AI large model access data security control method

Abstract

This invention discloses a method for controlling access security of large AI models, relating to the field of data security technology. The invention includes the following steps: Step S1: Data classification and grading preprocessing: Multi-dimensional feature extraction is performed on the original data to be accessed by the large AI model. Based on a preset classification and grading rule base, the data security level value is calculated using the analytic hierarchy process (AHP). The specific formula is: where n is the number of data feature dimensions, w i The weight is the i-th feature dimension. This invention has the function of multi-dimensional and accurate data security management. Through multi-dimensional feature extraction, such as data sensitivity and importance, the security level is calculated using the analytic hierarchy process, making data classification and grading more accurate. At the same time, it combines natural language processing to parse access requests and generate an initial access token with multiple parameters, laying the foundation for subsequent security control.

Description

Technical Field

[0001] This invention relates to the field of data security technology, and in particular to a method for controlling the access security of large AI models. Background Technology

[0002] With the rapid development of artificial intelligence technology, the demand for data access from large AI models is increasing, and data security issues are becoming increasingly prominent. Traditional data security control methods are gradually showing many shortcomings when faced with the complex access scenarios of large AI models.

[0003] Traditional methods for data classification and grading mostly rely on a single dimension or fixed rules, making it difficult to comprehensively and accurately reflect the security attributes of data. For example, determining a security level based solely on one aspect of data sensitivity or importance fails to consider factors such as data usage frequency and update cycle. This leads to inaccurate assessments of data security levels. Regarding access control mechanisms, traditional methods typically employ static permission management strategies. Once permissions are granted, they do not dynamically adjust based on changes in the access subject's behavior or environment for a certain period. This static approach struggles to cope with complex and ever-changing access scenarios and cannot promptly detect and address potential security risks.

[0004] Furthermore, traditional data interaction security control and auditing mechanisms also have significant shortcomings. In terms of data encryption and de-identification, traditional methods may not be standardized enough or specific enough to meet the protection needs of data with different security levels; the recording and storage of audit logs are also relatively simple, lacking effective integrity verification and traceability mechanisms, making it difficult to quickly and accurately track and locate security issues.

[0005] To address these issues, we provide a method for controlling the security of access to large AI models' data. Summary of the Invention

[0006] The purpose of this invention is to provide a method for controlling the access security of large AI models. By combining multi-dimensional data classification and grading with a dynamic permission adjustment mechanism, it solves the problems of single data security level assessment dimensions and static permission management being unable to cope with dynamic risks in the prior art.

