Data element driven AI original infrastructure safety operation method

Abstract

The invention discloses a data element driven AI native infrastructure safety operation method, and belongs to the technical field of artificial intelligence and data safety. Aiming at the technical problems of extensive authority management, difficulty in safety traceability and low compliance verification efficiency in the existing AI infrastructure safety operation, the intelligent agent authority dynamic binding is realized by establishing a data element safety label system; constructing a full-link safety traceability map, and combining an AI abnormal behavior recognition and fusing mechanism; a compliance intelligent verification module is integrated, and multi-industry supervision requirements are automatically met. According to the method, refinement of safety management and control, automation of operation and intelligentization of compliance are realized, and the safety operation and compliance cost of an enterprise is remarkably reduced. The method is mainly used for safety operation and data management of the AI native infrastructure.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, data security and compliance technology, and in particular to an intelligent security operation method applied to AI native infrastructure, specifically including a security tagging system based on data elements, dynamic permission management of intelligent agents, full-link security traceability, AI abnormal behavior recognition and compliance intelligent verification and other technologies. Background Technology

[0002] With the rapid development of artificial intelligence technology, enterprises are building AI-native infrastructure to support their intelligent transformation. Under this architecture, numerous AI agents operate autonomously or semi-autonomously, accessing and processing various enterprise data resources in real time, forming a complex data flow network. However, existing security operation methods face significant challenges in addressing this new architecture: 1. Rigid Access Control: Traditional access control models based on role-based access control (RBAC) or attribute-based access control (ABAC) cannot adapt to the dynamic task execution characteristics of intelligent agents. Access configurations are mostly statically preset and cannot be dynamically adjusted according to changes in data value or task context, leading to the dilemma of "over-authorization" or "under-authorization".

[0003] 2. Challenges in Security Tracing: When data breaches or unauthorized operations occur, existing log systems struggle to reconstruct complex data flow paths. This is especially true when data has been processed and transformed by multiple agents, making it difficult to trace its source, destination, and intended use, resulting in lengthy and costly investigations.

[0004] 3. Inefficient Compliance Verification: Different industries (finance, healthcare, manufacturing, etc.) have strict and differentiated regulatory requirements for data security. Enterprises typically rely on manual compliance audits, which are slow to respond and prone to compliance risks due to misunderstandings of rules. In scenarios with high-frequency interactions from AI agents, traditional manual audits can no longer meet real-time requirements.

[0005] 4. Delayed Anomaly Detection: Existing security monitoring is mostly based on rule matching, lacking the ability to intelligently identify new attack patterns and internal threats. Abnormal behavior of intelligent agents (such as data crawling and permission abuse) is often only discovered after losses have occurred. Existing technologies, such as CN114785532A which discloses a blockchain-based data security traceability method, focus on preventing data tampering itself and do not address the issue of dynamic permissions for intelligent agents; CN115396178A which discloses an AI model security assessment method, does not cover data element management and multi-agent collaborative scenarios. Summary of the Invention

[0006] Technical problems to be solved Based on the background technology analysis, the core technical problem this invention aims to solve is: how to achieve refined, automated, and intelligent operation and management of data security in a multi-agent collaborative environment within an AI-native infrastructure. Specifically, this includes: 1. How to break through static permission restrictions and establish an intelligent agent permission mechanism that dynamically matches data value; 2. How to achieve complete tracing and visualization of complex data flow paths; 3. How to translate cross-industry compliance requirements into automatically executable verification rules; 4. How to identify abnormal behavior of intelligent agents in real time and respond automatically.

[0007] Technical solution To address the aforementioned technical problems, this invention provides a data-element-driven method for the secure operation of AI-native infrastructure, characterized by comprising the following core technical modules and their collaborative relationships: 1. Data Element Security Tagging Subsystem Value Assessment Engine: Employing a multi-dimensional assessment model, this engine comprehensively considers factors such as the business value, privacy sensitivity, timeliness, and scarcity of data to automatically calculate the value rating of data elements. The assessment model uses a weighted scoring algorithm: V = α×B + β×P + γ×T + δ×S, where B is the business value coefficient, P is the privacy sensitivity coefficient, T is the timeliness coefficient, S is the scarcity coefficient, and α, β, γ, and δ are adjustable weight parameters.

[0008] Tag Generator: Automatically generates structured security tags based on value level scoring. The tag format is: {data_id}{security_level}{valid_period}{usage_constraints}, supporting a hierarchical tag system and allowing parent-child tag inheritance relationships.

