Data association sensitivity analysis method and system based on graph attention network

Abstract

The invention relates to the field of data security, in particular to a data association sensitivity analysis method and system based on a graph attention network, and the method comprises the steps: constructing a knowledge graph based on the features and relationships of various different types of data; stacking the multi-layer graph attention network to obtain dependency relationship weights of nodes in different ranges in the knowledge graph; obtaining sensitivity during data combination based on a preset rule and a correlation model between the data, and performing feature coding and classification on the data combination based on the sensitivity to obtain categories of sensitive risks under different data combinations; performing similarity aggregation on the continuously increased data, and evaluating the uncertainty and sensitivity of data association; comparing the sensitivity with a sensitivity threshold value under a corresponding sensitive risk category to realize sensitivity analysis; according to the method, the complex association relationship of the data can be accurately captured, and compared with a traditional method, the sensitive association discovery rate is increased by 30%-50%.

Description

Technical Field

[0001] This invention relates to the field of data security, and specifically to a data association sensitivity analysis method and system based on graph attention networks. Background Technology

[0002] With the rapid development of big data, IoT and AI technologies, the scale of data collection and application is constantly expanding, and data security issues are becoming increasingly complex and critical. Traditional data security protection mainly focuses on the identification and protection of directly sensitive data, using rule-based methods to identify and protect obviously sensitive information, such as personal identification information, financial data, health records, etc.

[0003] However, a more insidious risk exists in practical applications: the combination of multiple data items that are inherently insensitive or of low sensitivity may generate new and unpredictable sensitive information. For example, combining anonymized location data with publicly available time information may lead to the re-identification of individuals; the combination of multiple business transaction records may reveal business strategies; and the aggregation of logs from different systems may expose network architecture and security policies. This kind of correlation sensitivity is covert, complex, and dynamic, making it difficult for traditional static rules and simple models to effectively identify and assess.

[0004] The existing technology has the following shortcomings:

[0005] 1. Insufficient identification of associated risks: Traditional methods mainly focus on the sensitivity of a single data point, making it difficult to deal with the complex sensitivity issues arising from data association.

[0006] 2. Data scale challenge: With the explosive growth of data volume, the possibilities of data combination increase exponentially, making it impossible to exhaustively list all possible combinations.

[0007] 3. Complex processing of multi-category data: Different categories of data (such as structured, semi-structured and unstructured data) have different structural, semantic and relational characteristics, making it difficult to represent and analyze them using a unified method.

[0008] 4. Sensitivity changes dynamically: The sensitivity of data is not static, but changes with the environment, context, and time. Static analysis methods are difficult to remain effective in the long run.

[0009] 5. Lack of standardized assessment methods: The assessment of data correlation sensitivity lacks unified standards and methods, making it difficult to compare and verify the results.

[0010] Therefore, there is an urgent need to design a method for comprehensive analysis of the correlation sensitivity of data. Summary of the Invention

[0011] To address the problem that existing technologies cannot identify and assess the sensitivity risks arising from data association in multi-source heterogeneous data environments, this invention proposes a data association sensitivity analysis method based on graph attention networks, including:

[0012] Knowledge graphs are constructed based on the features and relationships of various data types.

[0013] Stacking multiple layers of graph attention networks yields dependency weights for nodes at different ranges within the knowledge graph.

[0014] The sensitivity of data combinations is obtained based on the correlation model between preset rules and data, and the data combinations are feature-encoded and classified based on the sensitivity to obtain the category of sensitivity risk under different data combinations.

[0015] Perform similarity aggregation on the continuously increasing data and assess the uncertainty and sensitivity of data associations;

[0016] Sensitivity analysis is achieved by comparing the sensitivity with the sensitivity threshold under the corresponding sensitivity risk category.

[0017] The graph attention network uses a multi-head attention mechanism to calculate the attention weights between nodes in the knowledge graph; the association model is constructed based on the knowledge graph and the corresponding dependency weights.

[0018] Optionally, the construction of the knowledge graph based on the features and relationships of various data types specifically includes:

[0019] Feature extraction and feature fusion are performed on different types of data.

[0020] To obtain key entities and relationships between entities in the feature-fused data;

[0021] The key entities are used as nodes and the relationships are used as edges to construct a knowledge graph.

[0022] Optionally, after obtaining the key entities and relationships between entities in the feature fusion data and before constructing the knowledge graph, the process further includes at least one of the following operations: mapping the key entities and relationships to a predefined ontology model to ensure semantic consistency; merging key entities with the same concept; handling conflicts in the relationships based on preset conflict resolution rules; and inferring implicit relationships based on prior knowledge to expand the relationships.

[0023] Optionally, after constructing the knowledge graph, one or more of the following processes may be included:

[0024] The knowledge graph is then pruned, indexed, and partitioned.

[0025] Optionally, the graph attention network employs a type-aware attention mechanism.

[0026] Optionally, the attention coefficient of the type-aware attention mechanism is:

[0027]

[0028] Among them, α ij τ(i),τ(j),τ(e) Let τ be the attention coefficients for nodes i and j, τ(i) be the type of node i, τ(j) be the type of node j, τ(e) be the type of edge, and e be the type of edge. ij τ(i),τ(j),τ(e) Let be the attention coefficients when node i is of type τ(i) and node j is of type τ(j). Let N be the attention coefficients when node i is of type τ(i) and node k is of type τ(k). i Let i be the set of neighbors of node i.

