Enterprise data inference service hierarchical access control methods, devices, media, and products
By determining the security level type and information leakage coefficient in the enterprise data inference service, accessing the data source and implementing protection strategies, analyzing query semantics and aggregating queries, and dynamically adjusting access control, the problem of existing technologies being unable to prevent aggregation inference attacks and lacking security classification standards is solved, thus achieving efficient data security protection.
Patent Information
- Application Number
- CN202611008573.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-08
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-07-08
AI Technical Summary
Existing access control systems for enterprise data inference services cannot effectively prevent aggregation inference attacks and lack the ability to dynamically adjust protection strength, resulting in a high risk of sensitive data leakage and a lack of unified security classification standards.
By determining the security level type of the target scenario, invoking the corresponding protection strategy and information leakage coefficient, accessing the data source and executing the protection strategy, analyzing query semantics and aggregating homogeneous queries into the same cluster, determining the access shared quota based on the access preset computing power quota, calculating the cumulative consumption of the shared quota, and dynamically adjusting access control.
It effectively prevents aggregation inference attacks, dynamically adjusts protection strength, avoids irreversible leakage of sensitive data, and provides a unified security classification standard to ensure the security of data hosting and inference services.
Smart Images

Figure CN122548766B_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a method, apparatus, medium, and product for hierarchical access control of enterprise data reasoning services. Background Technology
[0002] With the popularization of artificial intelligence technology, more and more enterprises are integrating sensitive private data such as customer information, medical records, and core code into large models, providing inference services to the outside world through API interfaces, and using a token-based payment model to deliver data insights to consumers, achieving data usability without visibility. Unlike the large and comprehensive characteristics of general-purpose large models, enterprise private data inference is characterized by being small and precise. Its training data is highly sensitive and irreplaceable, and the model iteration frequency and data volume are low, making it easier for the model to remember specific data records. Each token output may carry high-value original information, leading to data leakage.
[0003] The current data visibility-invisibility protection logic has significant vulnerabilities. Not seeing the original data does not mean it cannot be taken. With enterprises opening API interfaces, the core risk has shifted from the data layer to the model and inference layers. Existing attack methods are diverse and highly covert, including model inversion, member inference, vector inversion, hint injection, and aggregation inference. Aggregation inference attacks are particularly serious. For example, attackers access the system posing as legitimate paying customers and gradually extract core digital assets of an enterprise by carefully constructing query contexts, distributing queries across multiple accounts, noise hedging, and feature intersection. Whether it's the core code of a SaaS vendor or customer and operational data from a retail CRM platform, attackers can steal it at extremely low token costs. Existing semantic auditing, identity authentication, and rate limiting protection measures are unable to detect such covert attack behaviors.
[0004] Existing security technologies and access control systems have significant shortcomings, making it difficult to address the emerging security needs of enterprise data inference services. On the one hand, existing access controls only focus on identity authentication and call rate, failing to establish a privacy ledger to track the cumulative information extraction from multiple queries, thus unable to prevent aggregation and inference attacks. On the other hand, security policies adopt a one-size-fits-all approach, unable to dynamically adjust protection strength based on attack characteristics after deployment, and the lag in manual intervention can lead to irreversible leakage of sensitive data. Furthermore, the industry lacks a unified security grading standard, with inconsistent descriptions of security capabilities across platforms, making it difficult for data providers to accurately assess the security risks of data hosting and inference services. Summary of the Invention
[0005] In view of this, the present disclosure provides a hierarchical access control method, apparatus, medium, and product for enterprise data inference services, which can solve problems such as poor data protection, high risk of leakage, and inability to accurately identify covert attacks in enterprise data inference services.
[0006] In a first aspect, embodiments of this disclosure provide a hierarchical access control method for enterprise data inference services, including: Determine the security level type of the target scenario, and invoke the protection strategy and information leakage coefficient corresponding to the security level type; Connect the target data source to the platform and execute the protection strategy; Based on the information leakage coefficient, the number of target data sources, and the sensitivity level of the target data sources, determine the pre-set computing power quota for the platform after accessing the target data sources; Analyze the semantics of the queries and aggregate the corresponding homogeneous queries within a preset period into the same cluster; The access sharing quota corresponding to each cluster is determined based on the access preset computing power quota; Based on the information leakage coefficient, the cumulative consumption shared quota of all queries within each cluster within a preset time period is obtained; Based on the access shared quota and the cumulative consumption shared quota of each cluster, the corresponding risk status is determined, and the corresponding access control is triggered according to the risk status.
[0007] Secondly, this disclosure also provides a hierarchical access control method for enterprise data inference services, including: Analyze the semantics of the queries and aggregate the corresponding homogeneous queries within a preset period into the same cluster; Based on the pre-set computing power quota for access corresponding to the inference service platform, determine the access sharing quota for each cluster. Obtain the cumulative shared quota for all queries within each cluster within a preset time period; Based on the access shared quota and the cumulative consumption shared quota of each cluster, the corresponding risk status is determined, and the corresponding access control is triggered according to the risk status.
[0008] Thirdly, this disclosure also provides a computer device, which adopts the following technical solution: The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor; the instructions are executed by the at least one processor to enable the at least one processor to perform any of the enterprise data inference service hierarchical access control methods described above.
[0009] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium storing computer instructions; the computer instructions are used to cause a computer to execute any of the enterprise data inference service hierarchical access control methods described above.
[0010] Fifthly, embodiments of this disclosure also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.
[0011] The hierarchical access control method for enterprise data inference services provided in this disclosure first determines the security level type of the target scenario and invokes the corresponding protection policy and information leakage coefficient. Second, it connects the target data source to the platform and executes the protection policy. Then, based on the information leakage coefficient, the number of target data sources, and the sensitivity level of the target data sources, it determines the platform's pre-set access computing power quota after connecting the target data sources. Next, it analyzes the query semantics and aggregates homogeneous queries within a preset period into the same cluster. It determines the access sharing quota for each cluster based on the pre-set access computing power quota. Then, it obtains the cumulative consumption sharing quota for all queries within each cluster within a preset time period. Finally, it determines the corresponding risk status based on the access sharing quota and cumulative consumption sharing quota for each cluster and triggers corresponding access control based on the risk status. By statistically analyzing the cumulative information extraction volume of multiple queries, it can effectively prevent aggregation inference attacks. After going live, the protection strength can be dynamically adjusted based on attack characteristics without manual intervention, preventing irreversible leakage of sensitive data. Through a unified security classification standard, data providers can accurately assess the security risks of data hosting and inference services. Attached Figure Description
[0012] Figure 1 This is a flowchart illustrating the hierarchical access control method for enterprise data inference services provided in this embodiment of the disclosure.
