A method and system for handling ciphertext execution exceptions of a privacy computing algorithm

Abstract

The present invention relates to the field of information technology, and discloses a method and system for handling ciphertext execution exceptions of a privacy computing algorithm. The method quantifies the data query sensitivity of each tenant by collecting differential privacy query parameters of several tenants, and uses an ARIMA model to perform time series analysis on the sensitivity quantification results to predict their future trends, and calculates the residual between the actual sensitivity and the predicted sensitivity to identify abnormal tenants; a sliding window method is used to extract statistical features of abnormal tenants to construct a feature set, and abnormal patterns and abnormal fingerprint feature clusters corresponding to each abnormal tenant are identified by clustering analysis; a model is trained based on the positive and abnormal fingerprint feature clusters of abnormal tenants, and the model is used to quantify the degree of influence of each abnormal pattern on the privacy protection level of the abnormal tenant, so as to adjust the privacy budget allocation ratio of each abnormal tenant according to the obtained quantitative analysis results and execute, thereby realizing effective handling of exceptions in the ciphertext execution of the privacy computing algorithm.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method and system for handling ciphertext execution exceptions of a privacy computing algorithm. Background Art

[0002] With the continuous development of privacy algorithms, differential privacy algorithms in privacy computing algorithms are often used in multi-tenant cloud environments. At present, differential privacy queries submitted by tenants are usually executed in ciphertext to protect privacy, and the global sensitivity of the query function is a key parameter to measure the level of privacy protection. However, since the query process is executed in ciphertext, the sensitivity calculation process cannot be directly observed. If the sensitivity is underestimated, the added noise is insufficient and the privacy protection level decreases. If the sensitivity is overestimated, although the privacy is protected, the introduced noise is too large and the practicality of the query results is reduced. Therefore, it is necessary to dynamically identify and handle sensitivity calculation anomalies during the execution of tenants' differential privacy queries.

[0003] It can be seen that how to detect and handle the sensitivity anomalies of tenants during the execution of differential privacy queries in a multi-tenant cloud environment has become a technical problem that technical personnel in this field need to solve urgently. Summary of the invention

[0004] The present invention provides a method and system for handling ciphertext execution exceptions of a privacy computing algorithm, which solves the problem of how to detect and handle sensitivity exceptions of tenants during the execution of differential privacy queries in a multi-tenant cloud environment.

[0005] In order to solve the above technical problems, the present invention provides a method for handling ciphertext execution exceptions of a privacy computing algorithm, comprising:

[0006] Acquire differential privacy query parameters of several tenants, and quantify the data query sensitivity of each tenant according to the differential privacy query parameters to obtain sensitivity calculation feature data;

[0007] Performing time series analysis on the sensitivity calculation feature data through the ARIMA model to obtain sensitivity prediction data, and quantifying the residual between the sensitivity calculation feature data and the sensitivity prediction data to evaluate whether each of the tenants has sensitivity deviation anomalies, and obtain abnormal 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 a cluster analysis is performed on the feature set to obtain an abnormal pattern and an abnormal fingerprint feature cluster corresponding to each abnormal tenant;

[0009] Acquire the normal fingerprint feature cluster of each abnormal tenant, build an abnormal pattern impact mechanism model in combination with the abnormal fingerprint feature cluster, and quantify the impact of each abnormal pattern on the privacy protection level of each abnormal tenant through the abnormal pattern impact mechanism model to obtain a quantitative analysis result;

[0010] The privacy budget allocation ratio of each abnormal tenant is adjusted according to the quantitative analysis result, and a privacy budget allocation plan is obtained and executed to achieve the processing of anomalies in the ciphertext execution of the privacy computing algorithm.

[0011] Compared with the prior art, the embodiments of the present invention have the following advantages:

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

[0013] (2) By constructing an abnormal mode impact mechanism model and quantifying the impact of each abnormal mode on the privacy protection level, we provide a scientific basis for the rational allocation of the privacy budget, effectively detect sensitivity anomalies in a multi-tenant environment, and dynamically adjust the privacy protection strategy to improve the overall privacy protection level. We effectively handle anomalies in the algorithm ciphertext execution process, thereby improving the robustness and practicality of the privacy computing algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the implementation mode will be briefly introduced below. Obviously, the drawings described below are only some implementation modes of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 It is a flowchart of a method for handling ciphertext execution exceptions of a privacy computing algorithm provided by an embodiment of the present invention;

[0016] Figure 2 It is a structural diagram of a ciphertext execution exception handling system for a privacy computing algorithm provided by a certain embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the 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 ordinary technicians in this field without creative work 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 are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first", "second", "third", etc. may explicitly or implicitly include one or more of the feature. In the description of this application, unless otherwise specified, "plurality" means two or more.

[0019] In the description of the present application, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the 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, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0020] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as those commonly understood by those skilled in the art. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood by specific circumstances.

[0021] In one embodiment, if Figure 1 As shown, the first aspect of the present invention provides a method for handling ciphertext execution exceptions of a privacy computing algorithm, comprising:

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

[0023] In one embodiment, step S1 includes:

[0024] Acquire data differential values ​​and weight coefficients of several tenants, and obtain a first tenant isolation boundary value by performing a dot product operation and normalization processing on the data differential values ​​and the weight coefficients to construct a tenant differential parameter matrix; each row in the tenant differential parameter matrix contains a tenant identifier and its corresponding data query sensitivity;

[0025] A data query sensitivity threshold is set according to the tenant differential parameter matrix, and each of the tenants is grouped and clustered according to the data query sensitivity threshold to obtain a plurality of tenant groups with different data query sensitivity levels, and a differential privacy query parameter of each of the tenant groups is quantized and encoded according to a preset desensitization rule to obtain a first feature space;

[0026] Using a gradient boosting tree to associate and map the parameter range interval in the first feature space with the data query sensitivity level, output a parameter mapping vector, and perform random forest feature extraction on the parameter mapping vector to obtain a second feature space;

[0027] Obtaining an intermediate indicator of global sensitivity calculation in the second feature space, processing the intermediate indicator of global sensitivity calculation by an adaptive clustering algorithm to obtain a set of cluster center points, and constructing a feature vector dimension matrix based on the set of cluster center points;

[0028] An orthogonal transformation is performed on the feature vector dimension matrix to obtain a standardized feature vector, and a weighted sum is performed on the standardized feature vector and the first tenant isolation boundary value to generate a differential privacy query parameter set as the sensitivity calculation feature data output.

[0029] Specifically, in the differential privacy algorithm, the present invention uses the tenant access control library to store real-time differential privacy query requests of different tenants and their related data, such as data differential values ​​and weight coefficients, etc.; wherein the data differential value reflects the degree of dependence of the query result on a single record in the data set, and the weight coefficient is used to adjust the privacy protection level of different queries. When the number of tenants is 300, the data differential value of each tenant is distributed between 0 and 1, and the weight coefficient is determined to be between 0.1 and 0.9 according to the data sensitivity of the tenant. Then, the present invention accesses the control library through a differential privacy query request to obtain the differential privacy query parameters; then, a dot product operation is performed on the data differential value and the weight coefficient, and the operation result is processed by the maximum and minimum value normalization method to obtain the initial tenant isolation boundary value to construct a 300×2-dimensional tenant differential parameter matrix, and each row in the tenant differential parameter matrix contains a tenant identifier and its corresponding data query sensitivity; wherein, the data query sensitivity can be understood as the maximum change in the query result when the same query is executed on adjacent data sets (i.e., data sets that only differ by one record).

[0030] Then, the data query sensitivity threshold is set according to the distribution characteristics of the data in the tenant differential parameter matrix, preferably 0.8, and the tenants are grouped and clustered based on the data query sensitivity threshold. The Euclidean distance can be used to measure the similarity between tenants, and the hierarchical clustering method is used to divide them into multiple tenant groups with different data query sensitivity levels. For example, 300 tenants are divided into 25 groups, each group corresponds to a data query sensitivity level, and the number of tenants in each group is 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 fuzzification, replacement, encryption, and masking to map the character parameters into numerical features, and then use them to construct the first feature space.

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

[0032] Several intermediate indicators of global sensitivity calculation are extracted from the second feature space, such as the change of the query function, the scale parameter of the noise, the change of the data distribution, the sensitivity boundary, etc., and a density-based adaptive clustering algorithm is used to process these indicator data, and the neighborhood radius of the algorithm is set to 0.3 and the minimum number of samples is set to 5. The cluster center point set can be obtained, and then the cluster center points are organized into an 8×10-dimensional feature vector dimension matrix.

[0033] The eigenvector dimension matrix is ​​decomposed and orthogonally transformed through singular value decomposition, and the eigenvector with variance explanation rate of 95% is retained as the standardized eigenvector. The standardized eigenvector and the tenant isolation boundary value are weighted and summed, and the weights of the two are preferably 0.25 and 0.75. Other relevant information (such as query type, query timestamp, query frequency, data access scope, operation type, etc.) are also included in the feature data. A set of differential privacy query parameters for 300 tenants is generated and output 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, which reflects the adaptive characteristics of differential privacy protection.

