An ai behavior monitoring method and system based on virtualized hardware root of trust

Abstract

The application discloses an AI behavior monitoring method and system based on a virtualized hardware root of trust, and belongs to the technical field of cloud computing and artificial intelligence security. A virtual trust root certificate is derived and issued for a virtualized trusted execution environment by a physical hardware root of trust, and a private key is stored in a secure memory area and cannot be exported, so that an AI behavior monitoring agent in the virtualized environment obtains a trusted signature capability of the same level as the physical root of trust. The monitoring agent collects multi-dimensional behavior data of the AI system and calculates the behavior complexity cumulative entropy value, and the dynamic threshold is adaptively determined based on statistical distribution. The monitoring agent is periodically dynamically measured and protected during runtime, and the signature data is accompanied by a complete physical root certificate chain. The system cooperates with an AI computing power scheduling system and a privacy computing compliance proof system to form a cross-protocol cooperative protection system. The application provides hardware-level behavior monitoring and evidence storage capability with reliable electronic signature efficiency for cloud and end-side AI deployment.

Description

Technical Field

[0001] This invention relates to the fields of cloud computing and artificial intelligence security technology, specifically to a method and system that extends the trust chain of physical hardware trust roots to virtualized environments, providing hardware-level behavioral monitoring and judicial evidence preservation capabilities for cloud and edge AI systems. Background Technology

[0002] With the large-scale deployment of AI technology on cloud servers and edge devices, real-time monitoring and security auditing of AI system behavior has become a fundamental requirement for regulatory compliance. Hardware roots of trust (such as TPMs and security chips) are the cornerstone of building a trusted computing environment, providing unforgeable signatures and timestamps for data. However, in cloud computing environments, AI systems typically run in virtualized environments (such as virtual machines and containers), making direct access to the underlying physical hardware roots of trust impossible. Existing solutions either require each virtual instance to have its own dedicated physical chip, resulting in extremely high costs, or resort to pure software signatures, which lack legal evidentiary value and fail to provide hardware-level trusted monitoring while ensuring cost-effectiveness. Existing vTPM (Virtual Trusted Platform Module) solutions employ a "real-time forwarding mode," where all signature requests from virtual instances must be forwarded to the physical TPM via the hypervisor for execution. This approach has four drawbacks: the physical TPM becomes a performance bottleneck, limiting the number of concurrent virtual instances it can support; high signature latency fails to meet the response requirements of real-time AI monitoring; virtual instances are highly coupled with the physical TPM, so if the physical TPM fails, the signature function of all virtual instances is interrupted; and when virtual instances migrate online, the trust chain breaks, resulting in gaps in monitoring records. Summary of the Invention