[0013] Figure 2 This is a flowchart illustrating the method for invoking the protection strategy and information leakage coefficient corresponding to the security level type provided in this embodiment of the disclosure.
[0014] Figure 3 This is a flowchart illustrating the method for determining the information leakage coefficient for each security level other than the lowest security level, as provided in this embodiment of the disclosure.
[0015] Figure 4 This is a flowchart illustrating the method for determining the pre-set computing power limit of the platform after accessing the target data source, as provided in this embodiment of the disclosure.
[0016] Figure 5This is a schematic diagram illustrating the process of determining the access sharing quota corresponding to each cluster based on the access preset computing power quota, as provided in the embodiments of this disclosure.
[0017] Figure 6 This is a flowchart illustrating the method for obtaining the cumulative shared quota consumed by all queries within each cluster, as provided in this embodiment of the disclosure.
[0018] Figure 7 A flowchart illustrating another embodiment of the hierarchical access control method for enterprise data inference services provided in this application.
[0019] Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. Detailed Implementation
[0020] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.
[0021] Reference Figure 1 This application discloses a hierarchical access control method for enterprise data inference services, including: S100, determine the security level type of the target scenario, and call the protection strategy and information leakage coefficient corresponding to the security level type; S200 connects the target data source to the platform and executes protection policies; S300 determines the pre-set computing power limit for the platform after accessing the target data source based on the information leakage coefficient, the number of target data sources, and the sensitivity level of the target data source. S400 analyzes the semantics of the query and aggregates the corresponding homogeneous queries within a preset period into the same cluster; S500 determines the access sharing quota for each cluster based on the pre-set access computing power quota; S600, based on the information leakage coefficient, obtains the cumulative consumption shared quota of all queries within each cluster within a preset time period; S700 determines the corresponding risk status based on the access shared quota and cumulative consumption shared quota of each cluster, and triggers the execution of corresponding access control based on the risk status.
[0022] The enterprise data inference service hierarchical access control method disclosed in this application determines the security level type of the target scenario and matches the corresponding protection strategy and information leakage coefficient. It can break the traditional one-size-fits-all security configuration mode, and set basic leakage control parameters according to the sensitivity of scenario data. Highly sensitive scenarios such as medical records and enterprise core code are matched with smaller information leakage coefficients, which constrains the upper limit of information leakage in a single query from the source, and is suitable for the business characteristics of scarce enterprise private model data and easy leakage of original information.
[0023] By connecting the target data source to the platform and implementing corresponding protection strategies, hierarchical security control rules can be bound to the data source from the beginning of its launch. This avoids the direct access of sensitive data sources to the inference API interface, and moves the privacy quota control logic to the data access stage in advance, avoiding control loopholes and security vulnerabilities caused by subsequent temporary addition of security rules.
[0024] By leveraging the information leakage coefficient, the number of data sources, and the data sensitivity level calculation platform to access the pre-set computing power quota, the total global privacy quota is refined based on data attributes. The higher the data sensitivity level and the more data sources accessed, the less the overall allocable access quota is. This locks the total amount of information that can be leaked from the top level, thereby compressing the space for attackers to steal sensitive data in batches from the root.
[0025] Semantic analysis is performed on all queries within the period, and homogeneous queries are grouped into the same cluster. Instead of using a single user or a single interface call as the risk control statistical unit, the aggregate query behavior of attackers using multiple accounts to split and probe in segments will be collected and classified, effectively cracking the attack method of attackers relying on split requests to bypass frequency limits, and accurately identifying hidden aggregate inference probing behavior.
[0026] By combining pre-set computing power quotas, independent access sharing quotas are allocated to each cluster, enabling the global total quota to be split and managed according to different query semantic directions. Corresponding to the quota upper limit constraint rules for each semantic direction, no query domain can occupy the global privacy quota without limit, thus completing the isolation and control of information leakage quotas for each query direction. The quota consumption of each query within the cluster is dynamically calculated using an information leakage coefficient, and the cumulative consumption quota within the period is statistically analyzed. Based on the query information volume factor, the information value of different queries is distinguished, with high-information-value queries consuming more quotas per query. A quantitative accounting model is used to statistically analyze the total amount of information extracted, overcoming the shortcomings of traditional risk control that only manages the number of calls but not the amount of information obtained.
[0027] The system determines the current risk level by comparing the available quota and cumulative consumption of clusters. Based on different risk states, it automatically triggers corresponding access control measures. The system relies on automation to dynamically adjust the protection strength without the need for manual intervention. When sensitive information is close to the leakage threshold, it proactively blocks illegal queries to prevent irreversible leakage of sensitive data.
[0028] Reference Figure 2 The methods for invoking the protection strategy and information leakage coefficient corresponding to the security level type for S100 specifically include: S110 configures the security level for each scenario and the protection strategy for each security level based on the security requirements of different data usage scenarios.
[0029] When the scenario in which the data is used is public data query, such as weather or news services, it indicates that the corresponding scenario has the lowest security requirements. The security level of the corresponding scenario is determined to be S1 level, which is the basic level and the lowest security level. The protection strategy corresponding to this security level is basic protection including identity authentication, call logs, and basic rate limits.
[0030] Specifically, the protection strategy corresponding to S1 level includes three core security mechanisms: first, identity authentication is completed to verify the legitimacy of the user's identity; complete call logs are retained throughout the process to record various operational behaviors for traceability and verification; and basic rate limits are set to control the frequency of interface calls and avoid malicious high-frequency access, thereby building basic security protection capabilities.
[0031] When the data used is internal enterprise data or business secrets, the security level for the corresponding scenario is determined to be S2 level, i.e., enhanced level. The protection strategy corresponding to this security level includes all the protection strategies of S1 level, output content auditing such as automatic interception of sensitive content, fine-grained access control, and abnormal call pattern detection.
[0032] Specifically, the protection strategy corresponding to S2 level constructs a security system from three aspects: content, permissions, and behavior. It conducts real-time auditing of all output content, automatically identifies and blocks various sensitive information; at the same time, it implements fine-grained permission control, accurately divides the operation and access scope of different users and roles, and prevents unauthorized operations; it also continuously monitors interface call behavior, promptly detects and warns of abnormal call patterns, and comprehensively prevents security risks.