[0034] In actual application scenarios, a user information query system contains sensitive data such as power plant output power, substation load conditions, user power consumption patterns, historical energy consumption records, etc. The differential privacy query parameters calculated by the above method can automatically adjust the amount of noise added according to the sensitivity of different types of data, while protecting user privacy and 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, and only adds noise with a standard deviation of 0.3 to the low-sensitivity tenant group, fully realizing the dynamic adjustment of the differential privacy protection strength.

[0035] The present invention can more accurately evaluate the data query sensitivity of tenants by comprehensively considering the data differential value and weight coefficient in the differential privacy query; use machine learning algorithms such as gradient boosting tree and random forest for feature extraction and mapping, which further improves the accuracy of data query sensitivity assessment; group and cluster tenants according to the data query sensitivity threshold, which helps to implement different privacy protection measures for tenants with different sensitivity levels; quantize and encode differential privacy query parameters through preset desensitization rules, thereby enhancing the privacy protection capability of data; combine multiple algorithms and technical means to make the data processing process more efficient and flexible, and adjust parameters and algorithms according to actual needs to adapt to different data processing scenarios and needs.

[0036] S2. Performing time series analysis on the sensitivity calculation feature data through an ARIMA model to obtain sensitivity prediction data, and quantifying the residual between the sensitivity calculation feature data and the sensitivity prediction data to evaluate whether each of the tenants has sensitivity deviation anomalies, and obtaining abnormal tenants;

[0037] In one embodiment, the time series analysis of the sensitivity calculation feature data is performed by using an ARIMA model to obtain sensitivity prediction data, and the residual between the sensitivity calculation feature data and the sensitivity prediction data is quantified, including:

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

[0039] Performing stationarity processing on the time series feature matrix to obtain a stationary time series data set to construct an ARIMA model, estimating the autoregressive term coefficient and the moving average term coefficient of the ARIMA model by the least squares method, and performing a significance test on the parameters of the ARIMA model to obtain a goodness of fit index;

[0040] The optimal parameter combination is screened by the goodness of fit index, and the ARIMA model with the optimal parameter combination is used for multi-step iterative prediction on the validation data set to calculate the confidence interval of the prediction result and obtain the sensitivity prediction sequence;

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

[0042] Specifically, in tenant sensitivity assessment, the present invention selects a fixed-size time window, slides the window on the sensitivity calculation feature data, and moves a fixed time interval each time, thereby obtaining a series of sampled data to smooth the data and reduce noise, while retaining the trend and periodicity of the time series; then, the autocorrelation function and partial autocorrelation function of the sampled data are calculated (to reveal the autocorrelation and pattern in the time series data), and the values ​​of ACF and PACF are quantified to construct a time series feature matrix. Taking the cross-tenant data access scenario of a certain institution 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 is obtained, and it can be found that the sensitivity calculation feature data reaches a significant correlation when the lag period is 12, the autocorrelation coefficient is 0.85, and the partial autocorrelation coefficient is 0.72, indicating that the data has obvious periodic characteristics.

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

[0044] The optimal ARIMA model parameter combination is selected according to the goodness of fit index, and the ARIMA model with the optimal parameter combination is used to perform rolling prediction of 24 points on the validation data set. The upper and lower limits of the 95% confidence interval of the prediction result are the predicted value ±0.12, respectively, to obtain the sensitivity prediction sequence; the sensitivity prediction sequence is then used to construct a Kalman filter. The state transfer matrix of the Kalman filter adopts a 2×2 unit matrix, the observation matrix is ​​[1,0], the diagonal elements of the state noise covariance matrix are set to 0.01, and the observation noise variance is set to 0.04. These matrices are used to describe the changes in the system state and the generation process of the observation data. Then, the prediction-update cycle is used to enable the constructed Kalman filter to effectively track the changing trend of the sensitivity sequence. Finally, the prediction information and observation information are fused to obtain more accurate sensitivity prediction data for residual calculation with the sampling data.

[0045] The present invention can extract the inherent laws and patterns of time series data through sliding time window sampling and time series feature construction, thereby improving the accuracy of prediction; the ARIMA model can be used to capture the long-term dependence and periodicity of the time series, while the Kalman filter can fuse the prediction information and real-time observation data to further improve the accuracy of prediction; the optimal parameter combination can be screened through significance test, thereby avoiding the problems of overfitting and underfitting and enhancing the robustness of the model; and the Kalman filter can update the prediction results in real time and adapt to the changes in time series data, so that the scheme can realize real-time prediction and dynamic adjustment in practical applications, thereby improving the timeliness and practicality of the prediction.

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

[0047] By quantifying the Pearson autocorrelation coefficient and the partial autocorrelation coefficient of the sampled data, a multidimensional sensitivity initial sequence is obtained, and a stationarity test is performed on the multidimensional sensitivity initial sequence to obtain a multidimensional stationary sensitivity sequence;

[0048] An ARIMA model is constructed according to the multidimensional stationary sensitivity sequence, and a log-likelihood function is optimized by a maximum likelihood estimation method to calculate an autoregressive coefficient and a moving average coefficient of the ARIMA model to obtain a fitting parameter set;

[0049] Performing residual analysis on the multidimensional stationary sensitivity sequence according to the fitting parameter set, and selecting an optimal model order combination based on the Akaike information criterion and the Bayesian information criterion, so as to obtain a prediction parameter set based on the ARIMA model prediction with the optimal model order combination;

[0050] A radial basis kernel support vector regression model is constructed based on the prediction parameter set, and a kernel function bandwidth parameter and a penalty factor are set, and the optimal hyperparameters of the radial basis kernel support vector regression model are determined by cross-validation, so as to predict a multidimensional sensitivity prediction sequence according to the radial basis kernel support vector regression model with the optimal hyperparameters;

[0051] 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 sampling data is calculated to obtain the residual.

[0052] Specifically, the present invention processes the sampled data, calculates the Pearson autocorrelation coefficient (ACF) and the partial autocorrelation coefficient (PACF) of each dimension to identify patterns in the sampled data, such as periodicity, trend, etc., and generates a multidimensional sensitivity initial sequence; then the multidimensional sensitivity initial sequence is subjected to a stationarity test, such as an ADF (Augmented Dickey-Fuller) test or an augmented Dickey-Fuller test, to ensure that the sequence is stationary, thereby ensuring the accuracy of the ARIMA model prediction. The multidimensional sensitivity time series analysis takes the cross-departmental data access of an organization as a specific scenario. It monitors the sampled data formed by the three dimensions of query frequency, data volume and computational complexity in the sensitivity calculation feature data. The query frequency of a tenant under normal conditions 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 sliding window sampling, the autocorrelation coefficient of the query frequency dimension is calculated to be 0.82, the autocorrelation coefficient of the data volume dimension is 0.75, and the autocorrelation coefficient of the complexity dimension is 0.68, indicating that all three dimensions have significant time series correlation.

[0053] Then, based on the stable multidimensional sensitivity sequence, an ARIMA (p, d, q) model is constructed, where p is the autoregressive order, d is the number of differences (usually determined by 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 coefficient and the moving average coefficient are calculated to obtain the fitting parameter set. When constructing the autoregressive difference moving average model, the log-likelihood function value obtained by maximum likelihood estimation is preferably -856.4, the diagonal elements of the autoregressive coefficient matrix are preferably 0.72, 0.68 and 0.65, and the diagonal elements of the moving average coefficient matrix are preferably -0.35, -0.42 and -0.38.

[0054] The ARIMA model with a fitting parameter set is used to perform residual analysis on the multidimensional stationary sensitivity series, and then the Akaike information criterion (AIC) and the Bayesian information criterion (BIC) are used to select the optimal model order combination (p, d, q), and the combination of autoregressive order p=2, difference order d=1, and moving average order q=1 is obtained. The prediction parameter set is obtained based on the ARIMA model with the optimal model order combination.

[0055] The prediction parameter set is used as input to construct a radial basis kernel (RBF) support vector regression model, that is, the radial basis kernel function is used in the support vector regression, the kernel function bandwidth parameter is set to 0.8, the penalty factor is set to 10, and the root mean square error obtained by 5-fold cross validation is 0.086, so as to determine the optimal hyperparameters, and then use the RBF support vector regression model with the optimal hyperparameters to predict the multidimensional sensitivity prediction sequence; among them, the prediction results show that the predicted value of the query frequency in the next 4 hours is 45 to 55 times per hour, and the average deviation from the actual observation value is 3 times; the predicted value of the data volume is between 180 and 520 records, and the average deviation is 25; the predicted value of the query execution time is between 90 and 320 milliseconds, and the average deviation is 15 milliseconds.

[0056] After standardizing the features of each dimension and its prediction error in the multidimensional sensitivity prediction sequence, the sensitivity prediction data is obtained. The Euclidean distance between the sensitivity prediction data and the original sampled data is calculated to obtain the residual. In the data sharing scenario, the above scheme also shows good results; when using this method to monitor cross-partial communication archive queries, abnormal batch data export behavior was found. The frequency of the query suddenly increased to 150 times per hour, and the amount of data returned by a single query reached 3 times the normal value. The execution time exceeded the prediction interval by 2 standard deviations, and the residual vector norm reached 0.412, far exceeding the warning threshold of 0.2. Through timely discovery and processing, illegal access to sensitive data is avoided. In other application scenarios, the above scheme discovered an average of 12 sensitivity anomalies per month, with an accuracy rate of 91%, which shows its effectiveness.

