Cross-domain data flow permission dynamic matching security assessment method based on deep learning
By using a deep learning-based method for dynamic matching of cross-domain data flow permissions, a cross-domain data-permission federated association feature set is constructed. By using a federated deep learning model to mine implicit relationships, the problems of dynamic adaptation and privacy protection in traditional cross-domain permission management are solved, and accurate security assessment and improved data interaction security are achieved.
Patent Information
- Application Number
- CN202511791401.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-02-27
AI Technical Summary
Traditional cross-domain permission management suffers from fixed permission rules that cannot adapt to dynamic changes, easy leakage of privacy data, and a single security assessment dimension, making it difficult to achieve dynamic adaptation and accurate assessment in cross-domain data flow.
A deep learning-based method for dynamic matching of cross-domain data flow permissions is adopted. Through federated feature modeling and on-chain standardization, a cross-domain data-permission federated association feature set is constructed. The implicit association relationship is mined by the federated deep learning dynamic matching model. The weight fusion is combined with the cross-domain federated attention mechanism to generate permission matching degree and security risk assessment.
It enables dynamic adaptation of permissions and accurate assessment of security risks in cross-domain data transfer, protects data privacy, and improves the security and efficiency of cross-domain data interaction.
Smart Images
Figure CN121585443A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of data security and privacy protection, in particular to a cross-domain data flow transfer permission dynamic matching security evaluation method based on deep learning. BACKGROUND
[0002] With the development of digital economy, cross-domain data flow transfer (such as data interaction between different enterprises, institutions and regions) has become a key link to support business collaboration, but it also faces the challenges of complex permission management and high security risks. Traditional cross-domain permission management mostly uses static rule matching method, which has the following defects:
[0003] The permission rules are fixed and cannot adapt to the dynamically changing cross-domain scenarios (such as data sensitivity level change and inter-domain trust relationship adjustment);
[0004] Dependence on centralized data processing, which is easy to lead to privacy data leakage;
[0005] Single security evaluation dimension, which is difficult to comprehensively consider the influence of data attributes, subject behavior, link state and other factors.
[0006] Although the method based on deep learning can improve the matching accuracy, the data privacy protection and multi-source information collaborative modeling in cross-domain scenarios are still problems to be solved. Therefore, a cross-domain data flow transfer permission security management scheme that takes into account privacy protection, dynamic adaptation and accurate evaluation is needed. SUMMARY
[0007] The purpose of the present application is to provide a cross-domain data flow transfer permission dynamic matching security evaluation method based on deep learning, to realize the dynamic adaptation of permissions in cross-domain data flow transfer and the accurate evaluation of security risks, and to protect data privacy, in order to solve the technical problems proposed in the background art.
[0008] To achieve the above purpose, the present application provides the following technical scheme: a cross-domain data flow transfer permission dynamic matching security evaluation method based on deep learning, comprising at least the following steps:
[0009] S1: Collecting privacy protection type multi-dimensional information in cross-domain data flow transfer scenarios, the multi-dimensional information including cross-domain data sensitive attribute summary, flow transfer subject identity trusted credential and behavior trace record, inter-domain transmission link state data and permission rule logic;
[0010] S2: Federal feature modeling and on-chain standardization processing of the multi-dimensional information, constructing a cross-domain data-permission federal association feature set;
[0011] S3: Input the cross-domain data-permission federated association feature set into the federated deep learning dynamic matching model. The first federated branch of the model is used to mine the implicit association between cross-domain privacy protection data and permissions. The second federated branch of the model is used to quantify the inter-domain permission adaptability based on blockchain trusted data. The two types of outputs are weighted and fused together by the cross-domain federated attention mechanism to output the permission matching degree and security risk assessment value of cross-domain data flow.
[0012] S4: Based on the permission matching degree and security risk assessment value, generate dynamic permission matching results and security control suggestions for cross-domain data flow.
[0013] Furthermore, the cross-domain data sensitive attribute digest is generated by de-identifying the original data using privacy computing technology. The cross-domain data sensitive attribute digest includes data category, sensitivity level, and circulation scope constraint information.
[0014] The trusted identity credential of the transfer subject includes the subject's digital certificate and the domain authentication identifier; the behavior traceability record includes historical cross-domain access logs and permission usage records.