[0029] Optionally, the feature encoding and classification of the data combination based on the sensitivity specifically includes:

[0030] Based on the aforementioned sensitivity, a Transformer architecture is used to encode the features of the data combination, and then a Softmax multi-classifier is used to classify different types of sensitivity risks.

[0031] Optionally, the step of performing similarity aggregation on the continuously increasing data and assessing the uncertainty and sensitivity of data associations specifically involves:

[0032] The nearest neighbor algorithm is used to identify and aggregate similar data in the continuously increasing data. Then, a probabilistic model is used to obtain the uncertainty and sensitivity of data association. The probabilistic model includes at least one of Bayesian network, joint probabilistic analysis and conditional risk assessment.

[0033] Optionally, the sensitivity threshold is obtained using statistical analysis, expert evaluation, risk modeling, or regulatory benchmarking.

[0034] A second aspect of this application provides a data association sensitivity analysis system based on graph attention networks, comprising:

[0035] Knowledge graph construction module: used to build knowledge graphs based on the features and relationships of various data types;

[0036] Weight acquisition module: used to stack multi-layer graph attention networks to obtain the dependency weights of nodes of different ranges in the knowledge graph;

[0037] Category determination module: used to obtain the sensitivity of data combinations based on preset rules and the correlation model between data, and to perform feature encoding and classification of data combinations based on the sensitivity to obtain the category of sensitivity risk under different data combinations;

[0038] Sensitivity assessment module: used to perform similarity aggregation on continuously increasing data and assess the uncertainty and sensitivity of data associations;

[0039] Result acquisition module: used to compare the sensitivity with the sensitivity threshold under the corresponding sensitivity risk category to realize sensitivity analysis;

[0040] The graph attention network uses a multi-head attention mechanism to calculate the attention weights between nodes in the knowledge graph; the association model is constructed based on the knowledge graph and the corresponding dependency weights.

[0041] Optionally, the knowledge graph construction module constructs a knowledge graph based on the features and relationships of various data types, specifically including:

[0042] Feature extraction and feature fusion are performed on different types of data.

[0043] To obtain key entities and relationships between entities in the feature-fused data;

[0044] The key entities are used as nodes and the relationships are used as edges to construct a knowledge graph.

[0045] Optionally, after the graph construction module obtains the key entities in the feature-fused data and the relationships between entities, but before constructing the knowledge graph, it further includes at least one of the following operations:

[0046] The key entities and relationships are mapped to a predefined ontology model to ensure semantic consistency; key entities with the same concept are merged; conflicts in the relationships are handled based on preset conflict resolution rules; and implicit relationships are inferred based on prior knowledge to expand the relationships.

[0047] Optionally, after constructing the knowledge graph, the graph construction module may further include one or more of the following processes:

[0048] The knowledge graph is then pruned, indexed, and partitioned.

[0049] Optionally, the graph attention network in the weight acquisition module adopts a type-aware attention mechanism.

[0050] Optionally, the attention coefficient of the type-aware attention mechanism in the weight acquisition module is:

[0051]

[0052] Among them, α ij τ(i),τ(j),τ(e) Let τ be the attention coefficients for nodes i and j, τ(i) be the type of node i, τ(j) be the type of node j, τ(e) be the type of edge, and e be the type of edge. ij τ(i),τ(j),τ(e) Let be the attention coefficients when node i is of type τ(i) and node j is of type τ(j). Let N be the attention coefficients when node i is of type τ(i) and node k is of type τ(k). i Let i be the set of neighbors of node i.

[0053] Optionally, the category determination module performs feature encoding and classification on the data combinations based on the sensitivity, specifically as follows:

[0054] Based on the aforementioned sensitivity, a Transformer architecture is used to encode the features of the data combination, and then a Softmax multi-classifier is used to classify different types of sensitivity risks.

[0055] Optionally, the sensitivity acquisition module performs similarity aggregation on the continuously increasing data and assesses the uncertainty and sensitivity of data associations, specifically:

[0056] The nearest neighbor algorithm is used to identify and aggregate similar data in the continuously increasing data. Then, a probabilistic model is used to obtain the uncertainty and sensitivity of data association. The probabilistic model includes at least one of Bayesian network, joint probabilistic analysis and conditional risk assessment.

[0057] Optionally, the sensitivity threshold in the result acquisition module is obtained using statistical analysis, expert evaluation, risk modeling, or regulatory benchmarking.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] This invention provides a data association sensitivity analysis method and system based on graph attention networks, including: constructing a knowledge graph based on the features and relationships of various data types; stacking multiple layers of graph attention networks to obtain dependency weights of nodes in different ranges within the knowledge graph; obtaining the sensitivity of data combinations based on preset rules and data association models, and performing feature encoding and classification on the data combinations based on the sensitivity to obtain the categories of sensitivity risks under different data combinations; performing similarity aggregation on data in different data combinations and evaluating the uncertainty and sensitivity of data associations; comparing the sensitivity with the sensitivity threshold under the corresponding sensitivity risk category to achieve sensitivity analysis; the solution of this application has:

[0060] 1) High accuracy: By combining graph attention networks and feature encoding classification, it can accurately capture complex relationships in data, improving the sensitive relationship detection rate by 30% to 50% compared to traditional methods;

[0061] 2) High adaptability to dynamic changes in sensitivity: Real-time similarity aggregation of data and assessment of the uncertainty and sensitivity of data associations can enable sensitivity assessment to change with environment, context and time.