[0007] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0008] This invention relates to a method for controlling data security access to large AI models, comprising the following steps: Step S1: Data classification and grading preprocessing: Multi-dimensional feature extraction is performed on the original data to be accessed by the large AI model. Based on a preset classification and grading rule base, the data security level value is calculated using the analytic hierarchy process (AHP). The specific formula is as follows: Where n is the number of data feature dimensions, w i The weight of the i-th feature dimension (determined through expert scoring combined with historical access logs during training) Satisfy; C i To score the security level of the i-th feature dimension, a 5-level Likert scale is used for quantification (level 1 is the lowest security level, and level 5 is the highest security level). The scoring factors include concrete indicators such as the confidentiality of the data content and the scope of impact of data leakage. Step S2: Access request parsing and verification: The access request of the AI ​​big model is parsed using natural language processing technology to extract parameters such as the request subject identifier (MID), data object unique identifier (DID), operation type (OT), timestamp accurate to milliseconds (TS), and device fingerprint (DF). Based on the S value in step S1, an initial access token T containing timestamp verification and device fingerprint binding is generated. init The token is valid for T years. valid =15×60+Δt (for request response time compensation); Step S3: Dynamic policy matching: Construct a three-dimensional policy space consisting of a subject security attribute space, a data security level axis, and an operational risk coefficient axis, where: the subject security attributes include the subject reputation value obtained through BP neural network training (range 0-1), the subject permission level based on the RBAC model (levels 1-10), and the compliance rate of access history in the past 30 days; the operational risk coefficient is calculated using the fuzzy comprehensive evaluation method: O r =γ·O t +δ·O s (γ+δ=1,γ≥0.6) Among them, O t Risk score for operation type (read = 3 points, write = 5 points, delete = 7 points), O s The operation frequency risk score is calculated based on the sliding window algorithm (linearly increasing when the number of operations per unit time exceeds a threshold); Step S4: Real-time risk assessment: Real-time collection of behavioral data through probes deployed on the data interaction link: including access frequency (times / minute), data query depth (number of nested SQL statements), data output volume (MB / time), and abnormal operation response time (ms); Environmental data: including network security level (classified into levels 1-5 according to NIST standards), device security status (integrity measurement value obtained through the TCG trusted computing module), and geofence coordinates (access area compliance based on GPS positioning); The improved DS evidence theory is used to fuse and calculate the real-time risk value R: R = α·R b +β·R e (α=0.6, β=0.4) where R b R represents the behavioral risk value (threshold range 0-100). e Environmental risk value (threshold range 0-100); Step S5: Dynamic adjustment of permissions: Establish a threshold R that includes permission contraction.th1 =80, Permission retention threshold R th2 =60, Permission Enhancement Threshold R th3 =40 Three-level dynamic adjustment mechanism: When R>R th1 When this occurs, permission contraction is triggered, generating a set of restricted permissions. At that time, R... th2 ≤R≤R th1 Maintain the current set of permissions, when R≤R th3 Temporary expansion is allowed; Step S6: Data interaction security control: Encryption processing: Data with security level S≥4 is asymmetric encrypted using the SM2 national cryptographic algorithm, and the key exchange process follows the IBC identifier cryptosystem; De-identification processing: For data fields containing personal information, dynamic masking algorithms (such as retaining the first 3 and last 4 digits and filling the middle with *), rule-based replacement methods (such as replacing the ID number with a virtual ID) or irreversible deletion methods are used; Format conversion: Adaptive conversion from data storage format to model input format is achieved through a pre-trained format mapping model (supporting JSON / Protobuf / Parquet format conversion). Transformation; Step S7: Access Log Recording and Auditing: Log Structure: Adopts a standardized log format that includes the digital certificate fingerprint of the access subject, the hash value of the data object, the semantic parsing tree of the operation, and the timestamp of the permission verification; Storage Mechanism: An audit log library is built based on the Hyperledger Fabric consortium blockchain. Each log block includes the hash value of the previous block, the Merkle tree root hash, and the timestamp; Verification System: Log file digests are generated using the SHA-256 algorithm, and log integrity verification is implemented using the bilinear pairing algorithm. The traceability marker includes the IP address, port number, and device certificate information of the entire link from the AI ​​large model to the data storage node.

[0009] The present invention is further configured such that the multi-dimensional feature extraction in step S1 specifically includes: data sensitivity (weight 0.3): classified into public data (1 point), restricted data (3 points), and sensitive data (5 points) according to the GDPR data classification standard; data importance (weight 0.4): scored according to the degree of impact of data loss on business (no impact = 1 point, serious impact = 5 points); data usage frequency (weight 0.2): quantified based on the number of accesses in the last 30 days (<10 times = 1 point, >100 times = 5 points); data update cycle (weight 0.1): scored according to the data update frequency (annual update = 1 point, real-time update = 5 points).

[0010] The present invention is further configured such that the initial access token generation formula in step S2 is optimized as follows: in, This indicates an XOR operation, where H is a hash function containing a salt value, which is dynamically obtained from a trusted security chip.

[0011] The present invention is further configured such that the environmental data in step S4 also includes: time access window: defining weekdays from 9:00 to 18:00 as the normal access period (risk coefficient 0.8), and other times as sensitive periods (risk coefficient 1.2); equipment safety status: calculated by the Hamming distance between the PCR register value obtained by the TPM chip and the benchmark value, and if the distance is >5, it is determined to be an unsafe device.

[0012] The present invention is further configured such that the dynamic adjustment of permissions in step S5 also includes a temporary permission enhancement mechanism: when it is detected that the AI ​​large model is performing a preset emergency task (such as disaster prediction model training), the validity period of the permission is automatically extended to T. valid ×2, and temporarily grant data write permissions.

