Conformance assessment method for a model

Abstract

The embodiment of the application provides a compliance evaluation method of a model. In view of the problem that in the prior art, model similarity analysis and license review are mutually separated, it is difficult to accurately identify the reuse relationship of a fine-tuned or modified model, and the evaluation is not comprehensive, consistency is poor, and efficiency is low, the method accesses a target model, an open source model and license information thereof, extracts multi-dimensional features of the target model and the open source model, establishes a similarity score between the models, divides constraint levels of license terms, and generates a compliance evaluation result by fusing the similarity and license constraints under unified rules and outputs the compliance evaluation result. The method can realize the association judgment of the model source, the similarity degree and the license restriction, and improve the comprehensiveness, consistency, automation degree and result traceability of the compliance evaluation.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method for evaluating the compliance of a model. Background Technology

[0002] In the current development of large AI models, reuse, fine-tuning, and compliance checks are usually carried out based on open source models, open source code frameworks, and relevant license terms.

[0003] However, existing technologies mostly rely on traditional code comparison or manual review methods, which are difficult to adapt to the structure and parameter characteristics of neural network models, resulting in incomplete model similarity assessment. At the same time, license verification and model analysis are disconnected, making it difficult to form a consistent risk assessment, which in turn limits the efficiency and reliability of large-scale detection.

[0004] Therefore, improving the comprehensiveness, consistency, and automation of AI large-scale model compliance testing to achieve accurate risk assessment has become a technical problem that needs to be solved. Summary of the Invention

[0005] The compliance assessment method for models provided in this application addresses the aforementioned technical issues. This method addresses the compliance testing needs of large AI models during open-source reuse, fine-tuning, and deployment. It comprehensively considers the correlation between the target model and the open-source model, as well as the differences in constraints imposed by the open-source license, to uniformly assess the potential compliance risks of the model, thereby improving the comprehensiveness, consistency, and automation of model compliance judgment.

[0006] Firstly, embodiments of this application provide a method for evaluating the compliance of a model, including:

[0007] Obtain the target model to be detected, multiple open-source models, and multiple open-source licenses;

[0008] Extract the first multi-dimensional feature data from the target model;

[0009] Extracting second-dimensional feature data from multiple open-source models;

[0010] Based on the first and second multi-dimensional feature data, the similarity score between the target model and multiple open-source models is determined.

[0011] Determine the license constraint level based on multiple open source licenses;

[0012] The compliance assessment results are determined based on similarity scores and license constraint levels.

[0013] Output the compliance assessment results.

[0014] In one possible embodiment, the first multidimensional feature data includes first structural feature data, which includes at least one of a structured topology graph, a layered structure feature vector, and a set of structural labels.

[0015] In one possible embodiment, the first multidimensional feature data includes first parameter feature data, which includes at least one of the following: parameter distribution feature vector, key layer parameter tensor, and parameter statistical index set.

[0016] In one possible embodiment, the first multi-dimensional feature data includes first data feature data, which includes at least one of data fingerprint hash value, data distribution feature vector, and training set overlap value.

[0017] In one possible embodiment, the first multi-dimensional feature data includes first behavioral feature data, which includes at least one of the following: inference result feature vector, behavioral index set, and output similarity value.

[0018] In one possible embodiment, it also includes:

[0019] The compliance assessment results are input into a pre-defined visualization report generation model to obtain and output a compliance report. The compliance report includes an inspection overview, model similarity analysis, license analysis, compliance assessment, rectification suggestions, and data traceability.

[0020] In one possible embodiment, it also includes:

[0021] Based on the compliance report, a risk visualization interface is generated, which is used to graphically display the compliance risks of the target model.

[0022] Displays a risk visualization interface.

[0023] In one possible implementation, the license constraint level is determined based on multiple open-source licenses, including:

[0024] Extract structured license information and license association rules from multiple open-source licenses;

[0025] Based on the structured information of the license and the rules associated with the license, the level of license constraint is determined.

