.65, and the preferred diagonal elements of the moving average coefficient matrix are -0.35, -0.42, and -0.38.

[0055] Residual analysis was performed on the multidimensional stationary sensitivity series using an ARIMA model with a fitted parameter set. Then, the optimal combination of model orders (p, d, q) was selected using the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). This yielded a combination of autoregressive order p = 2, difference order d = 1, and moving average order q = 1. The prediction parameter set was then predicted based on the ARIMA model with the optimal combination of model orders.

[0056] A radial basis function (RBF) support vector regression model was constructed using the predicted parameter set as input. Specifically, a radial basis function kernel was used in the support vector regression, with a kernel function bandwidth parameter set to 0.8 and a penalty factor set to 10. The root mean square error obtained through 5-fold cross-validation was 0.086, thus determining the optimal hyperparameters. The RBF support vector regression model with the optimal hyperparameters was then used to predict the multidimensional sensitivity prediction sequence. The prediction results showed that the predicted query frequency for the next 4 hours was 45 to 55 times per hour, with an average deviation of 3 times from the actual observations; the predicted data volume was between 180 and 520 records, with an average deviation of 25 records; and the predicted query execution time was between 90 and 320 milliseconds, with an average deviation of 15 milliseconds.

[0057] After standardizing the features and prediction errors of each dimension in the multidimensional sensitivity prediction sequence, sensitivity prediction data is obtained. The residual is then calculated by determining the Euclidean distance between the sensitivity prediction data and the original sampled data. In data sharing scenarios, the above scheme also demonstrated good results. When using this method to monitor cross-departmental communication file queries, abnormal batch data export behavior was detected. The frequency of these queries surged to 150 times per hour, the amount of data returned per query reached three times the normal value, the execution time exceeded the prediction interval by two standard deviations, and the residual vector norm reached 0.412, far exceeding the warning threshold of 0.2. Timely detection and handling prevented unauthorized access to sensitive data. Furthermore, in other application scenarios, the above scheme detected an average of 12 sensitivity anomaly events per month, achieving an accuracy rate of 91%, demonstrating its effectiveness.

[0058] This invention combines the ARIMA model and the RBF support vector regression model, making full use of the inherent regularity of time series data and the nonlinear fitting ability of machine learning models to improve the accuracy of prediction. The use of AIC and BIC for model selection and optimization ensures the simplicity and predictive performance of the selected model. By processing non-stationary sequences through stationarity tests and differencing operations, the data meets the modeling requirements of the ARIMA model and has a wide range of applications.

[0059] In this embodiment of the invention, the root mean square error value of the residual is quantified, and the error value is judged by a preset anomaly judgment threshold. Tenants with error values ​​not less than the anomaly judgment threshold are identified as abnormal tenants. These abnormal tenants are offset and marked from a pre-established fluctuation range comparison table. The current tenant's data query sensitivity evaluation is marked as abnormal and thus an abnormal tenant.

[0060] S3. Using a sliding window, extract statistical features from the sensitivity calculation feature data of each abnormal tenant to construct a feature set, and perform cluster analysis on the feature set to obtain the abnormal pattern and abnormal fingerprint feature cluster corresponding to each abnormal tenant.

[0061] In one embodiment, step S3 includes:

[0062] A fixed-length sliding window is used to segment the sensitivity calculation feature data of each abnormal tenant. The frequency domain mapping of each segment is performed by discrete Fourier transform to obtain frequency domain statistical features, and then the maximum and minimum values ​​are normalized to obtain the first feature set.

[0063] The first feature set is dimensionality reduced, and the dimensionality reduction result is subjected to singular value decomposition to obtain a feature vector matrix. Feature vector components with a cumulative contribution rate exceeding a preset contribution threshold are selected from the matrix and combined using Bayesian probability weighting to obtain the second feature set.

[0064] The second feature set is standardized, and the similarity between data in the standardized second feature set is calculated using Mahalanobis distance. Based on the similarity calculation results, the optimal number of clusters is determined using the silhouette coefficient, thus obtaining the clustering parameter set.

[0065] A Gaussian mixture clusterer is constructed based on the clustering parameter set, and the covariance matrix and random initialization parameters are set. The cluster centers of the Gaussian mixture clusterer are iteratively optimized using the expectation-maximization algorithm to obtain the first clustering result.

[0066] The centroids and covariance matrices of the clusters are extracted from the first clustering result. An anomaly measurement function is constructed based on the density index and isolation index obtained by quantifying the centroids and covariance matrices of each cluster. The anomaly score of each cluster is quantified through the anomaly measurement function.

[0067] Set the abnormal pattern division range, divide each cluster into multiple abnormal patterns based on the relationship between the abnormal score and the abnormal pattern division range, and mark the abnormal fingerprint feature cluster corresponding to each abnormal pattern.

[0068] First, in this embodiment of the invention, the sensitivity calculation feature data of each abnormal tenant is segmented using a sliding window of fixed length. The preferred window length is set to 120 minutes, and the sliding step size is 10 minutes, in order to capture local features in the data and make subsequent analysis more detailed. Then, a Discrete Fourier Transform (DFT) is applied to each segment of data to map the time domain data to the frequency domain, thereby extracting two frequency domain statistical features, the amplitude spectrum and the phase spectrum, and combining them with three time domain statistics, the mean, variance and kurtosis, to perform maximum and minimum value normalization processing to obtain the first feature set.

[0069] Based on the characteristics of the data and business requirements, a suitable dimensionality reduction method is selected, such as Principal Component Analysis (PCA) or Linear Discriminant Analysis (LDA), to reduce the dimensionality of the first feature set, resulting in five main feature vectors. Feature components with a cumulative contribution rate exceeding a preset contribution threshold (preferably 92%) are selected from these vectors. The selected feature vectors are then combined using a Bayesian probability weighting method with weight coefficients of 0.35, 0.28, 0.18, 0.12, and 0.07, respectively, to construct a multidimensional fingerprint feature that characterizes tenant behavior, which is the second feature set.

[0070] The second feature set is standardized to eliminate the dimensional differences between different features. Then, the similarity between the standardized data is calculated, and Mahalanobis distance is used to characterize the correlation between features. Based on the similarity calculation results, the silhouette coefficient is used to evaluate the clustering effect under different numbers of clusters to determine the optimal number of clusters, and thus obtain the clustering parameter set.

[0071] A Gaussian mixture clusterer was constructed, and the covariance matrix and randomly initialized parameters were set. The cluster centers were iteratively optimized using the expectation-maximization (EM) algorithm to obtain the first clustering result. The covariance matrix of the Gaussian mixture clusterer was in diagonal form. The expectation-maximization algorithm converged when the relative error was less than 0.001. A total of 15 iterations were performed. The clustering results showed that the average distance from the sample to the center of the normal tenant cluster was 0.28, and the standard deviation was 0.06.

[0072] Key information, namely the centroids and covariance matrices of the clusters, is extracted from the first clustering results. It is found that the average distance from the samples to the centroids of anomalous clusters reaches 0.85, and the intra-cluster distance distribution exhibits a significant skewness. Next, based on the key information of the clusters, anomaly measurement functions are constructed using a compactness index (the reciprocal of the average intra-cluster distance) and an isolation index (the geometric distance between the centroid of the current cluster and the centroids of the nearest neighboring clusters). A 95% confidence interval upper limit of 0.75 is set as the anomaly judgment threshold to quantify the anomaly score of each cluster. Then, the anomaly pattern division range is set according to actual needs. Based on the anomaly scores, the clusters are divided into multiple anomaly patterns, which means classifying the clusters corresponding to each anomalous tenant and labeling each anomaly pattern with a corresponding anomaly fingerprint feature cluster for subsequent analysis and processing.