[0013] The present invention is further configured such that, in step S6, the data integrity verification adopts the CRC-32 check algorithm, and the original data check value CRC is calculated before data interaction. src After interaction, calculate the CRC checksum of the received data. dst If CRC src ≠CRC dst This will trigger the retransmission mechanism.

[0014] The present invention is further configured such that, in step S7, the blockchain audit log library is equipped with a multi-node consensus mechanism, the consensus algorithm adopts the improved PBFT algorithm, and the nodes participating in the consensus need to pass two-factor authentication to ensure the security of the log uploading process.

[0015] The present invention is further configured such that, after step S7, abnormal access processing is performed, wherein the abnormal behavior detection model is trained using a Long Short-Term Memory (LSTM) network, and the detection indicators include: the daily access frequency exceeds three times the standard deviation of the historical average; the cross-security domain data transmission rate exceeds the bandwidth threshold; when an anomaly is detected, real-time warning information containing the threat level (level 1-5), the anomaly type code, and the occurrence timestamp is sent to the security management platform via REST API.

[0016] The present invention is further configured such that the dynamic update mechanism includes: a classification and grading rule base: the rules are iterated monthly based on the latest data leakage event database, and the sliding window duration of the frequency features is updated; a strategy matching algorithm: when the strategy matching failure rate exceeds 15% for 3 consecutive days, a genetic algorithm is triggered to optimize the weight parameters of the three-dimensional strategy space; and a risk assessment model: adopting an online learning method, absorbing newly generated behavioral data and environmental data in real time, and triggering model fine-tuning once every 100,000 data points.

[0017] The present invention is further configured such that the preset emergency task triggering conditions in the temporary permission enhancement mechanism include: the task tag contains a preset set of emergency task keywords (identified through natural language processing technology); and the security level of the task initiator is higher than the authorization threshold corresponding to the current data security level value.

[0018] The present invention has the following beneficial effects:

[0019] 1. This invention has the function of multi-dimensional and precise data security management. By extracting multi-dimensional features, such as data sensitivity and importance, it uses the hierarchical analysis method to calculate the security level, making data classification and grading more accurate. At the same time, it combines natural language processing to parse access requests and generate an initial access token with multiple parameters, laying the foundation for subsequent security control.

[0020] 2. This invention constructs a three-dimensional strategy space, combines BP neural networks and other methods to calculate operational risks, and then uses an improved DS evidence theory to fuse risk values ​​to achieve dynamic adjustment of permissions. Moreover, encryption and desensitization are performed during data interaction, and an audit log library is constructed in conjunction with blockchain to ensure data security and traceability. Attached Figure Description

[0021] 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.

[0022] Figure 1 The main flowchart of a method for secure data access control for large AI models;

[0023] Figure 2 This is a flowchart of the data classification and hierarchical preprocessing sub-process in a method for data security control of large AI models;

[0024] Figure 3 This is a flowchart of the access request parsing and verification sub-process in a method for secure data access control of large AI models;

[0025] Figure 4 A flowchart of the dynamic policy matching sub-process in a method for secure data access control of large AI models;

[0026] Figure 5 A flowchart of the real-time risk assessment sub-process in a method for data security control of large AI models;

[0027] Figure 6 A flowchart illustrating the dynamic permission adjustment sub-process in a data security control method for large AI models;

[0028] Figure 7 A flowchart illustrating the access log recording and auditing sub-process in a method for secure data access control of large AI models;

[0029] Figure 8 This is a flowchart of the dynamic update mechanism in a method for controlling data security access to large AI models. Detailed Implementation

[0030] The technical solutions of the present invention will be described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, and not all embodiments.