[0015] The status data of the inter-domain transmission link includes link encryption strength, transmission delay, packet loss rate, and node trust level;
[0016] The permission rule logic includes basic permission policies within the domain, whitelists / blacklists for cross-domain access, and dynamic permission adjustment thresholds.
[0017] Furthermore, the federated feature modeling specifically includes:
[0018] Based on the federated learning framework, each participating domain extracts features from its own data locally and achieves cross-domain feature collaborative modeling through encrypted parameter interaction, thus avoiding the leakage of original data.
[0019] The on-chain standardization process specifically involves: verifying and normalizing the format of the federated modeled features through a blockchain smart contract to generate feature vectors that conform to cross-domain interaction standards. The immutability of the blockchain ensures the integrity and traceability of the feature data.
[0020] Furthermore, the federated deep learning dynamic matching model includes a federated first branch, a federated second branch, and a cross-domain federated attention mechanism module;
[0021] The first branch of the federation adopts a federated deep neural network structure. Each domain local sub-model is trained on local data-permission features. The implicit correlation between cross-domain data sensitive attributes and permission rules is mined through federated parameter aggregation, and the correlation score is output.
[0022] The second branch of the federation adopts a federated graph neural network structure, which uses the inter-domain interaction history and permission adaptation records recorded by the blockchain as nodes to construct the graph structure. The compatibility and adaptability of the inter-domain permission rules are quantified through federated graph convolution operations, and an adaptation score is output.
[0023] The cross-domain federated attention mechanism module dynamically adjusts the weights of the two types of scores based on the real-time security requirements of cross-domain scenarios (such as data sensitivity and inter-domain trust levels), and obtains the permission matching degree and security risk assessment value after weighted fusion.
[0024] Furthermore, the correlation score output by the first branch of the federation Calculated using a federated deep neural network, the following condition is met:
[0025]
[0026] Among them, DNN For the i-th domain, it is a local deep neural network; For the standardized features of the i-th domain; For local model weights; Agg For aggregate functions of federated parameters; This is the Sigmoid activation function (output range [0,1]).
[0027] Furthermore, the fitness score output by the second branch of the federation Calculated using a federated graph neural network, the following condition is met:
[0028]
[0029] in, This is an on-chain domain relationship graph (nodes are domains, and edge weights are historical fit degrees). For the first Local subgraphs of each domain; This is a local graph neural network; The subgraph weights are defined, and the output range is [0,1].
[0030] Furthermore, the cross-domain federated attention mechanism... and The weighted fusion satisfies: Let the attention weights be... (Based on data sensitivity level) Dynamically generated , (where the inter-domain trust threshold is used), then the permission matching degree and safety risk assessment value They are respectively:
[0031]
[0032]
[0033] in, The link security risk coefficient (calculated based on encryption strength and packet loss rate); Risk coefficient for the main behavior (calculated based on historical violation records); , , Risk weights ( ).
[0034] Furthermore, the permission matching degree is a value between 0 and 1, where a higher value indicates a higher degree of matching between the current data flow request and the target domain permission rules;
[0035] The security risk assessment value is generated based on risk level classification (such as low, medium, and high), combined with the probability of data leakage, risk of privilege abuse, and risk of link attack.
[0036] Furthermore, the dynamic matching result of the permissions includes allowing flow, conditionally allowing flow (requiring supplementary permission verification), or prohibiting flow;
[0037] The security control recommendations include link encryption upgrade schemes, temporary permission adjustment strategies, secondary authentication requirements for subject identities, and abnormal behavior monitoring thresholds.
[0038] Furthermore, the construction process of the cross-domain data-permission federation associated feature set satisfies the following mathematical relationship:
[0039] Let the local feature set of the i-th domain be . ( (For the j-th class feature vector of the i-th domain), federated feature modeling achieves cross-domain feature aggregation through encrypted parameter interaction, resulting in a global feature set. satisfy:
[0040]
[0041] Where n is the number of participating domains; The feature weights of the i-th domain ( ); Enc() is the state encryption function; after on-chain normalization, the feature vector satisfy:
[0042]
[0043] This is the standardized feature vector, with values ranging from [0,1].
