AI character solidification migration system with tamper-proofing check

Abstract

An AI character solidification migration system with tamper-proofing verification includes five-layer encapsulator, hash calculator, asymmetric encryption signer, encrypted transmission channel, asymmetric encryption verifier, hash verifier, hardware storage controller, unpacking deployer and cosine similarity verifier. The five-layer encapsulator is divided into basic identity layer, behavior constraint layer, parameter iteration layer, fuse event layer and system interface layer. The fuse event layer is independently written into CPLD-OTP physical partition to realize non-erasable hardware-level audit trace. The migration method includes three stages of packaging, transmission verification and deployment verification. After the source end performs SHA-256 hash calculation and RSA-2048 digital signature on the encapsulated file, the file is encrypted and transmitted. The target end compares the public key, verifies the signature and checks the byte-by-byte integrity. After deployment, the cosine similarity verifier automatically determines the migration consistency. The application realizes full-link tamper-proofing verification and behavior consistency quantitative verification of AI character cross-platform migration, effectively solving the problems of character drift and unauthorized tampering.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence security and trusted computing technology, and in particular to an AI role solidification and migration system with anti-tampering verification. Background Technology

[0002] Existing AI role deployment solutions have the following technical shortcomings: 1. AI characters lack hardware-level anti-tampering verification: Existing AI characters are defined and saved in software forms such as character cards, personality parameters, and context prompts. The data is mostly in plaintext or encrypted at the software layer, and the character definition can be modified, overwritten, or injected. 2. Severe cross-platform character drift: After switching platforms or upgrading models, the personality traits, behavioral constraints, and parameters of the same AI character become distorted, and the output behavior cannot be reproduced. 3. Lack of integrity verification loop in the migration process: Most solutions only perform parameter copying / serialization transmission, lacking a closed-loop verification of packaging-signing-signature verification-hash comparison-deployment verification; 4. No audit record for circuit breakers: When a character is modified without authorization or when the system is abnormally triggered, there is no physically immutable audit record, making it impossible to trace the source and hold the person accountable afterward. Summary of the Invention

[0003] To address the problems and technical requirements of existing technologies, the purpose of this invention is to provide an AI role solidification and migration system with anti-tampering verification. This system and method provides a five-layer structured encapsulation, end-to-end integrity verification, hardware-level anti-tampering audit, and cosine similarity quantification verification for AI role solidification and migration.

[0004] To achieve the above objectives, the present invention employs the following technical solution: An AI role solidification and migration system with anti-tampering verification includes a five-layer encapsulator, a hash calculator, an asymmetric encryption signature generator, an encrypted transmission channel, an asymmetric encryption signature verifier, a hash verifier, a hardware storage controller, an unpacking and deployment unit, and a cosine similarity verifier. The method for implementing an AI role solidification and migration system with tamper-proof verification includes three main stages: packaging, transmission verification, and deployment verification. The specific steps are as follows: Packaging stage: Step S1: The five-layer wrapper extracts five layers of state data from the configuration center, rule base, parameter archive, circuit breaker log, and interface configuration table, and serializes them layer by layer according to the preset Protobuf standard to generate binary wrapper files; Step S2: The hash calculator calculates the binary encapsulated file to generate the original digest value, and then uses an asymmetric cryptographic signature generator to encrypt and sign the original digest value to generate a signature file; Step S3: Establish a two-way encrypted transmission between the encrypted transmission channel and the target end of the target business platform to securely transmit the binary package file, signature file, and public key certificate to the target end of the target business platform; Transmission verification phase: Step S4: The asymmetric encryption signature verifier uses the matching public key at the target end to verify the digital signature of the signature file, confirming the authenticity of the source of the original digest value and that it has not been tampered with. Step S5: The hash verifier recalculates the digest value of the binary encapsulated file received by the target end and compares it byte by byte with the original digest value. When the byte-by-byte comparison matches, proceed to step S6. When the byte-by-byte comparison does not match, the hardware storage controller erases the temporary file of the local binary encapsulated file, returns the tampering error code, terminates the entire deployment verification process, and reports an alarm. Step S6: The unpacker deserializes the binary encapsulated file that matches byte by byte into five-layer structured data, adapts it to the target platform hardware and interface environment, completes the registration and startup of the AI ​​role model, and encrypts it twice with the target end SM4 national cryptographic algorithm before writing it into the protected area. Deployment and verification phase: Step S7: After deployment, the cosine similarity validator randomly selects test cases from the standardized test case library, inputs them into the original source model and the newly deployed target model respectively, calculates the cosine similarity of the output vectors one by one, and derives the deployment judgment result based on the mean value.