[0031] Example 1

[0032] Please see Figure 1 A method for controlling data security access to large AI models includes the following steps: Step S1: Data classification and grading preprocessing: Multi-dimensional feature extraction is performed on the raw data to be accessed by the large AI model. Based on a preset classification and grading rule base, the data security level value is calculated using the analytic hierarchy process (AHP). The specific formula is as follows: Where n is the number of data feature dimensions, w i The weight of the i-th feature dimension (determined through expert scoring combined with historical access logs during training) Satisfy; C i To score the security level of the i-th feature dimension, a 5-level Likert scale is used for quantification (level 1 is the lowest security level, and level 5 is the highest security level). The scoring factors include concrete indicators such as the confidentiality of the data content and the scope of impact of data leakage. Step S2: Access request parsing and verification: The access request of the AI ​​big model is parsed using natural language processing technology to extract parameters such as the request subject identifier (MID), data object unique identifier (DID), operation type (OT), timestamp accurate to milliseconds (TS), and device fingerprint (DF). Based on the S value in step S1, an initial access token T containing timestamp verification and device fingerprint binding is generated. init The token is valid for T years. valid =15×60+Δt (for request response time compensation); Step S3: Dynamic policy matching: Construct a three-dimensional policy space consisting of a subject security attribute space, a data security level axis, and an operational risk coefficient axis, where: the subject security attributes include the subject reputation value obtained through BP neural network training (range 0-1), the subject permission level based on the RBAC model (levels 1-10), and the compliance rate of access history in the past 30 days; the operational risk coefficient is calculated using the fuzzy comprehensive evaluation method: O r =γ·O t +δ·O s (γ+δ=1,γ≥0.6) Among them, O t Risk score for operation type (read = 3 points, write = 5 points, delete = 7 points), O sThe operation frequency risk score is calculated based on the sliding window algorithm (linearly increasing when the number of operations per unit time exceeds a threshold); Step S4: Real-time risk assessment: Real-time collection of behavioral data through probes deployed on the data interaction link: including access frequency (times / minute), data query depth (number of nested SQL statements), data output volume (MB / time), and abnormal operation response time (ms); Environmental data: including network security level (classified into levels 1-5 according to NIST standards), device security status (integrity measurement value obtained through the TCG trusted computing module), and geofence coordinates (access area compliance based on GPS positioning); The improved DS evidence theory is used to fuse and calculate the real-time risk value R: R = α·R b +β·R e (α=0.6, β=0.4) where R b R represents the behavioral risk value (threshold range 0-100). e Environmental risk value (threshold range 0-100); Step S5: Dynamic adjustment of permissions: Establish a threshold R that includes permission contraction. th1 =80, Permission retention threshold R th2 =60, Permission Enhancement Threshold R th3 =40 Three-level dynamic adjustment mechanism: When R>R th1 When this occurs, permission contraction is triggered, generating a set of restricted permissions. At that time, R... th2 ≤R≤R th1 Maintain the current set of permissions, when R≤R th3 Temporary expansion is allowed; Step S6: Data interaction security control: Encryption processing: Data with security level S≥4 is asymmetric encrypted using the SM2 national cryptographic algorithm, and the key exchange process follows the IBC identifier cryptosystem; De-identification processing: For data fields containing personal information, dynamic masking algorithms (such as retaining the first 3 and last 4 digits and filling the middle with *), rule-based replacement methods (such as replacing the ID number with a virtual ID) or irreversible deletion methods are used; Format conversion: Adaptive conversion from data storage format to model input format is achieved through a pre-trained format mapping model (supporting JSON / Protobuf / Parquet format conversion). Transformation; Step S7: Access Log Recording and Auditing: Log Structure: Adopts a standardized log format that includes the digital certificate fingerprint of the access subject, the hash value of the data object, the semantic parsing tree of the operation, and the timestamp of the permission verification; Storage Mechanism: An audit log library is built based on the Hyperledger Fabric consortium blockchain. Each log block includes the hash value of the previous block, the Merkle tree root hash, and the timestamp; Verification System: Log file digests are generated using the SHA-256 algorithm, and log integrity verification is implemented using the bilinear pairing algorithm. The traceability marker includes the IP address, port number, and device certificate information of the entire link from the AI ​​large model to the data storage node.