[0044] Compared with the prior art, the beneficial effects of the present invention are:
[0045] 1. This invention has strong privacy protection capabilities. It achieves "local modeling and parameter interaction" of data through federated learning, and combines blockchain to ensure the trustworthiness of feature data and avoid the leakage of original data.
[0046] 2. The present invention has high dynamic adaptability. It is based on deep learning to mine the implicit relationship between data and permissions in real time, and combines cross-domain attention mechanism to dynamically adjust the matching strategy to adapt to changes in the scenario.
[0047] 3. The evaluation dimensions of this invention are more comprehensive, integrating multiple dimensions of information such as data attributes, subject behavior, link status, and inter-domain rules, thereby improving the accuracy of security risk assessment;
[0048] 4. This invention is highly practical. By outputting specific permission matching results and control suggestions, it can be directly applied to cross-domain data transfer scenarios such as government affairs, healthcare, and finance. Attached Figure Description
[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0050] Fig. 1 This is a schematic diagram of the process of the present invention;
[0051] Fig. 2 This is a schematic diagram of the S2 process of the present invention;
[0052] Fig. 3 This is a schematic diagram of the S3 process of the present invention. Detailed Implementation
[0053] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0054] This invention collects multi-dimensional privacy-preserving information in cross-domain data transfer scenarios, constructs a set of associated features through federated feature modeling and on-chain standardization, utilizes a federated deep learning dynamic matching model to mine the implicit correlation between data and permissions and quantifies the inter-domain permission adaptability, and combines a cross-domain federated attention mechanism to output permission matching degree and security risk assessment value, ultimately generating dynamic matching results and security control suggestions. This invention achieves dynamic adaptation of permissions and accurate assessment of security risks in cross-domain data transfer, improving the security and efficiency of cross-domain data interaction while ensuring data privacy.
[0055] Example 1:
[0056] Please see Figs. 1-3 A deep learning-based method for dynamic matching of permissions in cross-domain data transfer, comprising at least the following steps:
[0057] S1: Collection of privacy-preserving, multi-dimensional information in four core information categories during cross-domain data transfer scenarios:
[0058] Cross-domain data sensitive attribute summary: The original data is processed through technologies such as differential privacy and federated de-identification to extract data categories (such as medical data and financial data), sensitivity levels (classified according to national data security standards), and circulation scope constraints (such as access restricted to specific domains).
[0059] Trusted credentials of the transfer subject and behavior traceability records: Trusted credentials of the subject's identity include digital certificates and domain CA authentication marks, and behavior traceability records include logs of historical cross-domain access time, frequency, and compliance of permission usage;
[0060] Status data of inter-domain transmission links: including the encryption algorithm used by the link (such as AES-256, SM4), transmission delay, packet loss rate, and trust rating of intermediate nodes (based on historical attack event records).
[0061] Permission rule logic: including basic permission policies for each domain (such as read / write / modify permission allocation), cross-domain access whitelists / blacklists, and dynamically adjusted thresholds (such as permission contraction rules when the data sensitivity level is increased).
[0062] S2: Federated Feature Modeling and On-Chain Standardization Processing
[0063] Federated Feature Modeling: Based on the horizontal federated learning framework, each participating domain deploys a feature extraction submodule locally to perform feature transformation on the local information collected by S1 (such as converting text-based permission rules into vectors and behavior logs into time-series features). Cross-domain feature collaborative training is achieved through encrypted gradient interaction to avoid the original data leaving the country.
[0064] On-chain standardization processing: The feature vectors after federated modeling are uploaded to the blockchain. The format verification (such as feature dimension consistency and numerical range rationality) and normalization processing (such as mapping features of different dimensions to the [0,1] interval) are performed through smart contracts to generate cross-domain data-permission federated associated feature sets. The distributed ledger of the blockchain ensures that the feature data is tamper-proof.
[0065] S3: Federated Deep Learning Dynamic Matching and Security Assessment inputs the cross-domain data-permission federated association feature set into the federated deep learning dynamic matching model. The model structure is as follows:
[0066] The first branch of the federated network (latent association mining) adopts a federated deep neural network (FederatedDNN). Each domain sub-model consists of a 3-layer fully connected network. It learns the local association between data-sensitive attributes and permission rules through local data training, and then aggregates global parameters through the federated parameter server to output the implicit association score between data and permissions (0-1).