[0073] Applying the above scheme to user behavior analysis in telecom operators has demonstrated excellent identification results. For example, in a certain operator's 1000 enterprise tenants, three anomalous clusters involving 27 tenants were identified through multidimensional fingerprint feature clustering. These tenants' data access patterns exhibited frequent bursts of peaks, with a kurtosis of 5.2 in the time domain, significantly higher than the 2.3 of the normal clusters. The frequency domain features showed obvious high-frequency components, with the amplitude spectrum at 0.8 Hz being 3.5 times that of normal tenants. The average density index of the anomalous tenant clusters was 0.42, significantly lower than the 0.86 of the normal clusters, while the isolation index was as high as 0.95, indicating that the behavior patterns of these tenants deviated significantly from normal business processes. Through the comprehensive scoring of the anomaly measurement function, 12 risky tenant accounts were accurately identified, and the statistical feature distribution of their sensitivity sequences showed a significant deviation from historical patterns. This invention extracts more representative features and improves the accuracy of anomaly detection through steps such as frequency domain mapping, feature dimensionality reduction, and Bayesian probability weighting. It uses silhouette coefficients to determine the optimal number of clusters, constructs a Gaussian mixture clusterer, and iteratively optimizes the cluster centers, making the clustering results more reasonable and accurate. By constructing anomaly measurement functions and classifying anomaly patterns, it can clearly identify and classify different abnormal behavior patterns, providing convenience for subsequent processing.

[0074] S4. Obtain the normal fingerprint feature clusters of each of the abnormal tenants, and construct an abnormal pattern influence mechanism model by combining the abnormal fingerprint feature clusters. Then, quantify the influence of each abnormal pattern on the privacy protection level of each of the abnormal tenants through the abnormal pattern influence mechanism model to obtain quantitative analysis results.

[0075] In one embodiment, step S4 includes:

[0076] Obtain the normal fingerprint feature clusters of each of the abnormal tenants, and construct the overlap matrix by quantifying the Mahalanobis distance and cosine similarity of the normal fingerprint feature clusters and the abnormal fingerprint feature clusters in Euclidean space.

[0077] The minimum spanning tree algorithm is used to calculate the edge weights of the overlap matrix, and a feature association graph is constructed based on the edge weight calculation results. The node features in the feature association graph are extracted to construct a differential feature set.

[0078] Singular value decomposition is performed on the differential feature set, and the decomposition results are sorted to determine the main differential components. The importance of the main differential components is calculated by random forest to obtain the influence weight vector.

[0079] A directed acyclic graph is constructed based on the influence weight vector. The conditional dependencies between nodes in the directed acyclic graph are set, and the propagation probabilities between the nodes are extracted to construct a propagation strength matrix.

[0080] The propagation intensity matrix is ​​probabilistically inferred using a Bayesian network, and the conditional probability distribution between the nodes is calculated using the maximum likelihood estimation method. Based on the probabilistic inference results and the conditional probability distribution calculation results, an abnormal mode influence mechanism model is constructed.

[0081] The impact of the abnormal features corresponding to each abnormal pattern on the level of privacy protection is calculated based on the abnormal pattern impact mechanism model and used as an attenuation coefficient. A hierarchical structured privacy impact assessment index is constructed through the attenuation coefficient, and the hierarchical structured privacy impact assessment index is quantified to obtain a quantitative assessment vector.

[0082] Specifically, in this embodiment of the invention, fingerprint feature clusters of each abnormal tenant under normal conditions are collected and organized. These feature clusters represent the normal behavior patterns of the tenants, and the Mahalanobis distance and cosine similarity between normal and abnormal fingerprint feature clusters in Euclidean space are quantified. Finally, an overlap matrix is ​​constructed based on the calculated Mahalanobis distance (quantifying the correlation between features) and cosine similarity (measuring the directional similarity between feature vectors). The rows and columns of the matrix represent normal and abnormal fingerprint feature clusters, respectively, so that the constructed overlap matrix can more accurately reflect the similarity and difference between normal and abnormal fingerprint feature clusters. For example, in a cross-tenant data access scenario within an organization, the fingerprint features of a payment business tenant exhibit significant abnormal characteristics. During normal periods, the tenant's feature vector distribution shows low frequency and small fluctuations, with a mean standard deviation of 0.15 across the 12 dimensions of the feature space. However, during abnormal periods, the feature vectors show significant shifts in the three dimensions of frequency, amplitude, and duration, with the mean standard deviation rising to 0.42. The Mahalanobis distance between these two feature clusters yields a difference of 1.85, a cosine similarity of 0.62, and a feature cluster overlap of only 0.28.

[0083] Edge weights are calculated on the overlap matrix, with weights based on Mahalanobis distance or the inverse of cosine similarity, to reflect the degree of similarity between features. Then, the overlap matrix is ​​treated as an adjacency matrix of a weighted undirected graph, where nodes represent features and the calculated edge weights represent the overlap. The minimum spanning tree algorithm is then applied to this graph to obtain a minimum spanning tree, which is used to construct a feature association graph. In this graph, nodes represent features and edges represent the association strength determined by the minimum spanning tree. Next, the node features extracted from the feature association graph are combined to form a differential feature set, which contains key features that can distinguish between normal and abnormal behavior patterns.

[0084] Next, singular value decomposition is performed on the constructed differential feature set to obtain feature vectors and singular values. The feature vectors are then sorted according to the magnitude of the singular values, and the feature vectors with larger singular values ​​are selected as the main differential components. For example, if the top three feature values ​​after the differential feature set decomposition for the cross-tenant data access scenario account for 85% of the total variance, they can be identified as the main differential components. Then, the importance of these main differential components is calculated using random forest to obtain the influence weight vectors, which are 0.42, 0.35, and 0.23, respectively.

[0085] In power grid data sharing scenarios, the propagation of anomalous features exhibits a clear hierarchical structure. Therefore, this invention employs a directed acyclic graph (DAG) to characterize the dependencies between features, with the conditional probabilities between nodes reflecting the intensity of influence propagation. Specifically, a DAG is constructed based on influence weight vectors. Nodes in the graph represent anomalous features and their corresponding anomalous patterns, and edges represent the conditional dependencies between features. Based on prior knowledge, the conditional dependencies between nodes in the DAG (i.e., the influence relationships between anomalous features) are defined. Probability theory, statistics, or machine learning methods are used to extract the propagation probabilities between nodes from the DAG and organize them into a matrix. Rows and columns represent nodes, and matrix elements represent the propagation intensity (i.e., the propagation probability) between corresponding nodes, resulting in a propagation intensity matrix. Taking a power grid's communication archive query service as an example, anomalous query behavior first causes a change in frequency characteristics with a conditional probability of 0.85, then propagates to data volume characteristics with a conditional probability of 0.72, and finally affects the time distribution characteristics, with the conditional probability decreasing to 0.45.

[0086] This invention employs a Bayesian network to perform probabilistic inference on the propagation intensity matrix, calculates the conditional probability distribution among nodes, and uses the maximum likelihood estimation method to calculate the conditional probability distribution parameters among nodes. Combining the probabilistic inference results and the conditional probability distribution, an anomalous pattern influence mechanism model is constructed, which can also be considered an influence propagation model. Specifically, the probabilistic inference results from the Bayesian network show that anomalous features (such as frequency features and temporal distribution features) have a cumulative effect on the privacy protection level.