[0033] Example 2

[0034] Please see Figure 1-8 A method for controlling data security access to large AI models includes the following steps: Step S1: Data classification and grading preprocessing: Multi-dimensional feature extraction is performed on the raw data to be accessed by the large AI model. Based on a preset classification and grading rule base, the data security level value is calculated using the analytic hierarchy process (AHP). The specific formula is as follows: Where n is the number of data feature dimensions, w i The weight of the i-th feature dimension (determined through expert scoring combined with historical access logs during training) Satisfy; C i To score the security level of the i-th feature dimension, a 5-level Likert scale is used for quantification (level 1 is the lowest security level, and level 5 is the highest security level). Scoring factors include concrete indicators such as the degree of confidentiality of the data content and the scope of impact of data leakage. The multi-dimensional feature extraction in step S1 specifically includes: Data sensitivity (weight 0.3): classified according to the GDPR data classification standard into public data (1 point), restricted data (3 points), and sensitive data (5 points); Data importance (weight 0.4): scored according to the degree of impact of data loss on business (no impact = 1 point, severe impact = 5 points); Data usage frequency (weight 0.2). : Based on the number of visits in the last 30 days (<10 visits = 1 point, >100 visits = 5 points); Data update cycle (weight 0.1): scored according to data update frequency (annual update = 1 point, real-time update = 5 points); Step S2: Access request parsing and verification: parsing the access request of the AI ​​big model through natural language processing technology, extracting parameters such as request subject identifier (MID), data object unique identifier (DID), operation type (OT), timestamp accurate to milliseconds (TS), and device fingerprint (DF), and generating an initial access token T containing timestamp verification and device fingerprint binding based on the S value of step S1. init The token is valid for T years. valid =15×60+Δt (where Δt is the request response time compensation value), the initial access token generation formula in step S2 is optimized as follows: in, The XOR operation is represented by H, which is a hash function containing a salt value dynamically obtained from the trusted security chip. Step S3: Dynamic policy matching: Construct a three-dimensional policy space consisting of a subject security attribute space, a data security level axis, and an operational risk coefficient axis. The subject security attributes include the subject reputation value (range 0-1) obtained through BP neural network training, the subject permission level (levels 1-10) based on the RBAC model, and the compliance rate of access history over the past 30 days. The operational risk coefficient is calculated using the fuzzy comprehensive evaluation method: O r =γ·Ot +δ·O s (γ+δ=1,γ≥0.6) Among them, O t Risk score for operation type (read = 3 points, write = 5 points, delete = 7 points), O s The operation frequency risk score is calculated based on the sliding window algorithm (linearly increasing when the number of operations per unit time exceeds a threshold); Step S4: Real-time risk assessment: Real-time collection of behavioral data through probes deployed on the data interaction link: including access frequency (times / minute), data query depth (number of nested SQL statements), data output volume (MB / time), and abnormal operation response time (ms); Environmental data: including network security level (classified into levels 1-5 according to NIST standards), device security status (integrity measurement value obtained through the TCG trusted computing module), and geofence coordinates (access area compliance based on GPS positioning); The improved DS evidence theory is used to fuse and calculate the real-time risk value R: R = α·R b +β·R e (α=0.6, β=0.4) where R b R represents the behavioral risk value (threshold range 0-100). e For the environmental risk value (threshold range 0-100), the environmental data in step S4 also includes: Time access window: defining weekdays 9:00-18:00 as the normal access period (risk coefficient 0.8), and other times as sensitive periods (risk coefficient 1.2); Equipment safety status: calculated by the Hamming distance between the PCR register value obtained by the TPM chip and the baseline value, and if the distance is >5, it is determined to be an unsafe device; Step S5: Dynamic permission adjustment: establishing a permission contraction threshold R th1 =80, Permission retention threshold R th2 =60, Permission Enhancement Threshold R th3 =40 Three-level dynamic adjustment mechanism: When R>R th1 When this occurs, permission contraction is triggered, generating a set of restricted permissions. At that time, R... th2 ≤R≤R th1 Maintain the current set of permissions, when R≤R th3 In this case, temporary extensions are allowed. The dynamic adjustment of permissions in step S5 also includes a temporary permission enhancement mechanism: when it is detected that the AI ​​large model is performing a preset emergency task (such as disaster prediction model training), the validity period of the permission is automatically extended to T. valid×2, and temporarily grant data write permissions. The preset emergency task triggering conditions in the temporary permission enhancement mechanism include: the task tag contains a preset set of emergency task keywords (identified through natural language processing technology); the security level of the task initiator is higher than the authorization threshold corresponding to the current data security level value; Step S6: Data interaction security control: Encryption processing: Data with security level S≥4 is asymmetricly encrypted using the SM2 national cryptographic algorithm, and the key exchange process follows the IBC identifier cryptosystem; De-identification processing: For data fields containing personal information, dynamic masking algorithms (such as retaining the first 3 and last 4 digits and filling the middle with *), rule-based replacement methods (such as replacing the ID card number with a virtual ID) or irreversible deletion methods are used; Format conversion: through pre-training The training format mapping model (supporting JSON / Protobuf / Parquet format conversion) enables adaptive conversion between data storage formats and model input formats. The dynamic update mechanism includes: a classification and grading rule base: monthly rule iteration based on the latest data leakage event database, updating the sliding window duration of frequency features; a policy matching algorithm: when the policy matching failure rate exceeds 15% for three consecutive days, a genetic algorithm is triggered to optimize the weight parameters of the three-dimensional policy space; a risk assessment model: employing online learning, it absorbs newly generated behavioral and environmental data in real time, triggering model fine-tuning every 100,000 data points. In step S6, data integrity verification uses the CRC-32 checksum algorithm, calculating the CRC checksum of the original data before data interaction. src After interaction, calculate the CRC checksum of the received data. dst If CRC src ≠CRC dstThis triggers a retransmission mechanism; Step S7: Access Log Recording and Auditing: Log Structure: A standardized log format is adopted, including the digital certificate fingerprint of the access subject, the hash value of the data object, the semantic parsing tree of the operation, and the timestamp of the permission verification; Storage Mechanism: An audit log library is built based on the Hyperledger Fabric consortium blockchain. Each log block includes the hash value of the previous block, the Merkle tree root hash, and the timestamp; Verification System: Log file digests are generated using the SHA-256 algorithm, and log integrity verification is implemented using the bilinear pairing algorithm. The source traceability marker includes the full-link IP address, port number, and device information from the AI ​​large model to the data storage node. In step S7, a multi-node consensus mechanism is set up for the blockchain audit log library. The consensus algorithm adopts an improved PBFT algorithm. Nodes participating in the consensus need to pass two-factor authentication to ensure the security of the log upload process. After step S7, abnormal access is handled. The abnormal behavior detection model is trained using a Long Short-Term Memory (LSTM) network. The detection indicators include: the daily access frequency exceeds three times the standard deviation of the historical average; the cross-security domain data transmission rate exceeds the bandwidth threshold; when an anomaly is detected, a real-time warning message containing the threat level (level 1-5), the anomaly type code, and the occurrence timestamp is sent to the security management platform via REST API.