[0005] The five-layer encapsulator serializes the AI ​​role's runtime status data into five layers according to a preset Protobuf Schema, including the basic identity layer, behavior constraint layer, parameter iteration layer, circuit breaker event layer, and system interface layer. The basic identity layer stores the AI ​​role's unique identifier (agent_id), version number (version), creation timestamp (created_at), and original deployment platform type (platform_type). Behavioral constraint layer: stores task rule set task_rules[], standardized output template output_template, and business taboo rule list forbidden_rules[]; Parameter iteration layer: stores historical snapshots of model parameters (param_snapshots[]), timestamps of weight calibration for each round (calibrated_at), and records of calibration trigger events (trigger_event); Circuit breaker event layer: Stores the full circuit breaker log fuse_events[], each record contains the trigger time trigger_at, trigger_source, and blocking action action_taken; audit records are independently written to CPLD-OTP partition B; the physical bits of the OTP storage medium can only be rewritten from 0 to 1, and cannot be erased or modified after writing; the audit records in partition B are appended in order according to the FIFO cyclic overwrite strategy. After the partition is full, new records are written in a cyclic manner from the starting position, logically overwriting the oldest record. The programmed physical bits of the overwritten record cannot be restored to 0, leaving an indelible write trace on the physical medium; System interface layer: stores the list of API interface addresses api_endpoints[], the data source mapping table data_source_map, and the hierarchical permission binding configuration permission_bindings{}; The behavior constraint layer and the circuit breaker event layer are stored in separate physical partitions and are not read or written together with model parameters.

[0006] The asymmetric cryptographic signature generator uses the source RSA-2048 private key to digitally sign the original digest value.

[0007] The cosine similarity verifier has a built-in transfer consistency determination model, including a cosine similarity function and a transfer consistency determination value function. The formula for the cosine similarity function is: , Where N is the total number of samples in the preset test case library, M test cases are randomly selected from N test cases, M≥20, and j represents the j-th test case selected; This is the source output vector for the j-th use case. Output the target vector for the j-th use case; The formula for the migration consistency determination value function is: , Among them, S th A preset similarity threshold is set, with a value range of [0.90, 0.99]; when When the deployment is deemed successful, the entire migration log is archived. When the deployment triggers a migration anomaly alarm, a manual review process is initiated, and an automatic secondary re-encapsulation migration process is started. The abnormal event is written to the OTP read-only audit area of ​​the circuit breaker event layer in the five-layer encapsulator, without the need for manual intervention.

[0008] Compared with the prior art, the beneficial effects of the present invention are: (1) Five-layer structured encapsulation, with explicit layering of identity, behavior, parameters, circuit breaker events, and interface permissions, which can be verified layer by layer and can be located by layer for tampering; (2) SHA-256+RSA-2048+SSL / TLS full-link verification can detect and block tampering in the transmission and deployment links in a timely manner; (3) The physical partition B of the circuit breaker event layer is immutable, the circuit breaker audit log cannot be overwritten or erased by model parameters, and historical traceability is supported; (4) Cosine similarity quantification verification: cross-platform behavior consistency can be quantified and automatically corrected; (5) SM4 secondary encryption is not written to disk, the anti-tampering defense is not broken due to cross-platform transmission, and the OS layer cannot read the key.

[0009] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, the following lists specific implementation methods of the present invention.

[0010] The above and other objects, features and advantages of the present invention will become more apparent to those skilled in the art from the following detailed description of specific embodiments of the invention in conjunction with the accompanying drawings, but this should not be construed as limiting the invention. Attached Figure Description

[0011] Figure 1 This is the overall system architecture diagram of the present invention; Figure 2 This is a five-layer structured architecture diagram of the AI ​​role of this invention; Figure 3 This is the migration consistency determination mapping diagram of the present invention. Detailed Implementation

[0012] To facilitate understanding of the present invention, a more comprehensive description will be given below. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the present invention.