[0087] For each anomalous pattern, its impact on the level of privacy protection is calculated based on its path and node weights in the model. This can be achieved by simulating the propagation process of the anomalous pattern in the model and observing its impact on the privacy protection level assessment criteria. The degree of impact can be represented by a quantified value, such as a number between 0 and 1, where 0 represents no impact and 1 represents the maximum impact. This value is used as a decay coefficient to reflect the degree to which the anomalous pattern weakens the level of privacy protection. Then, based on multiple aspects of privacy protection (such as data collection, storage, processing, and transmission), a hierarchical structured privacy impact assessment index is designed. Each layer of index corresponds to one aspect of privacy protection, and the next layer of index is a refinement or specific manifestation of the previous layer's index. For each index, a value is assigned according to the decay coefficient of its corresponding anomalous pattern. All indicators are organized in a hierarchical structure to form a complete privacy impact assessment system. Finally, the quantified value of each index in the hierarchical structured privacy impact assessment index is used as an element of a vector to construct a quantified assessment vector. The dimension of the quantified assessment vector is equal to the number of indicators in the assessment system.

[0088] Statistical distribution analysis was performed on the quantitative assessment vectors to calculate the mean and variance of the impact degree of different tenant groups. A warning threshold based on confidence intervals was set to obtain risk level labels. Applying this method to power grid communication data access, the data anonymization intensity decreased by 15%, and the privacy protection level of query results decreased by 28%. This impact diffused to related tenants through the network structure, with the average protection level of first-level related tenants decreasing by 18% and second-level related tenants by 12%. The statistical distribution of the quantitative assessment vectors showed significant hierarchical characteristics: the mean impact degree of the directly related tenant group was 0.72, with a variance of 0.15; the mean impact degree of the indirectly related tenant group was 0.45, with a variance of 0.23. A warning threshold of 0.65 was set based on a 95% confidence interval. Tenants exceeding this threshold were marked as high-risk. This successfully identified a data access violation event involving 15 related tenants, with 8 tenants exceeding the warning threshold, forming a typical spreading risk propagation pattern.

[0089] This invention can more accurately identify the characteristics of anomalous tenants by constructing a feature association graph and a differential feature set; it can reveal the propagation path and intensity of anomalous patterns among different features by constructing a directed acyclic graph and a propagation intensity matrix; and it can quantify the overall impact of anomalous patterns on the level of privacy protection by constructing a hierarchical structured privacy impact assessment index and a quantitative assessment vector, so as to facilitate the subsequent adjustment of the privacy budget allocation ratio and thus solve the anomalous problem in the ciphertext execution of privacy computing algorithms.

[0090] S5. Adjust the privacy budget allocation ratio of each abnormal tenant according to the quantitative analysis results to obtain a privacy budget allocation scheme, and execute the privacy budget allocation scheme;

[0091] In one embodiment, step S5 includes:

[0092] A risk assessment function is constructed based on the quantitative analysis results to score the risk of each abnormal tenant, and the correlation between each abnormal tenant is determined based on the relationship between the risk score results and the preset risk threshold.

[0093] The influence propagation coefficient of each abnormal tenant is quantified using the exponential weighting method, and an association strength matrix is ​​constructed by combining the influence propagation coefficient and the association relationship. The association strength matrix is ​​then subjected to feature decomposition to obtain the tenant association index.

[0094] A linear decay function is set according to the tenant correlation index to classify the budget reduction ratio, and a normal distribution function is used to non-linearly reduce the privacy budget allocation ratio of each abnormal tenant to obtain the first budget adjustment set.

[0095] The compensation coefficients for each abnormal tenant are calculated using the least squares method based on the first budget adjustment set, and a compensation weight vector is set in combination with the data query sensitivity corresponding to each abnormal tenant to obtain the second budget adjustment set.

[0096] The second budget adjustment set is optimized using gradient descent to obtain a privacy budget allocation scheme for each of the abnormal tenants, and the privacy budget allocation scheme is then executed.

[0097] Specifically, this invention aims to identify the potential threats to privacy protection levels posed by each anomalous tenant and quantify the degree of such threats. It comprehensively reflects tenant performance in privacy protection, using metrics such as the number of privacy breaches, the scale of privacy breaches, and the impact of privacy breaches as evaluation indicators. A risk assessment function is constructed by weighted summation based on the quantitative analysis results (i.e., the overall impact of each anomalous pattern on the level of privacy protection) and these indicators. The weights can be determined through expert scoring, questionnaires, data analysis, etc., without specific limitations, allowing the risk assessment function to calculate a risk score for each anomalous tenant. Then, the calculated risk scores are compared with preset risk thresholds to determine the relationships between anomalous tenants. These relationships can be represented using graph theory, with anomalous tenants as nodes and relationships as edges. In cross-tenant data access scenarios, the impact of anomalous tenants is quantified through the risk assessment function. Taking a payment service tenant as an example, its risk score is 0.85, higher than the preset threshold of 0.7. Comparing the risk scores of each anomalous tenant with the preset risk thresholds determines that its impact involves 15 directly related tenants and 25 indirectly related tenants.

[0098] An exponential weighting method is used to quantify the influence propagation coefficient of each abnormal tenant. This method considers factors such as transaction frequency and volume between tenants. In the cross-tenant data access scenario, the influence propagation coefficients calculated by exponential weighting are 0.75, 0.45, and 0.25, corresponding to first-level, second-level, and third-level associated tenants, respectively. The influence propagation coefficients are then combined with their corresponding association relationships to construct an n×n association strength matrix A, where n is the number of abnormal tenants. The element a_ij in the association strength matrix A represents the influence strength of abnormal tenant i on abnormal tenant j, which can be determined based on the influence propagation coefficient and the tightness of the association relationship. This matrix reflects the association strength and influence degree between abnormal tenants. Next, the association strength matrix A is subjected to eigenvalue decomposition to obtain eigenvalues ​​and eigenvectors. Based on the magnitude and importance of the eigenvalues, the first k eigenvalues ​​(k is much smaller than n) and the corresponding eigenvectors are selected to calculate the association degree index of each tenant. This can be achieved by linearly combining the eigenvectors with the tenant's influence propagation coefficients. The combination coefficient can be an eigenvalue, and the association degree index reflects the position and importance of the abnormal tenant in the overall association network.

[0099] Based on the quantified correlation index, a linear decay function is set, such as setting a baseline budget reduction ratio. According to the correlation index score, the budget is linearly reduced according to a certain ratio to obtain the linear decay function. The budget reduction ratio is divided into different levels. For example, the budget reduction ratio of direct related tenants in the first interval is 35%, that of second-level related tenants in the second interval is 20%, and that of third-level related tenants in the third interval is 10%, with a total budget reduction of 0.65, of which 0.45 is used to compensate for the privacy protection of normal tenants, etc. The specific interval range is obtained according to actual needs or experience. Next, the parameters (mean and standard deviation) of the normal distribution function are set. These parameters can be determined according to business needs and historical data. The normal distribution function is used to non-linearly reduce the privacy budget allocation ratio set for each abnormal tenant according to its historical privacy protection performance, business scale, data sensitivity, etc., to obtain the first budget adjustment set. For example, the initial privacy budget of abnormal tenants in the cross-tenant data access scenario is reduced from 1.0 to 0.35. The specific reduction amount is distributed among related tenants at different levels through the normal distribution function.