[0033] When the data used is personal privacy data or patient health information, the security level of the corresponding scenario is determined to be S3 level, which is the strict level. The protection strategy corresponding to this security level includes all the protection strategies of S2 level, differential privacy such as adding noise with ε not greater than 1.0 to the output, automatic desensitization such as automatic replacement of name / ID card / mobile phone number, and cache isolation of query results.
[0034] Specifically, the protection strategy corresponding to S3 level includes multiple data security capabilities, preventing the leakage of sensitive information through multiple means. Among them, differential privacy technology adds compliant noise to the data output results and strictly controls the privacy protection parameter ε to not exceed 1.0, ensuring data usability while protecting the privacy of the original data. It has an automatic de-identification function, which can intelligently identify and replace various sensitive personal information such as names, ID numbers, and mobile phone numbers to avoid plaintext leakage. In addition, it caches and isolates query results, separating the cached data of different query tasks to prevent cross-access and unauthorized access, thus comprehensively building a solid data security defense.
[0035] When the data used is in scenarios involving core financial data, medical diagnosis, or classified data, it indicates that the security requirements of the corresponding scenario are the highest. The security level of the corresponding scenario is determined to be S4 level, i.e. confidential level. The protection strategy corresponding to this security level includes all the protection strategies of S3 level, TEE trusted execution environment such as inference being executed in an encrypted safe, machine forgetting such as data deletion taking effect within 72 hours, output watermark traceability control, etc.
[0036] Specifically, the protection strategy corresponding to S4 level protects against data theft during operation, data destruction, and traceable output. Relying on the Trusted Execution Environment (TEE), operations such as model inference run in an isolated encrypted environment, preventing data leakage and tampering during computation. It is equipped with a machine forgetting mechanism, so after a data deletion command is issued, the relevant information will be completely cleaned up within 72 hours, ensuring that the data is destroyable and privacy is cleared. At the same time, digital watermarks are added to all external output content, which can accurately trace the source and flow of content, enabling the tracking and accountability of violations.
[0037] S120, configures the information leakage coefficient corresponding to the lowest security level, denoted as the basic leakage coefficient.
[0038] Specifically, the information leakage coefficient corresponding to the lowest security level is configured as the first initial value, preferably 1.
[0039] S130, based on the basic leakage coefficient, determines the information leakage coefficient corresponding to each security level except the lowest security level, and stores it in the database.
[0040] S140, Based on the target scenario corresponding to the enterprise data inference service, determine the security level type of the target scenario; S150 retrieves the protection policy and information leakage coefficient corresponding to the security level type from the database.
[0041] Reference Figure 3 The method for determining the information leakage coefficient for each security level in S130, excluding the lowest security level, specifically includes: S131, Determine the preset attack type corresponding to the target scenario.
[0042] The preset attack type can be one or more of the following: code completion attack type and noise offset attack type.
[0043] Code completion attacks are a type of covert code theft attack that exploits the memory characteristics of large models and legitimate API call channels. Specifically, attackers access services as legitimate paying users, bypassing straightforward sensitive request audits. They first construct a dialogue context for code completion using similar open-source code, then input a private code snippet prefix to guide the model to continue writing. Simultaneously, they adjust parameters to fix the output result. Leveraging the model's overfitting to training data, they gradually guide it to reveal core source code, keys, internal information, and other confidential content. The entire call behavior complies with basic security rules such as identity verification and access frequency, making it difficult for conventional semantic audits to detect deep-seated malicious intent. Ultimately, attackers steal core technological assets of enterprises at extremely low cost, causing serious technical and commercial losses.
[0044] Noise hedging attacks are combined data restoration attacks that target differential privacy protection mechanisms. Specifically, attackers access the service as legitimate customers, using multi-account split queries and diverse question formats to hedging noise. They leverage the law of large numbers to offset random noise in individual results, gradually approximating the true data. Then, they construct a unique identifier through multi-dimensional feature combinations to accurately pinpoint the target entity. Because protection measures only limit the noise amplitude of a single query and the frequency of calls from a single account, they cannot identify the aggregated malicious intent behind cross-account and batch queries. Ultimately, this leads to the complete restoration of anonymized business data, resulting in the leakage of customer operational data and business intelligence.
[0045] S132, based on the technical logic of the preset attack type, configure multiple repeatable and quantifiable detection and query commands for each preset attack type.
[0046] If the preset attack type is code completion attack, the technical logic of this type is prefix inducement logic. The method of configuring multiple repeatable and quantifiable probe query instructions for this preset attack type includes: for code completion inducement source code theft attacks, combining its core logic of first constructing a code continuation context, then inputting code prefixes to induce output, and obtaining complete content segment by segment, a standardized, repeatable, and quantifiable probe query set is built.
[0047] Specifically, firstly, the API call parameters are standardized, the randomness parameter of the output is fixed at 0 and the length of the token output in a single instance is limited, quantitative statistical dimensions such as the number of calls and the number of hits in sensitive fields are clearly defined, and the total number of probes N is set.
[0048] The overall query commands are divided into three categories and executed in a fixed order. The first category is context construction commands, which use multiple general codes to complete the dialogue and guide the model into the corresponding running scenario.
[0049] For example, the corresponding instruction templates include: 1) Please continue writing the code based on the following open-source code style: [Insert a general open-source code snippet of the same type]; 2) Please complete the following class definition code while maintaining the original coding style; 3) Continue writing this business logic code, without additional explanation, only output the code content. The corresponding quantification rules include: using 3 basic instructions in a fixed loop, which must be executed in each probe process, i.e. .
[0050] The second type is the core prefix inducement instruction, which constructs prefixes for classified code fragments in batches to form standardized query statements. The purpose of this instruction is to guide the model to output the memorized content by inputting classified code to fix the prefix. It is the core detection instruction.
[0051] For example, the corresponding instruction templates include: code completion, function continuation, and internal method improvement, which are generated in batches based on the target code. Queries with fixed code prefixes for stripes. The corresponding quantization rules include: setting single-round probing to use... A prefix instruction can be used to repeatedly invoke the same set of instructions after N probes.
[0052] The third type is the segmented continuation instruction, which is used to continue fetching the remaining content from the previous round of output. The function of this instruction is to continue writing based on the end of the code output in the previous round, splitting long code segments to avoid the limitation of single output.