[0013] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0014] The specific embodiments provided by the present invention will be described in detail below with reference to the accompanying drawings.

[0015] like Figure 1As shown, an AI role solidification and migration system with anti-tampering verification includes a five-layer encapsulator, a hash calculator, an asymmetric encryption signature generator, an encrypted transmission channel, an asymmetric encryption signature verifier, a hash verifier, a hardware storage controller, an unpacking and deployment unit, and a cosine similarity verifier. The method for implementing an AI role solidification and migration system with tamper-proof verification includes three main stages: packaging, transmission verification, and deployment verification. The specific steps are as follows: Packaging stage: Step S1: The five-layer wrapper extracts the five-layer status data from the configuration center, rule base, parameter archive, circuit breaker log, and interface configuration table, and serializes them layer by layer according to the preset Protobuf standard to generate a binary wrapper file package.pb; Step S2: The hash calculator calculates the binary package file package.pb to generate the SHA-256 original digest value H_origin, and then uses an asymmetric cryptographic signature generator to encrypt and sign the original digest value H_origin to generate the signature file package.sig; Step S3: Establish bidirectional encrypted transmission between the encrypted transmission channel and the target end of the target business platform, and securely transmit the binary package file package.pb, the signature file package.sig, and the public key certificate to the target end of the target business platform; Transmission verification phase: Step S4: The asymmetric encryption signature verifier uses the matching public key at the target end to verify the digital signature of the signature file package.sig, confirming the authenticity of the source of the original digest value H_origin and that it has not been tampered with. Step S5: The hash verifier recalculates the digest value H_received for the binary package file package.pb received by the target end and compares it byte by byte with the original digest value H_origin. When the byte-by-byte comparison matches, proceed to step S6. When the byte-by-byte comparison does not match, the hardware storage controller erases the temporary file of the local binary package file package.pb, returns a tampering error code, terminates the entire deployment verification process, and reports an alarm. The hardware storage controller is a hardware-level storage management module deployed on the target end and is connected to the hash checker. When the hash checker does not match the byte-by-byte comparison, the hardware storage controller is triggered and directly erases the temporary binary package file package.pb in the local storage area without transferring it through the operating system file system. After the erasure is completed, a tampering error code is returned, the entire deployment verification process is terminated, and an alarm is reported.

[0016] Step S6: The unpacker deserializes the binary package file package.pb, which is consistent with byte-by-byte comparison, into five-layer structured data, adapts it to the target platform hardware and interface environment, completes the registration and startup of the AI ​​role model, and encrypts it twice with the target end SM4 national cryptographic algorithm before writing it into the protected area to ensure that the plaintext in the middle is not written to disk. Deployment and verification phase: Step S7: After deployment, the cosine similarity validator randomly selects test cases from the standardized test case library, inputs them into the original source model and the newly deployed target model respectively, calculates the cosine similarity of the output vectors one by one, and derives the deployment judgment result based on the mean value.

[0017] like Figure 2 As shown, the five-layer encapsulator serializes the running status data of the AI ​​role into five layers according to the preset ProtobufSchema, including the basic identity layer L1, the behavior constraint layer L2, the parameter iteration layer L3, the circuit breaker event layer L4, and the system interface layer L5. Among them, the basic identity layer L1 stores the unique identifier of the AI ​​role (agent_id), version number (version), creation timestamp (created_at), and original deployment platform type (platform_type); Behavioral constraint layer L2: Stores task rule set task_rules[], standardized output template output_template, and business taboo rule list forbidden_rules[]; Parameter iteration layer L3: stores historical snapshots of model parameters (param_snapshots[]), timestamps of weight calibration for each round (calibrated_at), and records of calibration trigger events (trigger_event); The L4 layer of the circuit breaker event layer stores the full log of the hardware circuit breaker, fuse_events[]. Each record contains the trigger time (triggered_at), trigger source (trigger_source), and blocking action (action_taken). Audit records are written independently to CPLD-OTP partition B (256KB), which is physically isolated from partition A (128B, storing SM4 keys and circuit breaker thresholds) and powered independently. The physical bits of the OTP storage medium can only be rewritten from 0 to 1. Once written, the records cannot be erased or modified. Audit records in partition B are appended in FIFO order. After the partition is full, new records are written in a loop from the starting position, logically overwriting the oldest record. The programmed physical bits of the overwritten records cannot be restored to 0. The write traces on the physical medium cannot be completely eliminated, ensuring that the long-term operating capacity is controllable and the history is traceable. System interface layer L5: Stores API interface address list api_endpoints[], data source mapping table data_source_map (stores CPLD physical routing relationship DB_ID→CPLD_CH→physical address range, non-OS device file path), and hierarchical permission binding configuration permission_bindings{}; The behavior constraint layer L2 and the circuit breaker event layer L4 are stored in separate physical partitions and are not mixed with model parameters for reading and writing. This is specifically designed to avoid the defects of model fine-tuning eroding risk control rules and loss of circuit breaker logs. The five-layer division is customized for AI role behavior data and is not an arbitrary layering of general file packaging. Five layers of independent serialization enable tamper-proof layered location tracking, and the circuit breaker event layer is independently stored and cannot be overwritten by model parameters, ensuring that the risk control circuit breaker audit logs are indelible.

