A method for identifying evidence in cybersecurity compliance assessments based on local AI large-scale models.

By adopting an evidence verification method based on a local AI large model, compiling an evidence verification automaton and using an out-of-model deterministic verifier, the problem of evidence accuracy and completeness in the information security level assessment is solved, achieving efficient evidence identification and verification, and reducing redundant processing and incorrect combinations.

CN123122934APending Publication Date: 2026-11-17XINJIANG QUANTUM COMM TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611345683.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-01
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for identifying evidence in cybersecurity compliance assessments struggle to effectively verify the accuracy and completeness of evidence when the amount and format of evidence increase. Furthermore, after version updates, unaffected evidence is repeatedly processed, and the lack of machine-executable state transition relationships leads to errors in information retrieval and evidence splicing.

Method used

We employ an evidence verification method based on a local AI large model. By compiling an evidence verification automaton, limiting the access rights of evidence, using an external deterministic verifier to reenact the evidence location and normalized facts, establishing a witness hypergraph, solving closed state transition and source independence constraints, and generating verifiable atomic evidence records, we ensure the accuracy and integrity of the evidence.

Benefits of technology

It reduces secondary uncertainty in evidence sources, prevents incorrect combinations of different assets, different time periods, and replication sources, clarifies the task of supplementing missing slots and insufficient sources, and improves the accuracy and efficiency of the assessment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123122934A_ABST
    Figure CN123122934A_ABST
Patent Text Reader

Abstract

This invention discloses a method for identifying evidence in cybersecurity level protection assessments based on a local AI large model, specifically involving cybersecurity level protection assessments and computer data verification technology. The invention compiles applicable assessment items into an evidence verification automaton. The local AI large model only proposes candidate evidence containing the relative location of the source, interpretation rules, and state transition identifiers. An external deterministic verifier recreates the source interval, transformation trajectory, and normalized facts from the locked original evidence version. A witness hypergraph is constructed using verifiable atomic evidence records. Candidate evidence packets are output only when the closed witness subgraph reaches the acceptance state. If it cannot reach the acceptance state, a targeted supplementary collection task is generated with a minimal unmet constraint kernel, and local recalculation is performed. This prevents model-generated content, refutation facts without scanning range, and evidence from different assets and time periods from directly forming assessment basis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of cybersecurity level protection assessment technology, and more specifically, to a method for identifying evidence in level protection assessment based on a local AI large model. Background Technology

[0002] Cybersecurity level protection assessment requires collecting configuration files, command echoes, logs, scan reports, screenshots, policy documents, and interview records, and mapping the collected content to applicable assessment items. When the amount of evidence increases and its formats become heterogeneous, a local large language model can assist in retrieving fragments and extracting fields.

[0003] However, the existing methods for identifying evidence in cybersecurity compliance assessments still have the following technical shortcomings in practical applications: Intelligent assessment schemes typically generate assessment item evidence and conclusions from models, while information retrieval schemes can also add source anchors, time information, evidence graphs, or confidence levels; however, citations, normalized values, and anchors are still model outputs; even if another model judges the accuracy, it still cannot prove that the normalized facts can be recalculated from the locked version according to the same rules. The quantity, similarity, or score of relevant evidence are not equivalent to sufficient evidence; the evaluation unit may simultaneously require object identity, configuration status, running results, and source independence; when there is a lack of machine-executable state transition relationships, different assets, different time periods, or conflicting segments may still be spliced ​​together, especially "not found" cannot be based solely on the fact that the document does not mention it, but also requires proof that the scan range is complete; Current incomplete processing methods mostly involve returning for modification, listing missing items, or performing a full re-examination. When missing items do not correspond to specific failure constraints, it is still necessary to manually determine the assets, time, interface, and fields to be collected. After a version update, a full recalculation will also repeatedly process unaffected evidence. Therefore, it is necessary to restrict the evidence access rights of the model, reconstruct the facts from the original version to verify whether the evidence combination can reach the acceptable state, and transform the reasons for failure into an executable re-collection task. To this end, a method for identifying evidence for information security level protection assessment based on a local AI large model is introduced. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for identifying evidence in information security compliance assessments based on a local AI large model.

[0005] To achieve the above objectives, this invention provides the following technical solution: Obtaining a confirmed standard rule table, project applicability information, and asset benchmark information; compiling the combination of applicable assessment items and the tested object into an evidence verification automaton. The automaton includes an initial state, intermediate states, an acceptance state, and a pending acquisition state; state transition rules record evidence slots, object domains, valid time domains, allowed source types, source independence, factual polarity, interpretation rules, and mandatory or alternative relationships.