[0057] The invention combines the ARIMA model and the RBF support vector regression model, makes full use of the inherent rules of time series data and the nonlinear fitting ability of the machine learning model, and improves the accuracy of prediction; uses AIC and BIC for model selection and optimization to ensure the simplicity and prediction performance of the selected model; processes the non-stationary series through stationarity test and difference operation, so that the data meets the modeling requirements of the ARIMA model and has a wide range of applications.

[0058] The present invention quantifies the root mean square error value of the residual, and judges the error value through a preset abnormality judgment threshold, and regards tenants whose error values ​​are not less than the abnormality judgment threshold as abnormal tenants, and marks these abnormal tenants for offset from a pre-established fluctuation range comparison table, marking the current tenant data query sensitivity as having an assessment offset abnormality and being an abnormal tenant.

[0059] S3, using a sliding window to extract statistical features from the sensitivity calculation feature data of each abnormal tenant to construct a feature set, and performing cluster analysis on the feature set to obtain abnormal patterns and abnormal fingerprint feature clusters corresponding to each abnormal tenant;

[0060] In one embodiment, step S3 includes:

[0061] The sensitivity calculation feature data of each abnormal tenant is segmented by using a fixed-length sliding window, and each segmented data is mapped to the frequency domain by discrete Fourier transform to obtain frequency domain statistical features and perform maximum and minimum value normalization processing to obtain a first feature set;

[0062] Performing dimensionality reduction on the first feature set, and performing singular value decomposition on the dimensionality reduction result to obtain a feature vector matrix, from which feature vector components whose cumulative contribution rates exceed a preset contribution threshold are selected for Bayesian probability weighted combination to obtain a second feature set;

[0063] The second feature set is standardized, and the similarity between the data in the standardized second feature set is calculated by Mahalanobis distance, so as to determine the optimal number of clusters based on the similarity calculation result using the silhouette coefficient to obtain a clustering parameter set;

[0064] A Gaussian mixture clusterer is constructed based on the clustering parameter set, and a covariance matrix and random initialization parameters are set, and the clustering center of the Gaussian mixture clusterer is iteratively optimized by using expectation maximization to obtain a first clustering result;

[0065] Extracting the center point and covariance matrix of the cluster from the first clustering result, and constructing an anomaly measurement function according to the compactness index and isolation index quantified by the center point and covariance matrix of each cluster, and quantifying the anomaly score of each cluster by the anomaly measurement function;

[0066] An abnormal pattern division range is set, each of the clusters is divided into a plurality of abnormal patterns based on the relationship between the abnormal score and the abnormal pattern division range, and an abnormal fingerprint feature cluster corresponding to each abnormal pattern is marked.

[0067] First, the present invention uses a fixed-length sliding window to segment the sensitivity calculation feature data of each abnormal tenant. The preferred window length is set to 120 minutes and the sliding step is 10 minutes to capture local features in the data, making subsequent analysis more detailed. Then, a discrete Fourier transform (DFT) is applied to each segmented data to map the time domain data to the frequency domain, thereby extracting two frequency domain statistical features, namely the amplitude spectrum and the phase spectrum, and combining them with the three time domain statistics of the mean, variance and kurtosis to perform maximum and minimum value normalization processing to obtain the first feature set.

[0068] According to the characteristics of the data and business needs, we select appropriate dimensionality reduction methods, such as principal component analysis (PCA) and linear discriminant analysis (LDA), to reduce the dimensionality of the first feature set and obtain five main eigenvectors. We then select feature components whose cumulative contribution rates exceed the preset contribution threshold (preferably 92%) and use the Bayesian probability weighted method to combine the selected feature vectors. The weight coefficients are 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.

[0069] The second feature set is standardized to eliminate the dimensional differences between different features. Then, the similarity between the standardized data is calculated, and the Mahalanobis distance is used to characterize the correlation between the features. Based on the similarity calculation results, the silhouette coefficient is used to evaluate the clustering effects under different numbers of clusters to determine the optimal number of clusters, and then the clustering parameter set is obtained.

[0070] A Gaussian mixture clusterer was constructed, the covariance matrix and random initialization parameters were set, and the cluster centers were iteratively optimized through the expectation-maximization (EM) algorithm to obtain the first clustering result. The covariance matrix of the Gaussian mixture clusterer was in diagonal form, and the expectation-maximization algorithm converged when the relative error was less than 0.001. A total of 15 iterations were performed, and the clustering results showed that the average sample-to-center distance of the normal tenant cluster was 0.28, and the standard deviation was 0.06.

[0071] By extracting key information from the first clustering result, namely the center point and covariance matrix of the cluster, it can be found that the average distance from the sample to the center of the abnormal cluster reaches 0.85, and the distance distribution within the cluster shows obvious skewness; then, the anomaly measurement function is constructed based on the closeness index (the inverse of the average distance within the cluster) and the isolation index (the distance to the nearest cluster) constructed according to the key information of the cluster, and the 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 mode division range is set according to actual needs, and the cluster is divided into multiple anomaly modes based on the anomaly score, that is, the cluster corresponding to each abnormal tenant is classified, and the corresponding anomaly fingerprint feature cluster is marked for each anomaly mode for subsequent analysis and processing.

[0072] The above scheme is applied to the user behavior analysis of telecom operators, and it shows good recognition effect. For example, among the 1,000 enterprise tenants of a certain operator, three abnormal clusters involving 27 tenants were identified through multi-dimensional fingerprint feature clustering. The data access patterns of these tenants showed frequent burst peaks, and the kurtosis of the time domain feature reached 5.2, which was much higher than the 2.3 of the normal cluster. The frequency domain feature showed the existence of obvious high-frequency components. The amplitude spectrum amplitude at 0.8 Hz was 3.5 times that of the normal tenant. The average compactness index of the abnormal tenant cluster was 0.42, significantly lower than the 0.86 of the normal cluster, and the isolation index was as high as 0.95, indicating that the behavior patterns of these tenants deviated significantly from the normal business process. Through the comprehensive scoring of the abnormal measurement function, 12 tenant accounts with risks were accurately identified, and the statistical feature distribution of their sensitivity sequence had a significant deviation from the historical pattern. The present 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. The optimal number of clusters is determined by using the silhouette coefficient, a Gaussian mixture clusterer is constructed, and cluster centers are iteratively optimized to make clustering results more reasonable and accurate. By constructing anomaly measurement functions and dividing anomaly modes, different abnormal behavior modes can be clearly identified and classified, which provides convenience for subsequent processing.

[0073] S4, obtaining a normal fingerprint feature cluster of each abnormal tenant, combining the abnormal fingerprint feature cluster to construct an abnormal pattern impact mechanism model, and quantifying the impact of each abnormal pattern on the privacy protection level of each abnormal tenant through the abnormal pattern impact mechanism model to obtain a quantitative analysis result;

[0074] In one embodiment, step S4 includes:

[0075] Obtaining the normal fingerprint feature cluster of each abnormal tenant, and constructing an overlap matrix by quantifying the Mahalanobis distance and cosine similarity between the normal fingerprint feature cluster and the abnormal fingerprint feature cluster in the Euclidean space;

[0076] The minimum spanning tree algorithm is used to calculate the edge weights of the overlap matrix, and a feature association graph is constructed according to the edge weight calculation results, and node features in the feature association graph are extracted to construct a difference feature set;

[0077] Performing singular value decomposition on the difference feature set, sorting the decomposition results to determine the main difference components, calculating the importance of the main difference components through random forest, and obtaining an influence weight vector;

[0078] Constructing a directed acyclic graph based on the influence weight vector, setting conditional dependencies between nodes in the directed acyclic graph, and extracting transmission probabilities between the nodes to construct a propagation intensity matrix;

[0079] Probabilistic reasoning is performed on the propagation intensity matrix through a Bayesian network, and conditional probability distribution between the nodes is calculated through maximum likelihood estimation, so as to construct an abnormal mode impact mechanism model according to the probabilistic reasoning result and the conditional probability distribution calculation result;

[0080] According to the abnormal pattern impact mechanism model, the degree of influence of the abnormal features corresponding to each abnormal pattern on the privacy protection level is calculated as the attenuation coefficient, and a hierarchical structured privacy impact assessment index is constructed and quantified through the attenuation coefficient to obtain a quantitative assessment vector.

[0081] Specifically, the present invention collects and organizes the fingerprint feature clusters of each abnormal tenant in a normal state, and these feature clusters represent the normal behavior patterns of the tenants, and quantifies the Mahalanobis distance and cosine similarity of the normal fingerprint feature clusters and the abnormal fingerprint feature clusters in the Euclidean space; finally, an overlap matrix is ​​constructed according to the calculated Mahalanobis distance (quantifying the correlation between features) and cosine similarity (measuring the directional similarity between feature vectors), and the rows and columns of the matrix represent the normal and abnormal fingerprint feature clusters, respectively, so that the constructed overlap matrix can more accurately reflect the similarities and differences between the normal and abnormal fingerprint feature clusters. For example, in the cross-tenant data access scenario of an institution, the fingerprint characteristics of a payment business tenant showed significant abnormal characteristics. The characteristic vector distribution of this tenant in normal period showed low frequency and small fluctuation characteristics, and the mean standard deviation in the 12 dimensions of the feature space was 0.15; while in abnormal period, the characteristic vector shifted significantly in the three dimensions of frequency, amplitude and duration, and the mean standard deviation rose to 0.42. By calculating the Mahalanobis distance of these two feature clusters, a difference of 1.85 was obtained, and the cosine similarity was 0.62, and the overlap of feature clusters was only 0.28.