[0009] 2. Intelligent Agent Permission Dynamic Binding Subsystem Permission Policy Engine: Stores a "data tag - agent permission" mapping policy library, supporting context-based adaptive policy adjustments. Policy rules are implemented using the Drools rule engine, supporting complex condition judgments. Real-time Binding Module: Calculates the set of permissions to be granted in real time when an agent initiates a data request. Specifically, it includes: a. Receive requests from intelligent agents and parse target data identifiers; b. Query the data security label to obtain the current value level; c. Calculate dynamic permission coefficients based on agent credit score, task type, and access context; d. Generate personalized permission tokens based on the mapping strategy, with the validity period bound to the task cycle.

[0010] 3. End-to-end security traceability subsystem Graph construction engine: Employs Neo4j graph database to store data flow relationships. Nodes include data entities, agents, users, and system components, while edges include operation type, timestamp, and context information. The construction algorithm uses real-time stream processing with latency less than 100 milliseconds. Path reconstruction module: Provides multi-dimensional query interfaces, supporting: a. Backward tracing: Starting from the specified data, query all downstream usage paths; b. Backward tracing: Reconstructing the complete historical flow chain from the current data state; c. Impact Analysis: Assess the scope of the impact of a single security issue on the overall situation.

[0011] 4. AI Abnormal Behavior Recognition Subsystem Behavior modeling module: Uses LSTM neural network to model the historical normal behavior sequence of each agent, and learns its access patterns, time distribution, data preferences and other features.

[0012] Real-time detection module: Based on the Isolation Forest and Local Outlier Factor algorithms, it calculates behavioral anomaly scores in real time. Detection metrics include: Abnormal access frequency: The deviation from the historical baseline exceeds the threshold; Permission usage error: Access exceeded the normal permission range; Abnormal time pattern: Active during unusual time periods; Data association error: Accessing unrelated data combinations.

[0013] Circuit breaker response module: Implements a tiered circuit breaker mechanism. Level 1 Alert: Logs the anomaly and notifies the administrator; Level 2 rate limiting: Reduces the frequency of agent access; Level 3 circuit breaker: Suspends all data operations of the agent; Automatic recovery: Normal access will be automatically restored after the anomaly is resolved.

[0014] 5. Compliance Intelligent Verification Subsystem Rule Knowledge Base: Employs ontology modeling to build an industry compliance rule base, transforming natural language legal provisions into executable machine rules. Supports over 5000 core compliance requirements across 10+ industries including finance, healthcare, and manufacturing. Automatic Validation Engine: Utilizes a combination of rule reasoning and pattern matching. a. Preprocessing: Extracting data operation context feature vectors; b. Rule matching: Use the RETE algorithm to quickly match applicable rules; c. Reasoning and Judgment: Performing logical reasoning on complex compliance requirements; d. Result generation: Output compliance / non-compliance judgment and basis.

[0015] Report Generator: Automatically generates structured compliance reports, supporting multi-dimensional summary analysis by time, business line, and compliance item. The collaborative working mechanism of each module is as follows: 1. When new data enters the system, the data element security tagging subsystem performs value assessment and generates tags for it; 2. When an agent requests access to data, the permission dynamic binding subsystem generates real-time permissions based on the data tag and the agent's context; 3. All data operations are recorded by the end-to-end security traceability subsystem, constructing a complete data flow map; 4. The AI ​​abnormal behavior recognition subsystem monitors the behavior of the intelligent agent in real time and triggers a circuit breaker when an anomaly is detected; 5. Key data flow nodes are automatically reviewed for compliance by the compliance intelligent verification subsystem; 6. All subsystems share a unified security policy management center to ensure policy consistency. Beneficial effects

[0016] Compared with the prior art, the present invention has the following significant advantages: 1. Improved granularity of access control: By binding data element tags with dynamic permissions, the granularity of permissions is refined from the traditional data table level to the data element level, improving the accuracy of permission matching by more than 85% while reducing the risk of over-authorization.

[0017] 2. Significantly improved security traceability efficiency: The full-chain traceability map reduces the security incident investigation time from an average of 3-5 days to less than 30 minutes, with a traceability completeness of 99.9%, and supports real-time visualization of complex data flow relationships.

[0018] 3. Significantly reduced compliance operating costs: Automated compliance verification reduces the workload of manual audits by more than 90%, compliance checks are upgraded from weekly / monthly to real-time at the second level, and the time to adapt to cross-industry rules is shortened from several months to several days.

[0019] 4. Enhanced intelligence in anomaly detection: AI-driven anomaly identification accuracy reaches 95%, with a false alarm rate of less than 5%, realizing the transformation from passive response to proactive defense, and shortening the average threat response time to within 2 minutes.

[0020] 5. System scalability and adaptability: The modular design supports flexible expansion and can be quickly adapted to new industry compliance requirements through a plug-in mechanism, supporting the secure operation of tens of millions of intelligent agents and petabytes of data. Attached Figure Description

[0021] Appendix Figure 1This is the overall system architecture and data flow diagram provided in the embodiments of the present invention.