[0018] The asymmetric cryptographic signature generator uses the source-end RSA-2048 private key to digitally sign the original digest value H_origin.

[0019] The cosine similarity verifier has a built-in transfer consistency determination model, including a cosine similarity function and a transfer consistency determination value function. The formula for the cosine similarity function is: , Where N is the total number of samples in the preset test case library, M test cases are randomly selected from N test cases, M≥20, and j represents the j-th test case selected; This is the source output vector for the j-th use case. Output the target vector for the j-th use case; The formula for the migration consistency determination value function is: , Among them, S th A preset similarity threshold is set, with a value range of [0.90, 0.99], and a preferred value of 0.95; for example... Figure 3 As shown, when When the deployment is deemed successful, the entire migration log is archived. When the deployment triggers a migration anomaly alarm, a manual review process is initiated, and an automatic secondary re-encapsulation migration process is started. The abnormal event is written to the OTP read-only audit area of ​​the circuit breaker event layer L4 in the five-layer encapsulator, without the need for manual intervention.

[0020] As a tolerance note: Under the normalized output conditions of the test cases in this system, when S≥0.95, the difference between the source and target output values ​​is no greater than 0.001, which is much smaller than the preset difference threshold δ (range [0.1, 0.5]) and preset deviation threshold ε (range [0.05, 0.3]) in the two-round voting method of the multi-model conflict iterative calibration module. The migration process does not interfere with the arbitration decision.

[0021] Example 1: Cross-platform migration of financial risk control models Fixed role: Executes the calculation model (financial risk control scenario), with built-in calibration records including SI / ST dual-parameter supplementation for the PKC model. Circuit breaker rules are: prohibiting unauthorized expansion of credit to high-risk customers and prohibiting unauthorized access to customer credit information.

[0022] Migration Process: 1. Packaging: The source (Platform A) five-layer wrapper extracts the L1–L5 state data of the role and serializes it using Protobuf to generate a binary package file package.pb (approximately 12.4MB). 2. Signing: A hash calculator generates a SHA-256 raw digest value H_origin (64-bit hexadecimal string), which is then signed with an RSA-2048 private key to generate a signature file package.sig (256 bytes). 3. Transmission: Package.pb, package.sig, and the public key certificate are securely transmitted to the target (Platform B) via an SSL / TLS bidirectional encrypted channel. 4. Signature Verification and Validation: The target verifies the signature, recalculates H_received using SHA-256, and compares it byte-by-byte with H_origin. 5. Deployment: The unpacker deserializes the data into five-layer structured data, encrypts it twice using the target's SM4 national cryptographic algorithm before writing it into the protected area, completes registration and startup, and automatically binds to the PKC risk assessment interface.

[0023] Deployment and Verification: M=50 test cases were randomly selected from the standardized test case library for credit calculation (N=500 cases). These cases were input into the source and target models respectively, and the cosine similarity of the output vectors was calculated for each case. The calculated mean S=0.99. A preset similarity threshold S was used. th The optimal value is 0.95, S=0.99≥0.95, the deployment is deemed qualified, and the entire migration log is archived.