[0006] The standard rule table maps the detection object, detection method, expected control state, and result recording requirements to object identity slots, allowed source types, fact slots, and state transition guard conditions, respectively. During compilation, inapplicable rules are trimmed according to the security protection level, general security requirements, and applicable extension requirements, and the standard version identifier, rule table summary, and project applicability summary are written into the automaton fingerprint of the evidence verification automaton.

[0007] Multi-source original evidence is appended to solidify the original evidence version and content digest, generating an addressable evidence version with character-to-byte mapping or media coordinate back-mapping. Each reproducible transformation record includes an operation identifier, operation version, input / output digest, parameters, and position back-mapping. Text position back-mapping points to the original byte range, while media position back-mapping points to page numbers and region coordinates or start and end timecodes. Reproducibility is rejected if the operation version or parameters are inconsistent.

[0008] Evidence fragments are retrieved based on the object domain, time domain, and allowed source type of the state transition rules. The evaluation unit identifier, allowed state transition identifiers and interpretation rule identifier set, fragment position, and fragment content are input into the AI ​​big model located within the local isolated inference boundary. The AI ​​big model only outputs the relative position within the fragment, direct citation, candidate assets and time, interpretation rule identifier, state transition identifier, rule parameters, and candidate normalized facts. It must not be written into the original evidence library, modify the automaton, or trigger the acceptance state.

[0009] The external deterministic validator first converts the half-open relative character range output by the model into the half-open source byte range of the same addressable evidence version based on the fragment's baseline byte position and character-byte mapping. The conversion relationship is as follows:

[0010] in, and These represent the start and end positions of the source byte range, respectively. This indicates the base byte position of the evidence fragment within the addressable evidence version. and These represent the start and end positions of the relative character range within the segment, respectively. This represents the cumulative mapping from the character boundary of a fragment to the relative byte offset. The interval is left-closed and right-open; a candidate proposal is rejected if the relative start position is less than zero, the relative end position is greater than the number of characters in the fragment, or the relative start position is not less than the relative end position.

[0011] The out-of-model deterministic validator rereads the original evidence version and calculates the content summary, recursively transforms the trajectory, and retrieves direct citations by source byte range; subsequently, it only calls the deterministic interpretation rules registered in the state transition rules to recalculate the normalized facts. The deterministic interpretation rules include key-value extraction, enumeration mapping, time parsing, address normalization, and threshold comparison. The input fields, parameter types, valid domains, and output types of the rules are fixed in the rule table.

[0012] A verifiable atomic evidence record is generated only when the content summary, transformed trajectory output summary, direct citation, recalculated normalized fact, and state transition guard condition all pass. The record fingerprint binds the source version summary, source byte range, interpretation rule, rule parameters, normalized fact, and state transition identifier, with the following calculation relationship:

[0013]

[0014] in, This represents the fingerprint of the i-th verifiable atomic evidence record. Represents a collision-resistant summary function. This indicates a summary of the source version. and Indicates the source range, Indicates the interpretation rule identifier, This represents the rule parameters that have undergone boundary checks. This represents a normalized fact that is recalculated with certainty. Indicates a state transition identifier.

[0015] When there are issues such as location out of bounds, summary change, inconsistency in direct citations, failure of transformation trajectory replay, disallowed interpretation rules, out-of-bounds interpretation rule parameters, inconsistent recalculated normalized facts, non-unique assets, or non-intersecting time domains, the out-of-model deterministic validator rejects candidate evidence proposals and refuses to record them into the witness hypergraph.

[0016] When the state transition rule requires that the absence of a target item be used as a negative fact, only verifiable atomic evidence records that simultaneously carry scan range credentials are accepted. The scan range credentials include at least the target asset set, the inspected field or address range, the acquisition command or scanner version, execution parameters, start time, end time, completion status, and original output version. If the completion status is abnormal, the target asset is not covered, or the scan range is smaller than required by the state transition rule, the state transition corresponding to the negative fact will not be triggered.

[0017] A witness hypergraph for the evaluation unit is established, using verifiable atomic evidence records that have passed deterministic verification as evidence nodes, a single record or group of records required for a state transition as witness hyperedges, and automata states as state nodes. File copies, exports, or transfers with the same original acquisition root identifier are counted as having the same source; conflict edges are established when fact values ​​are opposite and the object domain and valid time domain overlap.

[0018] A closed witness subgraph must contain all input evidence nodes, corresponding source version nodes, interpretation rule nodes, and continuous state transitions from the initial state to the accepting state required for each witness superedge. The subgraph is considered an accepting witness only when there are no dangling inputs, the number of source independence requirements meets the rule, there are no unresolved conflicts, and all state transition guard conditions are met. No confidence score or similarity score replaces the source independence constraint, conflict constraint, or state transition guard condition. Its main constraints can be expressed as:

[0019] in, This represents the set of required slots for the evaluation unit u. This represents the set of candidate records that have passed verification. This indicates whether record i can trigger the state transition corresponding to slot q. Indicates whether to select record i. This represents a set of conflicting record pairs. Let R represent the set of selected records, and let R represent the reachability operation performed according to the state transition rule. and These represent the initial state and the acceptance state, respectively. Alternate slots, source independence, and the range of negative facts scans are written into the solution model as hard constraints of the same level.