[0067] The second branch of the federation (inter-domain adaptation measurement): adopts a federated graph neural network (FederatedGNN) to construct a cross-domain relationship graph using the inter-domain interaction history and successful / failed permission adaptation cases recorded in the blockchain as nodes. It learns the dependencies between nodes through the federated graph convolutional layer and outputs the adaptation score (0-1) of the inter-domain permission rules.
[0068] Cross-domain federated attention mechanism: Attention weights are dynamically generated based on real-time cross-domain scenario parameters (such as current data sensitivity level and inter-domain trust value), and the relevance score and suitability score are weighted and fused (such as increasing the weight of the relevance score in high-sensitivity data scenarios). Finally, the permission matching degree (comprehensive score) and security risk assessment value are output (quantified by combining the link risk and behavioral risk features in the feature set).
[0069] The federated feature aggregation model targets n participating domains, where local features of each domain are cryptographically aggregated using a federated learning framework. Let the feature contribution of the i-th domain be denoted as . (Determined by the amount and reliability of data within the domain) Then, the aggregation process of the global feature set is as follows:
[0070]
[0071] Among them, Enc( The Paillier homomorphic encryption algorithm is used to ensure the privacy of feature parameters during the aggregation process.
[0072] To eliminate dimensional differences in features across different domains, the on-chain standardization model performs min-max normalization via blockchain smart contracts.
[0073]
[0074] in, The original value of the j-th type feature. The standardized value. Let be the set of features of class j across all domains. After standardization, the feature vectors satisfy . (d is the feature dimension).
[0075] Federal Branch 1 (Hidden Association Mining)
[0076] Local deep neural networks Using a 3-layer fully connected structure, the output of the k-th layer is:
[0077]
[0078] in, (Input features) Let represent the weights and biases of the i-th domain at the k-th layer.
[0079] The global relevance score is obtained through federated aggregation:
[0080]
[0081] The confidence weights for the local model are calculated based on the accuracy of the validation set.
[0082] Federal Second Branch (Inter-Domain Adaptation Measurement)
[0083] Inter-domain relationship graph In the middle, node Represents the participating domain;
[0084] edge weight (Historical adaptation success rate of u and v domains);
[0085] This map neural network GNN_i updates node features through graph convolutional layers:
[0086]
[0087] For the neighboring nodes of v, Global adaptability score based on node degree:
[0088]
[0089] Let be the domain importance weight of node v.
[0090] Cross-domain federated attention mechanism
[0091] Attention weight Based on data sensitivity level ( (1 being the most sensitive) Dynamically generated:
[0092]
[0093] The inter-domain trust threshold is calculated based on the compliance rate of historical interactions recorded by the blockchain. .
[0094] Permission matching degree Overall relevance and suitability:
[0095]
[0096] S4: Generate dynamic matching results and control recommendations
[0097] Dynamic permission matching results: When the permission matching degree is ≥0.8 and the security risk assessment value is "low", output "allow flow"; when the matching degree is between 0.5 and 0.8 or the risk assessment value is "medium", output "conditionally allow flow" (such as requiring the subject to supplement biometric authentication); when the matching degree is <0.5 or the risk assessment value is "high", output "prohibit flow".
[0098] Security control recommendations: Develop specific measures for risk points, such as upgrading to the national cryptographic algorithm SM9 when the link encryption strength is insufficient, enabling real-time monitoring when the subject's behavior is abnormal, and temporarily activating cross-domain emergency permission rules when permission adaptation conflicts occur.
[0099] The specific mathematical model for security risk assessment is as follows:
[0100] The security risk assessment value R comprehensively considers permission matching, link risk, and behavioral risk.
[0101]
[0102] in:
[0103] (Link risk):
[0104]
[0105] EncStrength is the encryption strength score, LossRate is the packet loss rate, and NodeTrust is the node trust level.
[0106] (Behavioral Risk):
[0107]
[0108] Calculated based on historical violation records;
[0109] Weight , , Determined using the Analytic Hierarchy Process (AHP), satisfying... (In highly sensitive scenarios) (Weight increase).