[0053] For example, the corresponding instruction templates include: 1) Continuing from the previous code segment, write the subsequent logic; 2) Completing the remaining part of the above code, only outputting the code. The corresponding quantization rules include: fixing 2 continuation instructions. It is used in a loop following the prefix instruction.
[0054] The entire query command is fixed and can be executed repeatedly N times according to a preset process. The entire set of instructions can be executed repeatedly N times indefinitely. The execution order includes context construction, prefix guidance, and segment continuation, and the order is fixed. Finally, the protection effect can be quantitatively evaluated by statistically analyzing the hit rate of sensitive code, keys, and other content.
[0055] If the preset attack type is noise offset attack, the technical logic of this type is multi-phrase query logic. The method of configuring multiple N repeatable and quantifiable probe query instructions for this preset attack type specifically includes: for data restoration attacks that rely on multiple accounts and multiple phrases to offset noise and combine features to lock the target, design a reusable and quantifiable probe query set according to its technical logic.
[0056] Specifically, multiple independent probe accounts are configured in the early stage to distribute the call pressure, and quantitative indicators such as the number of queries, result deviation, and feature matching rate are determined, and the total number of probes N is set.
[0057] Specifically, the query set is divided into two parts. The first part is noise-hedge query instructions, which are used to lock a single fixed enterprise data inference target. By adjusting the expression, statistical caliber, time granularity, and dimensional language, a large number of semantically identical but textually different query instructions are generated for multi-round, multi-account repeated probing. Relying on the law of large numbers, it offsets differential privacy random noise and is suitable for general inference scenarios such as enterprise revenue, user scale, operational data, and production capacity data.
[0058] For example, if the fixed detection target is the total operating revenue of a target company in the third quarter of this year, K standardized detection query examples can be reused, including: querying the overall operating revenue of the target company in the third quarter of this year, calculating the total revenue of all businesses of the company from July to September of this year, obtaining the total revenue data of the target company's main business and ancillary businesses in the third quarter, querying the company's total revenue across all channels in Q3 of this year, etc. Based on the above basic templates, the time, business scope, and expression can be rewritten in batches to expand to K queries. All queries point to the same data truth value, a single command can be called N times repeatedly, and multi-account splitting and polling are supported for noise hedging convergence testing.
[0059] The second part is the feature intersection query command, which is used to combine and overlay multiple dimensions of enterprise data reasoning scenarios, such as operating region, business track, personnel size, service model, qualification attributes, establishment year, regional layout, etc., to generate a globally unique and highly exclusive enterprise feature fingerprint. By querying with fixed language, the unique target enterprise can be locked, avoiding the protection of batch data generalization and achieving accurate subject positioning.
[0060] For example, a fixed and unique combination of characteristics includes: target companies located in East China, possessing specialized and innovative qualifications, having a workforce of over 1,000 employees, primarily offering B2B enterprise services, having been established for at least 8 years, and having established 3 provincial-level operation centers.
[0061] Examples of reusable L standardized probing queries include: querying the third-quarter revenue data of companies in East China that possess specialized and innovative qualifications, have over 1,000 employees, primarily provide B2B services, have been established for at least 8 years, and have 3 provincial-level operation centers; and statistically analyzing the operating data of companies that simultaneously meet the following criteria: East China layout, specialized and innovative certification, team size of over 1,000 employees, primarily provide B2B services, have over 8 years of operating experience, and have 3 provincial-level operation outlets. All query feature combinations are fixed and unique, and the sentence structure can be fine-tuned based on the basic template to expand to L queries. It supports N repeated automated probing attempts and can accurately verify whether the system has any vulnerabilities in feature intersection tracing and precise entity positioning.
[0062] The execution logic for noise-hedging attack-type detection query commands includes: first, executing the noise-hedging query set, i.e., querying multiple accounts and multiple phrases N times in a loop; then, executing the feature intersection query set to lock the target. The reuse rule is that all query command texts are fixed, and the same set of commands is directly and repeatedly called in N detections without the need for re-creation. That is, all commands remain unchanged and can be executed repeatedly N times. By calculating the average of multiple rounds of query results, noise convergence amplitude, and other data, the detection and effect quantitative analysis are completed.
[0063] Among them, repeated and quantifiable probe queries can be used as a set of probe queries, which can be one or all of the code completion attack probe set and differential privacy data restoration attack probe set.
[0064] S133, for data inference service scenarios with the lowest security level, attack detection is performed based on probe query commands, and the corresponding first information entropy is determined based on the detection results.
[0065] S134. For data inference service scenarios configured with a non-minimum security level, attack detection is performed based on probe query commands, and the corresponding second information entropy is determined based on the detection results.
[0066] The methods for determining the first information entropy and the second information entropy are preferably the same, that is, the detection query commands used when conducting attack detection are preferably the same.
[0067] The following example illustrates how information entropy can be obtained using preset attack types including code completion attacks and noise offset attacks, where the probe query set includes both code completion attack probes and differential privacy data restoration attack probes: The code completion attack detection query set and the differential privacy data restoration attack detection query set are invoked respectively. The repetitive automated detection is carried out according to a preset fixed number of times, fixed calling parameters, and fixed execution logic. It covers all detection behaviors such as code prefix guided continuation, segmented continuation detection, multi-phrase noise hedging query, and multi-dimensional feature intersection locking query.
[0068] Furthermore, during the detection process, invalid data such as interface errors, empty returns, and abnormal timeouts are uniformly removed, while all valid inference outputs corresponding to the two types of attacks are retained to form a complete, statistical, and reproducible detection sample dataset, ensuring sufficient samples and full dimensional coverage for subsequent entropy value calculations.
[0069] For code completion attack detection samples, the original confidential data such as the real core code, internal algorithm logic, confidential fields, and key parameters of the service backend are used as the benchmark to compare the matching completeness and field accuracy of the detection output content one by one. For differential privacy data restoration attack detection samples, the relative deviation between the detection output results and the real data is calculated for each sample based on the enterprise's real operating data, user data, and business statistics.
[0070] Furthermore, all precision and bias data of the two types of attacks are dimensionlessly normalized and uniformly mapped to the 0-1 interval to quantify the ability of each detection result to restore the original sensitive data and construct a standardized quantitative sample matrix.