[0100] Based on the first budget adjustment set, the compensation coefficient for each abnormal tenant is calculated using the least squares method, ranging from 0.2 to 0.8, to ensure the overall budget balance. Then, considering the data query sensitivity corresponding to each abnormal tenant (which can also be considered the data query sensitivity level), a compensation weight vector is set to adjust the compensation coefficient, resulting in the second budget adjustment set. For example, the compensation weight for high-sensitivity abnormal tenants (such as power plant output power, substation load status access, etc.) is 0.8, the compensation weight for medium-sensitivity abnormal tenants (such as user electricity consumption patterns, historical energy consumption records, etc.) is 0.5, and the compensation weight for low-sensitivity abnormal tenants (such as power company announcements, etc.) is 0.2, thus obtaining the second budget adjustment set.

[0101] The second budget adjustment set is optimized using gradient descent, with the objective of limiting compensation for a single tenant to no more than 50% of the original budget and the total compensation to no more than 70% of the total reduction. During iterative optimization, the convergence threshold of the objective function is set to 0.001, and the maximum number of iterations is 1000, ultimately yielding a compensation budget allocation vector. Next, a Lagrangian function is constructed based on this vector, with non-negativity and smoothing constraints (the compensation coefficient for adjacent abnormal tenants should not exceed 0.1). Quadratic programming is then used to solve the objective function to obtain the optimal compensation scheme. Residual analysis is then performed on the optimal compensation scheme to calculate the deviation distribution between the adjusted budget and the original budget. Compensation values ​​exceeding the threshold are corrected by setting a tolerance range. Finally, the tenant privacy budget allocation table is updated according to the corrected compensation scheme. Exponential smoothing with a smoothing coefficient of 0.3 is used to dynamically track the budget adjustment results, resulting in the final privacy budget allocation scheme, which is then executed.

[0102] In practical applications, abnormal tenant behavior occurred in a bank's credit data access scenario. After adopting the above solution, the abnormal tenant's original budget of 1.0 was reduced to 0.4, the budget of its 8 directly related tenants was reduced by an average of 0.3, and the budget of its 12 second-level related tenants was reduced by an average of 0.15. Through optimization calculation, 20 normal tenants received compensation, with compensation values ​​distributed between 0.02 and 0.08, and a total compensation amount of 0.42. Considering the data access frequency, high-frequency tenants received larger compensation to maintain query accuracy. The relative error after compensation was controlled within 5%. After continuous monitoring for 30 time windows, the fluctuation range of the final budget allocation scheme was controlled within ±0.05, demonstrating good stability. This achieved precise suppression of abnormal tenants and differentiated compensation for normal tenants, ensuring a dynamic balance of differential privacy protection in a multi-tenant environment.

[0103] This invention, through the construction of a risk assessment function and a correlation strength matrix, can more accurately assess the risks and correlations of abnormal tenants, providing a reliable basis for subsequent budget adjustments. Employing a non-linear reduction and compensation mechanism, it can allocate the privacy budget more rationally, ensuring improved computational efficiency and accuracy while protecting data privacy. Optimizing the privacy budget allocation scheme using gradient descent can handle various anomalies, improving the algorithm's robustness and adaptability. While protecting data privacy, it can also effectively handle abnormal tenants, preventing privacy leaks and data misuse, thus contributing to the maintenance of data security and user rights.

[0104] In one embodiment, after optimizing the second budget adjustment set using gradient descent to obtain and execute the privacy budget allocation scheme for each of the abnormal tenants, the process includes:

[0105] The sensitivity sequence during the execution of the privacy budget allocation scheme is extracted using the differential privacy query function. The sensitivity sequence is then subjected to frequency domain analysis using fast Fourier transform to obtain the feature fingerprint and singular value decomposition to obtain the first feature vector.

[0106] The similarity between the first feature vector and the pre-stored abnormal fingerprint feature cluster is calculated using Mahalanobis distance to construct a similarity matrix. The similarity matrix is ​​then normalized by setting a distance threshold to obtain the second feature vector.

[0107] Spectral clustering is performed on the second feature vector, and the feature space of the second feature vector is divided into clusters by setting the cluster radius and the minimum number of samples, so as to calculate the local density and relative distance of each cluster and obtain the density distribution matrix.

[0108] Extract the cluster center point set from the density distribution matrix, construct a feature topology graph based on the cluster center point set and inter-cluster connectivity, and perform hierarchical labeling on the feature topology graph to obtain a hierarchical structure matrix;

[0109] The budget adjustment amount for each abnormal tenant is calculated based on the hierarchical structure matrix, and the budget adjustment amount is limited by an exponential decay function. The budget adjustment amount is weighted by the tenant correlation index to obtain the budget adjustment plan.

[0110] A compensation resource allocation network is constructed based on the budget adjustment scheme, and network traffic constraints and capacity constraints are set. The compensation resource allocation network is solved by the maximum flow minimum cut algorithm to obtain the optimal compensation allocation scheme, and the optimal compensation allocation scheme is executed to realize the handling of anomalies in the ciphertext execution of the privacy computing algorithm.

[0111] Specifically, in this embodiment of the invention, the sensitivity sequence during the execution of the privacy budget allocation scheme is extracted according to the differential privacy query function, that is, the sensitivity calculation feature data after the execution of the privacy budget allocation scheme. The specific calculation and extraction process can be referred to above, and will not be repeated here. Then, the sensitivity sequence is transformed in the frequency domain using Fast Fourier Transform (FFT) to identify the periodic or trend features. The output of FFT is the feature fingerprint, which contains the information of the sensitivity sequence in the frequency domain, namely the amplitude spectrum and phase spectrum. It is used as the feature fingerprint to perform singular value decomposition to obtain the first feature vector.

[0112] The similarity between the first feature vector and the pre-stored abnormal feature clusters is calculated by Mahalanobis distance to construct a similarity matrix, which is used to represent the closeness between the first feature vector and the feature cluster. After normalizing the similarity matrix by setting a distance threshold, a clear clustering structure is formed in the high-dimensional feature space, and the second feature vector is obtained.

[0113] Next, spectral clustering was performed on the second feature vector, and the feature space of the second feature vector was divided into clusters with a cluster radius of 0.3 and a minimum sample size of 5. The local density and relative distance of each cluster were calculated to obtain a density distribution matrix to reflect the compactness and distribution characteristics between clusters. The density distribution showed that the local density of the abnormal clusters was 0.85, which was significantly higher than that of the normal clusters (0.32), and the relative distance between clusters was 2.4, indicating that the abnormal features had significant clustering. In addition, a similar phenomenon was observed in the power grid communication data access scenario. When an anomaly occurred in a certain communication file query, the local density reached 0.92, forming a typical high-density abnormal cluster.