[0003] This invention aims to address the problem that physical hardware roots of trust cannot economically cover large-scale virtualized AI environments. This invention derives and issues unique virtual root of trust credentials for each instance of a virtualized trusted execution environment (TEX), enabling AI behavior monitoring agents running in the virtual environment to obtain trusted signature capabilities at the same level as physical roots of trust. The core improvement of this invention lies in: First, it achieves a "one-time derivation, offline availability" mechanism. After the virtual root trust credential is derived, the signing operation can be completed independently without relying on the physical root trust being online in real time. The physical root only needs periodic health checks. Compared with the traditional forwarding vTPM solution, the number of AI-monitored instances supported by a single physical TPM is increased by 3 to 5 times, and the signature response latency is reduced by more than 75%. The private key of the virtual root trust credential is generated in the secure memory area of ​​the virtualized trusted execution environment instance. The plaintext of the private key never leaves this secure memory area and can only be called to perform signing operations within the instance. It cannot be exported or read, fundamentally ensuring the non-repudiation of the signature. Second, the technical specifications for virtual signatures have been clarified. The signature data of the virtual root trust certificate is accompanied by a complete physical hardware root trust certificate chain, which can verify the certainty of signature ownership, the verifiability of content integrity, and the non-repudiation of signature behavior, thus aligning with the compliance requirements of the "Interim Measures for the Administration of Generative Artificial Intelligence Services" for traceable and auditable AI behavior. Third, a complete health verification closed loop is constructed. Health verification is proactively initiated by the physical hardware root of trust. It retrieves the runtime memory integrity metric of the corresponding virtualization instance through the virtualization trusted execution environment manager and compares it with the baseline metric stored during credential derivation. If the verification fails, the system automatically freezes the virtual root of trust credential, suspends all its signature operations, and marks the virtual instance as untrusted. Frozen credentials can only be unfrozen and restored after the physical root performs integrity measurement and health verification again. Fourth, an AI behavior entropy quantification mechanism is introduced. The monitoring agent collects behavioral data from multiple dimensions of the AI ​​system, such as instruction decomposition depth, cross-domain correlation, policy mutation magnitude, and resource call expansion rate, to calculate the cumulative entropy value of behavioral complexity, achieving an objective quantitative assessment of AI autonomy. The dynamic threshold determination method has multi-form adaptability, including but not limited to statistical distribution methods, quantile methods, and machine learning anomaly detection model methods, which can be flexibly selected according to the monitoring scenario. During the operation of the monitoring agent, the virtual root of trust periodically performs dynamic integrity measurement on the code segment and core data segment of the monitoring agent. When the measurement is abnormal, the signing permission is immediately terminated and an alarm is triggered to prevent attackers from hijacking the monitoring agent through memory tampering. Fifth, the loading location of the monitoring agent is flexible and adaptable. The monitoring agent can be deployed inside a virtualized trusted execution environment instance or in a secure isolation domain of the Hypervisor layer. It passively collects AI behavior data from multiple virtualized instances through memory images and system call bypass images, adapting to different deployment architectures. Sixth, the trustworthiness of the timestamp service is ensured. The timestamps affixed to the monitoring records are provided by an independent timestamp service, whose public key is pre-measured and solidified by the physical hardware root of trust during the secure boot phase. The timestamp public key supports trusted updates; when the public key is rotated, it must be re-measured and solidified by the physical hardware root of trust before it can take effect. Seventh, it supports online migration and elastic scaling of virtual instances. During migration, the target physical hardware root of trust re-derives virtual credentials, and the extended fields of the new credentials are written into the hash value of the old credentials. Continuous migration forms a credential hash chain, and monitoring records achieve continuous traceability throughout the entire lifecycle through the hash chain. In cloud-based elastic scaling scenarios, adding a new virtual instance automatically triggers credential derivation, and destroying an instance automatically revokes the credentials. Eighth, it supports multi-tenant isolation and asymmetric / symmetric hybrid key systems. Virtual instances of different tenants correspond to independent virtual root trust credentials, and the physical hardware root trust isolates and manages tenant permissions. It also supports a symmetric key system as an alternative, where the physical root derives an independent symmetric signature key for each instance using a key derivation algorithm. Signing is performed locally on the instance, while verification is uniformly executed by the physical root. Ninth, support blockchain-based evidence storage as an alternative method. Virtualized monitoring records can be synchronized to consortium blockchain nodes, achieving distributed evidence storage through multi-node consensus. Block time replaces centralized timestamp services, and the credibility of block time is verified by on-chain public keys pre-measured by the physical hardware root of trust. Tenth, a cross-protocol collaborative protection system has been constructed. The AI ​​behavior monitoring system of this invention works collaboratively with the AI ​​computing power scheduling system and the privacy computing compliance certification system: before executing a route, the scheduling system must verify the validity of the monitoring system's credentials; when generating a report, the compliance system must cross-verify the consistency between the monitoring records and the scheduling logs. Failure of verification by any system triggers a global alarm, forming an inseparable, three-in-one joint defense network that completely blocks attackers from exploiting inter-system collaborative vulnerabilities. Attached Figure Description

[0004] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 Flowchart for Virtual Trust Root Derivation and Guardian Agent Loading; Figure 3 Flowchart for the generation and storage of guardianship data. Detailed Implementation