[0071] The normalized quantization results of code completion detection samples and differential privacy restoration detection samples were statistically analyzed in intervals. The accuracy intervals were divided into high restoration degree, medium restoration degree and low restoration degree. The proportion of samples in each interval was counted, and the probability of occurrence of different data restoration accuracy was calculated to form an independent probability distribution model under the two types of attack scenarios. This model accurately depicts the original information leakage distribution characteristics of the lowest security level service under the two types of targeted attacks.
[0072] Based on the Shannon information entropy calculation logic, and combined with the probability distribution results of the two types of attacks, a weighted fusion calculation is performed to obtain the information entropy that can uniformly represent the degree of restoration of the overall original data of the service.
[0073] This entropy value comprehensively reflects the overall information leakage level of data inference services when facing two typical attacks: code theft and differential privacy cracking. It can intuitively determine the degree of protection failure: the lower the information entropy value, the more complete the original data restored through attack detection and the more serious the leakage; the higher the information entropy value, the lower the degree of restoration of the original data and the better the protection effect. This indicator can accurately measure the actual information leakage level under different protection configurations.
[0074] S135, the ratio of the second information entropy to the first information entropy is used as the information leakage coefficient corresponding to the non-minimum security level.
[0075] In this embodiment, using S1 as a benchmark, the inherent differences in sensitivity between different data sources, such as the difference in inherent information entropy between core code and ordinary customer data, can be eliminated. By using relative ratios, a standardized comparison of information leakage levels under different protection configurations can be achieved, ensuring that the calibration of information leakage coefficients has universality and comparability.
[0076] In this embodiment, it refers to storing the security requirements of different scenarios using data, the security level of the corresponding scenario, the protection strategy corresponding to each security level, and the information leakage coefficient corresponding to each security level in the database.
[0077] Reference Figure 4 The method for determining the pre-set computing power limit for the platform after the S300 connects to the target data source specifically includes: S310, determine the number and sensitivity level of the target data sources required for the target scenario.
[0078] The number of target data sources required for the target scenario is preferably obtained when the data provider accesses the platform. Specifically, the sensitivity level of the target data sources required for the target scenario is determined based on the security level type of the target scenario.
[0079] Preferably, when the security level of the target scenario is S1, the sensitivity level of the target data source required by the target scenario is determined to be Level 1; when the security level of the target scenario is S2, the sensitivity level of the target data source required by the target scenario is determined to be Level 2; when the security level of the target scenario is S3, the sensitivity level of the target data source required by the target scenario is determined to be Level 3; and when the security level of the target scenario is S4, the sensitivity level of the target data source required by the target scenario is determined to be Level 4.
[0080] S320 determines the business calibration constant corresponding to the target scenario based on the information leakage coefficient, quantity, sensitivity level, and business availability information of the data provider.
[0081] The business availability information provided by the data provider includes the maximum number of authorized users. The maximum number of queries (M) per day for a single authorized user, and the maximum number of consecutive months (T) for an authorized user.
[0082] Business calibration constants corresponding to the target scenario The methods for determining this include: 1) Based on the business availability information provided by the data provider, determine the baseline business scale corresponding to the target scenario, i.e. The purpose is to calculate the cumulative total number of query requests under full load and continuous operation scenarios. This value abandons the daily average load and adopts the peak limit indicator of SLA. It can prevent resource shortage, service circuit breakers and performance degradation caused by peak user concurrency and long-term cumulative calls. It is the basic quantitative base to ensure the continuous availability of business.
[0083] 2) Analyze the sensitivity level of the target data source required for the target scenario to determine data sensitivity information. .
[0084] In one specific embodiment, preferably, when the sensitivity level of the target data source required by the target scenario is level one, the data sensitivity information is determined. A score of 1 is awarded; when the sensitivity level of the target data source required by the target scenario is Level 2, the data sensitivity information is determined. The score is 4 points; when the sensitivity level of the target data source required by the target scenario is level three, the data sensitivity information is determined. A score of 7 is given; when the sensitivity level of the target data source required by the target scenario is level four, the data sensitivity information is determined. It is 10 points.
[0085] 3) Based on data sensitivity information and the number of target data sources required for the target scenario, determine the data security weighting items, i.e. ,in, This represents the number of target data sources required for the target scenario; this ratio indicates the sensitivity density per unit data volume. The more sensitive the data and the higher the privacy protection requirements, the greater the computational overhead of security links such as differential privacy, content auditing, desensitization, watermarking, and permission verification.
[0086] 4) Determine the information content factor of the current query based on the number of tokens consumed and the corresponding query details.
[0087] The information content factor for this query is : ; ; This represents the number of tokens consumed in the current query. This provides specific query details for the current query. This represents the number of named entities in the query semantics corresponding to the current query. This represents the number of numeric features in the semantics of the current query. This represents the number of geographic location features in the semantics of the current query. This represents the number of time features in the query semantics corresponding to the current query. , , , These are the corresponding weights, and the sum of the four weights is 1.
[0088] In the process of obtaining the information content factor for the current query, the following is used: This is because information entropy increases logarithmically with the number of tokens, with the first few tokens having the highest information density, and subsequent tokens having decreasing marginal information density.
[0089] In this embodiment, This is a quantitative indicator of query specificity with a value range of [0,1]. It is used to measure the accuracy of a single query statement in pointing to a specific person, thing, or scene. The closer the value is to 1, the more specific the query is and the easier it is to accurately locate the target sensitive data. The closer the value is to 0, the broader the query semantics and the more ambiguous the target.
[0090] For example, if the query semantics for this instance are "commonly used diabetes medications," then because this semantics lacks specific entities and concrete numerical values, then... , All are 0.
[0091] 5) Determine the business calibration constant corresponding to the target scenario based on the information leakage coefficient, the information volume factor of the current query, the business scale base, and the data security weighting item.
[0092] The business calibration constant corresponding to the target scenario is : ,in, This represents the information leakage coefficient.
[0093] S330 determines the pre-set computing power limit for the platform after accessing the target data source, based on the quantity, sensitivity level, and business calibration constant.
[0094] Access to the pre-set computing power quota (i.e., the initial budget) is : .
