Private data leakage risk control method and device, equipment, medium and product

Abstract

The invention discloses a privacy data leakage risk control method and device, equipment, a medium and a product, and the method comprises the steps: carrying out the topic clustering and extraction of leakage data in a privacy data life cycle according to a keyword of a privacy leakage risk topic, and obtaining a probability matrix of the topic distribution of different leakage events; by calculating the similarity of the probability matrixes of the different leakage events, the relevance between the different privacy leakage events is analyzed, and the common feature vector of the privacy leakage risk is determined; and according to the common feature vector, constructing a data set of privacy leakage risk classification to perform model leakage risk probability training, and obtaining a leakage risk identification model. The invention provides a recognition model capable of quickly and accurately realizing key privacy data.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method, device, equipment, medium and product for controlling risk of privacy data leakage. Background Art

[0002] In recent years, with the development of mobile communications, Internet of Things technology and geographic information systems, the application of location-based services has become more and more common. In layman's terms, location-based services (LBS) are that users obtain the location information of mobile terminals (such as longitude and latitude coordinates) through a set of positioning technologies, use the mobile communication network to make location-related service requests, and obtain various location-related services provided by location service providers to mobile users themselves or others and communication systems. Location-based services have a wide range of applications in location positioning, route navigation, query, identification, detection, etc., and are hailed as one of the most dazzling service models of mobile Internet.

[0003] However, while the commercialization of location-based services has brought great convenience to people's lives, the hidden security and privacy leakage issues have gradually emerged. In the control of privacy leakage risks during the life cycle of location data and other private data, facing massive data, how to quickly and accurately identify key privacy data from massive heterogeneous data has become a major technical bottleneck in privacy protection. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a privacy data leakage risk control method, device, equipment, medium and product, and provides an identification model that can quickly and accurately realize key privacy data.

[0005] An embodiment of the present invention provides a method for controlling the risk of privacy data leakage, the method comprising:

[0006] Based on the keywords of the privacy leakage risk theme, the leakage data within the privacy data life cycle is clustered and extracted to obtain the probability matrix of the theme distribution of different leakage events;

[0007] By calculating the similarity of the probability matrices of different leakage events, the correlation between different privacy leakage events is analyzed, and the common feature vectors of privacy leakage risks are determined;

[0008] According to the common feature vector, a data set for privacy leakage risk classification is constructed to train the model leakage risk probability, thereby obtaining a leakage risk identification model.

[0009] Preferably, based on the keywords of the leakage risk theme, the leakage data within the privacy data life cycle is clustered and extracted to obtain a probability matrix of the theme distribution of different leakage events, including:

[0010] Using the Latent Dirichlet Allocation topic model algorithm with preset extraction parameters to extract representative words of the topic from the leaked data, and calculating the probability distribution of keywords under different topics;

[0011] According to the semantic similarity of representative words, different topics are merged and optimized to obtain topics of different categories;

[0012] Calculate the probability matrix of the topic distribution of different leak events.

[0013] Preferably, by calculating the similarity of the probability matrices of different leakage events, analyzing the correlation between different privacy leakage events, and determining the common feature vector of privacy leakage risk, including:

[0014] Normalize the probability matrix of different events to obtain the unitized eigenvector;

[0015] By calculating the dot product between the probability evidences, the cosine similarity values ​​of the probability matrices of different leakage events are determined;

[0016] According to the calculated cosine similarity, a preset dimensionality reduction algorithm is used to extract the principal components of the privacy leakage risk characteristics to form a privacy leakage risk characteristic subspace;

[0017] The original data is mapped into the privacy leakage risk feature subspace to determine the common feature vector of the privacy leakage risk.

[0018] As a preferred solution, according to the common feature vector, a data set for privacy leakage risk classification is constructed to train the model leakage risk probability, and a leakage risk identification model is obtained, including:

[0019] Based on the common feature vectors, the support vector machine algorithm is used to classify and label the privacy leakage data within the privacy data life cycle, and the labeling information of a part of the privacy leakage data is obtained as the basis of the training sample set;

[0020] For unlabeled data, we use the transductive learning method of support vector machines to propagate the label information of labeled samples to unlabeled samples by utilizing the structural information of labeled samples and the similarity between samples, thus constructing a dataset for privacy leakage risk classification.

[0021] A cross-validation scheme is adopted to randomly divide the data set into several mutually exclusive subsets, one of the subsets is selected as the validation set each time, and the remaining subsets are used as the training set, and the validation is repeated, and the average value of several validation results is taken as the performance estimation of the support vector machine model;

[0022] The model parameters are traversed and optimized through grid search or random search, and the parameter combination with the highest expected value is selected as the optimal parameter to obtain the leakage risk identification model.

[0023] Preferably, the method further comprises:

[0024] The preset association rule mining algorithm is used to analyze the association rules between different categories of privacy leakage risks in the life cycle of location data, and to determine the propagation path and impact scope of privacy leakage risks in the life cycle of location data.

[0025] Preferably, a preset association rule mining algorithm is used to analyze the association rules between different types of privacy leakage risks in the life cycle of location data, and determine the propagation path and impact scope of privacy leakage risks in the life cycle of location data, including:

[0026] The Apriori association rule mining algorithm is used to determine the support and confidence thresholds through cross-validation;

[0027] By presetting the evaluation indicators, the association rules between different types of privacy leakage risks in the privacy data life cycle are analyzed, and frequent item sets and strong association rules are screened out to form a privacy leakage risk association rule set;

[0028] According to the privacy leakage risk association rule set, a privacy leakage risk association graph is constructed, and privacy leakage risk nodes of different categories are connected according to the association rules to form a directed acyclic graph;

[0029] A preset mining algorithm is used on the directed acyclic graph to calculate the importance, authority and centrality of each privacy leakage risk node, identify key privacy leakage risk nodes and propagation paths, and determine the impact scope of privacy leakage risks within the life cycle of privacy data.

[0030] Furthermore, the method further comprises:

[0031] A propagation network topology diagram of the privacy leakage risk is constructed according to the propagation path and the impact scope, and key nodes and vulnerable links of the privacy leakage risk are identified.

[0032] Furthermore, a propagation network topology diagram of the privacy leakage risk is constructed according to the propagation path and the impact scope, and key nodes and vulnerable links of the privacy leakage risk are identified, including:

[0033] Using graph database technology, taking privacy leakage risk events as nodes and the correlation between different events as edges, a directed weighted graph is formed as the topology graph of the propagation network;

[0034] According to the propagation network topology graph, a label propagation algorithm is used to perform community discovery, and the labels of nodes are updated iteratively until the label distribution converges;

[0035] Dividing the communication network topology graph into a number of privacy leakage risk communities;

[0036] According to the characteristics of the privacy leakage risk community, using preset evaluation indicators, calculate the importance index of each privacy leakage risk node within the community and between communities, determine the importance ranking of the nodes in the propagation network topology diagram, and determine the key nodes of privacy leakage risk;

[0037] The shortest path algorithm is used to solve the shortest path between any two nodes, and the frequency of occurrence of intermediary nodes is counted to identify structural holes in the network and determine the vulnerable links for the propagation of privacy leakage risks.

[0038] Preferably, the method further comprises:

[0039] According to the key nodes and vulnerable links, the information entropy risk assessment model is used to calculate the privacy leakage risk values ​​of different nodes and links in the life cycle of privacy data to obtain risk assessment results.

[0040] Furthermore, according to the key nodes and vulnerable links, the information entropy risk assessment model is used to calculate the privacy leakage risk values ​​of different nodes and links in the life cycle of the privacy data, and the risk assessment results are obtained, including:

[0041] The life cycle of the privacy data is divided into multiple stages, and the fault tree analysis method is used to gradually decompose the privacy leakage event into multiple intermediate events and basic events, and the causal relationship between the events is represented by logic gates to form an inverted tree structure diagram;

[0042] Filter out basic events with higher importance as the object set for privacy leakage risk assessment;

[0043] For each risk assessment object, we use the hierarchical analysis method to compare and score the importance of each pair of factors at the same level from the two dimensions of the probability of occurrence and impact of privacy leakage incidents, and obtain a pairwise comparison matrix;

[0044] Calculate the weight vector of each criterion, and comprehensively weight them to obtain the privacy leakage risk score of each assessment object at the target layer. Take the arithmetic mean of the scoring results to obtain the qualitative assessment value of the privacy leakage risk.

[0045] According to the qualitative evaluation value, the probability of occurrence and severity of consequences of privacy leakage are quantified as the expected value of the membership function using the fuzzy comprehensive evaluation method, and the information entropy of the privacy leakage risk of each risk assessment object is obtained by substituting it into the calculation formula of the information entropy model;

[0046] The information entropy is normalized to obtain privacy leakage risk values ​​of different nodes and links as the risk assessment result.

[0047] Preferably, the method further comprises:

[0048] According to the risk assessment results, a fuzzy comprehensive evaluation method is used to determine the priority of privacy leakage risks within the life cycle of privacy data, and match the corresponding privacy protection strategy.

[0049] Furthermore, according to the risk assessment results, a fuzzy comprehensive evaluation method is used to determine the priority of privacy leakage risks within the life cycle of the privacy data, and a corresponding privacy protection strategy is matched, including:

[0050] Sort the privacy leakage risk values ​​of different nodes in the risk assessment result from high to low to form a sorting vector of privacy leakage risk;

[0051] Each risk object in the sorting vector is grouped according to the life cycle stage, and matched one by one with different evaluation indicators to form a risk indicator matching matrix, and a fuzzy evaluation indicator set of privacy leakage risk is constructed;

[0052] The weight vector of each index is determined by using the hierarchical analysis method, and the maximum eigenvalue and corresponding eigenvector of the matrix are calculated by constructing a pairwise comparison judgment matrix.

[0053] Substitute the quantitative evaluation value of each risk object into the membership function, calculate its membership on each evaluation index, and form a fuzzy relationship matrix;

[0054] Using the fuzzy matrix synthesis operation principle, the weight vector and the fuzzy relationship matrix are synthesized to obtain the fuzzy comprehensive evaluation result vector of privacy leakage risk;

[0055] Normalizing the fuzzy comprehensive evaluation result vector, and determining the priority of the privacy leakage risk according to the normalized privacy leakage risk evaluation index;

[0056] According to the priority, a corresponding privacy protection strategy is matched in a preset privacy protection measure library.