[0062] 3) Expanding processing categories: It can realize the fusion analysis of data from different categories.

[0063] 4) High scalability: The modular system design supports the processing and expansion of different types of data, adapting to diverse business scenarios and data environments;

[0064] 5) Through algorithm optimization and system architecture design, efficient processing of large-scale data has been achieved, supporting real-time or near real-time sensitivity monitoring. Attached Figure Description

[0065] Figure 1 This is a detailed schematic diagram illustrating the steps of the data association sensitivity analysis method based on graph attention networks proposed in this invention;

[0066] Figure 2 The present invention proposes Figure 1 A detailed step diagram of step S1 in the diagram;

[0067] Figure 3 This is a schematic diagram illustrating the application process of the data association sensitivity analysis method based on graph attention networks proposed in this invention;

[0068] Figure 4 This is a schematic diagram of the data association sensitivity analysis system based on graph attention network proposed in this invention. Detailed Implementation

[0069] This invention proposes a data association sensitivity analysis method and system based on graph attention networks. It intelligently detects and assesses the sensitivity risks that may arise from data association in multi-source heterogeneous data environments. The method first constructs a sensitivity aggregation knowledge graph for various types of data through feature extraction. Then, it uses an improved graph attention network (GAT) to learn the graph structure and capture the complex relationships between data. Finally, it classifies and assesses the sensitivity risks generated by data aggregation. Furthermore, it combines the nearest neighbor algorithm and a probabilistic model to achieve a large-scale data sensitivity trigger detection function, capable of identifying sensitivity changes caused by data accumulation. Compared to traditional methods, this invention has higher detection accuracy, stronger interpretability, and better scalability, and can be widely applied to various scenarios such as data security risk assessment, data sharing security control, and data access management.

[0070] Example 1:

[0071] A data association sensitivity analysis method based on graph attention networks, such as Figure 1 As shown, the process includes the following steps S1 to S5:

[0072] S1: Construct a knowledge graph based on the features and relationships of various data types.

[0073] In further optimized solutions, such as Figure 2 As shown, step S1 includes the following steps S11 to S13.

[0074] S11: Extract features from different types of data and perform feature fusion;

[0075] S12: Obtain the key entities in the feature-fused data and the relationships between entities;

[0076] S13: Construct a knowledge graph by using the key entities as nodes and the relationships as edges.

[0077] In step S11, different types of data packets include structured data, text data, and multimedia data. Therefore, feature representations are extracted from different types of data, including structured data features, text semantic features, and multimedia features.

[0078] Different feature extraction methods are used for different types of data.

[0079] 1. Structured Data: Use rule-based methods to extract key fields and attributes. "Rule-based methods" typically refer to pre-defined feature extraction rules based on business needs or data types. For example, specifying which fields are key attributes, how to identify customer ID, amount, time, etc. This ensures consistency and reusability in the feature extraction process. A specific implementation method is as follows:

[0080] Algorithm 1: Feature Extraction from Structured Data

[0081] Input: Structured dataset D, field importance threshold θ

[0082] Output: Feature vector set F

[0083] Initialize feature vector set

[0084] For each record r∈D

[0085] Initialize the feature vector f r =[]

[0086] For each field field∈r

[0087] If importance (field) ≥ θ

[0088] f r .append(process(field))

[0089] F.add(f r )

[0090] Return F

[0091] 2. Text Data: Deep semantic features are extracted using pre-trained language models such as BERT. Text features can be represented as:

[0092] h i =BERT(x i )

[0093] Among them, x i For a text sequence, h i This is the extracted feature vector.

[0094] 3. Multimedia data: Visual features are extracted using models such as convolutional neural networks (CNN).

[0095] Feature fusion is mainly used to represent data containing multiple modalities using multimodal fusion techniques. The fusion formula is as follows:

[0096] h fused =MLP([h structured ||h text ||h multimedia ])

[0097] Among them, h fused The feature vector is the result of feature fusion. MLP stands for Multilayer Perceptron, used to learn the fused representation of features from different modalities. h structured For the characteristics of structured data (such as table fields, values, categories, etc.), h text For features of text data (such as text content, descriptive information, comments, etc.), h multimedia Features of multimedia data (such as images, audio, and video).

[0098] Step S12 mainly includes entity recognition and relation extraction. Entity recognition is used to identify key entities from the data, and relation extraction is used to discover various relationships between entities. Then, in step S13, the key entities are used as nodes and the relationships are used as edges to construct a knowledge graph.

[0099] In a further preferred embodiment, after obtaining the key entities and relationships between entities in the feature fusion data and before constructing the knowledge graph, at least one of the following operations is included: mapping the key entities and relationships to a predefined ontology model to ensure semantic consistency; merging key entities with the same concept; handling conflicts in the relationships based on preset conflict resolution rules; and reasoning about implicit relationships based on prior knowledge to expand the relationships.