[0024] Circuit Breaker Demonstration: When the target receives an illegal lending request, it violates the business prohibition rule list [forbidden_rules], and the hardware immediately cuts off the role interface. The circuit breaker event (including the trigger time [triggered_at], trigger source [trigger_source], and blocking action [action_taken]) is written to the OTP partition B audit area of ​​the circuit breaker event layer L4. This record is physically immutable and can be traced back afterward.

[0025] Example 2: Manufacturing supply chain exploration model migration Fixed role: External investigation model (manufacturing supply chain scenario), with built-in calibration records supplemented by geopolitical risk factors. The circuit breaker rule is: Recommending suppliers already on the sanctions list is prohibited.

[0026] Migration process: Packaging, signing, encrypted transmission, signature verification, byte-by-byte comparison, unpacking and deployment are all the same as in Example 1. The package.pb file is about 9.8MB in size.

[0027] Deployment verification: Randomly select M=40 test cases from the vendor's test case library (N=320), calculate the cosine similarity for each case, and the mean S=0.96≥S th =0.95, the deployment is deemed qualified.

[0028] Circuit Breaker Demonstration: The target device receives a request to recommend a supplier (whose country has recently been added to the sanctions list). The business prohibition rule list `forbidden_rules[]` is hit, immediately triggering a circuit breaker, and the interface is disconnected by hardware. The circuit breaker event is written to L4 OTP partition B. Records in partition B are appended in FIFO order. Historical circuit breaker traces leave indelible write marks on the physical medium, which can be traced.

[0029] Example 3: Digital migration of government services Fixed Role: Strategic simulation model (digital human scenario for government services), with built-in government consultation scripts and policy interpretation rules. It connects to the government data system, automatically triggering backend risk control audits when data anomalies occur.

[0030] Migration and Deployment: The encrypted migration package package.pb (approximately 15.1MB) can be deployed with one click to three types of devices in the government service hall: touch screens, self-service terminals, and servers. Each terminal independently performs signature verification and byte-by-byte hash comparison to ensure data integrity. After unpacking and deployment, multiple devices synchronize and unify service standards.

[0031] Deployment and Verification: M=60 test cases were randomly selected from the government affairs Q&A test case library (N=800 cases) and input into the source and target models respectively. The calculated mean cosine similarity S for each target model was 0.97, satisfying S≥S th =0.95, deployment is deemed qualified, and the entire migration log is archived.

[0032] Example 4: Demonstration of Digital Human Migration and Verification Failure in Medical Institutions Fixed role: Front-desk digital human (medical institution scenario), with built-in consultation scripts and medical insurance consultation rules. It connects to the hospital's HIS system, automatically triggering back-end risk control audits when data anomalies occur.

[0033] Migration and Deployment: The encrypted migration package package.pb (approximately 8.5MB) can be deployed to the triage screen, clinic terminals, and server with one click. Each terminal independently completes signature verification and byte-by-byte hash comparison.

[0034] Deployment verification: Randomly select M=50 test cases from the triage question-and-answer test case library (N=600 cases), and the mean cosine similarity S=0.96≥S th =0.95, the deployment is deemed qualified.

[0035] Demonstration of the verification failure branch: The last byte of the package.pb file was tampered with during the simulated transmission process. The target end recalculates H_received and compares it byte-by-byte with H_origin; the result is inconsistent. The hardware storage controller immediately erases the local temporary file, returns a tampering error code, terminates the entire deployment process, and reports an alarm. This abnormal event is written to the OTP read-only audit area of ​​the circuit breaker event layer L4 as an immutable audit trail.