[0020] When no witnesses are available, slot coverage, substitution relations, source independence, scan range, object domain, effective time domain, and conflict exclusion constraints are guarded with hypothesis literals. The SAT or SMT solver is then called to obtain the unsatisfiable kernel, which is then reduced to a minimal unsatisfiable constraint kernel through deletion checks. "Minimal" means that no proper subset of this kernel makes the constraints unsatisfiable, not that it claims to have the smallest cardinality among all unsatisfiable kernels. The relationship is as follows:

[0021] in, Let a represent the basic constraints on verified records and state transitions in the evaluation unit u, where U represents the minimum unsatisfied constraint kernel, and a and Indicates a hypothetical statement. This indicates that the constraint within the parentheses cannot be satisfied. This indicates that the constraints within the parentheses are satisfied. If deleted... If the condition is still not met, delete the character and continue testing; until the condition is met after deleting any character.

[0022] The minimal unmet constraints are transformed into supplementary collection tasks, each containing the target asset, target time domain, evidence type, collection interface, expected fields, and interpretation rule identifiers: missing slots correspond to evidence type and fields, insufficient sources correspond to new collection roots, insufficient range corresponds to expanded scanning, conflicts correspond to higher-priority measured sources within overlapping time periods, and missing time corresponds to sources with effective times or event times. Supplementary evidence undergoes reprocessing through steps S2 to S4 and must not bypass external deterministic verification.

[0023] Establish a dependency index from the original evidence version, reproducible change records, reverifiable atomic evidence records, witness superedges to the evaluation unit. When the original evidence version is replaced by an appended new version, the interpretation rules or automaton fingerprints change, or a new evidence is returned by a re-collection task, the subsequent object is marked as invalid according to the dependency index and recalculated; the witness subgraph region that does not depend on the changed object and whose re-verified fingerprint is valid is retained.

[0024] The candidate evaluation evidence package includes an automaton fingerprint, a selected verifiable atomic evidence record, a record fingerprint, a state transition sequence, a conflict resolution record, and a review reason. The review process does not invoke a large AI model; instead, it reproduces each verifiable atomic evidence record and its acceptance state based on the original evidence version pointed to by the candidate evaluation evidence package, the verifiable evolution trajectory, and the deterministic interpretation rules. If reproduction fails, the corresponding candidate evaluation evidence package is revoked and enters a state of pending supplementary collection or manual review.

[0025] The technical effects and advantages of this invention are as follows: By limiting the large AI model to a proposer without evidence admission rights, and having an external deterministic verifier reproduce the location, transformation, and normalization of facts from the locked source, accepted records can be repeatedly verified without calling the model again, reducing the secondary uncertainty of the evidence source caused by probabilistic citation verification. By compiling the evaluation rules into an evidence verification automaton and requiring closed state transitions, source independence, scan range, and conflict-free conditions in the witness hypergraph, "retrieving relevant content" is separated from "the evidence is sufficient to support the evaluation unit," preventing the incorrect combination of refutation facts from different assets, different time periods, duplicate sources, or no coverage. By obtaining the minimum unmet constraint kernel from the verification failure model and mapping it to the re-acquisition task that can be executed by the acquisition interface, missing slots, insufficient sources, insufficient ranges and conflict resolution can be assigned to specific objects respectively; the dependency index ensures that re-acquisition and version changes only trigger failures in the relevant areas, and retains the unaffected and re-verified accepted witnesses. Attached Figure Description

[0026] Figure 1 This is a flowchart of the evidence identification method for information security compliance assessment based on a local AI large model according to the present invention. Figure 2 This is a flowchart of the minimal unsatisfied constraint-driven supplementary sampling and dependency failure handling of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1 provides a reproducible software path. The system is deployed on the evaluation terminal or the institution's intranet and includes standard compilation, evidence versioning, local model inference, deterministic verification, graph storage, constraint solving, supplementary data collection scheduling, and verification services.

[0029] like Figures 1-2 As shown, the method is executed in the following order: standard compilation, evidence versioning, model candidate proposal, deterministic verification, witness status verification, constraint kernel solving, and targeted supplementary sampling. The output identifier of the previous step serves as the input of the next step; the candidate outputs of the AI ​​large model and the verified evidence are stored separately, and they do not share the admission status field.