[0100] After obtaining the key entities and relationships, we can first map them to a predefined ontology model to ensure semantic consistency. Then, the knowledge graph is constructed as follows:

[0101] Algorithm 2 Knowledge Graph Construction

[0102] Input: Feature vector set F, ontology model O

[0103] Output: Knowledge graph G(V,E)

[0104] Initialize node set Edge set

[0105] For each feature vector f∈F

[0106] entities = entity recognition(f)

[0107] V.addAll(entities)

[0108] relations = relation extraction(f, entities)

[0109] For each relation (e1,r,e2)∈relations

[0110] e1' = ontology mapping (e1, O)

[0111] r' = ontology mapping(r, O)

[0112] e2' = ontology mapping (e2, O)

[0113] E.add((e1',r',e2'))

[0114] Return G(V,E)

[0115] In a further preferred embodiment, key entities with the same concept are merged; conflicts in the relationships are handled based on preset conflict resolution rules; implicit relationships are inferred based on prior knowledge to expand the relationships; the purpose of the above processing is to integrate knowledge from different sources, resolve conflicts, and enhance the integrity of the graph.

[0116] 1. Merge key entities with the same concept for entity alignment;

[0117] 2. Conflict Detection and Resolution: Discovering and resolving knowledge conflicts, such as attribute conflicts and relationship conflicts. Specifically, this is implemented as follows:

[0118] Conflict resolution (v1, v2) = {v1, if credibility (v1) > credibility (v2); v2, if credibility (v1) < credibility (v2); f merge (v1,v2), otherwise}

[0119] Where v1 and v2 are conflicting values, f merge For merging functions, choose the appropriate merging strategy based on the specific situation.

[0120] 3. Knowledge Reasoning: Based on prior knowledge, reason about implicit relationships, infer implicit relationships and attributes, and fill in the gaps in the graph.

[0121] Prior knowledge generally refers to entities, attributes, and relationships that have been explicitly represented in the knowledge graph through steps such as feature extraction, entity recognition, relation extraction, and ontology mapping. In other words, the graph has stored and structured various factual information that has been identified by the system.

[0122] Reasoning methods include rule-based reasoning, logical reasoning, embedding reasoning, graph neural network reasoning, and joint reasoning. Rule-based reasoning infers relationships and attributes based on expert-defined rules (such as if-then statements); logical reasoning uses first-order logic, descriptive logic, and other methods to logically infer relationships between entities; embedding reasoning utilizes knowledge image embeddings (such as TransE, RotatE, etc.) to predict potential relationships through vector space calculations; graph neural network reasoning uses GNN models to automatically learn implicit dependencies between entities and relationships to infer unknown relationships; and joint reasoning combines rule-based and model-based methods to improve the accuracy and coverage of inference.

[0123] In a further optimized scheme, after constructing the knowledge graph, the following processing is also performed on the knowledge graph:

[0124] (1) Graph pruning, also known as graph trimming, is performed on the knowledge image by removing unimportant or redundant nodes or edges to reduce the size of the graph. The importance score of nodes is calculated as follows during the modification process:

[0125] Score(v)=α·Degree(v)+β·Centrality(v)+γ·Sensitivity(v)

[0126] Wherein, Score(v) is the importance score of node v, Degree(v) is the degree centrality of node v, Centrality(v) is the betweenness centrality of node v, Sensitivity(v) is the sensitivity relevance of node v, and α, β, and γ are weight coefficients.

[0127] The formula for calculating the importance score of an edge is similar to that for calculating the importance score of a node, and will not be repeated here.

[0128] (2) Index Building: Establish an efficient graph index structure to accelerate query and retrieval operations. The index building process is generally completed automatically during the knowledge graph construction or optimization phase. The index is dynamically updated when new nodes or relationships are added to ensure that query and retrieval operations are always efficient. Graph databases such as Neo4j and TigerGraph can be used with their built-in efficient indexing mechanisms, and their automatically generated and optimized graph index structures can be directly utilized.

[0129] (3) Partition optimization: For large-scale knowledge graphs, graph partitioning is performed to support distributed processing.

[0130] S2: Stack multi-layer graph attention networks to obtain the dependency weights of nodes in different ranges in the knowledge graph.

[0131] Graph Attention Networks (GATs) learn the relationships between nodes through a multi-head attention mechanism, capturing complex dependency patterns in the data.

[0132] The core of GAT is to calculate the attention weights between nodes, which represent the strength of the association between nodes, e. ij :

[0133] e ij =LeakyReLU(a T [Wh i ||Wh j ])

[0134] Among them, h i h is the feature vector of node i. j Let be the feature vector of node j, W be the weight matrix, a be the attention vector, and || denote the concatenation operation.

[0135] The attention coefficient α is obtained by normalizing the association strength. ij :

[0136] α ij =exp(e ij ) / Σ k∈Ni exp(e ik )

[0137] Among them, N iLet k be the set of neighbors of node i, k be the k-th node, and e be the set of neighbors of node i. ik Let be the association strength between node i and node k.