[0026] In one possible embodiment, the compliance assessment result is determined based on the similarity score and the license constraint level, including:

[0027] Based on similarity scores and license constraint levels, the infringement risk level of the target model is determined;

[0028] Based on the infringement risk level, similarity score, license constraint level, and the use case information of the target model, a compliance assessment result is generated.

[0029] In one possible embodiment, it also includes:

[0030] Based on the level of infringement risk, a corresponding early warning instruction is determined. The early warning instruction includes at least one of an early warning information push instruction and a model locking instruction.

[0031] Execute the warning command.

[0032] In one possible embodiment, based on first multi-dimensional feature data and second multi-dimensional feature data, a similarity score between the target model and multiple open-source models is determined, including:

[0033] Based on the first multi-dimensional feature data and the second multi-dimensional feature data, a first score for each dimension is determined. The first score is the similarity score of the same dimension feature data in the first multi-dimensional feature data and the second multi-dimensional feature data.

[0034] Obtain the first weight for each dimension;

[0035] Based on the first score and corresponding first weight of each dimension, the similarity score between the target model and multiple open source models is determined.

[0036] In one possible implementation, obtaining the first weights for each dimension includes:

[0037] Based on the model type of the target model, the preset second weights are modified to obtain the first weights for each dimension.

[0038] Secondly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0039] The memory stores the instructions that the computer executes;

[0040] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0041] Thirdly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0042] The compliance assessment method for models provided in this application obtains the target model to be tested, multiple open-source models, and multiple open-source licenses. It determines the similarity score by combining the multi-dimensional feature data of the target model and multiple open-source models, and determines the license constraint level based on multiple open-source licenses. This method can correlate and integrate model similarity analysis and license constraint analysis to form and output a compliance assessment result for the target model, thereby improving the comprehensiveness, consistency, accuracy, and efficiency of compliance detection for large AI models and risk assessment. Attached Figure Description

[0043] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0044] Figure 1 A flowchart illustrating the compliance assessment method for the model provided in this application embodiment;

[0045] Figure 2 A schematic diagram of the structure of the compliance assessment device for the model provided in the embodiments of this application;

[0046] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0047] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0048] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0049] In the field of AI model governance and intellectual property compliance control, when developing large language models, visual models, or autonomous driving models, companies typically access target models, open-source model resources, and corresponding license information to conduct compliance checks around model reuse, fine-tuning, deployment, and commercial launch. This scenario generally involves model asset management, open-source resource management, and risk auditing, requiring unified judgment on model source, similarity, and license restrictions throughout the R&D process.

[0050] In existing technologies, a common practice is to first manually review the model code, configuration files, or some training data, or use traditional code comparison tools to identify reuse relationships with open-source resources. Then, legal or compliance personnel separately verify the license terms to determine whether the model can be released, commercially viable, or requires rectification. The basic idea is to perform model analysis and license review separately, and then manually synthesize the conclusions from both parts to complete the final risk assessment.

[0051] However, the aforementioned technical solutions are mainly designed for traditional software code reuse scenarios and are difficult to effectively reflect the similarity of neural network models in terms of structural design, parameter distribution, and overall behavior. For models that have undergone fine-tuning, pruning, distillation, or local modification, even if there are significant differences on the surface of the code, they may still maintain a high degree of similarity to the open-source model in terms of the model itself. This makes the judgment results relying solely on code comparison or human experience incomplete and unstable.

[0052] Meanwhile, license verification and model similarity analysis are disconnected, lacking a unified mechanism for correlation assessment. As a result, even if a target model is found to be similar to an open-source model, it is difficult to reach a consistent assessment conclusion by combining the varying degrees of constraint of multiple open-source licenses. With an increasing number of models and complex sources, this also leads to problems such as low review efficiency, unclear risk classification, and insufficient traceability of conclusions, thereby affecting enterprise R&D decisions and the pace of commercialization.

[0053] In light of this, improving the comprehensiveness, consistency, and automation of model compliance assessment has become a pressing technical challenge. To address this, a model compliance assessment method is proposed. This method involves acquiring the target model to be tested, multiple open-source models, and multiple open-source licenses. Multi-dimensional feature data is extracted from both the target model and the multiple open-source models. Based on the extracted multi-dimensional feature data, a similarity score is determined between the target model and the multiple open-source models. Then, based on the multiple open-source licenses, the license constraint level is determined. Finally, the compliance assessment result is determined by combining the similarity score and the license constraint level, and the final compliance assessment result is output.