[0030] S100, Prepare a standard rule table. The assessment manager, based on the current basic requirements for graded protection adopted at the time of application, assessment requirements, and applicable extended requirements for the project, will create a confirmed structured rule table. Each rule row includes the rule version, assessment item identifier, object type, evidence slot identifier, permitted source type, factual polarity, effective time requirement, number of source independence criteria, set of interpretation rules, preceding status, and succeeding status.

[0031] The standard rule table is not directly written into the official rule base by the AI ​​model. The model can provide splitting suggestions, but only versions that pass field validation, clause verification, and confirmation by the responsible person will be compiled; unconfirmed suggestions will not participate in evidence verification.

[0032] S110: Read project applicability information, including security protection level, applicable general security requirements, applicable extended requirements, assessment time base, and type of the object being tested; read asset base information, including asset unique identifier, equipment name, address, region, business system, time zone, and alias. Eliminate inapplicable rules and use the combination of applicable assessment items and asset unique identifiers as the assessment unit.

[0033] S120 is a compilation evidence verification automaton. The initial state indicates that there is no verified evidence; the intermediate state is represented by the satisfied slot set, source count, valid time coverage, and conflict flag; the accept state indicates that all mandatory transfers have been completed, each alternative group has completed at least one allowed transfer, the source independence requirement has been met, and there are no unresolved conflicts; if the accept state cannot be reached, it does not directly output non-compliance, but enters the pending sampling state.

[0034] Each state transition has a unique state transition identifier. Facts that can be proven with a single piece of evidence correspond to unary witness conditions; facts that require both configuration statements and runtime results to support correspond to multivariate witness conditions; facts that allow either command echoing or configuration file support form alternative transitions. After the automaton is compiled, a summary is calculated for the standard version identifier, rule table specification sequence, and project applicability specification sequence to form the automaton fingerprint.

[0035] The S200 receives original evidence from multiple sources. The evidence access interface establishes collection events for device configuration files, command echoes, logs, scan reports, screenshots, scan documents, policy documents, asset tables, network topology diagrams, and interview audio, recording the collection interface, collection subject, collection time, target asset candidates, and original file content summaries.

[0036] The original evidence base employs an append-only versioning model. When the same logical file changes, a new version is added and a replacement relationship is established, without overwriting the original version. Content digests can use SM3 or SHA-256; after writing, the file is read again and the digest is calculated. If the two digests differ, the write operation is marked as failed. The digest is used to detect content changes, and the inability to overwrite versions is ensured by both storage access control and the append-only write strategy.

[0037] S210, for text files encoded in UTF-8, GB18030, or other registered encodings, first detects the encoding and uses the detection result as a transformation parameter; after decoding, it generates a standard text layer, while saving the cumulative mapping from each standard text character boundary to the original byte boundary. It only standardizes the registered newline representation, without deleting spaces, tabs, or punctuation that serve as field separators.

[0038] For DOCX, PDF, and table files, paragraphs, table cells, page numbers, and object relationships are parsed, and text layers are generated in reading order, retaining the source object identifier for each text block. For screenshots and scanned pages, text blocks are generated using a locked version of the OCR model and parameters, and page numbers, text block region coordinates, and the mapping of recognized characters to regions are recorded. For audio, sentences are generated using a locked version of the speech-to-text model, and the start timecode, end timecode, and speaker are recorded.

[0039] Each transformation step saves the operation identifier, software or model version, parameters, input summary, output summary, and location backmap. During replay, the input objects are read and the operations are performed in the same order, stopping if the output summary is different; OCR or speech-to-text can only prove that the text layer was generated by a specific media and a specific transformation version, and the final review interface still displays the corresponding media area or time period.

[0040] S220, Generate the retrieval index. Index items include evaluation unit candidate identifiers, source version identifiers, text block identifiers, baseline byte positions, asset candidates, collection time, and allowed source types. Vector similarity is only used to narrow down the candidate fragment range and does not form an evidence admission score; fragments are pre-filtered by object domain and time domain before being provided to the local model.

[0041] S300 calls upon a local AI model to propose candidates. It employs a local generative Transformer model already capable of following Chinese instructions, registering the output field pattern in the constraint decoder. Input fields include at least the evaluation unit identifier, a set of allowed state transition identifiers, a set of allowed interpretation rule identifiers, a source version identifier, a fragment identifier, a fragment baseline byte position, fragment text, and a set of asset aliases.

[0042] The model output fields must include at least the proposal identifier, state transition identifier, interpretation rule identifier, rule parameters, relative start position, relative end position, direct citation, asset candidate, effective time candidate, fact polarity, and candidate normalized fact. The source version identifier, fragment identifier, and fragment baseline byte position are filled into the read-only context by the calling program; corresponding fields after model rewriting are not accepted.

[0043] Temperature, sampling method, and maximum output length only affect the number of candidates, not whether the evidence is valid; parameters can be selected using a validation set that does not contain data on the target project, and no model confidence threshold that can change the acceptance status should be set.