[0138] To enhance the model's representational power and stability, a multi-head attention mechanism is employed. In this mechanism, node i's new feature h' is generated after multi-head attention aggregation. i Represented as:

[0139] h' i =|| k＝1 K σ(Σ j∈Ni α ij k W k h j )

[0140] Where K is the number of attention heads, α ij k W is the normalization coefficient for the k-th attention head. k σ is the weight matrix of the k-th attention head, and σ is a non-linear activation function, such as ELU.

[0141] In a further preferred embodiment, since the processing objects of this application include heterogeneous graphs with different types of nodes and edges, the graph attention network adopts a type-aware attention mechanism:

[0142]

[0143] Among them, α ij τ(i),τ(j),τ(e) Let τ be the attention coefficients for nodes i and j, τ(i) be the type of node i, τ(j) be the type of node j, τ(e) be the type of edge, and e be the type of edge. ij τ(i),τ(j),τ(e) Let be the attention coefficients when node i is of type τ(i) and node j is of type τ(j). Let be the attention coefficients when node i is of type τ(i) and node k is of type τ(k).

[0144] By stacking multiple layers of GAT, node dependencies of different ranges are captured, and therefore the feature output of each layer is:

[0145] h i (l+1) =|| k＝1 K σ(Σ j∈Ni α ij k,(l) W k,(l) h j (l) )

[0146] Where l is the layer index, hi (l+1) Let α be the feature vector representing the node in the (l+1)th layer. ij k,(l) W is the normalized coefficient of the k-th attention head in the l-th layer. k,(l) Let h be the weight matrix of the k-th attention head in the l-th layer. j (l) Let the node in the l-th layer represent the feature vector.

[0147] After the above GAT stacking process, a knowledge image with dependency weights can be obtained.

[0148] S3: Based on the preset rules and the correlation model between the data, obtain the sensitivity of the data combination, and perform feature encoding and classification on the data combination based on the sensitivity to obtain the category of sensitivity risk under different data combinations.

[0149] First, it is necessary to construct a relationship model between the data. A relationship model refers to a mathematical and machine learning model used to model and analyze the relationships between data items. It can be based on structured modeling methods such as graph neural networks and Ransformer to capture direct or indirect relationships between data items. The relationship model in this application includes direct relationships, indirect relationships, and contextual relationships; wherein:

[0150] Direct association: A direct join based on explicit fields or reference relationships;

[0151] Indirect association: Implicit relationships established through intermediate nodes or inference, such as associations established through third-party data;

[0152] Contextual association: Association based on contextual factors such as usage scenarios and access patterns.

[0153] Association strength can be quantified based on information theory:

[0154] I(X;Y)=Σ x∈X Σ y∈Y p(x,y)log[p(x,y) / (p(x)p(y))]

[0155] Where I(X;Y) is the mutual information of variables X and Y, p(x,y) is the joint probability distribution, p(x) is the marginal probability distribution of variable x, and p(y) is the marginal probability distribution of variable y.

[0156] The pre-defined rules mainly include business rules and security rules used to identify sensitive data aggregations. For example, the definitions and processing requirements for personal privacy and sensitive information in laws, regulations or industry standards; the rules for prohibiting the combination of sensitive data set within the organization (such as exposing certain fields at the same time); and security policies in the process of data flow, access, and sharing (such as access permissions, the principle of minimum necessity, etc.). These rules can be explicit logical judgments (such as the simultaneous appearance of information A and B is sensitive) or judgment criteria based on practical experience.

[0157] Furthermore, data combination refers to the combination of multiple related data items (such as different nodes and their relationships) in a weighted knowledge graph or data network according to certain business scenarios or analysis objectives for feature extraction and risk analysis.

[0158] In a further preferred approach, the Transformer architecture is used to encode the features of the data combinations:

[0159] 1. Input representation: Represent the data combinations as a sequence, adding positional encoding to preserve order information:

[0160] x i =h i +p i

[0161] Among them, x i For a data combination sequence, h i p is a feature representation of the data combination. i For position encoding;

[0162] 2. Self-attention computation: Capturing dependencies between data items through a self-attention mechanism:

[0163] Attention(Q,K,V) = softmax(QK) T / √d k V

[0164] Where Attention(Q,K,V) is the weighted sum of the attention weights calculated from the query matrix Q, key matrix K, and value matrix V, over the value matrix V, √d k Let be the dimension of the key vector, and softmax be a non-linear function.

[0165] 3. Multi-head attention: Using multiple attention heads to learn dependencies from different perspectives:

[0166] MultiHead(Q,K,V)=Concat(head1,...,head h W O

[0167] Among them, MultiHead(Q,K,V) represents multi-head attention. head i For the output of the i-th attention head, W i Q , W i V Let W be the learnable parameter matrix of the i-th attention head, and let W be the learnable parameter matrix of the output. O Concat() maps the multi-head concatenation result back to the original dimension, concatenating the outputs of multiple attention heads along the feature dimension. h This is the output of the h-th attention head.

[0168] 4. Feedforward Network: Each attention layer is followed by a positional feedforward network to enhance the model's representational power.

[0169] FFN(x) = max(0, xW1+b1)W2+b2

[0170] Where FFN(x) is the output of the feedforward network, x is the input of the feedforward network, i.e. the output of the attention layer, W1 is the weight matrix of the first layer, b1 is the bias of the first layer, W2 is the weight matrix of the second layer, b2 is the bias of the second layer, and W1, W2, b1 and b2 are all learnable parameters.