[0114] Based on the density distribution matrix, a certain strategy (such as selecting the point with the highest local density or using a comprehensive index of density and distance) is employed to determine the cluster center point of each cluster. These determined cluster centers are then combined to form a cluster center point set. Each cluster center point is then assigned a node. Edges are defined based on inter-cluster connectivity; that is, if two cluster centers satisfy a certain connectivity condition under a certain metric (such as distance, similarity, etc.), an edge is drawn between these two nodes. To describe the inter-cluster relationships more precisely, a weight can be assigned to each edge, which can be calculated based on the distance, similarity, or other relevant indices between cluster centers. Finally, based on the defined nodes and edges... The process involves constructing a feature topology graph using graph theory tools or graph data structures in programming languages, along with weights. A suitable hierarchical algorithm, such as hierarchical clustering or community detection, is then selected to identify the hierarchical structure in the feature topology graph and assign nodes to different levels. The feature topology graph is used as input to execute the chosen hierarchical algorithm, outputting the level information for each node. Based on the output of the hierarchical algorithm, a hierarchical structure matrix is ​​constructed, where rows and columns correspond to nodes in the graph, and elements represent the hierarchical relationships between nodes. Specifically, if two nodes belong to the same level, the corresponding element is 1 (or some non-zero value); otherwise, it is 0. Furthermore, more information about hierarchical relationships, such as level depth and parent-child relationships, can be recorded in the hierarchical structure matrix as needed. Finally, the hierarchical structure matrix is ​​verified and adjusted to ensure it accurately reflects the hierarchical structure in the feature topology graph, and may be refined or simplified as needed.

[0115] Based on the hierarchical structure matrix, a basic budget adjustment amount is set for tenants at different levels. The adjustment amount is determined according to the level of the tenant, the degree of anomaly, or other relevant factors. For each anomalous tenant, a specific budget adjustment amount is calculated based on its level and the basic adjustment amount. A suitable exponential decay function is selected, and the budget adjustment amount for each anomalous tenant is used as input. The selected exponential decay function is applied to calculate the final budget adjustment amount for each anomalous tenant. These adjustment amounts should reflect both the degree of impact of the tenant in the anomaly and meet the requirements of smoothness and controllability of budget constraints. Then, based on the tenant correlation index, a budget adjustment amount is calculated for each tenant... A weighting coefficient is determined for each tenant, or the quantified correlation value of the tenant correlation index can be used as the weighting coefficient to reflect the relative importance of the tenant in the budget adjustment process. For example, tenants closely associated with multiple abnormal tenants may face a larger weighting coefficient. The final budget adjustment amount for each abnormal tenant is multiplied by its weighting coefficient to obtain the weighted budget adjustment amount, which makes the budget adjustment more in line with the actual relationship between tenants and business needs. Finally, the weighted budget adjustment amounts for each abnormal tenant are summarized to obtain the overall budget adjustment plan. This plan should list in detail the budget adjustment amount for each tenant, the total adjusted budget amount, and other information. In cross-departmental data sharing within banks, this tiered adjustment strategy effectively controls the spread of risk; the budget for abnormal tenants decreased from 1.0 to 0.4, and the average budget for directly associated tenants was adjusted from 0.8 to 0.6.

[0116] A compensation resource allocation network is constructed based on the budget adjustment scheme. This network includes source nodes (representing resource providers or budget sources), abnormal tenant nodes (representing recipients of resources or budgets requiring compensation), sink nodes (representing the endpoint of resources or budgets, potentially used to indicate overall budget balance or constraints), and edges representing the flow paths of resources or budgets between nodes. Each edge has capacity and network traffic constraints, representing the maximum capacity and maximum transmission volume of resources or budgets along that path. Next, the compensation resource allocation network is solved using a maximum flow minimum cut algorithm (such as the Ford-Fulkerson algorithm or the Edmonds-Karp algorithm) to obtain a compensation allocation scheme. Residual analysis is performed on the compensation allocation scheme to calculate the budget change before and after compensation. A fluctuation constraint threshold is set based on business needs and risk tolerance to obtain a compensation execution sequence. The compensation scheme with the smallest budget change and compliance with fluctuation constraints is selected as the optimal compensation allocation scheme, and this optimal scheme is executed to handle anomalies during the ciphertext execution of the privacy computation algorithm.

[0117] In the process of accessing user data by telecommunications operators, this solution successfully handled anomalies caused by batch data export. After anomalies were detected through real-time feature matching, budget adjustments and compensation allocations were completed within 5 minutes, effectively protecting the privacy data of 3 million users. It achieved rapid identification and accurate response to abnormal behavior while ensuring the continuity of normal business operations. This invention, through multi-step processing including frequency domain analysis, feature extraction, and similarity calculation, can more accurately identify anomalies in the privacy budget allocation process. It uses an exponential decay function and tenant correlation indicators to limit and weight budget adjustments, making budget adjustments more reasonable and fair. By constructing a compensation resource allocation network and using the maximum flow minimum cut algorithm for solution, it ensures optimal resource allocation and improves resource utilization efficiency. It can promptly detect and handle anomalies in the privacy budget allocation process, thereby enhancing the encrypted execution capability of the privacy computation algorithm and improving the level of data privacy protection.

[0118] In one embodiment, after optimizing the second budget adjustment set using gradient descent to obtain and execute the privacy budget allocation scheme for each of the abnormal tenants, the method further includes:

[0119] The sensitivity sequence is decomposed at multiple scales. High-frequency coefficients, low-frequency coefficients and energy distribution features are extracted from the decomposition results to construct a decomposed feature vector. The normalized cosine distance between the decomposed feature vector and the pre-stored abnormal fingerprint feature cluster is quantified to obtain a matching degree matrix.

[0120] Anomaly classification of the matching degree matrix is ​​performed using Gaussian kernel support vector machine to obtain a first risk assessment value to construct a directed weighted graph. The second risk assessment value is obtained by calculating the in-degree and out-degree of the nodes in the directed weighted graph and obtaining the secondary influence propagation coefficient to quantify the information entropy loss of each of the abnormal tenants.

[0121] The secondary compensation weight value is calculated based on the second risk assessment value, and the secondary compensation weight value is optimized by the total budget conservation constraint and the single tenant compensation upper limit constraint to obtain the optimal budget adjustment scheme for each abnormal tenant. The optimal budget adjustment scheme is then executed to handle the anomalies in the ciphertext execution of the privacy computing algorithm.

[0122] Specifically, in this embodiment of the invention, the sensitivity sequence is decomposed using a db4 wavelet basis at four levels. High-frequency coefficients (representing detailed information), low-frequency coefficients (representing the overall trend), and energy distribution features (such as the energy proportion at each scale) are extracted from the decomposition results. These extracted data are then combined to form a decomposed feature vector. Next, a cosine similarity metric is used to calculate the similarity between the decomposed feature vector and pre-stored abnormal fingerprint feature clusters. Standardization is then performed to construct a matching degree matrix representing the degree of matching between the decomposed feature vector and the abnormal fingerprint feature clusters based on the calculated standardized cosine distance. Taking cross-tenant data access on a payment platform as an example, after wavelet decomposition, an abnormal high-frequency component is detected at level 3, with an energy proportion reaching 0.45, significantly higher than the normal 0.12. The mean of the low-frequency coefficients increases from 0.35 to 0.82, indicating a persistent abnormality in the query behavior. The standardized cosine distance between the feature vector and the known abnormal fingerprint is 0.15, far below the normal threshold of 0.65.

[0123] The kernel function parameter of the Gaussian kernel support vector machine was set to 0.8, and the penalty factor was set to 10. The classification accuracy obtained through cross-validation was 0.92. This accuracy was used to classify anomalies in the matching degree matrix. One or more risk assessment indicators were defined according to business requirements. These indicators can be determined based on classification results (such as the proportion of anomalous tenants), statistical properties of feature vectors (such as mean and variance), or in combination with other business information. Then, based on the defined risk assessment indicators, a first risk assessment value was calculated. This value can be a numerical value representing the risk level of the anomalous tenant. Next, a directed weighted graph was constructed with the first risk assessment value as an edge and each anomalous tenant as a node. By calculating the in-degree and out-degree of the nodes in the graph, the secondary influence propagation coefficient was obtained to quantify the influence of the node in the propagation of anomalies. Finally, the information entropy loss of each anomalous tenant was quantified using the secondary influence propagation coefficient and the specific characteristics of the anomalous tenant (such as data volume and sensitivity). The information entropy losses of all anomalous tenants were summarized or weighted averaged to obtain a second risk assessment value, which represents the overall risk level of the entire system when facing anomalous tenants.