[0054] Under this technical approach, a processing architecture that integrates data access, feature extraction, similarity analysis, and compliance assessment can be used to incorporate model information and license information into the same assessment process, enabling a unified assessment of the compliance risks of the target model and providing a more stable basis for decision-making for R&D, legal, and management personnel.

[0055] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0056] Figure 1 This is a flowchart illustrating the compliance assessment method for the model provided in this application embodiment. The method provided in this application embodiment is applicable to any electronic device. Figure 1 As shown, the method includes:

[0057] S101: Obtain the target model to be detected, multiple open-source models, and multiple open-source licenses.

[0058] In this application, the target model refers to the model object whose compliance is to be tested. This model object can be a large language model, a visual model, a speech model, a multimodal model, or an autonomous driving perception and decision-making model. The corresponding input content includes at least one or more of the following: model file, model configuration file, version identifier, training specification file, and deployment specification file. An open-source model refers to a set of reference models used for comparison. Each open-source model corresponds to at least one model identifier, one model entity, and license information associated with that model. An open-source license refers to the license text or license rule data associated with multiple open-source models, used to constrain the use, modification, distribution, commercial deployment, or release of derivative models of the open-source models. In specific implementations, the executing entity can be a compliance assessment server deployed on an enterprise's internal R&D platform, or an assessment module in a model asset management system. The acquisition process includes receiving the target model and multiple open-source models from model repositories, artifact repositories, code hosting platforms, object storage, model registries, or third-party open-source community mirror sources, and receiving multiple open-source licenses from license repositories, results scraped from model release pages, pre-stored rule tables, or structured legal knowledge bases. To ensure consistency in subsequent comparison processing, after receiving the target model and multiple open-source models, the system can record basic information such as model identifier, version information, source information, verification information, and receipt time. When receiving multiple open-source licenses, the system can retain the license texts and establish a relationship between the open-source models and the licenses. If an open-source model has a main license and additional usage conditions, the system can store both the main license text and the additional terms. If the target model only provides an inference interface and does not directly provide a weight file, the system constructs an analyzable object of the target model through the interface description file, inference log, configuration parameters, and external declaration materials. If multiple open-source models have different framework formats, the system first calls the format adapter for parsing, converting the model body, structural description, and parameter tensors into unified intermediate representation data so that the subsequent feature extraction module can process them directly. Based on this processing method, each open-source model can be used as both a similarity comparison object and a license constraint judgment object, and the entire processing chain has a unified input basis and traceable relationships. It should be understood that the above examples are for demonstration purposes only and are not limiting.

[0059] S102. Extract the first multi-dimensional feature data from the target model.

[0060] In this application, the first multi-dimensional feature data refers to the set of multi-dimensional features extracted from the target model, which is used to characterize the attributes of the target model and serve as input for subsequent similarity score calculation.

[0061] In its implementation, the system first parses the target model, combining available model files, configuration content, parameter content, training instructions, deployment information, running results, or other relevant information to extract the first multi-dimensional feature data representing the target model. If the target model contains a configuration file, corresponding features can be extracted based on the configuration file; if the target model does not contain a complete configuration file, partial descriptive information can be reconstructed by traversing the model computation graph, loading weight names, and analyzing operator call sequences. If a training data list, dataset declaration, or data fingerprint file exists, corresponding features can be extracted from it; if a complete list does not exist, corresponding feature representations can be generated based on the target model's performance under test inputs. The system can also feed test inputs into the target model, record the output results, and form feature data related to model performance. For missing dimensions, the system records the missing identifier and places null values ​​for that dimension, adjusting the weights according to the missing value handling rules in subsequent scoring. To make features of different dimensions comparable, the extracted feature data can undergo unified representation processing. Based on the above analysis, it can be seen that by incorporating multiple attributes of the target model into the first multi-dimensional feature data, the object to be detected is no longer limited to surface information, thus providing an executable data foundation for identifying the similarity between models. It should be understood that the above example is merely illustrative and not limiting.