[0171] The deep feature representation of each data combination obtained through the above feature encoding will serve as input for subsequent sensitivity classification and risk analysis, enabling automatic identification and assessment of data sensitivity.

[0172] In a further optimized approach, a Softmax multi-classifier is used to classify different types of sensitivity risks:

[0173]

[0174] Where P(y=j|x) is the probability that sample x belongs to class j, z=Wx+b is the output of the classification layer, and K is the total number of sensitive classes. Mapping the predicted score of sample x belonging to class j to the positive number space, e z k The predicted score of sample x belonging to category k is mapped to the positive number space.

[0175] For situations involving multiple sensitivities, a multi-label classification method is employed.

[0176]

[0177] Wherein, P(y j =1|x) means x belongs to category yj The probability, σ is the sigmoid function, y j ∈{0,1} represents whether it belongs to the j-th type of sensitivity, z j This is the linear output of the model for the sensitivity of the j-th class.

[0178] S4: Perform similarity aggregation on the continuously increasing data and assess the uncertainty and sensitivity of data associations.

[0179] This step uses the classification output from step S3 as input to further conduct sensitive trigger detection on a gradually expanding scale of data, achieving batch identification and early warning of sensitive events. The large-scale sensitive trigger detection uses quantity threshold determination, nearest neighbor analysis, and probabilistic models to identify changes in sensitivity caused by the accumulation of data.

[0180] (1) Determine the sensitivity thresholds for different types of data and define the boundary between "safe" and "sensitive".

[0181] Sensitivity thresholds can be determined using methods such as statistical analysis, expert evaluation, risk modeling, and regulatory benchmarking.

[0182] Statistical analysis method: By analyzing a large amount of historical data, we can discover the relationship between sensitivity and data volume.

[0183] Expert evaluation method: Combining the experience of domain experts, reasonable thresholds are set for different types of data;

[0184] Risk Model: Constructing a mathematical model of the relationship between data volume and risk level, such as:

[0185] R = f(V,T,S,C,...)

[0186] Where R is the risk score, V is the data volume, T is the data type, S is the data sensitivity baseline, C is the context factor, and f is the risk function.

[0187] Regulatory benchmarking method: Refer to the data volume restrictions imposed by relevant laws and regulations.

[0188] (2) To identify similar patterns in an ever-increasing large-scale dataset, the nearest neighbor algorithm can be used. The steps include:

[0189] Step 1: Feature space construction, representing data as high-dimensional vectors to support similarity calculation;

[0190] Step 2: Distance metric selection. Choose an appropriate distance metric based on the data characteristics, such as Euclidean distance or cosine similarity. Cosine similarity is calculated as follows:

[0191] cos(x,y)=(x·y) / (||x||·||y||)

[0192] Where cos(x,y) is the cosine similarity value between vectors x and y.

[0193] Step 3: Nearest Neighbor Search Optimization: Use efficient index structures and algorithms, such as KD-trees and Locality Sensitive Hash (LSH), to accelerate the search process. The core idea of ​​LSH is to design a hash function that gives similar data items a high probability of being hashed into the same bucket.

[0194] h(x) = sgn(w·x+b)

[0195] Where h(x) is the hash function, sgn(·) is the sign function, w is the random vector, and b is the random offset.

[0196] In a further preferred embodiment, a probabilistic model is further employed to obtain the uncertainty and sensitivity of data associations. The probabilistic model includes at least one of Bayesian networks, joint probabilistic analysis, and conditional risk assessment, wherein:

[0197] Bayesian networks: establish conditional dependencies between data items and support probabilistic inference.

[0198] Joint probability analysis: assessing the overall sensitivity risk of combining multiple data items.

[0199] I(X1; X2; ...; X n )=Σ i＝1 n H(X i )-H(X1,X2,...,X n )

[0200] Where I(X1; X2; ...; X n ) represents multivariate mutual information, X i For the i-th data item, H(X) i H(X1,X2,...,X) is a single variable. n ) represents the joint entropy, and n represents the total number of data items or variables.

[0201] Conditional risk assessment: Analyzing the sensitivity (P) of data combinations under specific conditions.

[0202] P(Sensitivity|Data Conditions, Context Conditions)

[0203] The Bayesian network, joint probabilistic analysis, and conditional risk assessment methods described in this application are multiple probabilistic analysis methods for the same dataset. They can be used independently or in combination according to actual needs. Conditional risk assessment is usually based on the output of the aforementioned probabilistic models for further analysis.

[0204] S5: Compare the sensitivity with the sensitivity threshold under the corresponding sensitivity risk category to perform sensitivity analysis.

[0205] Real-time monitoring of data volume changes reveals sensitivity changes during these changes. These changes are then compared to sensitivity thresholds within the corresponding sensitivity risk categories to perform sensitivity analysis and provide timely warnings. This process can include the following steps:

[0206] Real-time and accurate monitoring of the cumulative amount of different types of data;

[0207] Predict trends in data volume and assess future risks;

[0208] The sensitivity threshold is dynamically adjusted based on changes in context and environment.