[0057] Preferably, the method further comprises:

[0058] By continuously monitoring the privacy leakage data during the life cycle of privacy data, the topic distribution, association rules and propagation network of privacy leakage risks are dynamically updated.

[0059] Furthermore, by continuously monitoring the privacy leakage data in the privacy data life cycle, the topic distribution, association rules and propagation network of privacy leakage risks are dynamically updated, including:

[0060] A real-time computing framework is used to process and analyze the data stream within the life cycle of private data in real time, and to extract characteristic indicators related to privacy leakage;

[0061] The characteristic indicators are matched in real time with the pre-trained privacy leakage risk topic model, association rule model and propagation network model to determine whether it is necessary to trigger the incremental update of the corresponding model and dynamically update the topic distribution, association rules and propagation path.

[0062] An embodiment of the present invention further provides a privacy data leakage risk control device, the device comprising:

[0063] The extraction module is used to cluster and extract the leaked data within the life cycle of the privacy data according to the keywords of the privacy leakage risk theme, and obtain the probability matrix of the theme distribution of different leakage events;

[0064] The vector module is used to analyze the correlation between different privacy leakage events by calculating the similarity of the probability matrices of different leakage events, and determine the common feature vectors of privacy leakage risks;

[0065] The training module is used to construct a data set for privacy leakage risk classification based on the common feature vector to train the model leakage risk probability and obtain a leakage risk identification model.

[0066] An embodiment of the present invention also provides a terminal device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the privacy data leakage risk control method as described in any of the above embodiments is implemented.

[0067] An embodiment of the present invention also provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the privacy data leakage risk control method as described in any of the above embodiments.

[0068] An embodiment of the present invention further provides a computer program product, including a computer program / instruction, which implements the steps of any of the above methods when executed by a processor.

[0069] The present invention provides a privacy data leakage risk control method, device, equipment, medium and product. According to the keywords of the privacy leakage risk theme, the leakage data within the privacy data life cycle is clustered and extracted to obtain the probability matrix of the theme distribution of different leakage events; by calculating the similarity of the probability matrix of different leakage events, the correlation between different privacy leakage events is analyzed to determine the common feature vector of privacy leakage risk; according to the common feature vector, a data set for privacy leakage risk classification is constructed to train the model leakage risk probability to obtain a leakage risk identification model. The present application provides an identification model that can quickly and accurately realize key privacy data. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a flowchart of a privacy data leakage risk control method provided by an embodiment of the present invention;

[0071] Figure 2 It is a structural schematic diagram of a privacy data leakage risk control device provided by an embodiment of the present invention;

[0072] Figure 3 It is a structural diagram of a terminal device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0073] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. 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.

[0074] In response to the above technical issues, this case provides a privacy data leakage risk control method, see Figure 1 , is a flow chart of a privacy data leakage risk control method provided by an embodiment of the present invention, the method comprising:

[0075] Step S1, based on the keywords of the privacy leakage risk theme, the leakage data within the privacy data life cycle is clustered and extracted to obtain the probability matrix of the theme distribution of different leakage events;

[0076] Step S2, by calculating the similarity of the probability matrices of different leakage events, analyzing the correlation between different privacy leakage events, and determining the common feature vector of privacy leakage risk;

[0077] Step S3: construct a data set for privacy leakage risk classification based on the common feature vectors to train the model leakage risk probability and obtain a leakage risk identification model.

[0078] In the specific implementation of this embodiment, based on the keywords of the privacy leakage risk theme, a preset model algorithm is used to perform topic clustering and extraction on the massive privacy leakage data within the life cycle of the location data to obtain a topic distribution probability matrix of different leakage events.

[0079] It should be noted that in this embodiment, leakage risk control is performed on privacy data. In actual application, it can be applied to the full-cycle risk control of location data. In this embodiment, location data is used as an example to illustrate the specific process of controlling privacy data. In other embodiments, it can be applied to the control of other privacy data.

[0080] By calculating the similarity of the distribution probability matrices of different leakage events, the correlation between privacy leakage events at different stages of the location data life cycle is determined, and the common feature vector of privacy leakage risk is obtained.

[0081] It should be noted that when performing similarity analysis, a cosine similarity analysis method or other similarity analysis methods may be used to determine the correlation between different events.

[0082] According to the common feature vector, a data set for privacy leakage risk classification is constructed to train the model leakage risk probability, thereby obtaining a leakage risk identification model.

[0083] According to the common feature vectors, the support vector machine algorithm is used to classify and label the privacy leakage data within the life cycle of location data. The labeling information of a part of the privacy leakage data is obtained through manual labeling or semi-supervised learning methods as the basis of the training sample set.

[0084] A statistical analysis is performed on the privacy leakage risk classification annotation set. According to the characteristics of the privacy leakage data, an appropriate prior distribution is selected, such as multinomial distribution, Dirichlet distribution, etc. The prior probability, conditional probability and other parameters are estimated through maximum likelihood estimation or Bayesian estimation methods. A probability model for privacy leakage risk classification is constructed to quantify the probability of occurrence and degree of harm of different types of privacy leakage events. According to the costs and benefits of different types of privacy leakage risks, a decision matrix is ​​constructed. The optimal decision threshold is obtained using criteria such as minimizing expected risk or maximizing expected utility. When the posterior probability of a privacy leakage event exceeds the preset threshold, it is judged as a high-risk event and the corresponding early warning and disposal mechanism is triggered.

[0085] It should be noted that in this case, when constructing a probability model for privacy leakage risk classification, it can be assumed that the privacy leakage risk category follows a multinomial distribution, the maximum likelihood estimation is used to calculate the prior probability, and the naive Bayesian assumption is used to calculate the conditional probability to obtain the joint distribution and the posterior distribution. Assume that there are three types of privacy leakage risks, namely low risk, medium risk, and high risk.

[0086] Through automated topic clustering, classification, and association rule mining, this case enables enterprises to efficiently identify and understand privacy leakage risk patterns in the life cycle of location data, reduce the workload of manual review, and improve the efficiency and response speed of security management.

[0087] In this case, leakage data throughout the entire life cycle is extracted and analyzed to form a long-term mechanism for classifying privacy leakage risks, thereby improving the accuracy and timeliness of privacy leakage risk assessments.

[0088] In another embodiment provided by the present invention, when step S1 is specifically performed, according to the keywords of the privacy leakage risk theme, the latent Dirichlet allocation topic model algorithm is used to perform topic clustering and extraction on the massive privacy leakage data within the life cycle of the location data, and the main types and characteristics of the privacy leakage risk are determined by analyzing the relevance and frequency of the topic keywords. Data mining technology is used to perform semantic analysis and association analysis on the topic keywords, and the implicit relationship and common characteristics between different leakage events are mined. Through the knowledge graph technology, the topic keywords are mapped and associated with the privacy leakage risk field knowledge to build a complete privacy leakage risk knowledge base.

[0089] Based on the privacy leakage risk knowledge base, we use machine learning algorithms such as support vector machines and random forests to establish a privacy leakage risk prediction model through feature engineering and model tuning. By analyzing the characteristic attributes and behavior patterns of location data, we predict potential privacy leakage risks.

[0090] By analyzing the data's timestamp, location tag and other metadata, the data is divided into stages such as data collection, transmission, storage, processing, and application. The privacy leakage risk prediction results are associated with the different stages of the location data life cycle, the privacy leakage risk level of each stage is determined, and targeted data security protection strategies and measures are formulated.

[0091] Adopt data desensitization and encryption technology, select appropriate desensitization strategies such as data encryption, data obfuscation, and differential privacy according to the sensitivity of the data and the risk level of privacy leakage, desensitize and encrypt the location data in the high-risk stage, reduce the risk of data leakage and illegal access, and protect the privacy and security of users. Through real-time monitoring and auditing of the access and use of location data, abnormal behavior and illegal operations can be discovered and blocked in a timely manner, and violations can be held accountable and punished, forming a long-term mechanism for privacy protection.

[0092] Specifically, when using the Latent Dirichlet Allocation (LDA) topic model algorithm to cluster and extract topics from location data, the number of topics can be set to 20 and the number of iterations to 1000. By calculating the probability distribution of keywords under each topic, keywords with a probability greater than 0.05 are used as representative words of the topic, and based on the semantic similarity of the representative words, the topics are merged and optimized using indicators such as cosine similarity. Ultimately, about 10 topic categories are obtained, and the spatiotemporal distribution characteristics of different topics are analyzed based on the timestamps and location tags of the data under each topic.

[0093] When constructing a knowledge graph of privacy leakage risk, you can use knowledge representation learning algorithms such as TransE to map topic keywords into a low-dimensional semantic space. By setting appropriate embedding dimensions and loss functions, you can optimize the vector representation of entities and relationships so that similar topic keywords are closer in the semantic space and dissimilar topic keywords are farther apart in the knowledge graph. When training a privacy leakage risk prediction model, you can choose the support vector machine (SVM) algorithm, use the Gaussian kernel function, set the penalty coefficient C to 1.0, and optimize the model's hyperparameters through grid search and cross-validation methods to improve the accuracy and generalization of predictions.

[0094] When desensitizing location data at high risk stages, differential privacy technology can be used, and the privacy budget ε can be set to 0.1. By adding Laplace noise to the original data, the data availability can be guaranteed while preventing the leakage of private information. When conducting real-time monitoring and auditing, rule-based and anomaly detection methods can be used to set appropriate thresholds and rules to analyze and judge data access and operation behaviors in real time. Once abnormal behaviors such as illegal access and data leakage are found, early warning and blocking mechanisms are immediately triggered, and violations are recorded, reported and punished according to preset strategies, forming a complete privacy protection closed loop.