[0035] Example 3

[0036] Data security access control for AI big data models in the financial industry

[0037] I. Application Scenarios

[0038] A commercial bank deployed a large AI model for credit risk control modeling, which required access to customer basic information, transaction records, credit reports and other data. The data contained sensitive personal information and had a significant impact on the business, so access security needed to be strictly controlled.

[0039] II. Specific Implementation Steps

[0040] (I) Data classification and grading preprocessing

[0041] Data Sensitivity: According to GDPR standards, customer ID numbers, account balances, etc. are classified as sensitive data (5 points), transaction types are classified as restricted data (3 points), and publicly available product information is classified as public data (1 point), with a weight of 0.3.

[0042] Data Importance: Loss of customer credit reports will cause the risk control model to fail and affect credit decisions, scoring 5 points; loss of ordinary transaction records has a smaller impact, scoring 3 points, with a weight of 0.4.

[0043] Data usage frequency: Customer credit data was accessed 120 times in the past 30 days, with a score of 5; historical transaction data was accessed 8 times, with a score of 1 and a weight of 0.2.

[0044] Data update cycle: Credit data is updated daily with a score of 5 points; historical transaction data is archived monthly with a score of 2 points and a weight of 0.1.

[0045] (II) Access Request Parsing and Verification

[0046] When the AI ​​model initiates a request to "obtain a customer's transaction history for the past six months":

[0047] Extract MID (Model Unique Identifier), DID (Customer Transaction Data ID), OT (Read Operation), TS (Precise Timestamp), and DF (Server Device Fingerprint).

[0048] Generate an initial access token with an expiration period of 15 x 60 seconds + 0.5 seconds response compensation, and bind it to a timestamp and device fingerprint.

[0049] (III) Dynamic Strategy Matching

[0050] Main security attributes: The reputation value of the model trained by the BP neural network is 0.85, the permission level based on RBAC is 7, and the access compliance rate in the past 30 days is 100%.

[0051] Operational risk coefficient: The reading operation risk score is 3 points. Based on the sliding window algorithm, the model accesses the device 2 times per minute (not exceeding the threshold). The operation frequency risk score is 1 point. The comprehensive risk coefficient is calculated using the fuzzy evaluation method.

[0052] (iv) Real-time risk assessment

[0053] Behavioral data: Access frequency 2 times / minute, SQL query depth 1 level, data output volume 5MB / time, exception response time 100ms.

[0054] Environmental data: Network security level NIST 4; the device's PCR register value detected by the TPM chip has a Hamming distance of 2 from the baseline value, indicating a secure state; access time is 10:00 AM on a weekday, which is within the normal access period.

[0055] (V) Dynamic Adjustment of Permissions

[0056] The real-time risk value is assessed at 55, which is in the 60-80 range. Maintain the current read permissions and do not trigger shrinkage or expansion.

[0057] (vi) Data Interaction Security Control

[0058] Encryption processing: Sensitive fields such as customer ID number (security level S=5) are encrypted using the SM2 national cryptographic algorithm.

[0059] Anonymization process: The transaction account number retains the first 3 and last 4 digits (e.g., 622****1234), with * filling in the middle.

[0060] Format conversion: Convert the Parquet format transaction data stored in the database into the JSON format required by the model.

[0061] Integrity verification: CRC-32 checksums are calculated before and after the interaction to ensure that the data has not been tampered with.

[0062] (vii) Access log recording and auditing

[0063] The logs contain the model's digital certificate fingerprint, data hash value, semantic parsing tree of the "query transaction history" operation, and permission verification timestamp.

[0064] Based on Hyperledger Fabric consortium blockchain storage, each block contains the hash of the previous block and the Merkle tree root, and log digests are generated through SHA-256, supporting full-link IP tracing.

[0065] (viii) Handling Abnormal Access

[0066] If the daily access frequency of the model exceeds three times the historical average standard deviation (e.g., if the daily average access is 50 times, the threshold is 50 + 3 × 10 = 80 times), the LSTM model will send a level 3 threat warning to the security platform via REST API after detection.

[0067] (ix) Dynamic update mechanism

[0068] The classification and grading rules are updated monthly based on the latest financial data breach incidents, such as adding a new sensitive category for "cross-border transaction data".

[0069] If the strategy matching failure rate exceeds 15% for three consecutive days (e.g., due to permission verification failure caused by adding new data types to the model), the genetic algorithm is triggered to optimize the weights of the three-dimensional strategy space.

[0070] The preferred embodiments of the present invention disclosed above are only for the purpose of illustrating the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation described herein. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can better understand and utilize the present invention.

Claims

1. An AI large model access data security control method, characterized in that: Includes the following steps: Step S1: Data classification and grading preprocessing, extracting multi-dimensional features, and calculating the security level value using the analytic hierarchy process based on the classification and grading rule base; Step S2: Parse the access request to extract the MID and DID parameters, generate an initial access token with timestamp verification and device fingerprint binding based on the security level value, and set the expiration time; Step S3: Construct a three-dimensional policy space to calculate the operational risk coefficient. The three-dimensional policy space consists of a subject security attribute space, a data security level axis, and an operational risk coefficient axis. Subject security attributes include the subject reputation value obtained through BP neural network training, the subject permission level (levels 1-10) based on the RBAC model, and the compliance rate of access history over the past 30 days. The operational risk coefficient is calculated using the fuzzy comprehensive evaluation method. ,in, Risk scoring for operation type The operation frequency risk score is calculated based on the sliding window algorithm; Step S4: Collect behavioral and environmental data to calculate real-time risk values; Step S5: Establish a three-level dynamic adjustment mechanism, which triggers permission contraction, maintenance, or temporary expansion based on the comparison between the real-time risk value and the thresholds for permission contraction, maintenance, and enhancement. Step S6: Encrypt data with a security level ≥4 using the SM2 national cryptographic algorithm, desensitize data containing personal information, and convert the format using a format mapping model; Step S7: Adopt a standardized log format, build an audit log library based on the Hyperledger Fabric consortium blockchain, and implement log integrity verification and traceability through SHA-256 and bilinear pairing algorithms.