[0209] Set three threshold levels: warning, restriction, and prohibition, to provide progressive risk alerts.

[0210] In summary, the overall process of this application is as follows: Figure 3 As shown, firstly, multi-source heterogeneous data is input to construct a knowledge graph. Then, GAT is used to learn graph features and assign weights to nodes and edges in the graph. Next, data association sensitivity analysis is performed to obtain sensitivity classification. For the sensitivity of a large amount of data that gradually increases, risk events and risk trends are detected, and early warnings are issued and appropriate measures are taken.

[0211] Example 2

[0212] Based on the same inventive concept, a second aspect of this application provides a data association sensitivity analysis system based on graph attention networks, such as... Figure 4 Shown, including:

[0213] Knowledge graph construction module: used to build knowledge graphs based on the features and relationships of various data types;

[0214] Weight acquisition module: used to stack multi-layer graph attention networks to obtain the dependency weights of nodes of different ranges in the knowledge graph;

[0215] Category determination module: used to obtain the sensitivity of data combinations based on preset rules and the correlation model between data, and to perform feature encoding and classification of data combinations based on the sensitivity to obtain the category of sensitivity risk under different data combinations;

[0216] Sensitivity assessment module: used to perform similarity aggregation on continuously increasing data and assess the uncertainty and sensitivity of data associations;

[0217] Result acquisition module: used to compare the sensitivity with the sensitivity threshold under the corresponding sensitivity risk category to realize sensitivity analysis;

[0218] The graph attention network uses a multi-head attention mechanism to calculate the attention weights between nodes in the knowledge graph; the association model is constructed based on the knowledge graph and the corresponding dependency weights.

[0219] In a further preferred embodiment, the knowledge graph construction module constructs a knowledge graph based on the features and relationships of various data types, specifically including:

[0220] Feature extraction and feature fusion are performed on different types of data.

[0221] To obtain key entities and relationships between entities in the feature-fused data;

[0222] The key entities are used as nodes and the relationships are used as edges to construct a knowledge graph.

[0223] In a further preferred embodiment, after the graph construction module obtains the key entities in the feature-fused data and the relationships between entities, and before constructing the knowledge graph, it further includes at least one of the following operations:

[0224] The key entities and relationships are mapped to a predefined ontology model to ensure semantic consistency; key entities with the same concept are merged; conflicts in the relationships are handled based on preset conflict resolution rules; and implicit relationships are inferred based on prior knowledge to expand the relationships.

[0225] In a further preferred embodiment, after the graph construction module constructs the knowledge graph, it further includes one or more of the following processes:

[0226] The knowledge graph is then pruned, indexed, and partitioned.

[0227] In a further preferred embodiment, the graph attention network in the weight acquisition module adopts a type-aware attention mechanism.

[0228] In a further preferred embodiment, the attention coefficient of the type-aware attention mechanism in the weight acquisition module is:

[0229]

[0230] Among them, α ij τ(i),τ(j),τ(e) Let τ be the attention coefficients for nodes i and j, τ(i) be the type of node i, τ(j) be the type of node j, τ(e) be the type of edge, and e be the type of edge. ij τ(i),τ(j),τ(e) Let be the attention coefficients when node i is of type τ(i) and node j is of type τ(j). Let be the attention coefficients when node i is of type τ(i) and node k is of type τ(k), and Ni be the total number of nodes.

[0231] In a further preferred embodiment, the category determination module performs feature encoding and classification on the data combinations based on the sensitivity, specifically as follows:

[0232] Based on the aforementioned sensitivity, a Transformer architecture is used to encode the features of the data combination, and then a Softmax multi-classifier is used to classify different types of sensitivity risks.

[0233] In a further preferred embodiment, the sensitivity acquisition module performs similarity aggregation on the continuously increasing data and assesses the uncertainty and sensitivity of data associations, specifically as follows:

[0234] The nearest neighbor algorithm is used to identify and aggregate similar data in the continuously increasing data. Then, a probabilistic model is used to obtain the uncertainty and sensitivity of data association. The probabilistic model includes at least one of Bayesian network, joint probabilistic analysis and conditional risk assessment.

[0235] In a further preferred embodiment, the sensitivity threshold in the result acquisition module is obtained using statistical analysis, expert evaluation, risk modeling, or regulatory benchmarking.

[0236] Example 3

[0237] This application also provides a data association sensitivity analysis system based on graph attention networks, including a data access layer, a processing layer, a storage layer, an interface layer, and a security layer, wherein:

[0238] The data access layer is responsible for securely and efficiently collecting data from various data sources;

[0239] The processing layer is used to implement core algorithms and analysis functions, including knowledge graph construction, GAT analysis, sensitivity calculation, etc., as shown in Example 1.

[0240] The storage layer is used to manage the storage of various types of data and models, including raw data, knowledge graphs, model parameters, etc.

[0241] The interface layer provides APIs and user interfaces, supporting integration with other systems and user interaction;

[0242] The security layer is used to implement security mechanisms such as access control and data encryption to protect system and data security.

[0243] The system supports multiple deployment modes, including local deployment, cloud deployment, and hybrid deployment, to meet the needs and environmental conditions of different users.