[0095] In another embodiment provided by the present invention, when the above step S2 is specifically performed:

[0096] By calculating the cosine similarity of the probability matrix of the topic distribution of different leakage events, the correlation between privacy leakage events at different stages of the location data life cycle is determined, and the common feature vector of privacy leakage risk is obtained. Dimensionality reduction algorithms such as principal component analysis, linear discriminant analysis (LDA), and t-SNE are used to reduce the dimensionality of the common feature vectors. According to the distribution of the data, the appropriate algorithm is selected to extract the principal component that best represents the privacy leakage risk characteristics and form a privacy leakage risk feature subspace. According to the privacy leakage risk feature subspace, clustering algorithms such as k-means and DBSCAN are used to perform cluster analysis on privacy leakage events at different stages. By adjusting the key parameters of the algorithm, high-risk privacy leakage event clusters are identified, and the privacy leakage scenarios that need to be focused on are determined.

[0097] We conduct in-depth analysis of typical events in the high-risk privacy leakage event cluster, extract the key features and behavior patterns of the events, build a privacy leakage risk profile, and use the key features in the risk profile as the discrimination conditions for anomaly detection and embed them into the training data of the anomaly detection model. According to the privacy leakage risk profile, we use anomaly detection algorithms such as the statistical 3σ principle, the distance-based LOF algorithm, and the density-based LOCI algorithm to monitor the real-time data stream within the life cycle of the location data. By building the feature space and distance measurement function of anomaly detection, we match it with the risk profile in real time to timely discover and warn suspected privacy leakage events that exceed the risk threshold. If a suspected privacy leakage event is found, the event response mechanism is triggered, and measures such as data desensitization, access control, and data destruction are taken to block the further expansion of privacy leakage, and trace and investigate the event.

[0098] Through continuous monitoring and analysis of privacy leakage incidents, using incremental learning methods, regularly adding new privacy leakage event data to the training set, retraining clustering models and anomaly detection models, and dynamically updating privacy leakage risk feature subspaces and risk portraits based on the model's performance indicators, we can form real-time perception and early warning capabilities for privacy leakage risks, optimize data security protection strategies, and improve the level of privacy protection throughout the life cycle of location data.

[0099] Specifically, when calculating the cosine similarity of the probability matrix of the topic distribution of different leakage events, the probability matrix can be normalized first, the probability value of each topic is divided by the L2 norm of the matrix to obtain the unitized eigenvector, and then the dot product of the two eigenvectors is calculated to obtain the cosine similarity value. For example, for two leakage events A and B, their topic distribution probability matrices are [0.2, 0.3, 0.5] and [0.1, 0.4, 0.5] respectively, the normalized eigenvectors are [0.36, 0.55, 0.91] and [0.18, 0.73, 0.91], and the cosine similarity is 0.360.18+0.550.73+0.91*0.91=0.96, indicating that the topic distributions of the two events are very similar and are likely to belong to the same type of privacy leakage.

[0100] When using the PCA algorithm for dimensionality reduction, you can select principal components with a variance explanation rate greater than 0.95, that is, the sum of the variance contribution rates of the first k principal components exceeds 95%, as the basis of the privacy leakage risk feature subspace. For a 100-dimensional common feature vector, after PCA dimensionality reduction, it is found that the variance contribution rates of the first 5 principal components are 0.5, 0.3, 0.1, 0.05, and 0.03, respectively, and the cumulative variance contribution rate is 0.98. Then, you can select the first 5 principal components to form the privacy leakage risk feature subspace, and map the original data to this subspace for subsequent analysis.

[0101] It should be noted that when using the DBSCAN algorithm for clustering, the eps parameter can be set to 0.5 and the min_samples parameter can be set to 5, that is, the neighborhood with a radius of 0.5 contains at least 5 sample points as core points, and the core points that can be reached by density are connected to form clusters, and the unlabeled sample points are regarded as noise points or isolated points. By adjusting the eps and min_samples parameters, the granularity of clustering and the tolerance to noise can be controlled. Increasing eps can form larger clusters, but some outliers may be included; increasing min_samples can improve the resistance to noise, but some small clusters may be divided into finer parts.

[0102] When constructing a privacy leakage risk profile, we can extract key features of the event, such as the type, quantity, sensitivity of the leaked data, the time, location, cause of the leakage, the impact and loss caused by the leakage, etc., and quantify these features into a high-dimensional vector as a risk profile for this type of event.

[0103] When performing anomaly detection, the Mahalanobis distance between each data point in the real-time data stream and the risk profile can be calculated. When the distance exceeds a threshold, such as 10, it is determined to be a suspected leakage event. The threshold can be set and adjusted based on historical data to balance the sensitivity and false alarm rate of detection. For a user's location data stream, extract the time, location, trajectory and other features, calculate the Mahalanobis distance with different types of privacy leakage risk profiles, and if the distance with a certain profile exceeds 10, trigger an early warning, mark the user as a high-risk user, and take corresponding security measures.

[0104] When dynamically updating the feature subspace and portrait of privacy leakage risk, the data of new privacy leakage events that have occurred in the recent period can be added to the historical data set at regular intervals, such as one month, and feature selection, dimensionality reduction, clustering and other steps can be re-performed to generate updated feature subspace and portrait.

[0105] By setting the appropriate update frequency and data window, we can keep the model stable while adapting to new privacy leakage risk scenarios in a timely manner. In the past month, 10 new privacy leakage incidents have occurred. The feature vectors of these 10 incidents were added to the original event data set, and the PCA model was retrained to obtain a new 5-dimensional feature subspace. The corresponding clustering results and risk profiles were updated, which improved the accuracy and coverage of anomaly detection.

[0106] In another embodiment provided by the present invention, when the above step S3 is specifically performed:

[0107] According to the common feature vectors, the support vector machine algorithm is used to classify and label the privacy leakage data within the life cycle of location data. The labeling information of a part of the privacy leakage data is obtained through manual labeling or semi-supervised learning methods as the basis of the training sample set. For unlabeled data, the transductive learning method of the support vector machine is used to infer and classify the unlabeled samples using the structural information of the labeled samples, and to construct a training sample set for privacy leakage risk classification.

[0108] A k-fold cross-validation scheme is adopted to randomly divide the training set into k mutually exclusive subsets. One of the subsets is selected as the validation set each time, and the remaining k-1 subsets are used as the training set. The validation is repeated k times, and the average of the k results is taken as the performance estimate of the support vector machine model. The penalty coefficient C, kernel function type, kernel function parameters, etc. are traversed and optimized through grid search or random search methods, and the parameter combination with the highest F1 value or AUC value is selected as the optimal parameter to improve the generalization ability and accuracy of the classifier.

[0109] According to the classification results of the support vector machine model, multiple prediction results are integrated by majority voting or soft voting to obtain the final privacy leakage risk category labeling. By setting a consistency threshold, the labeling results with low confidence are filtered out to improve the accuracy and reliability of the labeling and form a privacy leakage risk classification labeling set.

[0110] A statistical analysis is performed on the privacy leakage risk classification annotation set. According to the characteristics of the privacy leakage data, an appropriate prior distribution is selected, such as multinomial distribution, Dirichlet distribution, etc. The prior probability, conditional probability and other parameters are estimated through maximum likelihood estimation or Bayesian estimation methods. A probability model for privacy leakage risk classification is constructed to quantify the probability of occurrence and degree of harm of different types of privacy leakage events. According to the costs and benefits of different types of privacy leakage risks, a decision matrix is ​​constructed. The optimal decision threshold is obtained using criteria such as minimizing expected risk or maximizing expected utility. When the posterior probability of a privacy leakage event exceeds the preset threshold, it is judged as a high-risk event and the corresponding early warning and disposal mechanism is triggered.

[0111] Through continuous optimization and updating of the privacy leakage risk classification system, when the new privacy leakage event data reaches a certain scale, or when the performance indicators of the support vector machine model show a significant decline, the model retraining and parameter tuning are triggered, and the model is quickly updated and optimized using incremental learning or online learning methods. At the same time, according to the latest privacy leakage data distribution, the decision threshold and rule parameters are adjusted to form a long-term mechanism for privacy leakage risk classification and improve the accuracy and timeliness of privacy leakage risk determination. The privacy leakage risk classification system is associated with each stage of the location data life cycle, and differentiated prevention and control strategies and measures for different stages and types of privacy leakage risks are formulated. Starting from data collection, transmission, storage, processing, application and other links, the privacy protection level of location data is comprehensively improved.

[0112] Specifically, when manually labeling privacy-leaking data, an active learning method can be used to select samples with the largest amount of information from unlabeled data, such as outliers or samples close to the classification boundary, and present them for labeling, so as to obtain the maximum classification performance improvement with the minimum labeling cost. For example, using the minimum confidence strategy, the results predicted by the support vector machine model are sorted from low to high by confidence, and the top 10% samples with the lowest confidence are selected for manual labeling and added to the training set for iterative training. When performing transductive learning, the label propagation algorithm can be used to propagate the label information of the labeled samples to the unlabeled samples based on the similarity between the samples, and a fully connected graph is constructed. The weight of the edge represents the similarity between the samples. By minimizing the energy function, the labels of similar samples are made as consistent as possible, and the labels of dissimilar samples are made as different as possible.

[0113] Use the Gaussian kernel function to calculate the similarity matrix between samples, set the regularization parameter of the energy function to 0.1, and iterate the propagation of label information until the label distribution converges. When performing k-fold cross-validation, you can choose k=10, that is, divide the training set into 10 parts, select 1 part as the validation set each time, and 9 parts as the training set, repeat 10 times, and take the average performance as the performance evaluation of the support vector machine model.

[0114] For the penalty coefficient C, you can search on the logarithmic grid of [0.1, 1, 10, 100]; for the kernel function type, you can try the linear kernel, polynomial kernel, and radial basis kernel; for the kernel function parameters, such as the degree d of the polynomial kernel and the scale coefficient γ of the radial basis kernel, you can search on the logarithmic grid of [0.01, 0.1, 1, 10]. When labeling the privacy leakage risk category, you can use the majority voting method to vote on the 10 cross-validation prediction results of the support vector machine model to obtain the final category label, and set the consistency threshold to 0.7, that is, at least 7 prediction results are consistent before accepting the label, otherwise it is marked as rejected.