[0005] The present invention will now be described in detail with reference to the accompanying drawings and embodiments: Example 1: Virtualized Monitoring of Cloud-based AI Inference Clusters An AI cloud service provider manages an AI inference cluster comprising 5,000 physical servers, each equipped with a TPM 2.0 security chip as the physical hardware root of trust. Each physical server runs dozens of virtual machine instances, and each virtual machine deploys an independent AI inference service. During the secure server boot process, TPM performs integrity measurements on the BIOS, bootloader, and virtual machine manager. Upon successful measurement, TPM uses its internally stored unique root key to derive and issue a unique virtual root trust credential for each virtual machine instance. This credential is an X.509 format attribute certificate containing the virtual machine's identity, security attribute statements, and TPM's signature. The credential's private key is generated within the virtual machine's secure memory area; the plaintext private key never leaves this secure memory area and can only be accessed within the instance to perform signing operations; it cannot be exported or read. TPM also measures and stores the timestamp server's public key to ensure the reliability of timestamps in subsequent monitoring records. After the virtual root trust credential is derived, even if the TPM enters a dormant state, the AI ​​behavior monitoring agent within each virtual machine can still independently use its own virtual root trust credential for signing operations. The TPM proactively initiates a health check every 24 hours, pulling runtime memory integrity metrics for each virtualization instance through the Virtualization Trusted Execution Environment Manager and comparing them with the baseline metrics stored during credential derivation. The same physical TPM derives independent virtual root trust credentials for all virtual machine instances on the same physical node, with each credential's key and permissions completely isolated. The TPM performs health checks on all virtual roots trust in batches. If an anomaly is detected in a virtual credential, the TPM automatically freezes that credential, suspends its signing operations, and reports an alarm to the management node. Within each virtual machine, before loading, the AI ​​behavior monitoring agent's code hash value is verified by the virtual root of trust credential. After successful verification, the monitoring agent obtains a memory snapshot and system call sequence of the AI ​​inference service runtime through the security monitoring interface provided by the virtualization environment, and collects behavioral data across multiple dimensions, including instruction decomposition depth, cross-domain correlation, policy mutation magnitude, and resource call expansion rate. The monitoring agent can also be deployed in a security isolation domain at the Hypervisor layer, passively collecting AI behavior data from multiple virtualization instances via memory mirroring and system call bypass mirroring, and then calling the virtual root of trust credential of the corresponding instance to complete the signature. During the monitoring agent's operation, the virtual root of trust periodically performs dynamic integrity measurements on the monitoring agent's code segment and core data segment; if an anomaly is detected, signing privileges are immediately terminated and an alarm is triggered. After normalizing the observations for each dimension, the monitoring agent calculates the comprehensive entropy value using a weighted summation method. The comprehensive entropy value is equal to the sum of the products of the entropy value of each dimension and its corresponding weight, and the weights of each dimension are dynamically adjusted according to the monitoring scenario. The dynamic threshold is adaptively determined based on the statistical distribution of historical behavioral data, using the baseline mean plus three times the standard deviation as the upper limit of the threshold. In another implementation, the dynamic threshold is determined through a machine learning anomaly detection model: historical AI behavior entropy value sequences are collected as training data, and an unsupervised anomaly detection algorithm is used to train the dynamic threshold model. The model outputs the dynamic threshold based on real-time entropy value data. This method does not require manual parameter preset and can adapt to the behavioral characteristics of different AI systems. Behavioral data also includes at least one of the following: token generation volatility, tool call sequence entropy, and output semantic deviation, which, together with the above four dimensions, participate in the calculation of the cumulative entropy value of behavioral complexity. The entropy value calculation result is stored after being signed by a virtual trust root credential. The monitoring agent uses the virtual root of trust credential to sign the behavioral data, generating a signed monitoring data packet. The monitoring data generation module packages the signed behavioral data, the virtual root of trust credential, and the certificate chain of the physical TPM together to generate a virtualized monitoring record. This record is timestamped using a timestamp service, stored in a local tamper-proof storage medium, and synchronized to other monitoring nodes to form a distributed evidence storage network. Alternatively, the virtualized monitoring record can be synchronized to consortium blockchain nodes, achieving distributed evidence storage through multi-node consensus, replacing the centralized timestamp service with block time, and the credibility of the block time is verified by the on-chain public key pre-measured by the physical TPM. When regulatory agencies need to audit the historical behavior of an AI inference service, they can obtain the corresponding virtualized monitoring records. Because these records are signed with virtual root of trust credentials and come with a complete physical TPM certificate chain, the attribution of the signature can be verified, the integrity of the content can be verified, and the signing behavior can be non-repudiated. Therefore, these records can be used as valid evidence in courts, arbitration institutions, and administrative regulatory proceedings. When the timestamp service needs to update the public key, the physical TPM receives the new public key through a secure channel, verifies its signature validity, updates the solidified storage, and generates an updated storage record containing the hash values ​​of the old and new public keys. In cloud-based elastic scaling scenarios, once a new virtual machine instance is created, the physical TPM automatically triggers integrity measurement and virtual root trust credential derivation processes without manual intervention. When an instance is destroyed, the physical TPM automatically revokes the corresponding credential and records the event. In multi-tenant cloud scenarios, virtual machine instances for different tenants correspond to independent virtual root trust credentials. Tenants can only access monitoring records associated with their own instances, and the physical TPM isolates and manages tenant permissions. Example 2: Online Hot Migration Scenario for Virtual Instances A virtual machine instance needs to be migrated from physical server A to physical server B for load balancing. After the migration, the TPM on server B re-performs integrity measurements on the migrated virtual machine instance and derives a new virtual root of trust credential for it. The extended fields of the new credential are written to the hash value of the old credential on server A. Continuous migration forms a credential hash chain, and monitoring records achieve continuous traceability throughout the entire lifecycle through the hash chain. At the same time, the TPM on server A automatically revokes the original virtual root of trust credential to prevent credential abuse. Example 3: Alternative Scheme for Symmetric Key Systems In another embodiment, the virtual root of trust employs a symmetric key system. The physical TPM derives an independent symmetric signature key for each virtual machine instance using the HKDF key derivation algorithm. Signing operations are performed locally on the instance, while verification operations are uniformly executed by the physical TPM. This scheme eliminates the need for a certificate chain, making it suitable for scenarios sensitive to certificate management overhead, while maintaining the same offline signing capabilities as the independent root of trust scheme. Example 4: Handling Health Verification Failure A virtual machine instance was attacked, and its runtime code was tampered with, but the virtual root trust credential itself was not modified. During periodic health checks of this credential, the Trusted Execution Environment (TPM) proactively retrieved the runtime memory integrity metric of the virtual machine instance through the Trusted Execution Environment Manager (TPM) and compared it with the baseline metric stored during credential derivation. A discrepancy was found. The TPM determined the health check failed, automatically froze the virtual root trust credential, suspended all signature operations, and marked the virtual instance as untrusted. The system reported an alarm to the management node, notifying the administrator to intervene. After the administrator confirmed the attack had been cleared, the TPM re-performed the integrity metric and health check of the virtual instance. Upon successful verification, the virtual root trust credential was unfrozen, and normal monitoring functions were restored. Example 5: Cross-protocol collaborative protection In a complete cloud-based AI service deployment, the AI ​​behavior monitoring system, the AI ​​computing power scheduling system, and the privacy computing compliance certification system work collaboratively. When the AI ​​computing power scheduling system is about to route a user task to a cloud node, it must first query the AI ​​behavior monitoring system to see if the node's virtual root of trust credential is valid. If the credential has been frozen or revoked, the scheduling system automatically excludes that node and selects another available node. Simultaneously, when generating a compliance report, the privacy computing compliance certification system must cross-verify the consistency of the monitoring records from the AI ​​behavior monitoring system and the scheduling logs from the AI ​​computing power scheduling system in terms of timestamps and execution causality. If inconsistencies are found between the monitoring records and the scheduling logs, the compliance report automatically becomes invalid, the system triggers a global alert, and suspends all cross-system task execution. This cross-protocol collaborative protection system ensures that attackers cannot bypass the overall security defenses by attacking a single system.