[0022] Appendix Figure 2 This is a schematic diagram of data element security label classification and permission mapping provided in an embodiment of the present invention.

[0023] Appendix Figure 3 This is an example diagram of the end-to-end security traceability graph structure provided in the embodiments of the present invention.

[0024] Appendix Figure 4 This is a flowchart of the compliance intelligent verification module provided in this embodiment of the invention.

[0025] Appendix Figure 5 This is a flowchart of the AI ​​abnormal behavior recognition and circuit breaker mechanism provided in the embodiments of the present invention. Detailed Implementation The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0026] Example 1: Data Security Operation Scenario in the Financial Industry Step 1: System Initialization and Configuration Reference Figure 1 Deploy the security operation system of this invention. First, configure the basic environment: 1.1 Install the data element tagging engine and configure financial industry-specific valuation parameters: Business value weight α = 0.4 Privacy sensitivity weight β = 0.3 Timeliness weight γ = 0.2 Scarcity weight δ = 0.1 1.2 Deploy the graph database cluster (Neo4j 4.x), initialize the traceability graph schema, and define node types including: customer data, transaction data, risk data, report data, etc.

[0027] 1.3 Load the financial industry compliance rule library, including the Commercial Banking Law, the Data Security Law, GDPR financial provisions, etc., with a total of 1200+ machine-executable rules.

[0028] 1.4 Configure the intelligent agent registration center to register the identities and collect baseline behaviors of existing AI intelligent agents.

[0029] Step 2: Data Feature Tag Generation Reference Figure 2 When new customer transaction data enters the system: 2.1 The value assessment engine automatically analyzes data characteristics: Data Identifier: TX20230715001 Data content: Customer A's large cross-border transfer Business Value: High (involves cross-border business) Privacy sensitive: Extremely high (includes customer identity and transaction details) Timeliness: High (real-time transactions) Scarcity: Medium (approximately 100 similar transactions per day) Overall score: V = 0.4×0.9 + 0.3×1.0 + 0.2×0.8 + 0.1×0.6 = 0.86 2.2 Generate tags based on the rating mapping to security level L1 (confidential): { "data_id": "TX20230715001", "security_level": "L1", "valid_period": "2023-07-15 to 2023-07-22", "usage_constraints": { "allowed_agents": ["risk_control_agent", "aml_agent"], "max_access_freq": "10 times / hour", "retention_policy": "Encrypted storage for 180 days" } } Step 3: Dynamically grant permissions to the agent When an anti-money laundering (AML) agent requests access to the transaction data: 3.1 The permission policy engine receives requests: { "agent_id": "aml_agent_v2", "agent_trust_score": 0.92, "request_data": "TX20230715001", "operation_type": "read_analyze", "task_context": "Real-time anti-money laundering monitoring" } 3.2 Query the data tags to confirm that the security level is L1, and check whether the agent is in the allowed list.

[0030] 3.3 Generate permission tokens based on the dynamic permission calculation formula: Permission coefficient = agent_trust_score × task_priority × time_factor = 0.92 × 0.9 × 1.0 = 0.828 Permission scope = Basic permissions × Permission coefficient Grant permissions: {read: true, write: false, export: false, share: false} Validity period: Synchronized with the task execution cycle (this task is expected to last 30 minutes). 3.4 Return the permission token; the AML agent uses the token to access the data.

[0031] Step 4: End-to-end security traceability record Reference Figure 3 Record the complete operation chain: 4.1 Create a traceability record node: Node 1: Data Source { "id": "source_001", "type": "transaction_data", "content_ref": "TX20230715001", "create_time": "2023-07-15 14:30:00", "creator": "core_banking_system" } Node 2: Processing Agent { "id": "agent_aml", "type": "ai_agent", "name": "aml_agent_v2", "trust_level": 0.92 } 4.2 Creating operation relationship edges: Edge 1: Access Operation { "from": "agent_aml", "to": "source_001", "relation": "accessed", "time": "2023-07-15 14:31:15", "permission_token": "perm_xyz123", "operation": "read_analyze" } Edge 2: Result { "from": "agent_aml", "to": "result_001", "relation": "generated", "result_type": "risk_assessment", "risk_score": 0.75 } 4.3 Real-time map updates, supporting instant queries. In the event of a security incident, a Cypher query can be executed: MATCH path = (d:transaction_data {id:"TX20230715001"})-[*1..5]-(n) RETURN path Return all related nodes and relationships within 3 layers of this data within 30 milliseconds.