[0115] When constructing a probability model for privacy leakage risk classification, we can assume that the privacy leakage risk category follows a multinomial distribution, use maximum likelihood estimation to calculate the prior probability, and use the naive Bayesian assumption to calculate the conditional probability to obtain the joint distribution and posterior distribution. Assuming there are three types of privacy leakage risks, namely low risk, medium risk, and high risk, the prior probabilities in the training set are 0.2, 0.5, and 0.3, respectively. For a certain feature vector x, the conditional probabilities in the three categories are 0.1, 0.4, and 0.5, respectively. According to the Bayesian formula, the posterior probabilities can be calculated to be 0.1, 0.625, and 0.5, respectively, and the privacy leakage event is judged to be medium risk.

[0116] When optimizing the model, you can set the update to be triggered when the amount of new data reaches 20% of the original training set or the F1 value of the model drops by 5%. Use the gradient descent method to perform incremental training on the support vector machine model on the new data, set the learning rate to 0.01, and iterate 100 times. At the same time, adjust the threshold of the Bayesian decision based on the new prior distribution and cost matrix, such as increasing the threshold of high-risk events from 0.6 to 0.8 to control the false alarm rate.

[0117] In another embodiment provided by the present invention, the method further comprises the following steps:

[0118] The preset association rule mining algorithm is used to analyze the association rules between different categories of privacy leakage risks in the life cycle of location data, and to determine the propagation path and impact scope of privacy leakage risks in the life cycle of location data.

[0119] Specifically, the Apriori association rule mining algorithm is used to select appropriate support and confidence thresholds through cross-validation and other methods to determine the propagation path and impact scope of privacy leakage risks during the life cycle of location data.

[0120] Targeting key privacy leakage risk nodes and propagation paths, we combine various privacy protection technologies such as ABAC, MPC, DP, and FL to develop differentiated privacy protection strategies, and deploy customized privacy protection mechanisms in key links of the location data lifecycle, such as data collection, storage, transmission, processing, and application, to minimize privacy leakage risks while meeting business needs. By updating the privacy leakage risk association rule set and propagation model in real time, we continuously evaluate the correlation between emerging privacy leakage risks and existing risk clusters, dynamically adjust the priority and strength of privacy protection strategies, form a closed loop of privacy leakage risk analysis and prevention, and continuously improve the intelligence level of location privacy protection.

[0121] In another embodiment provided by the present invention, when determining the propagation path and the influence range, the following is specifically performed:

[0122] The Apriori association rule mining algorithm is adopted to select appropriate support and confidence thresholds through cross-validation and other methods to strike a balance between coverage and reliability. Multiple evaluation indicators such as rule support, confidence, improvement, novelty and simplicity are comprehensively considered to analyze the association rules between different categories of privacy leakage risks in the life cycle of location data, screen out frequent item sets and strong association rules, and form a set of privacy leakage risk association rules.

[0123] According to the privacy leakage risk association rule set, a privacy leakage risk association graph is constructed, and different categories of privacy leakage risk nodes are connected according to the association rules to form a directed acyclic graph, which comprehensively reflects the association strength and importance between privacy leakage risks. Based on the privacy leakage risk association graph, PageRank, HITS, Katzcentrality and other graph mining algorithms are used to calculate the importance, authority and centrality of each privacy leakage risk node according to the characteristics of the privacy leakage risk association graph, identify key privacy leakage risk nodes and propagation paths, and determine the impact range of privacy leakage risks in the life cycle of location data.

[0124] By constructing a time series model for the propagation of privacy leakage risk, such as ARMA and HMM, the privacy leakage risk propagation path is analyzed in time series, and combined with the attributes of risk nodes, the strength of association rules and other factors, a multivariate time series classification or regression model is trained to predict the probability of occurrence, impact range and loss value of privacy leakage risk. According to the impact range and propagation path of privacy leakage risk, community discovery algorithms such as LPA, Modularity and Walktrap are used to divide the privacy leakage risk association graph into several overlapping risk communities, identify closely related privacy leakage risk clusters, and update the community division results in real time as the association graph changes dynamically. For key privacy leakage risk nodes and propagation paths, a variety of privacy protection technologies such as ABAC, MPC, DP and FL are combined to formulate differentiated privacy protection strategies, and customized privacy protection mechanisms are deployed in key links of the location data life cycle, such as data collection, storage, transmission, processing and application, to minimize privacy leakage risks while meeting business needs. By updating the privacy leakage risk association rule set and propagation model in real time, continuously evaluating the correlation between emerging privacy leakage risks and existing risk clusters, dynamically adjusting the priority and intensity of privacy protection strategies, forming a closed loop of privacy leakage risk analysis and prevention and control, and continuously improving the intelligence level of location privacy protection.

[0125] Specifically, when using the Apriori algorithm to mine association rules for privacy leakage risks, you can first set the support threshold to 0.05 and the confidence threshold to 0.8, and iteratively generate frequent item sets and association rules. Then, use 10-fold cross validation to randomly divide the data set into 10 subsets, select one subset as the test set each time, and the remaining 9 subsets as training sets to evaluate the coverage and reliability of the association rules.

[0126] For each association rule, its support, confidence and lift are calculated. Support reflects the universality of the rule, confidence reflects the credibility of the rule, and lift reflects the effectiveness of the rule. For example, for the rule "GPS tracking in the data collection phase → unauthorized access in the data storage phase", if its support is 0.1, confidence is 0.9, and lift is 3.0, it means that the rule appears in 10% of privacy leakage events. In the case of GPS tracking, there is a 90% probability of unauthorized access, and the risk of unauthorized access is 3 times that of random access.

[0127] By balancing coverage and reliability, a combination of a support threshold of 0.08 and a confidence threshold of 0.85 is selected as the final parameter setting. When constructing a time series model for the propagation of privacy leakage risks, the ARMA (p, q) model can be used, where p is the order of the autoregressive term and q is the order of the moving average term.

[0128] By analyzing the autocorrelation coefficient and partial autocorrelation coefficient of the time series of privacy leakage risk, we determine the values ​​of p and q, p = 2, q = 1, which means that the privacy leakage risk at the current moment is related to the risks of the previous two moments and the random disturbance of the previous moment. Then, we use the maximum likelihood estimation method to estimate the parameters of the model, and use the fitted model to predict the privacy leakage risk in the future.

[0129] It is predicted that the probability of a privacy leak incident occurring every day in the next 7 days is 0.1, 0.15, 0.2, 0.18, 0.22, 0.25, and 0.28 respectively, and corresponding prevention and control measures can be taken in time. When discovering communities, the Walktrap algorithm can be used to calculate the distance between nodes by random walks, and the hierarchical clustering method can be used to merge the nodes with the closest distance from bottom to top until the modularity maximization criterion is met. The privacy leak risk association graph is divided into 5 communities. The privacy leak risks within each community are closely connected, while the connection between communities is relatively weak. By calculating the average loss of privacy leaks within each community, the risk community to be focused on is determined. When deploying a privacy protection mechanism, differential privacy technology can be used to add Laplace noise to the original data to ensure that the probability ratio of obtaining the same output results on any two adjacent data sets does not exceed exp(ε), where ε is the privacy budget, which controls the intensity of the noise. In the aggregation analysis of location data, ε=1.0 is set, and Laplace noise with a mean of 0 and a standard deviation of 1 / ε is added to the original location coordinates. The obtained aggregation result can ensure that within a radius of 100 meters, the probability of inferring whether a user is in the area does not exceed exp(1)≈2.72 times. By dynamically adjusting the privacy budget, the impact on data availability can be minimized while ensuring privacy security.

[0130] In another embodiment of the present invention, the method further comprises the following steps:

[0131] Through the propagation path and impact scope of privacy leakage risks in the life cycle of location data, a propagation network topology map of privacy leakage risks is constructed, and the PageRank algorithm is used to identify the key nodes and vulnerable links of privacy leakage risks in the life cycle of location data.

[0132] It should be noted that, in this embodiment, the PageRank algorithm is used to identify key nodes and vulnerable links. In other embodiments, other algorithms may be used to identify key nodes and vulnerable links.

[0133] Through methods such as incremental learning and online learning, the privacy leakage risk propagation network topology map and key node identification results are continuously updated to adapt to the dynamic changes of privacy leakage risks, and interpret and analyze key nodes and vulnerable links based on business scenarios, monitor privacy leakage risk events in real time, continuously optimize network models and algorithm parameters, and build an adaptive and explainable privacy leakage risk analysis and prevention and control system to guide the optimization and adjustment of privacy protection strategies.

[0134] Constructing a risk propagation network topology map and identifying key nodes can help accurately locate weak links in privacy protection, take preemptive measures, effectively prevent the occurrence and spread of risks, and reduce the impact and losses of actual leakage incidents.

[0135] In another embodiment provided by the present invention, when identifying key nodes and vulnerable links, the following steps are specifically performed:

[0136] Through the propagation path and impact scope of privacy leakage risk in the life cycle of location data, a propagation network topology of privacy leakage risk is constructed. Using graph database technology, privacy leakage risk events are used as nodes, and the associations between different events are used as edges to form a directed weighted graph. Based on the privacy leakage risk propagation network topology, the label propagation algorithm (LPA) is used for community discovery. By iteratively updating the labels of nodes, the labels of adjacent nodes are made as consistent as possible until the label distribution converges. The topology is divided into several closely related privacy leakage risk communities, and the differences and similarities of different communities in propagation paths and impact scopes are explored.

[0137] The LPA algorithm is simple to calculate and easy to implement, but the algorithm is highly random and sensitive to initial labels. According to the characteristics of the privacy leakage risk community, we use indicators such as degree, betweenness centrality, and eigenvector centrality to calculate the importance of each privacy leakage risk node within and between communities. These indicators are normalized and used as the initial values ​​of the PageRank algorithm to identify key privacy leakage risk nodes.

[0138] The PageRank algorithm simulates the behavior of web browsers through random walks, calculates the probability of each node being visited, and then measures the importance of the node. Let PR(i) represent the importance score of node i, d is the damping factor, N is the total number of nodes, M(i) is the set of nodes pointed to by the outgoing edges of node i, and L(j) is the number of incoming edges of node j. The iterative formula of PageRank is: PR(i) = (1-d) / N+d*∑{j∈M(i)}PR(j) / L(j), where d is generally taken as 0.85. Iterate until the PR value converges or reaches the maximum number of iterations, and obtain the importance ranking of nodes in the entire privacy leakage risk propagation network, and determine the key nodes of privacy leakage risk in the life cycle of location data.