[0082] The edge weights of the overlap matrix are calculated, and the weights can be based on the inverse of the Mahalanobis distance or cosine similarity, etc., to reflect the similarity between features; then the overlap matrix is ​​regarded as an adjacency matrix of a weighted undirected graph, in which the 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 then used to construct a feature association graph. The nodes in the graph represent features, and the edges represent the association strength determined based on the minimum spanning tree; then the node features extracted from the feature association graph are combined into a difference feature set, which contains key features that can distinguish normal from abnormal behavior patterns.

[0083] Then, singular value decomposition is performed on the constructed difference feature set to obtain eigenvectors and singular values. The eigenvectors are sorted according to the size of the singular values, and eigenvectors with larger singular values ​​are selected as the main difference components. For example, after the decomposition of the difference feature set corresponding to the cross-tenant data access scenario, the first three eigenvalues ​​account for 85% of the total variance, which can be determined as the main difference component. Then, the importance of these main difference components is calculated through random forest to obtain the influence weight vectors, which are 0.42, 0.35 and 0.23 respectively.

[0084] Since the propagation of abnormal features in the power grid data sharing scenario shows obvious hierarchy, the present invention uses a directed acyclic graph to depict the dependency between features, and the conditional probability between nodes reflects the intensity of influence transmission: that is, a directed acyclic graph is constructed based on the influence weight vector, the nodes in the graph represent abnormal features and their corresponding abnormal modes, the edges represent the conditional dependency between features, the conditional dependency between nodes in the directed acyclic graph (i.e., the influence relationship between abnormal features) is set according to prior knowledge, and the transmission probability between nodes is extracted from the directed acyclic graph using probability theory, statistics or machine learning methods, and organized into a matrix form, where rows and columns represent nodes, and the matrix elements represent the transmission intensity (i.e., transmission probability) between corresponding nodes, so as to construct a transmission intensity matrix. Taking the communication archive query service of a certain power grid as an example, the abnormal query behavior first causes the change of frequency characteristics, with a conditional probability of 0.85, and then transfers to the data volume characteristics, with a conditional probability of 0.72, and finally affects the time distribution characteristics, with the conditional probability reduced to 0.45.

[0085] The present invention uses a Bayesian network to perform probabilistic reasoning on the propagation intensity matrix, calculates the conditional probability distribution between nodes, and uses the maximum likelihood estimation method to calculate the conditional probability distribution parameters between nodes. Combining the probabilistic reasoning results and the conditional probability distribution, the abnormal pattern impact mechanism model is constructed, which can also be considered as an impact transmission model. Among them, according to the probabilistic reasoning results of the Bayesian network, the impact of abnormal features (such as frequency features, time distribution features, etc.) on the privacy protection level has a cumulative effect.

[0086] For each abnormal pattern, the degree of its impact on the privacy protection level is calculated according to the weight of its path and node in the model. This can be achieved by simulating the propagation process of the abnormal pattern in the model and observing its impact on the privacy protection level assessment standard. The degree of impact can be represented by a quantitative value, such as a number between 0 and 1, where 0 represents no impact and 1 represents the maximum impact, and it is used as an attenuation coefficient to reflect the degree to which the abnormal pattern weakens the privacy protection level. Then, according to multiple aspects of privacy protection (such as data collection, storage, processing, transmission, etc.), a hierarchical structured privacy impact assessment indicator is designed. Each layer of indicators corresponds to an aspect of privacy protection, and the next layer of indicators is a refinement or specific manifestation of the previous layer of indicators. For each indicator, a value is assigned according to the attenuation coefficient of the corresponding abnormal pattern, and all indicators are organized in a hierarchical structure to form a complete privacy impact assessment system. Finally, the quantitative value of each indicator in the hierarchical structured privacy impact assessment indicator is used as an element of the vector to construct a quantitative assessment vector, and the dimension of the quantitative assessment vector is equal to the number of indicators in the assessment system.

[0087] The quantitative evaluation vectors are statistically analyzed for distribution, the mean and variance of the impact of different tenant groups are calculated, and the warning threshold based on the confidence interval is set to obtain the risk level mark. The method is applied to the access of power grid communication data, the data desensitization intensity is reduced by 15%, and the privacy protection level of the query results is reduced by 28%. This impact spreads to related tenants through the network structure. The average protection level of first-level associated tenants is reduced by 18%, and that of second-level associated tenants is reduced by 12%. The statistical distribution of the quantitative evaluation vectors shows significant stratification characteristics. The mean impact of the directly associated tenant group is 0.72, and the variance is 0.15; the mean impact of the indirectly associated tenant group is 0.45, and the variance is 0.23. The warning threshold set based on the 95% confidence interval is 0.65. Tenants exceeding this threshold are marked as high risk levels, and then a data unauthorized access incident involving 15 associated tenants is successfully identified, of which the impact of 8 tenants exceeds the warning threshold, forming a typical spreading risk propagation model.

[0088] By constructing feature association graphs and difference feature sets, the present invention can more accurately identify the characteristics of abnormal tenants; by constructing directed acyclic graphs and propagation intensity matrices, it can reveal the propagation paths and intensities of abnormal patterns among different features; by constructing hierarchical structured privacy impact assessment indicators and quantitative evaluation vectors, it can quantitatively evaluate the overall impact of abnormal patterns on the privacy protection level, so as to facilitate the subsequent adjustment of the privacy budget allocation ratio, thereby solving the abnormal problems in the ciphertext execution of the privacy computing algorithm.

[0089] S5. According to the quantitative analysis result, the privacy budget allocation ratio of each abnormal tenant is adjusted to obtain and execute a privacy budget allocation plan to handle the abnormality in the ciphertext execution of the privacy computing algorithm;

[0090] In one embodiment, step S5 includes:

[0091] Constructing a risk assessment function according to the quantitative analysis result to perform risk scoring on each of the abnormal tenants, and determining the association relationship between each of the abnormal tenants based on the relationship between the risk scoring result and a preset risk threshold;

[0092] An exponential weight method is used to quantify the influence propagation coefficient of each abnormal tenant, and an association strength matrix is ​​constructed in combination with the association relationship, and the association strength matrix is ​​subjected to characteristic decomposition to obtain a tenant association index;

[0093] A linear attenuation function is set according to the tenant association index to classify the budget reduction ratio, and a normal distribution function is used to nonlinearly reduce the privacy budget allocation ratio of each abnormal tenant to obtain a first budget adjustment set;

[0094] Based on the first budget adjustment set, the compensation coefficient of each abnormal tenant is calculated by using the least square method, and a compensation weight vector is set in combination with the data query sensitivity corresponding to each abnormal tenant to obtain a second budget adjustment set;

[0095] The second budget adjustment set is optimized by the gradient descent method to obtain and execute the privacy budget allocation plan for each abnormal tenant, so as to handle the anomalies in the ciphertext execution of the privacy computing algorithm.

[0096] Specifically, the present invention aims to identify the potential threat of each abnormal tenant to the privacy protection level and quantify the degree of such threat. It can comprehensively reflect the tenant's performance in privacy protection, such as the number of privacy leaks, the scale of privacy leaks, the impact of privacy leaks, etc. as evaluation indicators, and based on the quantitative analysis results (i.e., the overall impact of each abnormal mode on the privacy protection level) combined with these indicators for weighted summation to construct a risk assessment function. Each weight can be determined by expert scoring, questionnaire survey, data analysis and other methods, without specific limitation, so that the risk assessment function can calculate the risk score for each abnormal tenant; then compare the calculated risk score with the preset risk threshold to determine the association relationship between each abnormal tenant. The relationship can be represented by a graph theory method, with abnormal tenants as nodes and association relationships as edges. In the cross-tenant data access scenario, the impact of abnormal tenants is quantified by the risk assessment function. Taking a certain payment service tenant as an example, its risk score is 0.85, which is higher than the preset threshold of 0.7, and the risk score of each abnormal tenant is compared with the preset risk threshold to determine that its impact range involves 15 directly associated tenants and 25 indirectly associated tenants.

[0097] The exponential weight method is used to quantify the influence propagation coefficient of each abnormal tenant, which takes into account factors such as transaction frequency and transaction volume between tenants. In the cross-tenant data access scenario, the influence propagation coefficients calculated by exponential weight are 0.75, 0.45 and 0.25, corresponding to the first-level, second-level and third-level associated tenants respectively. The influence propagation coefficients are 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 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 closeness of the association relationship. The matrix is ​​used to reflect the association strength and influence degree between abnormal tenants. Then, the association strength matrix is ​​eigendecomposed to obtain eigenvalues ​​and eigenvectors. According to the size 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 index of each tenant. This can be achieved by linearly combining the eigenvector with the tenant's influence propagation coefficient. The combination coefficient can be the eigenvalue, and the association index reflects the position and importance of the abnormal tenant in the overall association network.