Claims

1. An AI behavior monitoring method based on a virtualized hardware root of trust, applied to a computing environment with a physical hardware root of trust, wherein the computing environment includes a cloud server or an edge computing device, characterized in that, Includes the following steps: Virtual root of trust derivation steps: During the secure boot process, the physical hardware root of trust performs an integrity measurement on the hardware platform on which it resides; After the measurement is passed, the physical hardware root of trust uses its internally embedded unique root key to derive and issue a unique virtual root of trust credential for at least one virtualized trusted execution environment instance. The virtual root of trust credential contains the identity identifier, attribute declaration, and physical hardware root of trust signature of the virtualized trusted execution environment instance. After the virtual root of trust is derived, the signing operation can be completed independently without relying on the physical hardware root of trust being online in real time. The physical hardware root of trust periodically performs health checks on the virtual root of trust. If the health check fails, the virtual root of trust credential is automatically revoked, its subsequent signing operation is blocked, and an alarm is reported to the management node. Monitoring Agent Loading Steps: An AI behavior monitoring agent is loaded within the virtualized trusted execution environment instance or in a secure isolation domain bound to the virtualized trusted execution environment instance. The code hash value of the AI ​​behavior monitoring agent is verified by the virtual root of trust credential before loading; it can only run after successful verification. The AI ​​behavior monitoring agent is used to collect behavioral data from the AI ​​system in real time and to sign the behavioral data using the virtual root of trust credential. The behavioral data includes at least two dimensions from the AI ​​system's instruction decomposition depth, cross-domain correlation, policy mutation magnitude, and resource call expansion rate, used to calculate the cumulative entropy value of behavioral complexity. Monitoring Data Generation and Storage Steps: The AI ​​behavior monitoring agent packages the signed behavioral data, the virtual root of trust credential, and the certificate chain of the physical hardware root of trust together to generate a virtualized monitoring record. The virtualized monitoring record is timestamped using a timestamp service and stored in an tamper-proof storage medium, forming AI behavior audit evidence with immutable characteristics. The public key of the timestamp service is pre-measured and solidified by the physical hardware root of trust during the secure startup phase to ensure the trustworthiness of the timestamp itself. The signature data of the virtual root of trust credential includes a complete physical hardware root of trust certificate chain, which can verify the certainty of signature ownership, the integrity of the content, and the non-repudiation of the signature behavior.