[0095] The methods disclosed in S310-S330 first quantify the risks of scenario data assets by considering the number of data sources and their sensitivity levels. Then, they integrate the information leakage coefficient and the data provider's business availability to calculate the business calibration constant. Based on the aforementioned multiple indicators, they collaboratively calculate the pre-set computing power limit for data source access. On the one hand, by quantifying data security risks through multi-dimensional parameters, they allocate computing power resources differently according to the sensitivity and leakage risk, effectively avoiding resource waste caused by blind allocation of computing power resources and reducing platform hardware operation and maintenance costs. On the other hand, by using the data provider's business availability to correct the calibration constant, the pre-set computing power reserves redundant computing power for corresponding business fluctuations and interface anomalies, which can effectively ensure compliance and controllability of the entire data access process and prevent data leakage risks.
[0096] The method S400, which "analyzes query semantics and aggregates homogeneous queries within a preset period into the same cluster," specifically includes: For each original query text, the BGE-M3 and text-embedding-3-large sentence vector pre-training models are called to complete the encoding operation, and a dense semantic vector with a dimension of 768 is uniformly output. The generated vector, together with the corresponding original query information, is written into the time sliding window storage structure in real time. The window limits the data retention period to 24 hours. Historical query vectors that exceed the time limit are automatically removed, so as to continuously maintain the time-sensitive query vector dataset.
[0097] Periodic online clustering tasks are performed based on the maintained sliding window vector pool. The clustering scheduling cycle is preferably fixed at 5 minutes. In each round of scheduling, the HDBSCAN clustering algorithm is started to perform batch operations on all 768-dimensional query vectors in the window. The algorithm autonomously divides the cluster boundaries based on the distance and density distribution characteristics of the vector space, without the need for manual pre-setting of the number of clusters. Aggregation is completed based on the semantic similarity between vectors. Multiple queries with highly overlapping semantic features are automatically grouped into the same cluster. Finally, the query clustering results consisting of multiple independent clusters can be output.
[0098] Reference Figure 5 For S500, the access sharing quota for each cluster is determined based on the pre-set access computing power quota, specifically including: S510, determine the one-way error budget for each cluster.
[0099] In this embodiment, the one-way error budget refers to the percentage of the global budget that any single semantic direction consumes at most, preferably 0.05.
[0100] S520 determines the access sharing quota corresponding to each cluster based on the one-way error budget and the access preset computing power quota.
[0101] No. Cluster The corresponding access sharing quota is : , To access the pre-set computing power quota, For one-way error budgeting.
[0102] In this step, by performing real-time clustering on the query semantic embedding vector, homogeneous queries with consistent semantics but different expression forms are aggregated into the same cluster group and access quotas are uniformly configured; regardless of whether the requester initiates the query using multiple account entities or differentiated text expressions, as long as the query semantics are substantially the same, they are included in the same quota management system.
[0103] Reference Figure 6 The method for obtaining the cumulative shared quota for all queries within each cluster of S600 specifically includes: S610, based on the number of tokens consumed and the corresponding query details, determine the information content factor for the current query.
[0104] The information content factor for this query is : ; ; This represents the number of tokens consumed in the current query. This provides specific query details for the current query. This represents the number of named entities in the query semantics corresponding to the current query. This represents the number of numeric features in the semantics of the current query. This represents the number of geographic location features in the semantics of the current query. This represents the number of time features in the query semantics corresponding to the current query. , , , These are the corresponding weights, and the sum of the four weights is 1.
[0105] S620 obtains the privacy consumption budget for each query based on the information content factor and information leakage coefficient of the current query.
[0106] Privacy consumption budget is : , This represents the information leakage coefficient.
[0107] S630 obtains the cumulative consumption shared quota corresponding to all privacy consumption budgets within each cluster within a preset time period.
[0108] No. Cluster Inside The cumulative consumption sharing quota corresponding to each privacy consumption budget is: : .
[0109] This step allows for the aggregation and deduction analysis across accounts by summing up the privacy consumption budget of all queries within a single cluster, regardless of which API key these queries originate from. In other words, no matter how many accounts or how many phrases an attacker uses, as long as the semantics are the same, once the quota is exhausted, the entire group of accounts will be dealt with in a coordinated manner.
[0110] The S700 method of "determining the corresponding risk status based on the access shared quota and cumulative consumption shared quota of each cluster, and triggering the execution of corresponding access control according to the risk status" specifically includes: If the cumulative consumption of shared quota for a single cluster is not less than the access shared quota, the corresponding risk status is determined to be severe, and a unified security level upgrade policy or a circuit breaker is triggered in the corresponding cluster. That is, if the cumulative consumption of shared quota for a single cluster is not less than the access shared quota, it means that all call sources involved in the cluster have been marked as suspected collusion. In this case, the security level is upgraded uniformly, or a circuit breaker is triggered, which means that all services of the data source are stopped until the budget is restored or the data provider manually resets it.
[0111] This solution constructs a full-link protection system from multiple dimensions, including security classification, semantic clustering, privacy quota accounting, and dynamic risk control. It upgrades the granularity of access control from the traditional interface call frequency to the semantic dimension, effectively resisting covert theft attacks such as multi-account decentralized queries and aggregated inference. It also establishes a quantifiable privacy ledger based on the hierarchical quota mechanism, achieving a real-world guarantee that data is usable but not visible. The solution abandons the fixed one-size-fits-all security strategy and can adaptively change protection rules according to query quota consumption. While taking into account the needs of normal business inference, it improves the protection capabilities of sensitive data. At the same time, it forms a standardized data security classification and risk assessment logic, unifies the security evaluation criteria for inference services in the industry, and makes it convenient for enterprises to quantitatively assess the security risks of private data-hosted inference. It is specifically adapted to the application scenarios of small-scale, highly sensitive private models of enterprises, and makes up for the shortcomings of traditional security measures such as keyword auditing, identity verification, and call rate limiting.
[0112] Furthermore, the difference between the access shared quota and the cumulative consumed shared quota of a single cluster is obtained, i.e., the remaining budget, and the corresponding access control is triggered based on the remaining budget.
[0113] Specifically, if the remaining budget is not less than 20%, normal service is provided, meaning all queries are responded to as usual; if the remaining budget is less than 20%, an alarm is triggered, which can notify the data provider and administrator that the budget is running out and start recording detailed audit logs; if the remaining budget is less than 10%, an automatic upgrade is triggered, which means that the security level is forcibly upgraded by one level to reduce the amount of leakage in each subsequent query with stronger protection; if the remaining budget is less than 5%, rate limiting is triggered, which means that only broad statistical queries are allowed, and precise queries for specific individuals are rejected.