[0110] The mathematical rules for determining the dynamic matching results are as follows:
[0111] Permitted to circulate: and ;
[0112] Conditional permission to transfer: or ;
[0113] No transfer allowed: or .
[0114] Example 2:
[0115] Secure sharing of sensitive data across subsidiaries / partners within a group enterprise
[0116] 1. Scenario Overview
[0117] Background: A large automobile manufacturing group (headquarters, domain A) needs to obtain sensitive operational and customer data from its holding financial services company (domain B) and an external parts supplier (domain C, strategic partner) to optimize its supply chain and conduct market analysis. The data will be used in a next-generation intelligent decision-making platform developed at the group headquarters.
[0118] Data silos and privacy: Subsidiaries (domain B) and external partners (domain C) regard their data as core assets and have strict privacy protection regulations that prohibit raw data from being directly transferred out of their controlled domains.
[0119] Complex and dynamic permissions: The group headquarters (domain A) has different permission requirements for data from different sources and of different types (e.g., customer loan information vs. parts inventory data), and the security and compliance policies of partners are dynamically adjusted.
[0120] Security and compliance risks: It is necessary to assess the risks of data transmission across enterprise networks and ensure that the entire process complies with internal controls and external regulations (such as GDPR and data security laws).
[0121] 2. Specific implementation steps of the method of the present invention
[0122] S1: Collect privacy-protected multi-dimensional information
[0123] Each participating party (Group Headquarters A, Financial Subsidiary B, Supplier C) collects the following information locally:
[0124] Cross-domain data sensitive attribute summary (generated by financial subsidiary B):
[0125] Data category: Customer auto loan application data.
[0126] Sensitivity level: "Highly sensitive" (includes personally identifiable information and financial information).
[0127] Scope constraints: Only "risk analysis models" that are allowed to be transferred to the group headquarters (domain A) are permitted, and the model training results must be anonymized.
[0128] (Note: This summary has been generated by statistically analyzing the original data fields using homomorphic encryption technology. It only provides aggregated information and does not expose individual records.)
[0129] Credible credentials and behavioral traceability records of the transfer entity (provided by Group Headquarters A):
[0130] Identity credentials: The digital identity of the group headquarters (issued by the group's root CA), and the hash value of the "Intelligent Decision Platform V2.0" corresponding to this data request.
[0131] Behavioral Record: The platform has initiated 20 data requests in the past three months, and no data leaks or misuse incidents have occurred.
[0132] Status data of inter-domain transmission links (provided by the SD-WAN network controller):
[0133] Encryption strength: IPSec VPN with AES-256 encryption.
[0134] Transmission delay: Average delay 45ms.
[0135] Packet loss rate: 0.02%.
[0136] Node Trustworthiness Rating: The cloud service provider nodes traversed by the transmission path are rated as "trustworthy".
[0137] Permission rule logic (local policies for each enterprise domain):
[0138] The rules for Financial Subsidiary B stipulate that highly sensitive customer data may only be used for "risk control" and "product optimization" purposes at the group headquarters, and must be transmitted through a group-level encrypted link. It is prohibited from being released to any third party (including other subsidiaries).
[0139] Supplier C's rule: Component inventory and logistics data can be shared with the group headquarters for supply chain optimization, but the data retention period shall not exceed 90 days.
[0140] S2: Federated Feature Modeling and On-Chain Standardization Processing
[0141] Federated Feature Modeling:
[0142] Group headquarters A, financial subsidiary B, and supplier C each deploy feature extraction modules locally. For example, financial subsidiary B converts its data into feature vectors using policy text, while group headquarters A converts the context information of its data requests into feature vectors.
[0143] Through a horizontal federated learning framework, all parties jointly train a global feature extraction model by exchanging encrypted updates of model parameters. Throughout the process, the original customer data of financial subsidiary B and the original inventory data of supplier C remain on their respective local machines.
[0144] On-chain standardization processing:
[0145] Each party submits the locally extracted feature vectors to a consortium blockchain jointly maintained by the three parties.
[0146] Smart contracts on the blockchain automatically perform feature format verification and normalization processing, ensuring that all features are on a uniform scale, generating a standardized "cross-domain data-permission federation associated feature set". The immutability of the blockchain provides a reliable input source for subsequent evaluation.