2. The method according to claim 1, characterized in that, The physical hardware root of trust is at least one of a trusted platform module, a hardware security module, or a physically unclonable function module built into the processor; the virtualized trusted execution environment instance is a virtual machine, container, or secure enclave created based on hardware-assisted virtualization technology.

3. The method according to claim 1, characterized in that, The issuance process of the virtual root of trust certificate includes: the physical hardware root of trust uses its private key to sign the metric and attributes of the virtualized trusted execution environment instance, and generates an attribute certificate in X.509 format.

4. The method according to claim 1, characterized in that, The private key corresponding to the virtual root of trust is generated within the secure memory area of ​​the virtualized trusted execution environment instance. The plaintext of the private key never leaves this secure memory area and can only be called to perform signature operations within the instance. It cannot be exported or read.

5. The method according to claim 1, characterized in that, The specific process of the health verification is as follows: the physical hardware trust root actively pulls the runtime memory integrity metric value of the corresponding virtualization instance through the virtualization trusted execution environment manager and compares it with the baseline metric value stored when the credential is derived; if the metric values ​​are inconsistent, the health verification is determined to fail.

6. The method according to claim 1, characterized in that, During the operation of the AI ​​behavior monitoring agent, the virtual root of trust periodically performs dynamic integrity measurement on the code segment and core data segment of the monitoring agent; if the measurement value deviates abnormally, the signing permission is immediately terminated and an alarm is triggered.

7. The method according to claim 1, characterized in that, The public key of the timestamp service supports trusted updates: when the public key is rotated, it must be re-completed and solidified by the physical hardware root of trust before it can take effect. The entire process of public key updates is synchronously stored in a tamper-proof storage medium.

8. The method according to claim 1, characterized in that, The same physical hardware trust root is derived from multiple virtualized trusted execution environment instances on the same physical server node, and the keys and permissions of each credential are completely isolated; the physical hardware trust root performs health checks on all virtual trust roots in batches.

9. The method according to claim 1, characterized in that, When the virtualized trusted execution environment instance is hot-migrated online across physical nodes, the physical hardware root of trust of the target node re-measures the integrity of the migrated instance and derives a new virtual root of trust credential. The extended field of the new credential is written into the hash value of the old credential, and the continuous migration forms a credential hash chain. The monitoring record achieves continuous traceability throughout the entire lifecycle through the hash chain.

10. The method according to claim 1, characterized in that, In cloud-based elastic scaling scenarios, once a new virtualized trusted execution environment instance is created, the physical hardware root of trust automatically triggers integrity measurement and virtual root of trust credential derivation processes without manual intervention; when an instance is destroyed, the physical hardware root of trust automatically revokes the corresponding credential and records and archives it.

11. The method according to claim 1, characterized in that, In a multi-tenant cloud scenario, each tenant's virtualized trusted execution environment instance corresponds to an independent virtual root of trust credential. Tenants can only access the monitoring records associated with their own instance, and the physical hardware root of trust isolates and controls tenant permissions.

12. The method according to claim 1, characterized in that, The virtual root of trust adopts a symmetric key system: the physical hardware root of trust derives an independent symmetric signature key for each virtualization instance through a key derivation algorithm. The signing operation is completed locally on the instance, and the signature verification operation is uniformly executed by the physical hardware root of trust.

13. The method according to claim 1, characterized in that, The AI ​​behavior monitoring agent is deployed in the security isolation domain of the Hypervisor layer. It passively collects AI behavior data from multiple virtualization instances through memory mirroring and system call bypass mirroring, and calls the virtual trust root credential of the corresponding instance to complete the signature.