[0098] A linear attenuation function is set according to the quantified association index. For example, a baseline budget reduction ratio is set, and linear attenuation is performed according to a certain ratio according to the score of the association index to obtain a linear attenuation function. The budget reduction ratio is divided into different levels. For example, the budget reduction ratio of directly associated tenants in the first interval of the tenant association index is 35%, the budget reduction ratio of second-level associated tenants in the second interval is 20%, and the budget reduction ratio of third-level associated tenants in the third interval is 10%. The total budget reduction amount is 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; then 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, and the normal distribution function is used to nonlinearly reduce the privacy budget allocation ratio set by each abnormal tenant according to its historical privacy protection performance, business scale, data sensitivity and other factors to obtain the first budget adjustment set. For example, the initial privacy budget of the abnormal tenant in the cross-tenant data access scenario is reduced from 1.0 to 0.35. The specific reduction amount is distributed among associated tenants at different levels through the normal distribution function.

[0099] Based on the first budget adjustment set, the least squares method is used to calculate the compensation coefficient of each abnormal tenant, which is between 0.2 and 0.8 to ensure the balance of the overall budget; combined with the data query sensitivity corresponding to each abnormal tenant, which can also be considered as the data query sensitivity level corresponding to each abnormal tenant, the compensation weight vector is set to adjust the compensation coefficient to obtain the second budget adjustment set. For example, the compensation weight of abnormal tenants with high sensitivity levels (such as power plant output power, substation load situation access, etc.) is 0.8, the compensation weight of abnormal tenants with medium sensitive data levels (such as users' power usage patterns, historical energy consumption records, etc.) is 0.5, and the compensation weight of abnormal tenants with low sensitive data levels (such as power company announcements, etc.) is 0.2, and the second budget adjustment set is obtained.

[0100] The second budget adjustment set is optimized and allocated by the gradient descent method, and the upper limit of compensation for a single tenant is set to no more than 50% of the original budget, and the total compensation amount is set to no more than 70% of the total reduction amount. In the iterative optimization process, the convergence threshold of the objective function is set to 0.001, and the maximum number of iterations is 1000, and finally the compensation budget allocation vector is obtained; then, a Lagrangian function is constructed based on the compensation budget allocation vector, and non-negative constraints and smoothing constraints are set (the compensation coefficient of adjacent abnormal tenants is not greater than 0.1), and the objective function is solved by quadratic programming to obtain the optimal compensation plan; then, the optimal compensation plan is subjected to residual analysis, and the deviation distribution between the adjusted budget and the original budget is calculated, and the compensation value exceeding the threshold is corrected by setting a tolerance range; finally, the tenant privacy budget allocation table is updated according to the corrected compensation plan, and the budget adjustment results are dynamically tracked by the exponential smoothing method with a smoothing coefficient set to 0.3, and the final privacy budget allocation plan is obtained and executed.

[0101] In actual applications, abnormal tenant behavior occurred in a bank's credit data access scenario. After adopting the above solution, the abnormal tenant with an original budget of 1.0 was reduced to 0.4, the budgets of its 8 directly associated tenants were reduced by an average of 0.3, and the budgets of 12 secondary associated tenants were reduced by an average of 0.15; through optimization calculation, 20 normal tenants received compensation, and the compensation values ​​ranged from 0.02 to 0.08, with a total compensation of 0.42; and considering the frequency of data access, high-frequency tenants received larger compensation to maintain query accuracy, and the relative error after compensation was controlled within 5%. After continuous monitoring of 30 time windows, the fluctuation range of the final budget allocation plan was controlled within the range of ±0.05, which demonstrated good stability, achieved precise suppression of abnormal tenants and differentiated compensation for normal tenants, and ensured the dynamic balance of differential privacy protection in a multi-tenant environment.

[0102] By constructing a risk assessment function and an association strength matrix, the present invention can more accurately assess the risks and association relationships of abnormal tenants, providing a reliable basis for subsequent budget adjustments; by adopting a nonlinear reduction and compensation mechanism, the privacy budget can be allocated more reasonably, ensuring that while protecting data privacy, the computing efficiency and accuracy are improved; by optimizing the privacy budget allocation scheme through the gradient descent method, various abnormal situations can be dealt with, and the robustness and adaptability of the algorithm can be improved; while protecting data privacy, abnormal tenants can also be effectively handled to prevent privacy leakage and data abuse, which helps to maintain data security and user rights.

[0103] In one embodiment, the step of optimizing the second budget adjustment set by the gradient descent method to obtain and execute the privacy budget allocation scheme for each abnormal tenant includes:

[0104] Extracting a sensitivity sequence during the execution of the privacy budget allocation scheme according to a differential privacy query function, performing frequency domain analysis on the sensitivity sequence by fast Fourier transform to obtain a feature fingerprint and performing singular value decomposition to obtain a first feature vector;

[0105] The similarity between the first feature vector and the pre-stored abnormal fingerprint feature cluster is calculated by using the Mahalanobis distance to construct a similarity matrix, and the similarity matrix is ​​normalized by setting a distance threshold to obtain a second feature vector;

[0106] Performing spectral clustering on the second eigenvector, and setting a clustering radius and a minimum number of samples to group the feature space of the second eigenvector, so as to calculate the local density and relative distance of each cluster, and obtain a density distribution matrix;

[0107] Extracting a cluster center point set from the density distribution matrix, constructing a feature topology map based on the cluster center point set and inter-cluster connectivity, and hierarchically marking the feature topology map to obtain a hierarchical structure matrix;

[0108] Calculating the budget adjustment amount of each abnormal tenant according to the hierarchical matrix, and using an exponential decay function to impose a budget limit on each budget adjustment amount, and weighting the budget adjustment amount by the tenant association index to obtain a budget adjustment plan;

[0109] Based on the budget adjustment plan, a compensation resource allocation network is constructed, 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 plan and execute it to achieve the handling of anomalies in the ciphertext execution of the privacy computing algorithm.

[0110] Specifically, the present invention extracts the sensitivity sequence during the execution of the privacy budget allocation scheme 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 elaborated here; then the fast Fourier transform (FFT) is used to perform frequency domain conversion on the sensitivity sequence to identify the periodic or trend characteristics therein. The output result of FFT is the feature fingerprint, which contains the information of the sensitivity sequence in the frequency domain, namely the amplitude spectrum and the phase spectrum, and uses it as the feature fingerprint for singular value decomposition to obtain the first eigenvector.

[0111] The similarity between the first eigenvector and the pre-stored abnormal feature cluster is calculated by the Mahalanobis distance to construct a similarity matrix, which is used to represent the closeness between the first eigenvector and the feature cluster. A distance threshold is set to normalize the similarity matrix, and an obvious clustering structure is formed in the high-dimensional feature space to obtain the second eigenvector.

[0112] Then, spectral clustering is performed on the second eigenvector, and the clustering radius is set to 0.3 and the minimum number of samples is set to 5 to group the feature space of the second eigenvector, so as to calculate the local density and relative distance of each cluster, and obtain the density distribution matrix to reflect the compactness and distribution characteristics between clusters; among them, the density distribution shows that the local density of the abnormal cluster is 0.85, which is significantly higher than 0.32 of the normal cluster, and the relative distance between clusters is 2.4, indicating that the abnormal characteristics have significant clustering; in addition, a similar phenomenon is also observed in the power grid communication data access scenario. When an abnormality occurs in a communication archive query, the local density reaches 0.92, forming a typical high-density abnormal cluster.

[0113] Based on the density distribution matrix, a certain strategy (such as selecting the point with the largest local density or based on a comprehensive index of density and distance) is used to determine the cluster center of each cluster, and the determined cluster center points are combined to form a cluster center point set, and each cluster center point corresponds to a node. The edge is defined according to the connectivity between clusters, that is, if two cluster center points meet certain connectivity conditions under a certain metric (such as distance, similarity, etc.), an edge is drawn between the two nodes. In order to describe the relationship between clusters in more detail, a weight can be set for each edge, which can be calculated based on the distance, similarity or other relevant indicators between the cluster center points; then according to the defined nodes and edges, and weights, use graph theory tools or graph data structures in programming languages ​​to build a feature topology map; finally, select a suitable hierarchical algorithm, such as hierarchical clustering, community discovery algorithm, etc., identify the hierarchical structure in the feature topology map and assign nodes to different levels, use the feature topology map as input to execute the selected hierarchical algorithm, and output the hierarchical information to which each node belongs; then, based on the output results of the hierarchical algorithm, build a hierarchical matrix, the rows and columns of the matrix correspond to the nodes in the graph, and the elements in the matrix represent the hierarchical relationship between the nodes; specifically, if two nodes belong to the same level, the corresponding element is 1 (or a non-zero value), otherwise it is 0. In addition, more information about the hierarchical relationship can be recorded in the hierarchical matrix as needed, such as hierarchical depth, parent-child relationship, etc.; finally, verify and adjust the hierarchical matrix to ensure that the matrix correctly reflects the hierarchical structure in the feature topology map, and refine or simplify the matrix as needed.