[0244] This invention can be widely applied in the following scenarios:

[0245] 1. Data security risk assessment: comprehensively assess the organization's data security risks, with particular attention to the potential sensitivities arising from data associations.

[0246] 2. Data sharing security controls: Assess combined sensitivity risks and determine appropriate security control measures before sharing data.

[0247] 3. Data access control optimization: Based on the sensitivity analysis results, optimize access control strategies to prevent information leakage caused by excessive access.

[0248] 4. Data Breach Investigation: Analyze the sensitivity and potential impact of breached data to support incident response and remediation.

[0249] 5. Privacy Compliance: Helps organizations achieve data privacy compliance and meet the requirements of regulations such as GDPR and PIPL.

[0250] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A data association sensitivity analysis method based on graph attention networks, characterized in that, include: Knowledge graphs are constructed based on the features and relationships of various data types. Stacking multiple layers of graph attention networks yields dependency weights for nodes at different ranges within the knowledge graph. The sensitivity of data combinations is obtained based on the correlation model between preset rules and data, and the data combinations are feature-encoded and classified based on the sensitivity to obtain the category of sensitivity risk under different data combinations. Perform similarity aggregation on the continuously increasing data and assess the uncertainty and sensitivity of data associations; Sensitivity analysis is achieved by comparing the sensitivity with the sensitivity threshold under the corresponding sensitivity risk category. The graph attention network uses a multi-head attention mechanism to calculate the attention weights between nodes in the knowledge graph; the association model is constructed based on the knowledge graph and the corresponding dependency weights.

2. The data association sensitivity analysis method based on graph attention networks according to claim 1, characterized in that, The construction of the knowledge graph based on the features and relationships of various data types specifically includes: Feature extraction and feature fusion are performed on different types of data. To obtain key entities and relationships between entities in the feature-fused data; The key entities are used as nodes and the relationships are used as edges to construct a knowledge graph.

3. The data association sensitivity analysis method based on graph attention networks according to claim 2, characterized in that, After obtaining the key entities and relationships between entities in the feature-fused data, and before constructing the knowledge graph, at least one of the following operations is also included: The key entities and relationships are mapped to a predefined ontology model to ensure semantic consistency; key entities with the same concept are merged; and conflicts in the relationships are handled based on preset conflict resolution rules. Reasoning about implicit relations based on prior knowledge is used to expand the relations.

4. The data association sensitivity analysis method based on graph attention networks according to claim 2, characterized in that, After constructing a knowledge graph, one or more of the following processes are also involved: The knowledge graph is then pruned, indexed, and partitioned.

5. The data association sensitivity analysis method based on graph attention networks according to claim 1, characterized in that, The graph attention network employs a type-aware attention mechanism.

6. The data association sensitivity analysis method based on graph attention networks according to claim 5, characterized in that, The attention coefficient of the type-aware attention mechanism is: Among them, α ij τ(i),τ(j),τ(e) Let τ be the attention coefficients for nodes i and j, τ(i) be the type of node i, τ(j) be the type of node j, τ(e) be the type of edge, and e be the type of edge. ij τ(i),τ(j),τ(e) Let be the attention coefficients when node i is of type τ(i) and node j is of type τ(j). Let N be the attention coefficients when node i is of type τ(i) and node k is of type τ(k). i Let i be the set of neighbors of node i.

7. The data association sensitivity analysis method based on graph attention networks according to claim 1, characterized in that, The feature encoding and classification of data combinations based on the aforementioned sensitivity specifically involves: Based on the aforementioned sensitivity, a Transformer architecture is used to encode the features of the data combination, and then a Softmax multi-classifier is used to classify different types of sensitivity risks.

8. The data association sensitivity analysis method based on graph attention network according to claim 1, characterized in that, The process of performing similarity aggregation on continuously increasing data and assessing the uncertainty and sensitivity of data associations specifically includes: The nearest neighbor algorithm is used to identify and aggregate similar data in the continuously increasing data. Then, a probabilistic model is used to obtain the uncertainty and sensitivity of data association. The probabilistic model includes at least one of Bayesian network, joint probabilistic analysis and conditional risk assessment.

9. The data association sensitivity analysis method based on graph attention networks according to claim 1, characterized in that, The sensitivity thresholds are obtained using statistical analysis, expert evaluation, risk modeling, or regulatory benchmarking.

10. A data association sensitivity analysis system based on graph attention networks, characterized in that, include: Knowledge graph construction module: used to build knowledge graphs based on the features and relationships of various data types; Weight acquisition module: used to stack multi-layer graph attention networks to obtain the dependency weights of nodes of different ranges in the knowledge graph; Category determination module: used to obtain the sensitivity of data combinations based on preset rules and the correlation model between data, and to perform feature encoding and classification of data combinations based on the sensitivity to obtain the category of sensitivity risk under different data combinations; Sensitivity assessment module: used to perform similarity aggregation on continuously increasing data and assess the uncertainty and sensitivity of data associations; Result acquisition module: used to compare the sensitivity with the sensitivity threshold under the corresponding sensitivity risk category to realize sensitivity analysis; The graph attention network uses a multi-head attention mechanism to calculate the attention weights between nodes in the knowledge graph; the association model is constructed based on the knowledge graph and the corresponding dependency weights.

Images