14. The method according to claim 1, characterized in that, The behavioral data also includes at least one of token generation volatility, tool call sequence entropy, and output semantic deviation, which together with instruction decomposition depth, cross-domain correlation, strategy mutation magnitude, and resource call expansion rate participate in the calculation of the cumulative entropy value of behavioral complexity. The entropy value calculation result is stored after being signed by the virtual trust root credential.

15. The method according to claim 1, characterized in that, The monitoring data generation and storage steps also include: synchronizing the virtualized monitoring records to the consortium blockchain nodes, achieving distributed storage through multi-node consensus, replacing the centralized timestamp service with block time, and verifying the credibility of the block time with the on-chain public key pre-measured by the physical hardware root of trust.

16. The method according to claim 1, characterized in that, The monitoring data generation and storage steps also include: synchronizing the virtualized monitoring records to at least one other monitoring node to form a distributed storage network.

17. The method according to claim 1, characterized in that, The AI ​​behavior monitoring agent obtains at least one behavioral data from memory snapshots, system call sequences, and network communication data packets during AI system runtime through a security monitoring interface provided by the virtualization environment. The security monitoring interface is provided by the virtualization trusted execution environment manager and cannot be shut down or bypassed by the main operating system.

18. The method according to claim 1, characterized in that, The code hash value of the AI ​​behavior monitoring agent is measured and signed by the physical hardware root of trust during deployment, and this signature value is managed as part of the virtual root of trust credentials.

19. The method according to claim 1, characterized in that, The calculation of the cumulative entropy value of behavioral complexity includes: the instruction decomposition depth entropy value is the logarithm of the number of AI execution steps decomposed into each human instruction with a preset base; the cross-domain association entropy value is calculated based on the number of different knowledge domains or data sources that the AI ​​autonomously associates in a single task; the strategy mutation amplitude entropy value is calculated based on the frequency and amplitude of mutations in the AI ​​behavior pattern under the condition of no external stimulus; the resource call expansion entropy value is calculated based on the cumulative expansion rate of computing resources, network resources and data resources autonomously called by the AI ​​on the time axis; the comprehensive entropy value is equal to the sum of the products of the entropy values ​​of each dimension and their corresponding weights, and the weights of each dimension are dynamically adjusted according to the monitoring scenario; the dynamic threshold is adaptively determined based on the statistical distribution of historical behavioral data, including using the baseline mean plus a preset multiple of the standard deviation as the upper limit of the threshold, or using the preset quantile of historical data as a reference value.

20. The method according to claim 1, characterized in that, The security isolation domain includes at least one of a Hypervisor layer, a stand-alone security chip, or a protected execution environment.

21. The method according to claim 1, characterized in that, When the physical hardware root of trust fails the health check of the virtual root of trust, the system automatically freezes the virtual root of trust credential, suspends all its signature operations, and marks the virtual instance as untrusted. The frozen virtual root of trust credential can only be unfrozen and restored after the physical hardware root of trust has completed the integrity measurement and health check again.

22. An AI behavior monitoring system based on a virtualized hardware root of trust, characterized in that, The system includes a physical hardware root of trust, a virtualized trusted execution environment (TEX) manager, an AI behavior monitoring agent, and a tamper-proof storage module. The physical hardware root of trust is connected to the TEX manager and is used to derive and issue virtual root of trust credentials for TEX instances. The AI ​​behavior monitoring agent runs inside the TEX instance or in a secure isolation domain bound to the TEX instance, and is connected to the tamper-proof storage module. It is used to collect behavioral data of the AI ​​system in real time, call the virtual root of trust credentials to sign the behavioral data, and store the signed monitoring data in the tamper-proof storage module. The system is configured to perform the method described in any one of claims 1 to 21.

23. A cloud server cluster or edge computing device, characterized in that, The system of claim 22 is integrated.

24. The system according to claim 22, characterized in that, The AI ​​behavior monitoring system, AI computing power scheduling system, and privacy computing compliance certification system work together to form cross-protocol collaborative protection: Before executing multi-platform intelligent computing power routing, the AI ​​computing power scheduling system must verify the validity of the virtual trust root credentials of all candidate computing power nodes with the AI ​​behavior monitoring system, and the scheduling process log is synchronized to the behavior monitoring system for chain-based evidence storage; when generating a compliance report, the privacy computing compliance certification system must cross-verify the consistency of the monitoring records of the AI ​​behavior monitoring system and the scheduling logs of the AI ​​computing power scheduling system in terms of timestamps and execution causal chains; if the verification of any system fails, a global alarm will be triggered and all cross-system task execution will be suspended.

Images