[0139] Structural hole analysis is performed on the privacy leakage risk propagation network, and indicators such as the network constraint coefficient are used to measure the redundancy of node connections in the entire network and the characteristics of structural holes. The network constraint coefficient C is defined as C = ∑{i, j}(p_ij+∑{q}p_iq·p_qj)^2, i≠j, where p_ij represents the proportion of node i investing in node j, and q represents all intermediary nodes between nodes i and j. The larger the network constraint coefficient, the closer the connection between nodes, the higher the redundancy, and the fewer structural holes.

[0140] The calculation of network constraint coefficients needs to consider the direct connection strength and indirect connection path between nodes. The Floyd shortest path algorithm is used to solve the shortest path between any two nodes, and the frequency of occurrence of intermediary nodes is counted to identify structural holes in the network and determine the bottlenecks and vulnerable links of privacy leakage risk propagation. According to the key nodes and vulnerable links of privacy leakage risk, the DeepWalk algorithm is used for random walks. By performing a depth-first search on the graph, a node sequence is generated as the trajectory of the random walk, and then the Skip-gram model is used to learn the embedded representation of the nodes.

[0141] The transition probability of DeepWalk is f(v|w)=g(w|v)=1 / |N(v)|, if (v, w) are connected by an edge; otherwise it is 0, where |N(v)| represents the number of neighbors of node v. After obtaining the embedded representation of the node, machine learning algorithms such as logistic regression and support vector machine are used to train a classification model with node embedding as features and whether the node is infected as a label to predict the propagation probability and impact range of privacy leakage risks in the network. Monte Carlo simulation methods are used to conduct a large number of random experiments on the risk propagation network, count the frequency and impact range of different nodes being infected, and evaluate the effectiveness of different prevention and control strategies (such as cutting off certain key edges, isolating certain key nodes, etc.).

[0142] The results of privacy leakage risk prediction and prevention and control assessment are used as the basis and trigger conditions for dynamically updating the network topology and key nodes. When the infection probability of a node is predicted to exceed the threshold, the PageRank value of the node is recalculated to determine whether it becomes a new key node; when the influence range of a community is predicted to expand significantly, the community is re-divided to determine whether the community structure has changed; when it is evaluated that the effect of a certain prevention and control strategy is not ideal, the weights of the corresponding nodes and edges are adjusted to generate a new network topology.

[0143] Through methods such as incremental learning and online learning, the privacy leakage risk propagation network topology map and key node identification results are continuously updated to adapt to the dynamic changes of privacy leakage risks, and interpret and analyze key nodes and vulnerable links based on business scenarios, monitor privacy leakage risk events in real time, continuously optimize network models and algorithm parameters, and build an adaptive and explainable privacy leakage risk analysis and prevention and control system to guide the optimization and adjustment of privacy protection strategies.

[0144] Specifically, when constructing the privacy leakage risk propagation network topology diagram, the Neo4j graph database can be used, and the Cypher query language can be used to abstract privacy leakage risk events into nodes. The attributes of the nodes include event number, occurrence time, leakage data type, number of affected users, etc. The association between events is abstracted into edges, and the attributes of the edges include association type, association strength, propagation delay, etc.

[0145] For example, when applying the LPA algorithm for community discovery, each node is initially assigned a unique label. During the iteration process, each node selects the label with the highest frequency among its neighbors as its new label until the labels of all nodes no longer change. In the first round of iteration, 3 of node A's neighbors have labels of 1 and 2 have labels of 2, so A's new label is 1; in the second round of iteration, 4 of node A's neighbors have labels of 1 and 1 has a label of 2, so A's new label is still 1; after multiple rounds of iterations, the label of the community where node A is located converges to 1.

[0146] When calculating the PageRank value, each node is initially assigned an equal PR value of 1 / N. In each round of iteration, each node evenly distributes its current PR value to the nodes pointed to by the outgoing edge, collects PR values ​​from the incoming edge nodes, and adds a random jump factor (1-d) / N until the PR value converges. Assume that the damping factor d = 0.85, there are 4 nodes in total, the initial PR value is 0.25, and after 5 rounds of iterations, the PR value of A is 0.38, the PR value of B is 0.28, the PR value of C is 0.21, and the PR value of D is 0.13. The importance of ABCD is ranked as A>B>C>D.

[0147] When calculating the network constraint coefficient, the Floyd algorithm is first used to calculate the shortest path between any two nodes, and then the direct connection strength p between each node pair is counted. ij and indirect connection strength∑ q p iq ·p qj , finally add the two together and square them, and then sum them to get the constraint coefficient C of the entire network. In a network with 10 nodes, each node has an average of 3 direct edges and 2 indirect edges. The weight of the direct edge is 0.6, and the weight of the indirect edge is 0.3. Then the network constraint coefficient C = 10*[(0.6+20.30.3)^2+(0.6+20.30.3)^2] = 16.2, indicating that the connection between the nodes in the network is relatively close and lacks structural holes.

[0148] When using the DeepWalk algorithm to learn node embedding, the network is first randomly sampled. Each node is transferred from the current node to the neighboring node with equal probability, generating a random walk sequence of length 10. Each node is sampled 100 times to obtain 1000 node sequences. Then, the Skip-gram model is used, with the target node as the center, the window size is 5, the number of negative sampling is 5, the embedding dimension is 128, and training is performed for 50 epochs to obtain the embedding vector of each node.

[0149] For the node sequence A->B->C->D->E, the current central node is C, the context nodes are B and D, and the negative sampling nodes are F, G, and H. One-hot encoding is input, and the embedding vector of node C is output through a feedforward neural network. When predicting the probability of privacy leakage risk propagation, the node embedding vector is used as a feature, and whether the node has privacy leakage is used as a label to build a binary classification model, such as the logistic regression model h(x) = sigmoid(wTx+b), where x is the node embedding, w and b are the parameters to be learned, and the cross entropy loss function is used to train the model using the gradient descent method to obtain the optimal parameters w and b. For a new node x, h(x) is calculated to obtain its infection probability.

[0150] When evaluating the effectiveness of prevention and control strategies, Monte Carlo simulation is used to randomly select a certain proportion (such as 1%) of nodes on the privacy leakage risk propagation network as the initial infection nodes, set the infection rate (such as 5%) and recovery rate (such as 2%), iterate the propagation process until the infection scale reaches a stable state, repeat the experiment 1000 times, and calculate the average infection scale. Then, different prevention and control strategies are implemented, such as cutting off key edges, isolating key nodes, etc., repeat the above experiments, compare the average infection scale under different strategies, and evaluate the prevention and control effect. When dynamically updating the network topology, set a time window (such as 30 days), rebuild the network topology every other window, add new privacy leakage events as new nodes, add new associations as new edges, and delete expired nodes and edges to form a dynamic network.

[0151] According to the new network topology, the structural characteristics of the network, such as average degree, diameter, clustering coefficient, etc., are recalculated to judge the changing trend of the network structure. If certain characteristic indicators change significantly (such as exceeding the threshold), the community discovery, node importance assessment, risk prediction and other algorithms are triggered to re-run, and the corresponding results are updated to form a real-time privacy leakage risk analysis and prevention and control mechanism.

[0152] In another embodiment provided by the present invention, after determining the key nodes and the vulnerable links, the method further includes:

[0153] According to the key nodes and vulnerable links of privacy leakage risks in the life cycle of location data, the information entropy risk assessment model is adopted to calculate the privacy leakage risk values ​​of different nodes and links in the life cycle of location data, and form a quantitative assessment result of privacy leakage risk.

[0154] According to the quantile of the privacy leakage risk index, the quantile method is used to formulate a quantitative risk classification standard, and the 25% quantile, 50% quantile and 75% quantile are taken as the thresholds for low, medium and high risk levels, respectively, to form a privacy leakage risk level map. Combined with the privacy leakage risk level map, for each privacy leakage risk prevention and control measure, its input cost and expected benefit are calculated to obtain the cost-benefit ratio. According to the cost-benefit ratio, it is sorted from high to low, and the prevention and control measures with low input cost and significant risk reduction effect are given priority until the cumulative cost reaches the budget ceiling. A privacy leakage risk prevention and control roadmap in the short-term, medium-term and long-term time dimensions is formulated, the implementation effect of the prevention and control measures is continuously tracked, the priority and resource allocation are dynamically adjusted, the rational allocation of organizational resources and the effective implementation of prevention and control measures are guided, a closed-loop management of privacy leakage risk prevention and control work is formed, and the level of privacy security management of location data throughout its life cycle is continuously improved.

[0155] Through quantitative risk assessment and priority sorting, enterprises can reasonably allocate security resources according to the severity and urgency of risks, avoid blind investment, ensure that resources are used effectively, and improve the input-output ratio.

[0156] In another embodiment provided by the present invention, when evaluating the privacy leakage risk values ​​of different nodes and links, the following steps are specifically performed:

[0157] According to the key nodes and vulnerable links of privacy leakage risk in the life cycle of location data, the information entropy risk assessment model is used to calculate the privacy leakage risk values ​​of different nodes and links in the life cycle of location data, and form a quantitative assessment result of privacy leakage risk. The life cycle of location data is divided into multiple stages, such as data collection, storage, transmission, processing, application, etc. The fault tree analysis method (FTA) is used to gradually decompose the top-level event of privacy leakage into multiple intermediate events and basic events through top-down logical deduction, and the causal relationship between events is represented by logic gates to form an inverted tree structure diagram, and the basic events with higher importance are screened out as the object set of privacy leakage risk assessment.

[0158] For each risk assessment object, the analytic hierarchy process (AHP) is used to design a nine-point scale from the two dimensions of the probability of occurrence and the degree of impact of privacy leakage incidents. Each pair of factors at the same level is compared and scored in importance to obtain a pairwise comparison matrix, and a consistency test is performed. The weight vector of each criterion is calculated, and the privacy leakage risk score of each assessment object at the target layer is obtained by comprehensive weighting. The arithmetic mean of the scoring results is taken to obtain a qualitative assessment value of the privacy leakage risk.