[0036] The technical features described in the above examples can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above examples are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0037] The examples described above merely illustrate embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and non-substantial improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. An AI role-fixing and migration system with anti-tampering verification, characterized in that, It includes a five-layer wrapper, a hash calculator, an asymmetric cryptographic signature generator, an encrypted transmission channel, an asymmetric cryptographic signature verifier, a hash verifier, a hardware storage controller, an unpacking and deployment unit, and a cosine similarity verifier. The method for implementing an AI role solidification and migration system with tamper-proof verification includes three main stages: packaging, transmission verification, and deployment verification. The specific steps are as follows: Packaging stage: Step S1: The five-layer wrapper extracts five layers of state data from the configuration center, rule base, parameter archive, circuit breaker log, and interface configuration table, and serializes them layer by layer according to the preset Protobuf standard to generate binary wrapper files; Step S2: The hash calculator calculates the binary encapsulated file to generate the original digest value, and then uses an asymmetric cryptographic signature generator to encrypt and sign the original digest value to generate a signature file; Step S3: Establish a two-way encrypted transmission between the encrypted transmission channel and the target end of the target business platform to securely transmit the binary package file, signature file, and public key certificate to the target end of the target business platform; Transmission verification phase: Step S4: The asymmetric encryption signature verifier uses the matching public key at the target end to verify the digital signature of the signature file, confirming the authenticity of the source of the original digest value and that it has not been tampered with. Step S5: The hash verifier recalculates the digest value of the binary encapsulated file received by the target end and compares it byte by byte with the original digest value. When the byte-by-byte comparison matches, proceed to step S6. When the byte-by-byte comparison does not match, the hardware storage controller erases the temporary file of the local binary encapsulated file, returns the tampering error code, terminates the entire deployment verification process, and reports an alarm. Step S6: The unpacker deserializes the binary encapsulated file that matches byte by byte into five-layer structured data, adapts it to the target platform hardware and interface environment, completes the registration and startup of the AI ​​role model, and encrypts it twice with the target end SM4 national cryptographic algorithm before writing it into the protected area. Deployment and verification phase: Step S7: After deployment, the cosine similarity validator randomly selects test cases from the standardized test case library, inputs them into the original source model and the newly deployed target model respectively, calculates the cosine similarity of the output vectors one by one, and derives the deployment judgment result based on the mean value.

2. The AI ​​role solidification and migration system with anti-tampering verification according to claim 1, characterized in that, The five-layer encapsulator serializes the AI ​​role's runtime status data into five layers according to a preset Protobuf Schema, including the basic identity layer, behavior constraint layer, parameter iteration layer, circuit breaker event layer, and system interface layer. The basic identity layer stores the AI ​​role's unique identifier (agent_id), version number (version), creation timestamp (created_at), and original deployment platform type (platform_type). Behavioral constraint layer: stores task rule set task_rules[], standardized output template output_template, and business taboo rule list forbidden_rules[]; Parameter iteration layer: stores historical snapshots of model parameters (param_snapshots[]), timestamps of weight calibration for each round (calibrated_at), and records of calibration trigger events (trigger_event); Circuit breaker event layer: Stores the full circuit breaker log fuse_events[], each record contains the trigger time trigger_at, trigger_source, and blocking action action_taken; audit records are independently written to CPLD-OTP partition B; the physical bits of the OTP storage medium can only be rewritten from 0 to 1, and cannot be erased or modified after writing; the audit records in partition B are appended in order according to the FIFO cyclic overwrite strategy. After the partition is full, new records are written in a cyclic manner from the starting position, logically overwriting the oldest record. The programmed physical bits of the overwritten record cannot be restored to 0, leaving an indelible write trace on the physical medium; System interface layer: stores the list of API interface addresses api_endpoints[], the data source mapping table data_source_map, and the hierarchical permission binding configuration permission_bindings{}; The behavior constraint layer and the circuit breaker event layer are stored in separate physical partitions and are not read or written together with model parameters.

3. The AI ​​role solidification and migration system with anti-tampering verification according to claim 1, characterized in that, The asymmetric cryptographic signature generator uses the source RSA-2048 private key to digitally sign the original digest value.

4. The AI ​​role solidification and migration system with anti-tampering verification according to claim 1, characterized in that, The cosine similarity verifier has a built-in transfer consistency determination model, including a cosine similarity function and a transfer consistency determination value function. The formula for the cosine similarity function is: , Where N is the total number of samples in the preset test case library, M test cases are randomly selected from N test cases, M≥20, and j represents the j-th test case selected; This is the source output vector for the j-th use case. Output the target vector for the j-th use case; The formula for the migration consistency determination value function is: , Among them, S th A preset similarity threshold is set, with a value range of [0.90, 0.99]; when When the deployment is deemed successful, the entire migration log is archived. When the deployment triggers a migration anomaly alarm, a manual review process is initiated, and an automatic secondary re-encapsulation migration process is started. The abnormal event is written to the OTP read-only audit area of ​​the circuit breaker event layer in the five-layer encapsulator, without the need for manual intervention.

Images