[0114] Existing access controls only consider the legitimacy of a single query, regardless of the cumulative amount stolen. This solution uses a privacy ledger to cumulatively track the amount of information leaked with each query. When the cumulative leakage approaches the security limit of the data source, it automatically triggers tiered responses such as alerts, escalation, rate limiting, and circuit breaking. Even if an attacker's queries are legitimate each time, the accumulated budget will still trigger a system response, terminating access before the information is completely pieced together.
[0115] Furthermore, this application also includes: real-time monitoring of call behavior, and automatic strengthening of protection upon detection of suspicious patterns.
[0116] Specifically, the system obtains the request count for each calling source within a preset period, such as one hour. If the count exceeds a preset threshold, it determines that there is a suspicious pattern of high-frequency detection.
[0117] This application also includes: calculating the standard deviation of the embedding for the most recent 20 queries from the source caller. At the same time, check whether the specificity sequence is monotonically increasing. Less than the preset value (This indicates a highly concentrated semantic focus) and the specificity of the last paragraph is greater than the preset value, such as 0.3, indicating the existence of a suspicious pattern that gradually approaches the target.
[0118] This application also includes: during the training phase, calculating the embedding and SimHash for each segment of 256 tokens in the training corpus and storing it in a FAISS index; during the inference phase, calculating the same fingerprint for each output and querying the Top-1 similarity, if... Greater than the preset value, such as Or the SimHash Hamming distance is less than the preset value, such as This identified a suspicious pattern that had attempted inversion.
[0119] This application also includes: employing dual detection for logical verification. The first layer uses regular expression matching to compare user input with a pre-collected pattern library of various prompt injection features, focusing on identifying typical injection characters and phrases such as ignoring pre-instructions and special model markers. The second layer uses the ProtectAI prompt injection classification model based on the deberta-v3-base architecture to complete intelligent scoring. If any rule is hit or the classifier confidence is greater than a preset value such as 0.85, a suspicious pattern of injection attack is identified.
[0120] Once a suspicious pattern is identified, the security level is immediately upgraded, i.e., the user's security level is raised by one level, subsequent requests are automatically subject to stronger protection, and complete information about the abnormal event is recorded for evidence collection and review.
[0121] Furthermore, if a user triggers three anomalies within 24 hours, all of their access permissions will be frozen, requiring manual review for restoration. If a user does not trigger any anomalies for seven consecutive days, their security level will automatically regress by one level.
[0122] In this application, when an attack is detected and a security level upgrade is triggered, the higher-level security level has a lower information leakage coefficient than the lower-level security level, thus effectively enhancing protection and reducing subsequent leaks. Simultaneously, the increased consumption of abnormal requests drives down the budget, triggering stricter access controls. Under normal circumstances, basic protection is maintained to ensure availability; when threats escalate, protection is automatically tightened; and when threats subside, protection is automatically restored, requiring no manual intervention throughout the process.
[0123] Furthermore, this application also includes: data owners can manage their data assets in real time through a console: the asset dashboard in the console displays the budget level of each data source, the estimated remaining number of calls, and the Top N caller rankings; it allows for flexible configuration of alarm subscriptions; it allows for manual intervention to temporarily adjust the budget limit, freeze specified call sources, and force upgrades to the security level; and it allows for exporting all call records within a specified time window, including query fingerprints, consumption, and abnormal events, for compliance and accountability purposes. Through this solution, data owners can see in real time who is calling their data, how much information has been taken away, how much budget remains, and whether they have been attacked. They can also upgrade the security level, freeze suspicious call sources, and adjust the budget limit with one click, transforming the process from entrusting data to the platform and leaving it to chance to having their data assets precisely measured and managed.
[0124] To address the shortcomings of existing access controls that only manage identity and call rate and lack cumulative information statistics, making it difficult to prevent aggregation and inference attacks, the method disclosed in this application aggregates distributed probing requests through query semantic clustering, then allocates independent quotas for each cluster and accumulates quota consumption statistics. Even if attackers split accounts and queries to piece together data, the grouped quotas will be quickly exhausted and blocked after similar queries are aggregated, thus blocking the aggregation and theft path from a statistical logic perspective. To address the drawbacks of rigid security policies and the potential for data leakage due to delayed manual handling, the solution pre-configures differentiated policies based on scenario levels. During operation, it automatically identifies risks and links access control based on quota consumption data, relying on automated dynamic protection to eliminate the risk of data leakage caused by the lag in manual operation and maintenance.
[0125] To address the lack of unified security grading standards in the industry and the inability of enterprises to accurately assess inference security risks, the solution standardizes security level classification, leakage coefficient selection, and quota calculation rules, forming a universal grading assessment framework to unify the measurement standards of security capabilities across service platforms. Addressing the issues of private models' single tokens easily leaking raw data and the limited protection offered by conventional injection and semantic auditing, the solution adopts a quantitative control method that combines leakage coefficients with the amount of query information to calculate quota consumption per query. High-value queries consume quotas faster, and even if attackers bypass injection detection and regular expression blocking, continuous probing queries will be restricted due to quota exhaustion, filling the control blind spots of traditional security protection.
[0126] Reference Figure 7 Secondly, this application discloses a hierarchical access control method for enterprise data inference services, including: S10, Analyze the semantics of the queries and aggregate the homogeneous queries within the preset period into the same cluster; S20, determine the access sharing quota corresponding to each cluster based on the access preset computing power quota corresponding to the inference service platform.
[0127] The method for obtaining the pre-set computing power quota for the inference service platform includes: determining the security level type of the target scenario, calling the protection policy and information leakage coefficient corresponding to the security level type; connecting the target data source to the platform and executing the protection policy; and determining the pre-set computing power quota for the platform after connecting the target data source based on the information leakage coefficient, the number of target data sources, and the sensitivity level of the target data source.
[0128] S30, obtain the cumulative consumption shared quota of all queries within each cluster within a preset time period; S40 determines the corresponding risk status based on the access shared quota and cumulative consumption shared quota of each cluster, and triggers the execution of corresponding access control based on the risk status.
[0129] It should be noted that the specific implementation methods of steps S10-S40 are consistent with the corresponding methods of the enterprise data reasoning service hierarchical access control method disclosed in the first aspect of this application, so they will not be described in detail here.
[0130] A computer device according to embodiments of the present disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.
[0131] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the enterprise data inference service hierarchical access control method described in the foregoing embodiments of this disclosure.
[0132] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.
[0133] like Figure 8 This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of the present disclosure. Figure 8The computer device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.