[0159] According to the qualitative evaluation value, the fuzzy comprehensive evaluation method is used to quantify the probability of privacy leakage and the severity of consequences as the expected value of the membership function, and substitute it into the calculation formula of the information entropy model H i =-∑ j (P ij logP ij +C ij logC ij ), obtain the privacy leakage risk information entropy of each risk assessment object, and use linear normalization or logarithmic normalization to convert the risk information entropy into a privacy leakage risk index in the interval [0, 1]. The higher the risk index, the more likely the node or link is a key node or vulnerable link for privacy leakage and needs to be prevented.

[0160] Specifically, when applying the fault tree analysis method to identify the objects of privacy leakage risk assessment, "location data leakage" can be used as the top-level event, "excessive data collection", "authentication failure", "lack of access control", "weak encryption algorithm", "transmission link monitoring" and so on can be used as intermediate events, and "plaintext storage of location data", "third-party SDK abuse", "improper GPS accuracy setting", "HTTPS certificate error" and so on can be used as basic events. "AND gate" is used to indicate that the simultaneous occurrence of intermediate events leads to the top-level event, and "OR gate" is used to indicate that any occurrence of intermediate events leads to the top-level event. The structural importance of each event is calculated through Boolean algebra, and they are sorted from high to low according to the Fussell-Vesely importance, and the top 80% of the events with the cumulative contribution rate are taken as the evaluation objects.

[0161] When applying the analytic hierarchy process for scoring, "privacy leakage risk assessment" can be used as the target layer, "threat occurrence probability" and "vulnerability impact degree" as the criterion layer, and each assessment object as the solution layer. For the criterion layer, a 1-9 scale is used for scoring, 9 points represent equal importance, 1 point represents absolutely unimportant, and 5 3×3 pairwise comparison matrices are formed. For the solution layer, a 1-5 scale is used for scoring, 1 point represents very low risk, 5 points represent very high risk, and 5 10×2 pairwise comparison matrices are formed. YAAHP software is used for consistency test and weight calculation. If CR<0.1, the test is passed, and the arithmetic mean of the score is taken as the qualitative evaluation result.

[0162] When quantifying probability and severity, a membership function can be defined. For example, the "threat occurrence probability" can be divided into five levels: "very low", "low", "medium", "high", and "very high", and the values ​​are 0.1, 0.3, 0.5, 0.7, and 0.9 respectively; the "vulnerability impact degree" can be divided into five levels: "mild", "small", "medium", "large", and "severe", and the values ​​are 2, 4, 6, 8, and 10 respectively. The expected values ​​of Pij and Cij are the sum of the products of each membership and the assigned values.

[0163] When calculating the privacy leakage risk information entropy, taking the privacy leakage risk assessment of a location service APP as an example, the probability of the APP normally accessing the user's location is "high" through the hierarchical analysis method, and the value P is assigned 1j =0.7, the severity of the consequence is "medium", assigned a value of C 1j =6; the probability of abnormal frequent acquisition location is "high", assigned value P2j = 0.5, the severity of the consequence is "large", assigned value C 2j =8; ij and C ijSubstituting into the information entropy calculation formula, we have H1 = -[0.7log(0.7)+6log(6)]≈5.59, H2 = -[0.5log(0.5)+8log(8)]≈7.39.

[0164] Normalize the risk information entropy of the 10 assessment objects and take the maximum value H max =8, minimum value H min =2, using the linear normalization formula H i '=(H i -H min ) / (H max -H min ), the risk index of APP's normal access to user location is H1'=(5.59-2) / (8-2)=0.60, and the risk index of APP's abnormal and frequent collection of location is H2'=(7.39-2) / (8-2)=0.90, indicating that the latter has a higher risk of privacy leakage.

[0165] When formulating the risk grading standard, the 25% quantile Q1=0.25, 50% quantile Q2=0.50, and 75% quantile Q3=0.75 of the risk index of the 10 assessment objects are taken, then [0, 0.25] is low risk, [0.25, 0.50] is medium risk, [0.50, 0.75] is high risk, and [0.75, 1] ​​is extremely high risk. Based on this, a four-color risk level map is drawn to intuitively display the level of privacy leakage risks of different assessment objects.

[0166] When formulating a risk prevention and control roadmap, we sorted out common privacy leakage risk prevention and control measures, such as "data desensitization", "minimization of permissions", "multi-factor authentication", "homomorphic encryption", etc., estimated the investment cost (such as 1,000 yuan) and risk reduction benefit (such as 0.2) of each measure, calculated the cost-benefit ratio (1,000 / 0.2=5,000), and sorted them from low to high according to the cost-benefit ratio. If the annual prevention and control budget is 50,000 yuan, then "homomorphic encryption" (2,000 / 0.5=4,000) and "minimization of permissions" (3,000 / 0.4=7,500) are short-term priority prevention and control measures, "multi-factor authentication" (5,000 / 0.3=16,667) is a medium-term prevention and control measure, and "data desensitization" (10,000 / 0.2=50,000) is a long-term prevention and control measure.

[0167] An evaluation of the effectiveness of prevention and control is conducted every quarter. For example, "homomorphic encryption" reduces the risk index of the location data transmission link from 0.8 to 0.5. Based on the latest privacy leakage risk situation, the priority of various prevention and control measures is dynamically updated to form a PDCA continuous improvement closed loop.

[0168] In another embodiment provided by the present invention, the method further comprises:

[0169] Based on the quantitative assessment results of privacy leakage risks, the fuzzy comprehensive evaluation method is used to determine the priority of privacy leakage risks during the life cycle of location data, and formulate targeted privacy protection strategies and measures.

[0170] Build a full-life cycle, multi-level, customizable location privacy protection solution, and optimize and adjust risk priorities and protection strategies in real time according to the dynamic changes in privacy leakage risks to form a continuous improvement mechanism.

[0171] In another embodiment provided by the present invention, the privacy protection policy matching process specifically includes:

[0172] Based on the quantitative assessment results of privacy leakage risks, the fuzzy comprehensive evaluation method is used to determine the priority of privacy leakage risks during the life cycle of location data, and formulate targeted privacy protection strategies and measures.

[0173] The results of the quantitative assessment of privacy leakage risk are sorted from high to low according to the risk value to form a privacy leakage risk sorting vector. Each risk object in the sorting vector is grouped according to the life cycle stages such as data collection, transmission, storage, processing, and application, and matched one by one with evaluation indicators such as risk probability, impact scope, and degree of harm to form a risk-indicator matching matrix as the input data for fuzzy comprehensive evaluation.

[0174] A set of fuzzy evaluation indicators for privacy leakage risk is constructed, and the analytic hierarchy process is used to determine the weight vector of each indicator. By constructing a paired comparison judgment matrix, the maximum eigenvalue of the matrix and the corresponding eigenvector are calculated and a consistency test is performed, and the weight is modified until it passes the test.

[0175] According to the characteristics of the evaluation indicators, select appropriate membership functions, such as trapezoidal functions, Gaussian functions, etc., determine the parameter values ​​of the membership function, substitute the quantitative assessment value of each risk object into the membership function, calculate its membership on each evaluation indicator, form a fuzzy relationship matrix, and perform consistency check and correction. Apply the principle of fuzzy matrix synthesis operation, use the main factor prominent type or weighted average type operator, perform synthesis operation on the weight vector and the fuzzy relationship matrix, and obtain the fuzzy comprehensive evaluation result vector of privacy leakage risk. The evaluation result vector is normalized by the mean method. According to the normalized privacy leakage risk evaluation index, a reasonable threshold interval is set, and the risk is divided into three levels: high, medium, and low, and the corresponding privacy protection strategy is formulated.

[0176] For high-risk links, mandatory measures such as data desensitization, access control, and encrypted communication are taken; for medium-risk links, preventive measures such as anomaly monitoring, privacy auditing, and data traceability are taken; for low-risk links, general measures such as data classification, de-identification, and privacy training are taken to build a full-life cycle, multi-level, customizable location privacy protection solution, and according to the dynamic changes in privacy leakage risks, real-time optimization and adjustment of risk priorities and protection strategies are carried out to form a continuous improvement mechanism.

[0177] Specifically, after quantitatively evaluating the privacy leakage risk, a risk value list can be obtained, such as [0.9, 0.8, 0.7, 0.6, 0.5], indicating that there are 5 privacy leakage risk points, and the risk values ​​are arranged from high to low. These risk points are divided into three stages according to the location data life cycle: the data collection stage contains the 1st and 2nd risk points, the data storage and transmission stage contains the 3rd and 4th risk points, and the data application stage contains the 5th risk point, forming a 3×3 risk-indicator matching matrix.

[0178] The analytic hierarchy process was used to compare the three indicators of risk probability, impact range and degree of harm in pairs, assigning 1-9 points to form five 3×3 pairwise comparison matrices, calculate the maximum eigenvalue λ and eigenvector W of the matrix, and perform consistency test until CR=0.0084<0.1, and obtain the weight vector W=[0.62,0.24,0.14]. The Gaussian membership function exp(-((xa) / b)^2) was selected, and the mean a and standard deviation b were determined based on historical data. For the probability of risk, a=0.1, b=0.05; for the impact range, a=100, b=50; for the degree of harm, a=5, b=1.

[0179] Substitute the risk value list into the membership function and calculate the membership matrix R. For example, the membership vector of the first risk point is [0.92, 0.85, 0.78], the second is [0.88, 0.82, 0.74], and so on, forming a 5×3 fuzzy relationship matrix R. Use the weighted average synthesis operator M(·,⊕) to calculate the evaluation result vector B=W·R=[0.58, 0.53, 0.48, 0.44, 0.40], and then use the mean method to normalize B'=B / n=[0.097, 0.088, 0.080, 0.073, 0.067], where n=3. The threshold interval is set, and the normalized index greater than 0.09 is high risk, between 0.07 and 0.09 is medium risk, and less than 0.07 is low risk. The first risk point is high risk, the 2nd to 4th are medium risk, and the 5th is low risk.