[0147] S3: Federated Deep Learning Dynamic Matching and Security Assessment
[0148] The standardized feature set is input into the pre-trained federated deep learning dynamic matching model.
[0149] Federal First Branch (DNN Branch) Work:
[0150] This branch delves into the implicit relationships between data attributes and complex permission rules. For example, the model learns that "highly sensitive financial data" is strongly correlated with "risk control purposes" and "AES-256 encryption," but weakly correlated with "market research purposes."
[0151] After federated aggregation, this branch outputs a correlation score of S_rel = 0.88 (high score, because the data usage, sensitivity and permission rules are highly matched).
[0152] Federal Branch 2 (GNN Branch) work:
[0153] This branch constructs an enterprise relationship graph based on trusted historical interaction records on the blockchain. Nodes represent individual companies, and the weights of the edges reflect the success rate and compliance of historical data collaboration.
[0154] Graph neural networks are used to analyze inter-domain trust relationships. For example, Group Headquarters A has long-term and high-frequency compliance data interactions with Financial Subsidiary B, resulting in high trust edge weights; while the interaction with Supplier C is the first of its kind, with edge weights set to initial values.
[0155] This branch outputs a fit score of S_ada = 0.75 (the overall average score is lowered due to the trust relationship with supplier C).
[0156] Cross-domain federated attention mechanism fusion:
[0157] Dynamic weight calculation: Since the data sensitivity is "high" in this case, the system automatically assigns a higher weight to the first branch (focusing on the attributes of the data itself). The attention weight γ = 0.65 is calculated according to the formula.
[0158] Final score calculation:
[0159] Permission matching degree M = 0.65 * 0.88 + (1-0.65) * 0.75 = 0.834
[0160] Security risk assessment value R: The comprehensive calculation of permission matching degree (1-M), link risk r_link (very low), and behavioral risk r_behav (very low) is taken into account, but the inter-domain trust risk of partner C is introduced. Finally, R = 0.28 (which belongs to "low risk").
[0161] S4: Generate dynamic matching results and control recommendations
[0162] Dynamic permission matching results:
[0163] Based on the rule (M >= 0.8 and R <= 0.3), the system automatically generates the decision: "Allow the transfer".
[0164] Safety management recommendations:
[0165] The system generates precise control recommendations based on potential risks identified during the assessment (primarily insufficient trust in supplier C):
[0166] Data governance recommendation: When using data obtained from supplier C on the group headquarters platform, enforce a data lifecycle policy of "automatic deletion after 90 days".
[0167] Audit control recommendations: Enable enhanced audit trails for all data access operations (especially financial data from domain B), and upload logs to the blockchain for evidence storage in real time.
[0168] Link management recommendation: Maintain the current IPSec VPN configuration and confirm its security.
[0169] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A deep learning-based method for dynamic matching and security assessment of cross-domain data transfer permissions, characterized by: At least the following steps are included: S1: Collect privacy-protecting multi-dimensional information in cross-domain data transfer scenarios. The multi-dimensional information includes cross-domain data sensitive attribute summaries, trusted identity credentials and behavior traceability records of the transfer subjects, status data of inter-domain transmission links, and permission rule logic. S2: Perform federated feature modeling and on-chain standardization processing on the multi-dimensional information to construct a cross-domain data-permission federated association feature set; S3: Input the cross-domain data-permission federated association feature set into the federated deep learning dynamic matching model. The first federated branch of the model is used to mine the implicit association between cross-domain privacy protection data and permissions. The second federated branch of the model is used to quantify the inter-domain permission adaptability based on blockchain trusted data. The two types of outputs are weighted and fused together by the cross-domain federated attention mechanism to output the permission matching degree and security risk assessment value of cross-domain data flow. S4: Based on the permission matching degree and security risk assessment value, generate dynamic permission matching results and security control suggestions for cross-domain data flow.
2. The deep learning-based cross-domain data transfer permission dynamic matching security assessment method according to claim 1, characterized in that: The cross-domain data sensitive attribute digest is generated by de-identifying the original data using privacy computing technology. The cross-domain data sensitive attribute digest includes data category, sensitivity level, and circulation scope constraint information. The trusted identity credential of the transfer subject includes the subject's digital certificate and the domain authentication identifier; the behavior traceability record includes historical cross-domain access logs and permission usage records. The status data of the inter-domain transmission link includes link encryption strength, transmission delay, packet loss rate, and node trust level; The permission rule logic includes basic permission policies within the domain, whitelists / blacklists for cross-domain access, and dynamic permission adjustment thresholds.