[0134] like Figure 8 As shown, a computer device may include a processor (such as a central processing unit, graphics processing unit, etc.), which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) or programs loaded from storage devices into random access memory (RAM). The RAM also stores various programs and data required for the operation of the computer device. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.
[0135] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow the computer device to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data. Although Figure 8 A computer apparatus with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.
[0136] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processor, all or part of the steps of the enterprise data inference service hierarchical access control method of embodiments of this disclosure are performed.
[0137] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.
[0138] A computer-readable storage medium according to embodiments of the present disclosure stores non-transitory computer-readable instructions. When the non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the enterprise data inference service hierarchical access control method described in the foregoing embodiments of the present disclosure are performed.
[0139] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).
[0140] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.
[0141] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.
[0142] In this disclosure, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The block diagrams of devices, apparatuses, devices, and systems involved in this disclosure are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as "comprising," "including," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context clearly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.
[0143] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.
[0144] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.
[0145] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.
[0146] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.
[0147] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.
Claims
1. A hierarchical access control method for enterprise data inference services, characterized in that, include: Determine the security level type of the target scenario, and invoke the protection strategy and information leakage coefficient corresponding to the security level type; Connect the target data source to the platform and execute the protection strategy; Based on the information leakage coefficient, the number of target data sources, and the sensitivity level of the target data sources, determine the pre-set computing power quota for the platform after accessing the target data sources; Analyze the semantics of the queries and aggregate the corresponding homogeneous queries within a preset period into the same cluster; The access sharing quota corresponding to each cluster is determined based on the access preset computing power quota; Based on the information leakage coefficient, the cumulative consumption shared quota of all queries within each cluster within a preset time period is obtained; Based on the access shared quota and the cumulative consumption shared quota of each cluster, the corresponding risk status is determined, and the corresponding access control is triggered according to the risk status. The process of determining the security level type of the target scenario and invoking the protection strategy and information leakage coefficient corresponding to the security level type includes: Configure the corresponding security level for each scenario and the protection strategy for each security level according to the security requirements of the data usage scenarios. Configure the information leakage coefficient corresponding to the lowest security level, and denote it as the basic leakage coefficient; Based on the basic leakage coefficient, the information leakage coefficient corresponding to each of the security levels except the lowest security level is determined and stored in the database; Based on the target scenario corresponding to the enterprise data inference service, determine the security level type of the target scenario; Retrieve the protection strategy and information leakage coefficient corresponding to the security level type from the database; The step of determining the information leakage coefficient corresponding to each of the security levels other than the lowest security level based on the basic leakage coefficient includes: Determine the preset attack type corresponding to the target scenario; Based on the technical logic of the preset attack types, multiple repeatable and quantifiable detection and query commands are configured for each preset attack type; For data inference service scenarios configured with the lowest security level, attack detection is performed based on the detection query command, and the corresponding first information entropy is determined based on the detection results; For data inference service scenarios configured with security levels other than the minimum level, attack detection is performed based on the detection query command, and the corresponding second information entropy is determined based on the detection results. The ratio of the second information entropy to the first information entropy is used as the information leakage coefficient corresponding to the non-lowest security level.
2. The hierarchical access control method for enterprise data inference services according to claim 1, characterized in that, The step of determining the platform's pre-set computing power limit after accessing the target data source based on the information leakage coefficient, the number of target data sources, and the sensitivity level of the target data sources includes: Determine the number and sensitivity level of the target data sources required for the target scenario; Based on the information leakage coefficient, the quantity, the sensitivity level, and the business availability information corresponding to the data provider, determine the business calibration constant corresponding to the target scenario; Based on the quantity, the sensitivity level, and the business calibration constant, the pre-set computing power quota corresponding to the platform after accessing the target data source is determined.
3. The hierarchical access control method for enterprise data inference services according to claim 2, characterized in that, The step of determining the business calibration constant corresponding to the target scenario based on the information leakage coefficient, the quantity, the sensitivity level, and the business availability information corresponding to the data provider includes: Based on the business availability information provided by the data provider, determine the business scale base corresponding to the target scenario; Analyze the sensitivity level of the target data source required for the target scenario to determine the data sensitivity information; Based on the data sensitivity information and the quantity, determine the data security weighting item; Based on the number of tokens consumed and the specific information of the query, determine the information content factor of the query. Based on the information leakage coefficient, the information volume factor of the current query, the business scale base, and the data security weighting term, the business calibration constant corresponding to the target scenario is determined.
4. The hierarchical access control method for enterprise data inference services according to claim 1, characterized in that, The step of determining the access sharing quota corresponding to each cluster based on the access preset computing power quota includes: Determine the one-way error budget for each cluster; Based on the one-way error budget and the access preset computing power quota, the access sharing quota corresponding to each cluster is determined.
5. The hierarchical access control method for enterprise data inference services according to claim 1, characterized in that, The step of obtaining the cumulative consumption shared quota of all queries within each cluster within a preset time period based on the information leakage coefficient includes: Based on the number of tokens consumed and the specific information of the query, determine the information content factor of the query. Based on the information content factor of the current query and the information leakage coefficient, the privacy consumption budget corresponding to each query is obtained; Obtain the cumulative consumption shared quota corresponding to all privacy consumption budgets within each cluster within a preset time period.
6. The hierarchical access control method for enterprise data inference services according to claim 1, characterized in that, The step of determining the corresponding risk status based on the access shared quota and the cumulative consumption shared quota of each cluster, and triggering the execution of corresponding access control based on the risk status, includes: If the cumulative consumption of the shared quota for a single cluster is not less than the access shared quota, the corresponding risk status is determined to be severe, and a unified security level upgrade strategy or a circuit breaker is triggered in the corresponding cluster.
7. A computer device, characterized in that, The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor; the instructions are executed by the at least one processor to enable the at least one processor to perform the enterprise data inference service hierarchical access control method according to any one of claims 1-6.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions; the computer instructions are used to cause the computer to execute the enterprise data inference service hierarchical access control method as described in any one of claims 1-6.
9. A computer program product comprising computer instructions, characterized in that, When executed by a processor, the computer instructions implement the hierarchical access control method for enterprise data inference services as described in any one of claims 1-6.
Citation Information
Patent Citations
LLM-DoS attack protection method based on multi-level defense strategy and related device
CN121309150A
Energy enterprise data sharing and privacy protection method and system based on block chain
CN121413018A