[0044] If the model output cannot be parsed using the structured schema, parsing failure is recorded; if the state transition flag or interpretation rule flag is not in the allowed input set, it is immediately rejected; if the model service times out, it is retried a preset number of times after shortening the fragment, and if it still fails, the fragment is marked as requiring manual localization. Model anomalies do not modify the original evidence version, automaton, or verified records.

[0045] S400, Deterministic validation of the source byte range. For text-level fragments, the source byte range is calculated using the relative start and end positions as left-closed and right-open character boundaries, and the cumulative character-byte mapping and the fragment's base byte position. Proposals are rejected if the range is out of bounds, empty, or the boundary falls within a multi-byte character.

[0046]

[0047] The parameters in the formula have the same meaning as those in the invention, and the position units are character boundary number and byte offset, respectively. The external deterministic verifier rereads the version from the registered source version identifier, verifies the content summary and the transformation trajectory output summary, and then retrieves the text in reverse according to the calculation interval; if the retrieved text is inconsistent with the direct citation character by character, it is rejected. For the media text layer, the page number and region coordinates or start and end time codes are obtained simultaneously.

[0048] S410, Repetition Interpretation Rules. Each rule in the rule catalog includes a rule identifier, input type, output type, allowed parameters, implementation version, and test vector. For example, the key-value extraction rule retrieves values ​​from direct citations by registered field names and delimiters; the enumeration mapping rule maps "enabled" and "on" to Boolean true using a versioned mapping table; the time resolution rule converts timezoned time to a unified time base; and the address normalization rule resolves network addresses without changing their meaning.

[0049] The model can propose interpretation rules and parameters from the allowed set, but cannot upload executable code. The external deterministic validator checks the parameter types and ranges, performs the registration implementation, and obtains recalculated facts; it only proceeds if the recalculated facts are completely equal to the candidate normalized facts in the same canonical sequence. Thus, the accepted normalized facts are the results of the validator's calculations, not the model text itself, and are assigned a trust label.

[0050] S420, Verification of State Transition Guards. Asset candidates are mapped to unique asset identifiers via an asset alias table; if there is no match or multiple matches, manual review is required. The validity period is determined by the collection time, configuration effective time, log event time, or document validity period; if only the collection time is available and the rule requires an effective time, the two are not presumed to be the same. A verifiable atomic evidence record is generated only when the fact polarity, source type, object domain, and validity period domain all conform to the state transition rule.

[0051] The verifiable atomic evidence record stores the source version identifier, original acquisition root identifier, source start byte, source end byte, media location, transformation trajectory identifier, interpretation rule identifier, rule parameters, recalculated facts, asset identifier, valid time interval, state transition identifier, and validator version. A fingerprint is calculated and recorded after these fields are canonically serialized.

[0052]

[0053] Original files, converted files, and copies generated from the same original acquisition event share the original acquisition root identifier. Different filenames or different storage paths cannot automatically constitute independent sources; they are only registered as different original acquisition roots when they are generated by independent device interfaces, tool executions, or interview subjects, and there is no common parent acquisition event.

[0054] S430, Handling Negative Facts. For example, regarding the statement "The scan report does not list high-risk items," the system must not generate evidence of the absence of high-risk items solely from an empty list or summary statement. The data acquisition interface must also output scan scope credentials, including the target asset set, port or check item range, scanner version, policy parameters, start time, end time, exit status, and the original scan output version.

[0055] The out-of-model deterministic validator determines the inclusion relationship between the scan range and the set of objects and fields required by the state transition rules. If the target asset is missing, the task is canceled, the interface times out, the exit status is unsuccessful, the output is truncated, or the scan strategy does not cover the target check item, the denial fact transition is invalid; the system generates insufficient range constraints instead of interpreting empty results as compliant.

[0056] S500, such as Figure 2 As shown, a witness hypergraph is constructed. Original evidence version nodes and verifiable atomic evidence record nodes are connected via source edges, interpretation rule nodes and record nodes are connected via replay edges, and automaton state nodes are connected via witness hyperedges. The tail of each witness hyperedge contains all the records required to trigger the state transition, and the head contains the state after the transition.

[0057] Taking the evaluation unit that requires "device identity confirmed and password policy configuration consistent with actual authentication behavior" as an example, the object identity record first changes the initial state to the object confirmed state; the configuration file record and authentication log record together serve as inputs to the binary witness hyperedge, changing the state to the control state verified. If the two records correspond to different assets or the validity periods do not overlap, the guard condition of the hyperedge is false.