2. The AI ​​large-scale model access data security control method according to claim 1, characterized in that: The multi-dimensional feature extraction in step S1 specifically includes: Data sensitivity: According to the GDPR data classification standard, data is divided into public data, restricted data, and sensitive data; Data Importance: Score based on the degree of impact of data loss on business operations; Data usage frequency: quantified based on the number of visits in the last 30 days; Data update cycle: scored according to the frequency of data updates.

3. The method for secure data access control of large AI models according to claim 1, characterized in that: The formula for generating the initial access token in step S2 is optimized as follows: in, This represents an XOR operation, where H is a hash function containing a salt value, which is dynamically obtained from a trusted security chip. SHA-256 The hash function generates a hash value, which is then processed by the outer H hash to ultimately generate the initial access token. MID, DID, OT, TS, DF, and S represent module identifier, device identifier, operation type, timestamp, data feature representation, and salt value, respectively.

4. The AI ​​large model access data security control method according to claim 1, characterized in that: The environmental data in step S4 also includes: Time access window: Define weekdays from 9:00 to 18:00 as the normal access period, and other times as sensitive periods; Equipment safety status: The Hamming distance between the PCR register value obtained by the TPM chip and the baseline value is calculated. If the distance is greater than 5, the device is considered unsafe.

5. The method for secure data access control of large AI models according to claim 1, characterized in that: The dynamic permission adjustment in step S5 also includes a temporary permission enhancement mechanism: when it is detected that the AI ​​large model is executing a preset emergency task, the validity period of the permission is automatically extended to... And temporarily grant data write permissions. Indicates the token's validity period.

6. The AI ​​large model access data security control method according to claim 1, characterized in that: In step S6, the data integrity verification uses the CRC-32 checksum algorithm, and the original data checksum is calculated before data interaction. Calculate the verification value of the received data after interaction. ,like This will trigger the retransmission mechanism.

7. The AI ​​large model access data security control method according to claim 1, characterized in that: The audit log library is equipped with a multi-node consensus mechanism. The consensus mechanism adopts an improved PBFT algorithm. Nodes participating in the consensus must pass two-factor authentication to ensure the security of the log uploading process.

8. The method for secure data access control of large AI models according to claim 1, characterized in that: Following step S7, abnormal access processing is performed. The abnormal behavior detection model is trained using a Long Short-Term Memory (LSTM) network, and the detection metrics include: The daily visit frequency exceeded three standard deviations of the historical average. Cross-security domain data transmission rate exceeds bandwidth threshold; When an anomaly is detected, a real-time alert is sent to the security management platform via REST API, containing the threat level, anomaly type code, and occurrence timestamp.

9. The method for secure data access control of large AI models according to claim 1, characterized in that: The dynamic update mechanism includes: Classification and grading rule base: The rules are iterated monthly based on the latest data breach event database, and the sliding window duration of frequency features is updated. Strategy matching algorithm: When the strategy matching failure rate exceeds 15% for 3 consecutive days, the genetic algorithm is triggered to optimize the weight parameters of the three-dimensional strategy space. Risk assessment model: It adopts an online learning approach to absorb newly generated behavioral and environmental data in real time, and the model is fine-tuned every 100,000 data points.

10. The method for secure data access control of large AI models according to claim 5, characterized in that: The preset emergency task triggering conditions in the temporary permission enhancement mechanism include: Task tags contain a preset set of emergency task keywords; The security level of the task initiator is higher than the authorization threshold corresponding to the current data security level.

Images