[0180] For high-risk points in the data collection stage, the SHA-256 encryption algorithm is used to hash the original location data, converting the 128-bit longitude and latitude coordinates into 64-bit hash values ​​for transmission and storage. For medium-risk points in the storage and transmission stage, the k-anonymization algorithm is used, setting k=10, clustering 10 adjacent location points into one area, and retaining only the center coordinates of the area to replace the original precise location.

[0181] For low-risk points in the data application stage, the principle of least privilege is adopted. By default, applications are only allowed to access coarse-grained locations at the city level. When access to precise locations is required, user consent must be obtained. For all stages, programmers' privacy awareness training is continuously carried out, detailed data security management systems are compiled, and the entire process of data use is recorded to ensure auditability and traceability. At the same time, blockchain technology is used to build a unified location data sharing platform, and smart contracts are used to achieve fine-grained data permissions and access control, which not only promotes the open sharing of location data, but also protects the privacy and security of users.

[0182] In another embodiment provided by the present invention, the method further comprises:

[0183] By continuously monitoring the privacy leakage data within the life cycle of location data, dynamically updating the topic distribution, association rules and propagation network of privacy leakage risks, real-time early warning and dynamic control of privacy leakage risks within the life cycle of location data can be achieved.

[0184] In the privacy leakage risk assessment model, internal factors such as dynamically updated topic distribution, association rules, and propagation paths, as well as external factors such as data sensitivity, user privacy preferences, and attacker capabilities are comprehensively considered to improve the comprehensiveness and accuracy of risk assessment, and a dynamic threshold mechanism for risk assessment is established to dynamically adjust the risk threshold based on historical data. When the risk value exceeds the threshold, an early warning is automatically triggered, and corresponding dynamic prevention and control measures such as data desensitization, access restrictions, and isolation and blocking are taken according to the risk level. A visual monitoring platform for privacy leakage risks is established, and the dynamic changes in privacy leakage risks are intuitively displayed using real-time dashboards, maps, charts, etc., and interactive exploration and analysis tools are provided to allow users to customize monitoring views, drill-down queries, causal tracing, etc., to form an explainable privacy leakage risk analysis report; at the same time, a user feedback channel is set up to collect users' subjective evaluation of privacy leakage risks and improvement suggestions for countermeasures, which are used to correct model parameters and optimize prevention and control strategies, forming a dynamic prevention and control closed loop of privacy leakage risks in human-computer interaction, and continuously optimizing the privacy security of location data throughout its life cycle.

[0185] Through continuous monitoring and dynamic control, we ensure that enterprises can respond to regulatory requirements in a timely manner, effectively comply with data protection regulations, reduce legal risks and fines, and maintain corporate reputation and customer trust.

[0186] In another embodiment provided by the present invention, when dynamically updating the topic distribution, association rules and propagation network of privacy leakage risk, the following steps are specifically performed:

[0187] By continuously monitoring the privacy leakage data in the life cycle of location data, dynamically updating the topic distribution, association rules and propagation network of privacy leakage risks, real-time warning and dynamic control of privacy leakage risks in the life cycle of location data can be achieved. Real-time computing frameworks such as Apache Flink and Spark Streaming are used to process and analyze the data stream in the life cycle of location data in real time, extract feature indicators related to privacy leakage, and match and compare them with pre-trained privacy leakage risk topic models, association rule models, and propagation network models in real time to determine whether it is necessary to trigger incremental updates of the corresponding models to reduce unnecessary computing overhead.

[0188] According to the privacy leakage risk events discovered in real time, the hierarchical Dirichlet process and other methods are used to dynamically generate new sub-topics or merge similar topics based on the original topics to form a topic evolution tree; at the same time, a topic tracker is set up to map new topics to existing topics by calculating the JS divergence between topics, and allow the generation of new topics to achieve adaptive updates of topic distribution and maintain the continuity and interpretability of topics. Incremental association rule mining algorithms such as FUP and FPClose are used to perform association analysis on privacy leakage events generated in real time, mine new privacy leakage risk association rules, and use indicators such as support, confidence, and lift to evaluate the importance of rules. Prioritize the retention of rules with high stability, and dynamically prune and update existing association rules to achieve real-time maintenance of association rules.

[0189] Dynamic social network analysis methods such as time slice snapshots and sliding time windows are used to adaptively adjust the size of the time window according to the scale and density of the network. The multi-scale time window method is considered to extract the dynamic characteristics of the network at different time scales, and dynamically model and analyze the privacy leakage risk propagation network. The changes in indicators such as node centrality, community structure, and propagation path are used to evaluate the propagation trend and impact scope of privacy leakage risks in the network in real time, and dynamically adjust the propagation network model accordingly.

[0190] Specifically, when using real-time computing frameworks such as Apache Flink to process location data in real time, the frequency of data collection can be set to once every 5 seconds, the time window of data caching can be set to 1 minute, and Bloom filters can be used to deduplicate duplicate data, and the data ingestion rate can be controlled within 100,000 per second to ensure the real-time and accuracy of data processing. By matching real-time data with the privacy leakage risk topic model, if the similarity between new data and existing topics is less than 0.5, and the number of occurrences in the last 10 minutes exceeds 100 times, the incremental update of the topic model is triggered.

[0191] When dynamically updating the topic distribution, the HDP algorithm is used, the concentration parameter is set to 1.0, the corpus size is 10 million, the upper limit of the number of topics is 100, each new topic requires at least 100 associated documents, and the JS divergence with the existing topic is less than 0.3 to be merged. Otherwise, it is generated as a new topic and the topic evolution tree is updated. When mining incremental association rules, the minimum support is set to 0.01, the minimum confidence is set to 0.5, the maximum length of the rule is 3, the FPClose algorithm is used for mining, and the rule lift is calculated. Rules greater than 2.0 are considered strong association rules. At the same time, by counting the changes in the support and confidence of the rules in the last hour, day, and week, rules with a change rate of less than 10% are considered stable rules and are retained first. When conducting dynamic social network analysis, if the number of nodes in the network is less than 100, a 1-hour time window is used; if the number of nodes is between 100 and 1000, a 1-day time window is used; if the number of nodes is greater than 1000, a 1-week time window is used. At the same time, within each time window, a snapshot method is used to sample the network structure every 10 minutes.

[0192] By calculating the PageRank value of nodes, the modularity of communities and other indicators, the propagation trend of privacy leakage risks in the network is judged. If the number of nodes with PageRank values ​​exceeding 0.01 increases by more than 20%, or the modularity decreases by more than 0.1, it indicates that the risk of privacy leakage has a tendency to spread and needs to be controlled in time. In the privacy leakage risk assessment model, the fuzzy comprehensive evaluation method is used to set 9 evaluation indicators, including the sensitivity of data, the frequency of privacy leakage events, the strength of association rules, etc. The weight of each indicator is calculated by the hierarchical analysis method, and the risk threshold is set. When the comprehensive risk value exceeds 0.7, an orange warning is triggered, and when it exceeds 0.9, a red warning is triggered.

[0193] In the interactive visual risk monitoring platform, three levels of user permissions are set. Administrators can view all privacy leakage events, topic distribution, association rules, propagation paths and other detailed information. Ordinary users can only see their own privacy leakage risks, and visitors can only see desensitized statistical information. The platform provides multi-dimensional interactive operations, such as timeline dragging, area zooming, node clicking, etc. Users can customize monitoring topics, alarm methods, prevention and control measures, etc., and revise and feedback the risk analysis reports given by the system, continuously enhancing the system's self-learning ability.

[0194] See also Figure 2 , is a schematic diagram of the structure of a privacy data leakage risk control device provided by an embodiment of the present invention, the device comprising:

[0195] The extraction module is used to cluster and extract the leaked data within the life cycle of the privacy data according to the keywords of the privacy leakage risk theme, and obtain the probability matrix of the theme distribution of different leakage events;

[0196] The vector module is used to analyze the correlation between different privacy leakage events by calculating the similarity of the probability matrices of different leakage events, and determine the common feature vectors of privacy leakage risks;

[0197] The training module is used to construct a data set for privacy leakage risk classification based on the common feature vector to train the model leakage risk probability and obtain a leakage risk identification model.

[0198] The privacy data leakage risk control device provided in this embodiment can execute all the steps and functions executed in the privacy data leakage risk control method provided in any of the above embodiments, and the specific functions of the device are not described in detail here.

[0199] See also Figure 3 , is a schematic diagram of the structure of a terminal device provided in an embodiment of the present invention. The terminal device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a privacy data leakage risk control program. When the processor executes the computer program, the steps in each of the above-mentioned privacy data leakage risk control method embodiments are implemented, such as Figure 1 Alternatively, the processor implements the functions of each module in the above-mentioned device embodiments when executing the computer program.

[0200] Exemplarily, the computer program may be divided into one or more modules, and the one or more modules are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments that can perform specific functions, and the instruction segments are used to describe the execution process of the computer program in the privacy data leakage risk control device. For example, the computer program may be divided into several modules, and the specific functions of each module have been described in detail in a privacy data leakage risk control method provided in any of the above embodiments, and the specific functions of the device will not be repeated here.

[0201] The terminal device may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The terminal device may include, but is not limited to, a processor and a memory. A person skilled in the art may understand that the schematic diagram is only an example of a terminal device and does not constitute a limitation on a privacy data leakage risk control device. The terminal device may include more or fewer components than shown in the diagram, or a combination of certain components, or different components. For example, the terminal device may also include input and output devices, network access devices, buses, etc.

[0202] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the privacy data leakage risk control device, and uses various interfaces and lines to connect the various parts of the entire privacy data leakage risk control device.

[0203] The memory can be used to store the computer program and / or module, and the processor realizes various functions of the privacy data leakage risk control device by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0204] Wherein, if the module integrated in the privacy data leakage risk control device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0205] An embodiment of the present invention further provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of implementing the method by the functional network element in the above embodiment.

[0206] The computer program product provided in this embodiment can execute all the steps and functions of the privacy data leakage risk control method provided in any of the above embodiments, and the specific functions of the product are not described in detail here.