[0062] S103. Extract second multi-dimensional feature data from multiple open-source models.

[0063] In this application, the second multi-dimensional feature data refers to a set of multi-dimensional features extracted from multiple open-source models. Its dimensional system corresponds to the first multi-dimensional feature data and is used as a benchmark for comparison against the target model. For each open-source model, the system extracts the second multi-dimensional feature data corresponding to the first multi-dimensional feature data and generates a corresponding open-source model feature profile. To ensure the comparability of subsequent scoring, the second multi-dimensional feature data uses the same or compatible feature templates, normalization rules, encoding methods, and test sample sets as the first multi-dimensional feature data. Specifically, the system retrieves model files and license association information one by one according to the open-source model number. If there are many models, multiple open-source models are allocated to parallel processing nodes through a task scheduler, and each processing node executes a unified model parsing process. The system can combine the available model files, configuration content, parameter content, public instructions, running results, or other relevant information for each open-source model to extract the second multi-dimensional feature data for comparison. Considering that some open-source models may use different frameworks, different quantization methods, or different inference accuracies, this step performs format adaptation, precision alignment, and necessary processing on the models before extraction to reduce feature offsets caused by differences in the operating environment. For open-source models that cannot run completely, the system extracts the available second multi-dimensional feature data and marks the unextractable parts as missing. For external open-source models described only by API (Application Programming Interface), the system extracts the corresponding feature data through the interface call results and public documentation. After the second multi-dimensional feature data of each open-source model is generated, the system writes it into the feature index library and associates it with the corresponding model number, license number, feature completeness score, and extraction log to support subsequent similarity calculations and audit traceability. Based on the above analysis, this step establishes a unified feature profile for multiple open-source models, enabling the target model and the reference model to be compared in the same feature coordinate system, thereby ensuring that subsequent similarity scoring covers the model ontology and its related performance information. It should be understood that the above example is for demonstration purposes only and is not a limitation.

[0064] S104. Based on the first multi-dimensional feature data and the second multi-dimensional feature data, determine the similarity score between the target model and multiple open-source models.

[0065] In this application, the similarity score refers to the quantitative value of similarity obtained based on the comparison results of multi-dimensional features between the target model and each open-source model. This quantitative value can be a continuous score between 0 and 1, or a standardized score between 0 and 100. The system calculates a set of similarity scores for the target model and each open-source model, forming a model similarity list. In specific implementation, the system first establishes dimensional comparison rules, matching each dimension feature data in the first multi-dimensional feature data with the second multi-dimensional feature data of the corresponding open-source model in the same dimension. Let the similarity of each dimension be a1, a2, a3, and a4, and the corresponding weights be w1, w2, w3, and w4, then the overall similarity score Score can be expressed as Score = w1 × a1 + w2 × a2 + w3 × a3 + w4 × a4, where w1 + w2 + w3 + w4 = 1. This formula is used to compress similarity information from different dimensions into a unified score. The weights can be set by preset rules or dynamically adjusted according to the model type. When a dimension is missing, the system renormalizes the weights according to the remaining valid dimensions to prevent systematic underestimation caused by directly setting missing values ​​to zero. To further screen key risk targets, the system can extract the top N highest-scoring models from all open-source model similarity scores as candidate related models and record the detailed sub-dimensional scores of each candidate model. If multiple open-source models correspond to the same series of versions, the system can also cluster them according to version time, license change node, and similarity score to output the closest source model version. To reduce the impact of differences in the scales of different dimensions, the scores of each dimension are standardized before fusion. Based on the above analysis, this step, through a multi-dimensional score fusion mechanism, uniformly maps the degree of similarity between the target model and the open-source model into a sortable and thresholdable similarity score, thereby supporting subsequent linkage judgment with license constraint information. It should be understood that the above example is only for demonstration and not a limitation.

[0066] S105. Determine the license constraint level based on multiple open source licenses.