[0124] The second risk assessment value is standardized or normalized to fall within a reasonable range. An initial secondary compensation weight is assigned to each anomalous tenant based on this value. This weight can be a direct mapping of the second risk assessment value or adjusted appropriately according to business logic. Next, the total compensation budget is set to 120% of the original budget, and the maximum compensation each anomalous tenant can receive is set to 50% of the original budget. An optimization model is constructed using the secondary compensation weight as a variable and maximizing the overall satisfaction of all tenants (based on a combination of compensation amount and risk assessment value) as the objective. Then, an optimization algorithm is used to optimize the secondary compensation weight, finding the optimal weight combination that satisfies the constraints and optimizes the objective function through iterative calculation. The optimal secondary compensation weight for each anomalous tenant is extracted from the optimization results. Based on the calculated weight and total budget, the compensation amount each anomalous tenant should receive is calculated. The budget for each anomalous tenant is then adjusted based on the calculated compensation amount. This involves negotiation with tenants and signing compensation agreements, which are not the focus of this solution and will not be detailed here. This process yields the optimal budget adjustment plan for each anomalous tenant.

[0125] This invention, through the combined use of standardized cosine distance and Gaussian kernel support vector machine, can quickly and accurately identify anomalous data. Furthermore, by constructing a directed weighted graph and calculating the secondary influence propagation coefficient, the risk of anomalous data can be assessed more accurately. By optimizing the secondary compensation weight values ​​and obtaining the optimal budget adjustment scheme, the reasonable allocation and effective utilization of the budget can be ensured. Finally, by processing anomalous data and implementing the budget adjustment scheme, the security and stability of the privacy computing algorithm in the face of anomalous situations can be improved.

[0126] This application embodiment addresses the issue of detecting and handling sensitivity anomalies in tenants during differential privacy query execution in a multi-tenant cloud environment. It designs a method for handling encrypted execution anomalies in a privacy computation algorithm. This method involves acquiring differential privacy query parameters from multiple tenants and quantifying the data query sensitivity of each tenant based on these parameters to obtain sensitivity calculation feature data. Time series analysis is then performed on the sensitivity calculation feature data using an ARIMA model to obtain sensitivity prediction data. The residual between the sensitivity calculation feature data and the sensitivity prediction data is quantified to assess whether each tenant exhibits sensitivity shift anomalies, thus identifying anomalous tenants. Finally, a sliding window is used to extract each... The statistical features in the sensitivity calculation feature data of the abnormal tenants are used to construct a feature set, and cluster analysis is performed on the feature set to obtain the abnormal patterns and abnormal fingerprint feature clusters corresponding to each abnormal tenant. Normal fingerprint feature clusters of each abnormal tenant are obtained, and an abnormal pattern influence mechanism model is constructed by combining the abnormal fingerprint feature clusters. The influence mechanism model is then used to quantify the degree of impact of each abnormal pattern on the privacy protection level of each abnormal tenant, resulting in a quantitative analysis result. Based on the quantitative analysis result, the privacy budget allocation ratio of each abnormal tenant is adjusted to obtain a privacy budget allocation scheme, which is then executed to achieve effective detection and processing of anomalies during the ciphertext execution of the privacy computation algorithm.

[0127] In one embodiment, the encrypted execution exception handling method of the privacy computation algorithm further includes: in response to a query request for information stored by a tenant, determining target noise based on a differential privacy algorithm and the privacy budget corresponding to the tenant; generating a query result based on the query request and the target noise, and returning the query result.

[0128] This embodiment returns noisy query results when a tenant submits a differential privacy query, effectively protecting data privacy.

[0129] It should be noted that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order requirement for the execution of these steps, and they can be executed in other orders.

[0130] In another embodiment, as shown in FIG2, a second aspect of the present invention provides a ciphertext execution exception handling system for a privacy computing algorithm, comprising:

[0131] The feature data quantization module 10 is used to obtain differential privacy query parameters of multiple tenants, and quantify the data query sensitivity of each tenant according to the differential privacy query parameters to obtain sensitivity calculation feature data;

[0132] The abnormal tenant assessment module 20 is used to perform time series analysis on the sensitivity calculation feature data through the ARIMA model to obtain sensitivity prediction data, and quantify the residual between the sensitivity calculation feature data and the sensitivity prediction data to assess whether each tenant has a sensitivity shift anomaly and to identify abnormal tenants.

[0133] The abnormal pattern segmentation module 30 is used to extract statistical features from the sensitivity calculation feature data of each abnormal tenant using a sliding window to construct a feature set, and to perform cluster analysis on the feature set to obtain the abnormal pattern and abnormal fingerprint feature cluster corresponding to each abnormal tenant.

[0134] The impact degree quantification module 40 is used to obtain the normal fingerprint feature clusters of each of the abnormal tenants, to construct an abnormal pattern impact mechanism model by combining the abnormal fingerprint feature clusters, and to quantify the impact degree of each of the abnormal patterns on the privacy protection level of each of the abnormal tenants through the abnormal pattern impact mechanism model, so as to obtain the quantitative analysis results.

[0135] The allocation scheme execution module 50 is used to adjust the privacy budget allocation ratio of each abnormal tenant according to the quantitative analysis results, obtain a privacy budget allocation scheme, and execute the privacy budget allocation scheme.

[0136] It should be noted that each module in the aforementioned encrypted execution exception handling system for a privacy computation algorithm can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module. For specific limitations regarding the encrypted execution exception handling system for a privacy computation algorithm, please refer to the limitations regarding the encrypted execution exception handling method for a privacy computation algorithm described above; both have the same function and role, and will not be repeated here.

[0137] In summary, this invention relates to the field of information technology and discloses a method and system for handling anomalies in the encrypted execution of a privacy computing algorithm. It quantifies the data query sensitivity of each tenant by collecting differential privacy query parameters from multiple tenants, and uses an ARIMA model to perform time-series analysis on the quantified data query sensitivity results to predict future sensitivity trends. The residual between the actual sensitivity and the predicted sensitivity is calculated to identify anomalous tenants. A sliding window method is used to extract statistical features of anomalous tenants to construct a feature set, and cluster analysis is performed to identify the anomalous patterns and anomalous fingerprint feature clusters corresponding to each anomalous tenant. A model is trained based on the positive and anomalous fingerprint features of anomalous tenants, and this model is used to quantify the impact of each anomalous pattern on the privacy protection level of the anomalous tenant. The privacy budget allocation ratio of each anomalous tenant is adjusted according to the obtained quantitative analysis results and then executed, thereby achieving effective handling of anomalies in the encrypted execution of the privacy computing algorithm.