[0058] Establish conflict edges for records with opposite fact values ​​within the same asset, the same fact slot, and overlapping time domains. When a later version explicitly replaces a previous version and its effective date is determinable, the previous version can be considered historical evidence; conflicts that cannot be resolved based on version and time remain unresolved. Any subgraph containing both ends of an unresolved conflict cannot serve as accepted witness.

[0059] S510, Solve the closed witness subgraph. The solver only accepts records with a verification status of "passed"; for mandatory slots, at least one corresponding transfer is triggered; for alternative groups, at least one member is triggered; for source independence, a counting requirement for different original acquisition root identifiers is set; for conflict pairs, simultaneous selection is not allowed; and for negative facts, a scan range pass flag is set.

[0060]

[0061] If a state transition sequence exists from the initial state to the accepting state, then traverse backwards from the accepting state, checking all input records, source versions, and interpretation rules for each transition to obtain a closed witness subgraph. If any superedge lacks input, any record fingerprint is invalid, or any source version cannot be read, the closure check fails.

[0062] When multiple closed witness subgraphs exist simultaneously, this embodiment first selects the subgraph with fewer records; if the records are the same, it selects the subgraph with more direct measured sources and lower re-collection costs. The above sorting is only performed between subgraphs that have all reached the acceptance state, and confidence, similarity, or cost are not written as penalty terms to relax hard constraints.

[0063] S520 generates a candidate assessment evidence package. The evidence package includes the automaton fingerprint, assessment unit identifier, selected record, record fingerprint, state transition sequence, conflict resolution, media location, and review reason, and is marked as a candidate evidence package for assessment personnel to review. This method does not replace rating, personnel signature, or the final conclusion that should be made by the assessment organization according to law.

[0064] The review interface provides a model-free review button. The review process starts from the original evidence version, replays the text layer according to the transformation trajectory markers, retrieves direct citations by byte range, identifies the execution rules according to the interpretation rules, re-verifies the recorded fingerprints, and then replays the automaton according to the state transition markers. The entire review does not call the local AI large model, so model version updates will not change the fact that it has already passed the review by the same source and rules.

[0065] S600, no constraint kernel exists for accepting witnesses. Create a hypothesis literal for each required slot, substitution group, source independence, scan range, object domain, effective time domain, and conflict exclusion constraint; set all hypothesis literals to true and request the solver to reach the accept state. If unsatisfiable, read the initial unsatisfiable kernel returned by the solver.

[0066] The initial kernel is reduced through deletion: a hypothetical text is temporarily removed in stable order and the solution is recalculated. If the solution is still unsatisfactory, the text is permanently removed; if it becomes satisfactory, the text is restored. After one round, the kernel is checked again, and the deletion of any remaining text will make the constraint satisfactory. The output U is therefore a minimal kernel in the sense of inclusion relation, but it is not guaranteed to be the minimal kernel with the fewest elements among all minimal kernels.

[0067]

[0068] If the solver does not support natively unsatisfiable kernels, the minimum kernel can be obtained directly from all hypothesis texts using the deletion test described above. If a solution times out, the currently verified records and undecided states are retained, a solution timeout flag is output, and the solution is sent for manual review. Timeouts are not interpreted as a lack of evidence or non-compliance.

[0069] S610 converts the constraint kernel into a supplementary acquisition task. Each hypothesis text type registers a task template: the slot missing template obtains the allowed source and required fields for state transition; the source independence template requires different original acquisition root identifiers; the scan range template obtains missing assets or inspection items; the effective time template requires sources with effective time or event time; the conflict template requires obtaining higher-priority direct measurement records within conflict overlapping periods.

[0070] The fields for supplementary data collection tasks include task identifier, evaluation unit identifier, constraint identifier, asset identifier, target time interval, source type, collector identifier, required fields, set of allowed interpretation rule identifiers, completion conditions, and retry strategy. The collector identifier points to the registered local collection adapter; automatic retry stops and prompts for manual collection when the adapter is unauthorized, the interface is disconnected, or empty results are returned continuously.

[0071] The files obtained from the supplementary collection are re-entered into S200 as new versions of the original evidence. They cannot be directly written into the witness hypergraph without bypassing the external deterministic verification. The task is completed when the corresponding constraints disappear from the minimal kernel, not when only one file is received. If a new record introduces a conflict, the old task is closed and a conflict resolution task is generated.

[0072] S700, establish a dependency index. Each original version lists its derived transformation records, each transformation record lists its corresponding verifiable atomic evidence record, each record lists the participating witness hyperedges, and each witness hyperedge lists the associated evaluation units. When a new version is added, the successor object is marked as to be verified along the index from the replaced version.

[0073] For newly collected evidence that does not replace existing versions, only the evaluation units that intersect with its object domain, time domain, and allowed state transitions are added to the re-solution queue. When the interpretation rule version changes, records referencing that rule become invalid; when the automaton fingerprint changes, hyperedges referencing the changed state transitions become invalid. Graph regions that do not depend on the changed object are retained after re-verifying the record fingerprints.