[0114] According to the hierarchical matrix, the basic budget adjustment amount is set for tenants at different levels. The adjustment amount is determined according to the level of the level, the degree of abnormality or other relevant factors. For each abnormal tenant, the specific budget adjustment amount is calculated according to the level and the basic adjustment amount. A suitable exponential decay function is selected to take the budget adjustment amount of each abnormal tenant as input, and the selected exponential decay function is applied for calculation. The final budget adjustment amount of each abnormal tenant is determined by output. These adjustments should reflect the degree of influence of the tenant in the abnormality and meet the requirements of smoothness and controllability of the budget limit. Then, based on the tenant correlation index, the budget adjustment amount of each tenant is calculated. A weighting coefficient is determined for each tenant. The quantified correlation value of the tenant correlation index can also be used as a weighting coefficient to reflect the relative importance of the tenant in the budget adjustment process. For example, a tenant closely associated with multiple abnormal tenants may face a larger weighting coefficient. The final budget adjustment amount of each abnormal tenant is multiplied by its weighting coefficient to obtain a weighted budget adjustment plan, which will make the budget adjustment more in line with the actual relationship and business needs between tenants. Finally, the weighted budget adjustment amount of each abnormal tenant is summarized to obtain an overall budget adjustment plan, which should list in detail the budget adjustment amount of each tenant, the adjusted total budget, and other information. In cross-departmental data sharing in banks, this hierarchical adjustment strategy effectively controls the spread of risks. The budget of abnormal tenants is reduced from 1.0 to 0.4, and the average budget of directly associated tenants is adjusted from 0.8 to 0.6.

[0115] Based on the budget adjustment plan, a compensation resource allocation network is constructed, which includes source nodes (representing resource providers or budget sources), abnormal tenant nodes (representing resources or budget recipients that need to be compensated), and sink nodes (representing the end points of resources or budgets, which may be used to represent the balance or restriction of the overall budget), as well as edges: representing the flow path of resources or budgets between nodes. Each edge has capacity constraints and network traffic constraints, representing the maximum capacity and maximum transmission volume of resources or budgets on the path; then the compensation resource allocation network is solved by the maximum flow minimum cut algorithm (such as the Ford-Fulkerson algorithm, the Edmonds-Karp algorithm, etc.) to obtain the compensation allocation plan, and the residual analysis of the compensation allocation plan is performed to calculate the budget change before and after compensation. The fluctuation constraint threshold is set according to business needs and risk tolerance to obtain the compensation execution sequence, and the compensation plan with a small budget change and that meets the fluctuation constraint is selected as the optimal compensation allocation plan and executed, so as to achieve the handling of anomalies in the ciphertext execution of the privacy computing algorithm.

[0116] In the user data access of telecom operators, the implementation of this solution successfully handled the abnormal events caused by batch data export. After discovering the abnormality through real-time feature matching, the budget adjustment and compensation allocation were completed within 5 minutes, effectively protecting the privacy data of 3 million users, realizing the rapid identification and accurate response of abnormal behavior, and ensuring the continuity of normal business. The present invention can more accurately identify the abnormal situation in the privacy budget allocation process through multi-step processing such as frequency domain analysis, feature extraction and similarity calculation; the budget adjustment amount is limited and weighted by using exponential decay function and tenant association index, making the budget adjustment more reasonable and fair; by constructing a compensation resource allocation network and solving it using the maximum flow minimum cut algorithm, the optimality of resource allocation can be ensured and the resource utilization efficiency can be improved; the abnormal situation in the privacy budget allocation process can be discovered and handled in a timely manner, thereby enhancing the ciphertext execution capability of the privacy computing algorithm and improving the level of data privacy protection.

[0117] In one embodiment, after optimizing the second budget adjustment set by the gradient descent method to obtain and execute the privacy budget allocation scheme for each abnormal tenant, the method further includes:

[0118] Perform multi-scale decomposition on the sensitivity sequence, extract high-frequency coefficients, low-frequency coefficients and energy distribution features from the decomposition results to construct a decomposition feature vector, and quantify the standardized cosine distance between the decomposition feature vector and the pre-stored abnormal fingerprint feature cluster to obtain a matching degree matrix;

[0119] The matching degree matrix is ​​classified as abnormal by using a Gaussian kernel support vector machine to obtain a first risk assessment value to construct a directed weighted graph, and the in-degree and out-degree of the nodes in the directed weighted graph are calculated to obtain a secondary influence propagation coefficient to quantify the information entropy loss of each abnormal tenant to obtain a second risk assessment value;

[0120] The secondary compensation weight value is calculated according to the second risk assessment value, and the secondary compensation weight value is optimized through the total budget conservation constraint and the upper limit constraint of the single tenant compensation, so as to obtain the optimal budget adjustment plan for each abnormal tenant and execute it, so as to realize the handling of anomalies in the ciphertext execution of the privacy computing algorithm.

[0121] Specifically, the present invention performs a db4 wavelet base 4-layer decomposition on the sensitivity sequence, extracts high-frequency coefficients (representing detailed information) and low-frequency coefficients (representing overall trends) and energy distribution features (such as energy proportions at various scales, etc.) from the decomposition results, and combines these extracted data to form a decomposition feature vector; then the cosine similarity measurement method is used to calculate the similarity between the decomposition feature vector and the pre-stored abnormal fingerprint feature cluster, and a normalization process is performed to construct a matching degree matrix for representing the matching degree between the decomposition feature vector and the abnormal fingerprint feature cluster according to the calculated standardized cosine distance. Taking the cross-tenant data access of a payment platform as an example, after the sensitivity sequence is decomposed by wavelet, an abnormal high-frequency component is detected at the third layer, and the energy proportion reaches 0.45, which is significantly higher than 0.12 in the normal state. The mean value of the low-frequency coefficient increases from 0.35 to 0.82, indicating that the query behavior is persistently abnormal, and the standardized cosine distance between the feature vector and the known abnormal fingerprint is 0.15, which is much lower than the normal threshold of 0.65.

[0122] The kernel function parameter of the Gaussian kernel support vector machine is set to 0.8, the penalty factor is set to 10, and the classification accuracy obtained by cross-validation is 0.92. The matching degree matrix is ​​used to classify abnormalities. One or more risk assessment indicators are defined according to business needs. These indicators can be determined based on the classification results (such as the proportion of abnormal tenants), the statistical properties of the feature vector (such as the mean, variance), or in combination with other business information. Then, according to the defined risk assessment indicators, a first risk assessment value is calculated. This value can be a numerical value, indicating the risk level of the abnormal tenant; then, the first risk assessment value is used as an edge and each abnormal tenant is used as a node to construct a directed weighted graph, and the in-degree and out-degree of the nodes in the graph are calculated to obtain the secondary influence propagation coefficient to quantify the influence of the node in the abnormal propagation; finally, the secondary influence propagation coefficient and the specific characteristics of the abnormal tenant (such as the amount of data, sensitivity, etc.) are used to quantify the information entropy loss of each abnormal tenant, and the information entropy loss of all abnormal tenants is summarized or weighted averaged to obtain the second risk assessment value to indicate the overall risk level of the entire system when facing abnormal tenants.

[0123] The second risk assessment value is standardized or normalized to be within a reasonable numerical range, and an initial secondary compensation weight value is assigned to each abnormal tenant according to the second risk assessment value. The weight value can be set to be a direct mapping of the second risk assessment value, or adjusted appropriately according to the business logic; then the total budget amount for compensation is set to 120% of the original budget, and the maximum compensation amount that each abnormal tenant can obtain is set to 50% of the original budget. An optimization model is constructed with the secondary compensation weight value as a variable and the goal of maximizing the overall satisfaction of all tenants (based on a certain combination of the compensation amount and the risk assessment value); then the secondary compensation weight value is optimized using an optimization algorithm, and a weight value combination that satisfies the constraints and has the optimal objective function is found through iterative calculation; the optimal secondary compensation weight value of each abnormal tenant is extracted from the optimization result, and the compensation amount that each abnormal tenant should obtain is calculated according to the calculated weight value and the total budget, and then the budget of the abnormal tenant is adjusted according to the calculated compensation amount. This involves steps such as negotiation with the tenant and signing of a compensation agreement, which is not the focus of this plan and will not be described in detail here, and the optimal budget adjustment plan for each abnormal tenant can be obtained.

[0124] The present invention can quickly and accurately identify abnormal data through the combined use of standardized cosine distance and Gaussian kernel support vector machine; through the combined use of standardized cosine distance and Gaussian kernel support vector machine, abnormal data can be quickly and accurately identified; by constructing a directed weighted graph and calculating the quadratic influence propagation coefficient, the risk of abnormal data can be more accurately assessed; by optimizing the quadratic compensation weight value and obtaining the optimal budget adjustment plan, the reasonable allocation and effective utilization of the budget can be ensured; through the processing of abnormal data and the execution of the budget adjustment plan, the security and stability of the privacy computing algorithm in the face of abnormal situations can be improved.