[0207] It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A privacy data leakage risk control method, characterized in that: The method comprises: Based on the keywords of the privacy leakage risk theme, the leakage data within the privacy data life cycle is clustered and extracted to obtain the probability matrix of the theme distribution of different leakage events; By calculating the similarity of the probability matrices of different leakage events, the correlation between different privacy leakage events is analyzed, and the common feature vectors of privacy leakage risks are determined; According to the common feature vector, a data set for privacy leakage risk classification is constructed to train the model leakage risk probability, thereby obtaining a leakage risk identification model.

2. The privacy data leakage risk control method according to claim 1 is characterized in that: Based on the keywords of the leakage risk theme, the leakage data within the privacy data life cycle is clustered and extracted to obtain the probability matrix of the theme distribution of different leakage events, including: Using the Latent Dirichlet Allocation topic model algorithm with preset extraction parameters to extract representative words of the topic from the leaked data, and calculating the probability distribution of keywords under different topics; According to the semantic similarity of representative words, different topics are merged and optimized to obtain topics of different categories; Calculate the probability matrix of the topic distribution of different leak events.

3. The privacy data leakage risk control method according to claim 1, characterized in that: By calculating the similarity of the probability matrices of different leakage events, analyzing the correlation between different privacy leakage events, and determining the common feature vectors of privacy leakage risks, including: Normalize the probability matrix of different events to obtain the unitized eigenvector; By calculating the dot product between the probability evidences, the cosine similarity values ​​of the probability matrices of different leakage events are determined; According to the calculated cosine similarity, a preset dimensionality reduction algorithm is used to extract the principal components of the privacy leakage risk characteristics to form a privacy leakage risk characteristic subspace; The original data is mapped into the privacy leakage risk feature subspace to determine the common feature vector of the privacy leakage risk.

4. The privacy data leakage risk control method according to claim 1, characterized in that: According to the common feature vector, a data set for privacy leakage risk classification is constructed to train the model leakage risk probability, and a leakage risk identification model is obtained, including: Based on the common feature vectors, the support vector machine algorithm is used to classify and label the privacy leakage data within the privacy data life cycle, and the labeling information of a part of the privacy leakage data is obtained as the basis of the training sample set; For unlabeled data, we use the transductive learning method of support vector machines to propagate the label information of labeled samples to unlabeled samples by utilizing the structural information of labeled samples and the similarity between samples, thus constructing a dataset for privacy leakage risk classification. A cross-validation scheme is adopted to randomly divide the data set into several mutually exclusive subsets, one of the subsets is selected as the validation set each time, and the remaining subsets are used as the training set, and the validation is repeated, and the average value of several validation results is taken as the performance estimation of the support vector machine model; The model parameters are traversed and optimized through grid search or random search, and the parameter combination with the highest expected value is selected as the optimal parameter to obtain the leakage risk identification model.

5. The privacy data leakage risk control method according to claim 1, characterized in that: The method further comprises: The preset association rule mining algorithm is used to analyze the association rules between different categories of privacy leakage risks in the life cycle of location data, and to determine the propagation path and impact scope of privacy leakage risks in the life cycle of location data.

6. The privacy data leakage risk control method according to claim 5, characterized in that: The preset association rule mining algorithm is used to analyze the association rules between different types of privacy leakage risks in the life cycle of location data, and determine the propagation path and impact scope of privacy leakage risks in the life cycle of location data, including: The Apriori association rule mining algorithm is used to determine the support and confidence thresholds through cross-validation; By presetting the evaluation indicators, the association rules between different types of privacy leakage risks in the privacy data life cycle are analyzed, and frequent item sets and strong association rules are screened out to form a privacy leakage risk association rule set; According to the privacy leakage risk association rule set, a privacy leakage risk association graph is constructed, and privacy leakage risk nodes of different categories are connected according to the association rules to form a directed acyclic graph; A preset mining algorithm is used on the directed acyclic graph to calculate the importance, authority and centrality of each privacy leakage risk node, identify key privacy leakage risk nodes and propagation paths, and determine the impact scope of privacy leakage risks within the life cycle of privacy data.

7. The privacy data leakage risk control method according to claim 5, characterized in that: The method further comprises: A propagation network topology diagram of the privacy leakage risk is constructed according to the propagation path and the impact scope, and key nodes and vulnerable links of the privacy leakage risk are identified.

8. The privacy data leakage risk control method according to claim 7, characterized in that: A propagation network topology diagram of the privacy leakage risk is constructed according to the propagation path and the impact scope, and key nodes and vulnerable links of the privacy leakage risk are identified, including: Using graph database technology, taking privacy leakage risk events as nodes and the correlation between different events as edges, a directed weighted graph is formed as the topology graph of the propagation network; According to the propagation network topology graph, a label propagation algorithm is used to perform community discovery, and the labels of nodes are updated iteratively until the label distribution converges; Dividing the communication network topology graph into a number of privacy leakage risk communities; According to the characteristics of the privacy leakage risk community, using preset evaluation indicators, calculate the importance index of each privacy leakage risk node within the community and between communities, determine the importance ranking of the nodes in the propagation network topology diagram, and determine the key nodes of privacy leakage risk; The shortest path algorithm is used to solve the shortest path between any two nodes, and the frequency of occurrence of intermediary nodes is counted to identify structural holes in the network and determine the vulnerable links for the propagation of privacy leakage risks.

9. The privacy data leakage risk control method according to claim 7, characterized in that: The method further comprises: According to the key nodes and vulnerable links, the information entropy risk assessment model is used to calculate the privacy leakage risk values ​​of different nodes and links in the life cycle of privacy data to obtain risk assessment results.

10. The privacy data leakage risk control method according to claim 9, characterized in that: According to the key nodes and vulnerable links, the information entropy risk assessment model is used to calculate the privacy leakage risk values ​​of different nodes and links in the life cycle of privacy data, and the risk assessment results are obtained, including: The life cycle of the privacy data is divided into multiple stages, and the fault tree analysis method is used to gradually decompose the privacy leakage event into multiple intermediate events and basic events, and the causal relationship between the events is represented by logic gates to form an inverted tree structure diagram; Filter out basic events with higher importance as the object set for privacy leakage risk assessment; For each risk assessment object, we use the hierarchical analysis method to compare and score the importance of each pair of factors at the same level from the two dimensions of the probability of occurrence and impact of privacy leakage incidents, and obtain a pairwise comparison matrix; Calculate the weight vector of each criterion, and comprehensively weight them to obtain the privacy leakage risk score of each assessment object at the target layer. Take the arithmetic mean of the scoring results to obtain the qualitative assessment value of the privacy leakage risk. According to the qualitative evaluation value, the probability of occurrence and severity of consequences of privacy leakage are quantified as the expected value of the membership function using the fuzzy comprehensive evaluation method, and the information entropy of the privacy leakage risk of each risk assessment object is obtained by substituting it into the calculation formula of the information entropy model; The information entropy is normalized to obtain privacy leakage risk values ​​of different nodes and links as the risk assessment result.

11. The privacy data leakage risk control method according to claim 9, characterized in that: The method further comprises: According to the risk assessment results, a fuzzy comprehensive evaluation method is used to determine the priority of privacy leakage risks within the life cycle of privacy data, and match the corresponding privacy protection strategy.

12. The privacy data leakage risk control method according to claim 11, characterized in that: According to the risk assessment results, the fuzzy comprehensive evaluation method is used to determine the priority of privacy leakage risks within the life cycle of privacy data, and the corresponding privacy protection strategies are matched, including: Sort the privacy leakage risk values ​​of different nodes in the risk assessment result from high to low to form a sorting vector of privacy leakage risk; Each risk object in the sorting vector is grouped according to the life cycle stage, and matched one by one with different evaluation indicators to form a risk indicator matching matrix, and a fuzzy evaluation indicator set of privacy leakage risk is constructed; The weight vector of each index is determined by using the hierarchical analysis method, and the maximum eigenvalue and corresponding eigenvector of the matrix are calculated by constructing a pairwise comparison judgment matrix. Substitute the quantitative evaluation value of each risk object into the membership function, calculate its membership on each evaluation index, and form a fuzzy relationship matrix; Using the fuzzy matrix synthesis operation principle, the weight vector and the fuzzy relationship matrix are synthesized to obtain the fuzzy comprehensive evaluation result vector of privacy leakage risk; Normalizing the fuzzy comprehensive evaluation result vector, and determining the priority of the privacy leakage risk according to the normalized privacy leakage risk evaluation index; According to the priority, a corresponding privacy protection strategy is matched in a preset privacy protection measure library.

13. The privacy data leakage risk control method according to claim 11, characterized in that: The method further comprises: By continuously monitoring the privacy leakage data during the life cycle of privacy data, the topic distribution, association rules and propagation network of privacy leakage risks are dynamically updated.

14. The privacy data leakage risk control method according to claim 13, characterized in that: By continuously monitoring the privacy leakage data in the privacy data life cycle, the topic distribution, association rules and propagation network of privacy leakage risks are dynamically updated, including: A real-time computing framework is used to process and analyze the data stream within the life cycle of private data in real time, and to extract characteristic indicators related to privacy leakage; The characteristic indicators are matched in real time with the pre-trained privacy leakage risk topic model, association rule model and propagation network model to determine whether it is necessary to trigger incremental updates of the corresponding models and dynamically update the topic distribution, association rules and propagation paths.

15. A privacy data leakage risk control device, characterized in that: The device comprises: The extraction module is used to cluster and extract the leaked data within the life cycle of the privacy data according to the keywords of the privacy leakage risk theme, and obtain the probability matrix of the theme distribution of different leakage events; The vector module is used to analyze the correlation between different privacy leakage events by calculating the similarity of the probability matrices of different leakage events, and determine the common feature vectors of privacy leakage risks; The training module is used to construct a data set for privacy leakage risk classification based on the common feature vector to train the model leakage risk probability and obtain a leakage risk identification model.

16. A terminal device, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, and when the processor executes the computer program, it implements the privacy data leakage risk control method as described in any one of claims 1 to 14.

17. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the privacy data leakage risk control method as described in any one of claims 1 to 14.

18. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 14 are implemented.

Images