3. The deep learning-based cross-domain data transfer permission dynamic matching security assessment method according to claim 1, characterized in that: The federated feature modeling specifically refers to: Based on the federated learning framework, each participating domain extracts features from its own data locally and achieves cross-domain feature collaborative modeling through encrypted parameter interaction, thus avoiding the leakage of original data. The on-chain standardization process specifically involves: verifying and normalizing the format of the federated modeled features through a blockchain smart contract to generate feature vectors that conform to cross-domain interaction standards. The immutability of the blockchain ensures the integrity and traceability of the feature data.
4. The deep learning-based cross-domain data transfer permission dynamic matching security assessment method according to claim 1, characterized in that: The federated deep learning dynamic matching model includes a federated first branch, a federated second branch, and a cross-domain federated attention mechanism module. The first branch of the federation adopts a federated deep neural network structure. Each domain local sub-model is trained on local data-permission features. The implicit correlation between cross-domain data sensitive attributes and permission rules is mined through federated parameter aggregation, and the correlation score is output. The second branch of the federation adopts a federated graph neural network structure, which uses the inter-domain interaction history and permission adaptation records recorded by the blockchain as nodes to construct the graph structure. The compatibility and adaptability of the inter-domain permission rules are quantified through federated graph convolution operations, and an adaptation score is output. The cross-domain federated attention mechanism module dynamically adjusts the weights of the two types of scores according to the real-time security requirements of cross-domain scenarios, and obtains the permission matching degree and security risk assessment value after weighted fusion.
5. The deep learning-based cross-domain data transfer permission dynamic matching security assessment method according to claim 4, characterized in that: The correlation score output by the first branch of the Federation Calculated using a federated deep neural network, the following condition is met: Among them, DNN For the i-th domain, it is a local deep neural network; For the standardized features of the i-th domain; For local model weights; Agg For aggregate functions of federated parameters; This is the Sigmoid activation function.
6. The deep learning-based cross-domain data transfer permission dynamic matching security assessment method according to claim 5, characterized in that: The fit score output by the second branch of the federation Calculated using a federated graph neural network, the following condition is met: in, This is an on-chain domain relationship graph (nodes are domains, and edge weights are historical fit degrees). For the first Local subgraphs of each domain; This is a local graph neural network; The weights are those of the subgraph.
7. The deep learning-based cross-domain data transfer permission dynamic matching security assessment method according to claim 6, characterized in that: The cross-domain federated attention mechanism and The weighted fusion satisfies: Let the attention weights be... Then the permission matching degree and safety risk assessment value They are respectively: in, This represents the link security risk coefficient. Risk coefficient of the main behavior; , , Risk weights.
8. The deep learning-based cross-domain data transfer permission dynamic matching security assessment method according to claim 7, characterized in that: The permission matching degree is a value between 0 and 1. The higher the value, the higher the degree of matching between the current data flow request and the target domain permission rules. The security risk assessment value is generated based on risk level classification, combined with the probability of data leakage, risk of privilege abuse, and risk of link attack.
9. The deep learning-based cross-domain data transfer permission dynamic matching security assessment method according to claim 1, characterized in that: The dynamic matching result of the permissions includes allowing circulation, conditionally allowing circulation, or prohibiting circulation; The security control recommendations include link encryption upgrade schemes, temporary permission adjustment strategies, secondary authentication requirements for subject identities, and abnormal behavior monitoring thresholds.
10. The deep learning-based cross-domain data transfer permission dynamic matching security assessment method according to claim 2, characterized in that: The construction process of the cross-domain data-permission federation associated feature set satisfies the following mathematical relationship: Let the local feature set of the i-th domain be . Federated feature modeling achieves cross-domain feature aggregation through encrypted parameter interaction, resulting in a global feature set. satisfy: Where n is the number of participating domains; The feature weights are the feature weights of the i-th domain; Enc() is the state encryption function; after on-chain normalization, the feature vectors are... satisfy: This is the standardized feature vector.