[0074] Example 2 illustrates a processing path where a negative fact and a conflict coexist. An evaluation unit requires confirmation that a target asset has not enabled prohibited services within the evaluation time window. The model proposes the candidate fact "no prohibited services found" from the scan report summary, but the scan scope credentials only cover a portion of the addresses. Therefore, the out-of-model deterministic validator rejects the negative fact transition and generates an insufficient scan scope constraint.

[0075] Subsequently, the configuration file showed that the service was disabled, and the runtime log showed a connection event for the same service in a portion of the time window. The two records showed contradictory facts in the overlapping object and time domains, indicating a conflict edge was established by the system. The automaton could not reach the accept state, and the constraint kernel included insufficient scan range and conflict resolution; the supplementary acquisition task required performing a specified check on the complete address range of the target asset and obtaining the direct execution results for the overlapping time period.

[0076] Once the supplementary data collection results include successful completion scope credentials and the original output, the system generates a new verifiable atomic evidence record from that version and recalculates the relevant witness superedges. If the supplementary data collection test supports connection events, the candidate evidence package retains conflicting facts and enters manual judgment, without outputting compliance; if the test finds no service within the complete scope and the configuration version's effective time is later than the old log event, the version and time rules can resolve the old conflict, and then the closed witness subgraph determines whether to reach the acceptance state.

[0077] This embodiment does not use projected accuracy or time consumption ratios. Samples with out-of-bounds location issues, altered summary, duplicate source, inconsistent time zones, changed OCR version, scan aborted, missing slots, and conflicting facts can be used to check whether rejection, state blocking, constraint kernels, and local recalculation meet expectations.

[0078] Model candidate quality can be evaluated using field completeness rate, interval out-of-bounds rate, and validation pass rate; evidence validation quality can be evaluated using the consistency rate without model verification, the number of false acceptances, and the number of false rejections. Each dataset is isolated by project or asset root to avoid leakage of copy source.

[0079] It should be noted that the aforementioned byte range conversion, fingerprint recording, witness constraint, and minimal constraint kernel relationship are all data structures or logical definitions, not empirical formulas obtained from sample fitting, nor do they require dimensionless processing to eliminate physical dimensions; the descriptions of dimensionless processing, data simulation, and preset parameters in the following general implementation terms are only applicable when empirical evaluation formulas are used in alternative implementation methods, and are not used to interpret the aforementioned definitions.

[0080] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0081] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0082] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0083] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0084] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0085] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0086] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0087] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying evidence in information security compliance assessment based on a local AI large model, characterized in that, include: S1: Obtain the confirmed information security assessment standard rule table, project applicability information, and asset benchmark information, and compile the combination of applicable assessment items and the tested object into an evidence verification automaton; the state transition rules of the evidence verification automaton record the evidence slot, object domain, effective time domain, allowed source type, source independence, fact polarity, and mandatory or alternative relationships; S2: Obtain original evidence from multiple sources, solidify the original evidence version and content digest by appending, generate an addressable evidence version with character and byte mapping or media coordinate back mapping, and record the repeatable evolution trajectory from the original evidence version to the addressable evidence version. S3: Retrieve evidence fragments according to the state transition rules, input the evidence fragments and allowed state transition identifiers into the AI ​​big model located within the local isolated reasoning boundary, so that the AI ​​big model only outputs candidate evidence proposals containing the relative position within the fragment, direct citations, interpretation rule identifiers, state transition identifiers and candidate normalized facts; S4: The external deterministic verifier determines the source byte range based on the character and byte mapping, verifies the content summary of the original evidence version, retrieves the source content, recalculates the normalized fact according to the reproducible evolution trajectory and interpretation rule identifier, and verifies the object domain, effective time domain, source type and fact polarity of the corresponding state transition rule; Only when the recalculation result is consistent with the candidate normalized fact and the state transition guard condition is met, a verifiable atomic evidence record carrying the source byte range, content summary, transformation trajectory identifier, interpretation rule identifier, and state transition identifier is generated; The entire access-related fields of the verifiable atomic evidence record are standardized and serialized, and a record fingerprint is calculated so that when any source location information, asset identifier, effective time domain, fact polarity or state transition identifier changes, the corresponding record fingerprint becomes invalid. S5: Construct a witness hypergraph for the evaluation unit, using verifiable atomic evidence records that have passed deterministic verification as evidence nodes and one or more combinations of evidence nodes required to satisfy state transition as witness hyperedges; Only when a closed witness subgraph exists from the initial state to the acceptance state, and the closed witness subgraph satisfies source independence and does not contain unresolved conflicts, a candidate evaluation evidence package carrying a state transition sequence and a recorded fingerprint is output. S6: When the closed witness subgraph does not exist, set hypothetical text for the unsatisfied state transition constraints and solve the minimal unsatisfied constraint kernel. Convert the minimal unsatisfied constraint kernel into a supplementary acquisition task containing the target asset, target time domain, evidence type, acquisition interface, expected field, and interpretation rule identifier. After the supplementary evidence is processed by steps S2 to S4, only the witness hypergraph regions that depend on the corresponding original evidence version, verifiable atomic evidence records, or state transitions are invalidated and recalculated.