[0032] Step 5: Real-time monitoring of abnormal AI behavior Reference Figure 5 Continuously monitor the behavior of the intelligent agent: 5.1 Behavioral Modeling Phase (7-day training period): Collect historical behavior sequences of the AML agent and train an LSTM model: Input feature dimension: 10 Hidden layer units: 64 Time step: 24 (hours) Training accuracy: 96.3% 5.2 Real-time detection phase: An abnormal pattern was detected (example): Time: 2023-07-15 03:15:00 (Outside of working hours) Operation: High-frequency access to L1 level customer data (15 times / minute, baseline is 2 times / minute) Objective: Multiple unrelated customer accounts Anomaly score: 0.87 (threshold 0.75) 5.3 Triggering Level 2 Circuit Breaker Measures: Immediately log the abnormal event and send an alert to the security administrator. Rate limiting was applied to aml_agent_v2: access frequency was reduced to baseline level. Record detailed context for subsequent analysis. Step 6: Compliance Intelligent Automatic Verification Reference Figure 4 Conduct compliance checks at key operational points: 6.1 When an AML agent attempts to export analysis results, a compliance check is triggered: Verify context: - Data Category: Cross-border Transaction Data - Operation type: Data export - Target System: External Regulatory Reporting Platform - Relevant regulations: Article 15 of the People's Bank of China's Regulations on Cross-border Capital Flow Management 6.2 Rule Engine Matching Applicable Rules: Rule number: CBIRC_2023_015 Rule details: Cross-border transactions exceeding US$500,000 must be reported to the central bank within 24 hours. Triggering condition: amount > 500000 AND cross_border = true 6.3 Perform automatic verification: Transaction amount checked: $850,000 > $500,000 → Meets criteria Inspection and reporting time limit: From the time the transaction occurred to the current time = 2 hours < 24 hours → Compliant Check the reporting format: Compliant with the People's Bank of China's XML Schema standard → Compliant 6.4 Generate compliance reports and automatically submit them: Compliance Report: { "check_id": "comp_check_789", "timestamp": "2023-07-15 14:32:00", "data_ref": "TX20230715001", "applicable_rules": ["CBIRC_2023_015"], "compliance_status": "PASS", "details": "Meets the requirement to submit within 24 hours", "auto_action": "triggered_report_to_PBC" } Example 2: Implementation Optimization in Manufacturing Scenarios When implementing this in the manufacturing sector, it should be optimized to suit the specific characteristics of the industry. 1. Data Tagging System Adjustment: Increase the weight of the "Production Safety Relevance" dimension, add dedicated tag categories for equipment data, process parameters, and quality inspection data, and set higher timeliness requirements for real-time production line data. 2. Compliance rule base expansion: Integrates the "Guideline for Industrial Data Classification and Grading", adds ISO 27001 industrial control system security requirements, and configures industry-specific data retention and deletion rules. 3. Anomaly Detection Model Optimization: Train a dedicated detection model for equipment data access patterns, add correlation analysis of anomalies in production process parameters, and set specific circuit breaker strategies for production line anomalies. System performance indicators: After six months of operation in a production environment, the system of this invention has been implemented at manufacturing customers: Permission matching accuracy: 92.7% Anomaly detection accuracy: 94.1%, false alarm rate: 4.3%. Automatic compliance verification coverage: 98.5% Average response time for security incidents: 1.8 minutes Average latency for source tracing queries: 45 milliseconds System resource usage: CPU average 12%, memory average 3.2GB. Technical effectiveness verification: Comparison tests with the existing security system show that: 1. In simulated data breach scenarios, the time for tracing and locating the source was reduced from 4.5 hours to 8 minutes; 2. The workload for compliance audits has been reduced from 120 person-hours per month to 8 person-hours. 3. Data access violations caused by accidental operations decreased by 87%; 4. The system's ability to identify new threats has been improved by 3 times.

[0033] The above embodiments fully illustrate the technical feasibility and superiority of the present invention. Those skilled in the art can adjust the parameters and configurations according to specific application scenarios, and all such adjustments should be included within the scope of protection of the present invention.

Claims

1. A data-element-driven method for secure operation of AI-native infrastructure, characterized in that... Includes the following steps: Establish a data element security labeling system to classify and label data according to the value level of data elements; Based on the security tags of the data elements, a dynamic binding mechanism for agent permissions is implemented, which automatically adjusts the scope of calls and operation permissions of the agent according to the value level of the data. Construct a full-chain security traceability map to record the complete flow path of data from production to intelligent applications; automatically identify abnormal call behavior and trigger security circuit breakers through AI models; integrate a compliance intelligent verification module to automatically adapt compliance strategies according to the data supervision requirements of different industries and achieve second-level compliance verification.

Images