[0125] In the embodiment of the present application, based on the problem of how to detect and handle the sensitivity anomaly of tenants in the differential privacy query execution process in a multi-tenant cloud environment, a ciphertext execution anomaly handling method of a privacy computing algorithm is designed. The differential privacy query parameters of several tenants are obtained, and the data query sensitivity of each tenant is quantified according to the differential privacy query parameters to obtain sensitivity calculation feature data; the sensitivity calculation feature data is analyzed by time series through the ARIMA model to obtain sensitivity prediction data, and the residual between the sensitivity calculation feature data and the sensitivity prediction data is quantified to evaluate whether each tenant has a sensitivity shift anomaly, and obtain abnormal tenants; a sliding window is used to extract each tenant's sensitivity. The statistical features in the sensitivity calculation feature data of the abnormal tenants are used to construct a feature set, and the feature set is clustered and analyzed to obtain the abnormal patterns and abnormal fingerprint feature clusters corresponding to each abnormal tenant; the normal fingerprint feature clusters of each abnormal tenant are obtained to construct an abnormal pattern impact mechanism model in combination with the abnormal fingerprint feature clusters, and the abnormal pattern impact mechanism model is used to quantify the degree of influence of each abnormal pattern on the privacy protection level of each abnormal tenant to obtain a quantitative analysis result; the privacy budget allocation ratio of each abnormal tenant is adjusted according to the quantitative analysis result, and a privacy budget allocation plan is obtained and executed to achieve effective detection and processing of anomalies in the ciphertext execution of the privacy computing algorithm.

[0126] It should be noted that although the steps in the above flowchart are shown in sequence according to 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 restriction for the execution of these steps, and these steps can be executed in other orders.

[0127] In another embodiment, if Figure 2 As shown, the second aspect of the present invention provides a ciphertext execution exception processing system of a privacy computing algorithm, comprising:

[0128] A feature data quantification module 10 is used to obtain differential privacy query parameters of several tenants, and quantify the data query sensitivity of each tenant according to the differential privacy query parameters to obtain sensitivity calculation feature data;

[0129] The abnormal tenant evaluation 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 evaluate whether each of the tenants has sensitivity deviation anomalies, and obtain abnormal tenants;

[0130] The abnormal pattern classification module 30 is used to extract statistical features in the sensitivity calculation feature data of each abnormal tenant using a sliding window to construct a feature set, and perform cluster analysis on the feature set to obtain abnormal patterns and abnormal fingerprint feature clusters corresponding to each abnormal tenant;

[0131] The impact degree quantification module 40 is used to obtain the normal fingerprint feature cluster of each abnormal tenant, to build an abnormal pattern impact mechanism model in combination with the abnormal fingerprint feature cluster, and to quantify the impact degree of each abnormal pattern on the privacy protection level of each abnormal tenant through the abnormal pattern impact mechanism model to obtain a quantitative analysis result;

[0132] The allocation scheme execution module 50 is used to adjust the privacy budget allocation ratio of each abnormal tenant according to the quantitative analysis result, obtain and execute the privacy budget allocation scheme, so as to handle the anomalies in the ciphertext execution of the privacy computing algorithm.

[0133] It should be noted that each module in the above-mentioned ciphertext execution exception handling system of a privacy computing algorithm can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules. For the specific definition of the ciphertext execution exception handling system of a privacy computing algorithm, please refer to the definition of the ciphertext execution exception handling method of a privacy computing algorithm above. The two have the same functions and effects, which will not be repeated here.

[0134] In summary, the present invention relates to the field of information technology, and discloses a method and system for handling ciphertext execution exceptions of a privacy computing algorithm. The method collects differential privacy query parameters of several tenants to quantify the data query sensitivity of each tenant, and uses the ARIMA model to perform time series analysis on the quantification results of the data query sensitivity, predict future sensitivity trends, and calculate the residual between the actual sensitivity and the predicted sensitivity to identify abnormal tenants; a sliding window method is used to extract the statistical features of abnormal tenants to construct a feature set, and cluster analysis is performed to identify the abnormal patterns and abnormal fingerprint feature clusters corresponding to each abnormal tenant; a model is trained based on the positive and abnormal fingerprint features of the abnormal tenants, and the model is used to quantify the degree of influence of each abnormal pattern on the privacy protection level of the abnormal tenants, and the privacy budget allocation ratio of each abnormal tenant is adjusted and executed according to the obtained quantitative analysis results, so as to achieve effective handling of anomalies in the ciphertext execution of the privacy computing algorithm.

[0135] Each embodiment in this specification is described in a progressive manner, and the same or similar parts of each embodiment can be directly referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. It should be noted that the technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0136] The above-mentioned embodiments only express several preferred implementation modes of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in the technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be regarded as the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be based on the protection scope of the claims.

Claims

1. A method for handling ciphertext execution exceptions of a privacy computing algorithm, characterized in that: include: Acquire differential privacy query parameters of several tenants, and quantify the data query sensitivity of each tenant according to the differential privacy query parameters to obtain sensitivity calculation feature data; Performing time series analysis on the sensitivity calculation feature data through the ARIMA model to obtain sensitivity prediction data, and quantifying the residual between the sensitivity calculation feature data and the sensitivity prediction data to evaluate whether each of the tenants has sensitivity deviation anomalies, and obtain abnormal 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 a cluster analysis is performed on the feature set to obtain an abnormal pattern and an abnormal fingerprint feature cluster corresponding to each abnormal tenant; Acquire the normal fingerprint feature cluster of each abnormal tenant, build an abnormal pattern impact mechanism model in combination with the abnormal fingerprint feature cluster, and quantify the impact of each abnormal pattern on the privacy protection level of each abnormal tenant through the abnormal pattern impact mechanism model to obtain a quantitative analysis result; According to the quantitative analysis result, the privacy budget allocation ratio of each abnormal tenant is adjusted to obtain and execute a privacy budget allocation plan to achieve the processing of anomalies in the ciphertext execution of the privacy computing algorithm; The differential privacy query parameters include a data differential value and a weight coefficient; the differential privacy query parameters of a plurality of tenants are obtained, and the data query sensitivity of each tenant is quantified according to the differential privacy query parameters to obtain sensitivity calculation feature data, including: Acquire data differential values ​​and weight coefficients of several tenants, and obtain a first tenant isolation boundary value by performing a dot product operation and normalization processing on the data differential values ​​and the weight coefficients to construct a tenant differential parameter matrix; each row in the tenant differential parameter matrix contains a tenant identifier and its corresponding data query sensitivity; A data query sensitivity threshold is set according to the tenant differential parameter matrix, and each of the tenants is grouped and clustered according to the data query sensitivity threshold to obtain a plurality of tenant groups with different data query sensitivity levels, and a differential privacy query parameter of each of the tenant groups is quantized and encoded according to a preset desensitization rule to obtain a first feature space; Using a gradient boosting tree to associate and map the parameter range interval in the first feature space with the data query sensitivity level, output a parameter mapping vector, and perform random forest feature extraction on the parameter mapping vector to obtain a second feature space; Obtaining an intermediate indicator of global sensitivity calculation in the second feature space, processing the intermediate indicator of global sensitivity calculation by an adaptive clustering algorithm to obtain a set of cluster center points, and constructing a feature vector dimension matrix based on the set of cluster center points; Performing an orthogonal transformation on the feature vector dimension matrix to obtain a standardized feature vector, and performing a weighted summation on the standardized feature vector and the first tenant isolation boundary value to generate a differential privacy query parameter set as the sensitivity calculation feature data output; The time series analysis of the sensitivity calculation feature data is performed by the ARIMA model to obtain sensitivity prediction data, and the residual between the sensitivity calculation feature data and the sensitivity prediction data is quantified, including: The sensitivity calculation feature data is sampled at fixed intervals through a sliding time window to obtain sampled data, and a time series feature matrix is ​​constructed by quantizing an autocorrelation function and a partial autocorrelation function of the sampled data; Performing stationarity processing on the time series feature matrix to obtain a stationary time series data set to construct an ARIMA model, estimating the autoregressive term coefficient and the moving average term coefficient of the ARIMA model by the least squares method, and performing a significance test on the parameters of the ARIMA model to obtain a goodness of fit index; The optimal parameter combination is screened by the goodness of fit index, and the ARIMA model with the optimal parameter combination is used for multi-step iterative prediction on the validation data set to calculate the confidence interval of the prediction result and obtain the sensitivity prediction sequence; A Kalman filter is constructed according to the sensitivity prediction sequence, a state transfer matrix and an observation matrix are set, sensitivity prediction data are obtained through prediction update and measurement update, and the residual is calculated based on the sensitivity prediction data and the sampling data.

2. According to the method for handling ciphertext execution exceptions of a privacy computing algorithm according to claim 1, it is characterized in that: The method further includes performing time series analysis on the sensitivity calculation feature data through the ARIMA model to obtain sensitivity prediction data, and quantifying the residual between the sensitivity calculation feature data and the sensitivity prediction data. By quantifying the Pearson autocorrelation coefficient and the partial autocorrelation coefficient of the sampled data, a multidimensional sensitivity initial sequence is obtained, and a stationarity test is performed on the multidimensional sensitivity initial sequence to obtain a multidimensional stationary sensitivity sequence; An ARIMA model is constructed according to the multidimensional stationary sensitivity sequence, and a log-likelihood function is optimized by a maximum likelihood estimation method to calculate an autoregressive coefficient and a moving average coefficient of the ARIMA model to obtain a fitting parameter set; Performing residual analysis on the multidimensional stationary sensitivity sequence according to the fitting parameter set, and selecting an optimal model order combination based on the Akaike information criterion and the Bayesian information criterion, so as to obtain a prediction parameter set based on the ARIMA model prediction with the optimal model order combination; A radial basis kernel support vector regression model is constructed based on the prediction parameter set, and a kernel function bandwidth parameter and a penalty factor are set, and the optimal hyperparameters of the radial basis kernel support vector regression model are determined by cross-validation, so as to predict a multidimensional sensitivity prediction sequence according to 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 sampling data is calculated to obtain the residual.