[0138] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0139] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method for handling ciphertext execution exceptions in a privacy-preserving computation algorithm, characterized in that, include: Obtain differential privacy query parameters for multiple tenants, and quantify the data query sensitivity of each tenant based on the differential privacy query parameters to obtain sensitivity calculation feature data; The sensitivity calculation feature data is analyzed by time series analysis using the ARIMA model to obtain sensitivity prediction data. The residual between the sensitivity calculation feature data and the sensitivity prediction data is quantified to assess whether each tenant has sensitivity shift anomalies and to identify anomalous tenants. A sliding window is used to extract statistical features from the sensitivity calculation feature data of each abnormal tenant to construct a feature set, and cluster analysis is performed on the feature set to obtain the abnormal pattern and abnormal fingerprint feature cluster corresponding to each abnormal tenant. Obtain the normal fingerprint feature clusters of each of the abnormal tenants, and construct an abnormal pattern influence mechanism model by combining the abnormal fingerprint feature clusters. Then, quantify the impact of each abnormal pattern on the privacy protection level of each of the abnormal tenants through the abnormal pattern influence mechanism model to obtain quantitative analysis results. Based on the quantitative analysis results, the privacy budget allocation ratio of each abnormal tenant is adjusted to obtain a privacy budget allocation scheme, and the privacy budget allocation scheme is executed.

2. The method for handling encrypted execution anomalies in a privacy-preserving computation algorithm according to claim 1, characterized in that, The differential privacy query parameters include data difference values ​​and weight coefficients; wherein, The process of obtaining differential privacy query parameters for multiple tenants and quantifying the data query sensitivity of each tenant based on these parameters to obtain sensitivity calculation feature data includes: The data difference values ​​and weight coefficients of multiple tenants are obtained, and the first tenant isolation boundary value is obtained by performing a dot product operation and normalization on the data difference values ​​and the weight coefficients to construct a tenant difference parameter matrix; each row of the tenant difference parameter matrix contains the tenant identifier and its corresponding data query sensitivity; A data query sensitivity threshold is set according to the tenant differential parameter matrix. Each tenant is grouped and clustered according to the data query sensitivity threshold to obtain multiple tenant groups with different data query sensitivity levels. The differential privacy query parameters of each tenant group are quantized and encoded according to a preset desensitization rule to obtain a first feature space. Gradient boosting trees are used to associate and map the parameter range in the first feature space with the data query sensitivity level, outputting a parameter mapping vector. Random forest features are then extracted from the parameter mapping vector to obtain the second feature space. Obtain the intermediate index for global sensitivity calculation in the second feature space, process the intermediate index for global sensitivity calculation using an adaptive clustering algorithm to obtain a set of cluster center points, and construct a feature vector dimension matrix based on the set of cluster center points; The feature vector dimension matrix is ​​orthogonally transformed to obtain a standardized feature vector. The standardized feature vector is then weighted and summed with the first tenant isolation boundary value to generate a differential privacy query parameter set as the sensitivity calculation feature data.

3. The method for handling encrypted execution anomalies in a privacy computation algorithm according to claim 1, characterized in that, The step of performing time series analysis on the sensitivity calculation feature data using the ARIMA model to obtain sensitivity prediction data, and quantifying the residual between the sensitivity calculation feature data and the sensitivity prediction data, includes: The sensitivity calculation feature data is sampled at fixed intervals by a sliding time window to obtain sampled data, and the time series feature matrix is ​​constructed by quantifying the autocorrelation function and partial autocorrelation function of the sampled data. The time series feature matrix is ​​subjected to stationarity processing to obtain a stationary time series dataset for constructing an ARIMA model. The autoregressive term coefficients and moving average term coefficients of the ARIMA model are estimated by the least squares method, and the goodness-of-fit index is obtained by performing a significance test on the parameters of the ARIMA model. The optimal parameter combination is selected by the goodness-of-fit index, and the ARIMA model with the optimal parameter combination is subjected to multi-step iterative prediction on the validation dataset to calculate the confidence interval of the prediction results and obtain the sensitivity prediction sequence. A Kalman filter is constructed based on the sensitivity prediction sequence, a state transition matrix and an observation matrix are set, sensitivity prediction data is obtained through prediction update and measurement update, and the residual is calculated based on the sensitivity prediction data and the sampled data.

4. The method for handling encrypted execution anomalies in a privacy computation algorithm according to claim 1, characterized in that, The step of performing time series analysis on the sensitivity calculation feature data using the ARIMA model to obtain sensitivity prediction data, and quantifying the residual between the sensitivity calculation feature data and the sensitivity prediction data, includes: The sensitivity calculation feature data is sampled at fixed intervals using a sliding time window to obtain sampled data; By quantifying the Pearson autocorrelation coefficient and partial autocorrelation coefficient of the sampled data, a multidimensional sensitivity initial sequence is obtained, and the stationarity of the multidimensional sensitivity initial sequence is tested to obtain a multidimensional stationary sensitivity sequence. An ARIMA model is constructed based on the multidimensional stationary sensitivity sequence, and the log-likelihood function is optimized by the maximum likelihood estimation method to calculate the autoregressive coefficients and moving average coefficients of the ARIMA model, thereby obtaining the set of fitting parameters. Residual analysis is performed on the multidimensional stationary sensitivity sequence based on the fitted parameter set, and the optimal model order combination is selected based on the Akaike information criterion and the Bayesian information criterion, so as to obtain the prediction parameter set based on the ARIMA model with the optimal model order combination. A radial basis kernel support vector regression model is constructed based on the predicted parameter set, and the kernel function bandwidth parameter and penalty factor are set. The optimal hyperparameters of the radial basis kernel support vector regression model are determined through cross-validation, so as to predict the multidimensional sensitivity prediction sequence based on the radial basis kernel support vector regression model with the optimal hyperparameters. The features of each dimension in the multidimensional sensitivity prediction sequence are standardized to obtain sensitivity prediction data, and the Euclidean distance between the sensitivity prediction data and the sampled data is calculated to obtain the residual.

5. The method for handling encrypted execution anomalies in a privacy-preserving computation algorithm according to claim 1, characterized in that, The method involves extracting statistical features from the sensitivity calculation feature data of each abnormal tenant using a sliding window to construct a feature set, and then performing cluster analysis on the feature set to obtain the abnormal pattern and abnormal fingerprint feature cluster corresponding to each abnormal tenant, including: A fixed-length sliding window is used to segment the sensitivity calculation feature data of each abnormal tenant. The frequency domain mapping of each segment is performed by discrete Fourier transform to obtain frequency domain statistical features, and then the maximum and minimum values ​​are normalized to obtain the first feature set. The first feature set is dimensionality reduced, and the dimensionality reduction result is subjected to singular value decomposition to obtain a feature vector matrix. Feature vector components with a cumulative contribution rate exceeding a preset contribution threshold are selected from the matrix and combined using Bayesian probability weighting to obtain the second feature set. The second feature set is standardized, and the similarity between data in the standardized second feature set is calculated using Mahalanobis distance. Based on the similarity calculation results, the optimal number of clusters is determined using the silhouette coefficient, thus obtaining the clustering parameter set. A Gaussian mixture clusterer is constructed based on the clustering parameter set, and the covariance matrix and random initialization parameters are set. The cluster centers of the Gaussian mixture clusterer are iteratively optimized using the expectation-maximization algorithm to obtain the first clustering result. The centroids and covariance matrices of the clusters are extracted from the first clustering result. An anomaly measurement function is constructed based on the density index and isolation index obtained by quantifying the centroids and covariance matrices of each cluster. The anomaly score of each cluster is quantified through the anomaly measurement function. Set the abnormal pattern division range, divide each cluster into multiple abnormal patterns based on the relationship between the abnormal score and the abnormal pattern division range, and mark the abnormal fingerprint feature cluster corresponding to each abnormal pattern.