2. The method for identifying evidence in information security compliance assessment based on a local AI large model according to claim 1, characterized in that: The standard rule table maps the detection object, detection method, expected control state, and result recording requirements to object identity slots, allowed source types, fact slots, and state transition guard conditions, respectively. During compilation, inapplicable rules are trimmed according to the security protection level, general security requirements, and applicable extended requirements, and the standard version identifier, rule table summary, and project applicability summary are written into the automaton fingerprint of the evidence verification automaton.

3. The method for identifying evidence in information security compliance assessment based on a local AI large model according to claim 1, characterized in that: Each transformation in the reproducible transformation trajectory records the operation identifier, operation version, input summary, output summary, parameters, and position backmap; for text evidence, the position backmap points to the original byte range; for images, scanned documents, or audio, the position backmap points to the page number and region coordinates or start and end timecodes, respectively, and the reproducible result is rejected when the transformation operation version or parameters are inconsistent.

4. The method for identifying evidence in information security compliance assessment based on a local AI large model according to claim 1, characterized in that: The interpretation rule identifier points to a pre-registered deterministic interpretation rule, which includes key-value extraction, enumeration mapping, time parsing, address normalization, or threshold comparison. The out-of-model deterministic validator only allows the invocation of interpretation rules registered with the state transition rule, and rejects candidate evidence proposals when there are out-of-bounds location errors, summary changes, direct citation inconsistencies, transformation trajectory replay failures, out-of-bounds interpretation rule parameters, or inconsistencies in the normalized facts recalculated.

5. The method for identifying evidence in information security compliance assessment based on a local AI large model according to claim 1, characterized in that: When the state transition rule requires that the absence of a target item be used as a negative fact, only verifiable atomic evidence records that also carry scan range credentials are accepted. The scan range credentials include at least the target asset set, the inspected field or address range, the acquisition command or scanner version, execution parameters, start time, end time, and completion status. If the completion status is abnormal, the target asset is not covered, or the scan range is smaller than the state transition rule requirement, the state transition corresponding to the negative fact will not be triggered.

6. The method for identifying evidence in information security compliance assessment based on a local AI large model according to claim 1, characterized in that: The closed witness subgraph includes all input evidence nodes required for each witness superedge, the source version node corresponding to the input evidence node, and the state nodes that are continuous from the initial state to the acceptance state; copies with the same original acquisition root identifier are considered to be from the same source, and conflict edges are established between evidence nodes with opposite fact values ​​and overlapping object domains and effective time domains. No confidence or similarity score replaces the source independence constraint, conflict constraint, or state transition guard condition.

7. The method for identifying evidence in information security compliance assessment based on a local AI large model according to claim 1, characterized in that: Slot coverage, substitution relation, source independence, negative fact scan range, object domain, effective time domain, and conflict exclusion constraint are each guarded by a hypothetical literal. After the solver returns an unsatisfiable kernel, it attempts to delete one of the hypothetical literals in turn and solves again. If it is still unsatisfiable after deletion, the deletion result is retained until any hypothetical literal is deleted and becomes satisfiable, so as to obtain the minimal unsatisfiable constraint kernel in which no proper subset is unsatisfiable.

8. The method for identifying evidence in information security compliance assessment based on a local AI large model according to claim 1, characterized in that: Establish a dependency index from the original evidence version, reproducible change record, reverifiable atomic evidence record, witness superedge to evaluation unit; when the original evidence version is replaced by an appended new version, the interpretation rule or automaton fingerprint is changed, or the supplementary collection task returns new evidence, mark the subsequent object as invalid according to the dependency index, and retain the witness subgraph area that does not depend on the changed object and whose fingerprint is valid after re-verification.

9. The method for identifying evidence in information security compliance assessment based on a local AI large model according to claim 1, characterized in that: The candidate evaluation evidence package includes automaton fingerprints, selected verifiable atomic evidence records, record fingerprints, state transition sequences, conflict resolution records, and verification reasons. Without invoking the AI ​​big model, each verifiable atomic evidence record and acceptance state are reproduced according to the original evidence version pointed to by the candidate evaluation evidence package, the reproducible evolution trajectory, and the deterministic interpretation rules. If reproduction fails, the corresponding candidate evaluation evidence package is revoked.