3. According to the method for handling ciphertext execution exceptions of a privacy computing algorithm according to claim 1, it is characterized in that: The sliding window is used to extract statistical features in the sensitivity calculation feature data of each abnormal tenant to construct a feature set, and the feature set is clustered to obtain the abnormal mode and abnormal fingerprint feature cluster corresponding to each abnormal tenant, including: The sensitivity calculation feature data of each abnormal tenant is segmented by using a fixed-length sliding window, and each segmented data is mapped to the frequency domain by discrete Fourier transform to obtain frequency domain statistical features and perform maximum and minimum value normalization processing to obtain a first feature set; Performing dimensionality reduction on the first feature set, and performing singular value decomposition on the dimensionality reduction result to obtain a feature vector matrix, from which feature vector components whose cumulative contribution rates exceed a preset contribution threshold are selected for Bayesian probability weighted combination to obtain a second feature set; The second feature set is standardized, and the similarity between the data in the standardized second feature set is calculated by Mahalanobis distance, so as to determine the optimal number of clusters based on the similarity calculation result using the silhouette coefficient to obtain a clustering parameter set; A Gaussian mixture clusterer is constructed based on the clustering parameter set, and a covariance matrix and random initialization parameters are set, and the clustering center of the Gaussian mixture clusterer is iteratively optimized by using expectation maximization to obtain a first clustering result; Extracting the center point and covariance matrix of the cluster from the first clustering result, and constructing an anomaly measurement function according to the compactness index and isolation index quantified by the center point and covariance matrix of each cluster, and quantifying the anomaly score of each cluster by the anomaly measurement function; An abnormal pattern division range is set, each of the clusters is divided into a plurality of abnormal patterns based on the relationship between the abnormal score and the abnormal pattern division range, and an abnormal fingerprint feature cluster corresponding to each abnormal pattern is marked.

4. According to the method for handling ciphertext execution exceptions of a privacy computing algorithm according to claim 1, it is characterized in that: The normal fingerprint feature cluster of each abnormal tenant is obtained to build an abnormal pattern impact mechanism model in combination with the abnormal fingerprint feature cluster, and the impact degree of each abnormal pattern on the privacy protection level of each abnormal tenant is quantified by the abnormal pattern impact mechanism model to obtain a quantitative analysis result, including: Obtaining the normal fingerprint feature cluster of each abnormal tenant, and constructing an overlap matrix by quantifying the Mahalanobis distance and cosine similarity between the normal fingerprint feature cluster and the abnormal fingerprint feature cluster in the 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 according to the edge weight calculation results, and node features in the feature association graph are extracted to construct a difference feature set; Performing singular value decomposition on the difference feature set, sorting the decomposition results to determine the main difference components, calculating the importance of the main difference components through random forest, and obtaining an influence weight vector; Constructing a directed acyclic graph based on the influence weight vector, setting conditional dependencies between nodes in the directed acyclic graph, and extracting transmission probabilities between the nodes to construct a propagation intensity matrix; Probabilistic reasoning is performed on the propagation intensity matrix through a Bayesian network, and conditional probability distribution between the nodes is calculated through maximum likelihood estimation, so as to construct an abnormal mode impact mechanism model according to the probabilistic reasoning result and the conditional probability distribution calculation result; According to the abnormal pattern impact mechanism model, the degree of influence of the abnormal features corresponding to each abnormal pattern on the privacy protection level is calculated as the attenuation coefficient, and a hierarchical structured privacy impact assessment index is constructed and quantified through the attenuation coefficient to obtain a quantitative assessment vector.

5. According to the method for handling ciphertext execution exceptions of a privacy computing algorithm according to claim 1, it is characterized in that: The step of adjusting the privacy budget allocation ratio of each abnormal tenant according to the quantitative analysis result, obtaining and executing a privacy budget allocation plan to handle abnormalities in the ciphertext execution of the privacy computing algorithm includes: Constructing a risk assessment function according to the quantitative analysis result to perform risk scoring on each of the abnormal tenants, and determining the association relationship between each of the abnormal tenants based on the relationship between the risk scoring result and a preset risk threshold; An exponential weight method is used to quantify the influence propagation coefficient of each abnormal tenant, and an association strength matrix is ​​constructed in combination with the association relationship, and the association strength matrix is ​​subjected to characteristic decomposition to obtain a tenant association index; A linear attenuation function is set according to the tenant association index to classify the budget reduction ratio, and a normal distribution function is used to nonlinearly reduce the privacy budget allocation ratio of each abnormal tenant to obtain a first budget adjustment set; Based on the first budget adjustment set, the compensation coefficient of each abnormal tenant is calculated by using the least square method, and a compensation weight vector is set in combination with the data query sensitivity corresponding to each abnormal tenant to obtain a second budget adjustment set; The second budget adjustment set is optimized by the gradient descent method to obtain and execute the privacy budget allocation plan for each abnormal tenant, so as to handle the anomalies in the ciphertext execution of the privacy computing algorithm.

6. According to the method for handling ciphertext execution exceptions of a privacy computing algorithm according to claim 5, it is characterized in that: The step of optimizing the second budget adjustment set by the gradient descent method to obtain and execute the privacy budget allocation scheme for each abnormal tenant includes: Extracting a sensitivity sequence during the execution of the privacy budget allocation scheme according to a differential privacy query function, performing frequency domain analysis on the sensitivity sequence by fast Fourier transform to obtain a feature fingerprint and performing singular value decomposition to obtain a first feature vector; The similarity between the first feature vector and the pre-stored abnormal fingerprint feature cluster is calculated by using the Mahalanobis distance to construct a similarity matrix, and the similarity matrix is ​​normalized by setting a distance threshold to obtain a second feature vector; Performing spectral clustering on the second eigenvector, and setting a clustering radius and a minimum number of samples to group the feature space of the second eigenvector, so as to calculate the local density and relative distance of each cluster, and obtain a density distribution matrix; Extracting a cluster center point set from the density distribution matrix, constructing a feature topology map based on the cluster center point set and inter-cluster connectivity, and hierarchically marking the feature topology map to obtain a hierarchical structure matrix; Calculating the budget adjustment amount of each abnormal tenant according to the hierarchical matrix, and using an exponential decay function to impose a budget limit on each budget adjustment amount, and weighting the budget adjustment amount by the tenant association index to obtain a budget adjustment plan; Based on the budget adjustment plan, a compensation resource allocation network is constructed, 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 plan and execute it to achieve the handling of anomalies in the ciphertext execution of the privacy computing algorithm.

7. According to claim 6, a method for handling ciphertext execution exceptions of a privacy computing algorithm is characterized in that: After optimizing the second budget adjustment set by the gradient descent method to obtain and execute the privacy budget allocation scheme for each abnormal tenant, the method further includes: Perform multi-scale decomposition on the sensitivity sequence, extract high-frequency coefficients, low-frequency coefficients and energy distribution features from the decomposition results to construct a decomposition feature vector, and quantify the standardized cosine distance between the decomposition feature vector and the pre-stored abnormal fingerprint feature cluster to obtain a matching degree matrix; The matching degree matrix is ​​classified as abnormal by using a Gaussian kernel support vector machine to obtain a first risk assessment value to construct a directed weighted graph, and the in-degree and out-degree of the nodes in the directed weighted graph are calculated to obtain a secondary influence propagation coefficient to quantify the information entropy loss of each abnormal tenant to obtain a second risk assessment value; The secondary compensation weight value is calculated according to the second risk assessment value, and the secondary compensation weight value is optimized through the total budget conservation constraint and the upper limit constraint of the single tenant compensation, so as to obtain the optimal budget adjustment plan for each abnormal tenant and execute it, so as to realize the handling of anomalies in the ciphertext execution of the privacy computing algorithm.

8. A ciphertext execution exception handling system for a privacy computing algorithm, characterized in that: A ciphertext execution exception handling method applied to the privacy computing algorithm as claimed in claim 1, comprising: A feature data quantification module, used to obtain differential privacy query parameters of several tenants, and quantify the data query sensitivity of each tenant according to the differential privacy query parameters to obtain sensitivity calculation feature data; An abnormal tenant assessment module is used to perform time series analysis on the sensitivity calculation feature data through an 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 obtain an abnormal tenant; An abnormal pattern classification 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 perform cluster analysis on the feature set to obtain abnormal patterns and abnormal fingerprint feature clusters corresponding to each abnormal tenant; An impact degree quantification module is used to obtain the normal fingerprint feature cluster of each abnormal tenant, to build an abnormal pattern impact mechanism model in combination with the abnormal fingerprint feature cluster, and to quantify the impact degree of each abnormal pattern on the privacy protection level of each abnormal tenant through the abnormal pattern impact mechanism model to obtain a quantitative analysis result; An allocation scheme execution module is used to adjust the privacy budget allocation ratio of each abnormal tenant according to the quantitative analysis result, obtain and execute the privacy budget allocation scheme, so as to handle the anomalies in the ciphertext execution of the privacy computing algorithm.

Images