6. The method for handling encrypted execution anomalies in a privacy-preserving computation algorithm according to claim 1, characterized in that, The process involves obtaining normal fingerprint feature clusters for each of the anomalous tenants, constructing an anomalous pattern impact mechanism model by combining these clusters, and quantifying the impact of each anomalous pattern on the privacy protection level of each anomalous tenant using this model, thereby obtaining quantitative analysis results, including: Obtain the normal fingerprint feature clusters of each of the abnormal tenants, and construct the overlap matrix by quantifying the Mahalanobis distance and cosine similarity of the normal fingerprint feature clusters and the abnormal fingerprint feature clusters in Euclidean space. The minimum spanning tree algorithm is used to calculate the edge weights of the overlap matrix, and a feature association graph is constructed based on the edge weight calculation results. The node features in the feature association graph are extracted to construct a differential feature set. Singular value decomposition is performed on the differential feature set, and the decomposition results are sorted to determine the main differential components. The importance of the main differential components is calculated by random forest to obtain the influence weight vector. A directed acyclic graph is constructed based on the influence weight vector. The conditional dependencies between nodes in the directed acyclic graph are set, and the propagation probabilities between the nodes are extracted to construct a propagation strength matrix. The propagation intensity matrix is ​​probabilistically inferred using a Bayesian network, and the conditional probability distribution between the nodes is calculated using the maximum likelihood estimation method. Based on the probabilistic inference results and the conditional probability distribution calculation results, an abnormal mode influence mechanism model is constructed. The impact of the abnormal features corresponding to each abnormal pattern on the level of privacy protection is calculated based on the abnormal pattern impact mechanism model and used as an attenuation coefficient. A hierarchical structured privacy impact assessment index is constructed through the attenuation coefficient, and the hierarchical structured privacy impact assessment index is quantified to obtain a quantitative assessment vector.

7. The method for handling encrypted execution anomalies in a privacy-preserving computation algorithm according to claim 2, characterized in that, The step of adjusting the privacy budget allocation ratio of each abnormal tenant based on the quantitative analysis results to obtain a privacy budget allocation scheme, and then executing the privacy budget allocation scheme, includes: A risk assessment function is constructed based on the quantitative analysis results to score the risk of each abnormal tenant, and the correlation between each abnormal tenant is determined based on the relationship between the risk score results and the preset risk threshold. The influence propagation coefficient of each abnormal tenant is quantified using the exponential weighting method, and an association strength matrix is ​​constructed by combining the influence propagation coefficient and the association relationship. The association strength matrix is ​​then subjected to feature decomposition to obtain the tenant association index. A linear decay function is set according to the tenant correlation index to classify the budget reduction ratio, and a normal distribution function is used to non-linearly reduce the privacy budget allocation ratio of each abnormal tenant to obtain the first budget adjustment set. The compensation coefficients for each abnormal tenant are calculated using the least squares method based on the first budget adjustment set, and a compensation weight vector is set in combination with the data query sensitivity corresponding to each abnormal tenant to obtain the second budget adjustment set. The second budget adjustment set is optimized using gradient descent to obtain a privacy budget allocation scheme for each of the abnormal tenants, and the privacy budget allocation scheme is then executed.

8. The method for handling encrypted execution exceptions of a privacy computation algorithm according to claim 7, characterized in that, The process of optimizing the second budget adjustment set using gradient descent to obtain a privacy budget allocation scheme for each of the abnormal tenants, and then executing the privacy budget allocation scheme, includes: The sensitivity sequence during the execution of the privacy budget allocation scheme is extracted using the differential privacy query function. The sensitivity sequence is then subjected to frequency domain analysis using fast Fourier transform to obtain the feature fingerprint and singular value decomposition to obtain the first feature vector. The similarity between the first feature vector and the pre-stored abnormal fingerprint feature cluster is calculated using Mahalanobis distance to construct a similarity matrix. The similarity matrix is ​​then normalized by setting a distance threshold to obtain the second feature vector. Spectral clustering is performed on the second feature vector, and the feature space of the second feature vector is divided into clusters by setting the cluster radius and the minimum number of samples, so as to calculate the local density and relative distance of each cluster and obtain the density distribution matrix. Extract the cluster center point set from the density distribution matrix, construct a feature topology graph based on the cluster center point set and inter-cluster connectivity, and perform hierarchical labeling on the feature topology graph to obtain a hierarchical structure matrix; The budget adjustment amount for each abnormal tenant is calculated based on the hierarchical structure matrix, and the budget adjustment amount is limited by an exponential decay function. The budget adjustment amount is weighted by the tenant correlation index to obtain the budget adjustment plan. A compensation resource allocation network is constructed based on the budget adjustment scheme, and network flow constraints and capacity constraints are set. The compensation resource allocation network is solved by the maximum flow minimum cut algorithm to obtain the optimal compensation allocation scheme, and the optimal compensation allocation scheme is executed.

9. The method for handling encrypted execution exceptions of a privacy computation algorithm according to claim 8, characterized in that, After optimizing the second budget adjustment set using gradient descent to obtain a privacy budget allocation scheme for each of the abnormal tenants, and executing the privacy budget allocation scheme, the process further includes: The sensitivity sequence is decomposed at multiple scales. High-frequency coefficients, low-frequency coefficients and energy distribution features are extracted from the decomposition results to construct a decomposed feature vector. The normalized cosine distance between the decomposed feature vector and the pre-stored abnormal fingerprint feature cluster is quantified to obtain a matching degree matrix. Anomaly classification of the matching degree matrix is ​​performed using Gaussian kernel support vector machine to obtain a first risk assessment value to construct a directed weighted graph. The second risk assessment value is obtained by calculating the in-degree and out-degree of the nodes in the directed weighted graph and obtaining the secondary influence propagation coefficient to quantify the information entropy loss of each of the abnormal tenants. The secondary compensation weight value is calculated based on the second risk assessment value, and the secondary compensation weight value is optimized by the total budget conservation constraint and the single tenant compensation upper limit constraint to obtain the optimal budget adjustment scheme for each abnormal tenant, and the optimal budget adjustment scheme is executed.

10. A system for handling exceptions in the ciphertext execution of a privacy-preserving computation algorithm, characterized in that, include: The feature data quantization module is used to obtain differential privacy query parameters of multiple tenants, and quantify the data query sensitivity of each tenant according to the differential privacy query parameters to obtain sensitivity calculation feature data; The abnormal tenant assessment module is used to perform time series analysis on the sensitivity calculation feature data using the ARIMA model to obtain sensitivity prediction data, and to quantify the residual between the sensitivity calculation feature data and the sensitivity prediction data in order to assess whether each tenant has an abnormal sensitivity shift and to identify abnormal tenants. The abnormal pattern segmentation module is used to extract statistical features from the sensitivity calculation feature data of each abnormal tenant using a sliding window to construct a feature set, and to perform cluster analysis on the feature set to obtain the abnormal pattern and abnormal fingerprint feature cluster corresponding to each abnormal tenant. The impact degree quantification module is used to obtain the normal fingerprint feature clusters of each of the abnormal tenants, to construct an abnormal pattern impact mechanism model by combining the abnormal fingerprint feature clusters, and to quantify the impact degree of each of the abnormal patterns on the privacy protection level of each of the abnormal tenants through the abnormal pattern impact mechanism model, so as to obtain the quantitative analysis results. The allocation scheme execution module is used to adjust the privacy budget allocation ratio of each abnormal tenant according to the quantitative analysis results, obtain a privacy budget allocation scheme, and execute the privacy budget allocation scheme.

Images