[0067] In this application, the license constraint level refers to the strength or risk level of constraints determined by multiple open-source licenses, used to indicate the degree of restriction on the use, modification, distribution, sublicensing, model derivation, and commercial deployment of each open-source model. The system independently parses the license associated with each open-source model and generates the corresponding license constraint level. Specifically, the system first converts multiple open-source license texts into structured clause data. The clause data includes at least the license name, version number, scope of effect, authorized objects, permitted behaviors, restricted behaviors, triggering conditions, retention obligations, and disclaimer fields. If the license text is in natural language format, the system identifies key clauses by combining rule templates with a text parsing model, such as commercial restriction clauses, redistribution requirement clauses, attribution retention clauses, open source requirement clauses, network propagation triggering clauses, model weight sharing obligation clauses, and derivative works application clauses. When the same open-source model is simultaneously constrained by multiple open-source licenses, the system merges and parses the multiple open-source licenses, using the clause with the highest constraint strength as the final determination result for the corresponding dimension. License constraint levels can be divided into multiple levels, such as low constraint, medium constraint, and high constraint, or can be represented by numerical levels, such as levels 1 to 5. In specific classification, if a license allows commercial use, permits modification, and only requires the retention of the copyright notice, it is classified as a lower constraint level; if a license requires derivative models to continue to publicly disclose related content upon distribution, to be accompanied by the same license, or to meet specific open obligations, it is classified as a medium-high constraint level; if a license explicitly restricts commercial use, restricts closed-source deployment, requires stricter infectiousness obligations, or imposes restrictions on use in specific scenarios, it is classified as a high constraint level. To enhance the consistency of the rules, the system maintains a license rule mapping table. The mapping table sets at least commercial restriction scores, distribution obligation scores, infectiousness clause scores, attribution obligation scores, and model-specific clause scores, and forms the total license constraint score through weighted aggregation. Let the commercial restriction score be b1, the distribution obligation score be b2, the infectiousness clause score be b3, the attribution obligation score be b4, and the model-specific clause score be b5, with corresponding weights v1, v2, v3, v4, and v5, then the license constraint score Level can be expressed as Level = v1×b1 + v2×b2 + v3×b3 + v4×b4 + v5×b5. Based on the Level range, the license is mapped to the corresponding constraint level. If there are version differences in the license, the system uses the license version that matches the release time of the open-source model as the parsing object; if the license text contains custom additional clauses, the system extracts the additional clauses separately and includes them in the total constraint score. If the license name cannot be directly identified, the system calculates the text matching degree based on the clause keywords and the standard license template. If the matching degree exceeds a preset threshold, the corresponding template is used to complete the structured parsing; otherwise, it is marked as a level requiring manual review and is considered a high-uncertainty constraint in subsequent evaluations.Based on the above analysis, this step transforms the legal restrictions originally scattered across different open-source license texts into a unified hierarchical representation, enabling it to be directly combined with model similarity scores for processing, thereby forming a consistent basis for compliance evaluation. It should be understood that the above example is merely illustrative and not limiting.

[0068] S106. Determine the compliance assessment results based on similarity scores and license constraint levels.

[0069] In this application, the compliance assessment result refers to the final judgment result formed by combining the similarity between the target model and each open-source model and the strength of the corresponding license constraints. This result can be expressed as pass, restricted pass, rectification required, high risk, prohibited from going live, or pending manual review. The system takes each open-source model as an assessment unit, performs joint analysis on the similarity score and license constraint level corresponding to the open-source model, and then summarizes the results of multiple assessment units to obtain the overall compliance assessment result of the target model. In specific implementation, the system pre-establishes an assessment rule matrix. The horizontal axis of the matrix is ​​the similarity score range, and the vertical axis is the license constraint level range. Different intersection areas correspond to different risk conclusions. For example, when the similarity score is higher than the first threshold and the license constraint level is in the high-level range, the output is high risk or prohibited from going live; when the similarity score is higher than the first threshold but the license constraint level is in the medium-level range, the output is rectification required or restricted pass, with an accompanying reminder of performance obligations; when the similarity score is in the middle range and the license constraint level is high, the output is pending manual review; when the similarity score is lower than the second threshold, even if there is a high constraint level, a low correlation risk can be output, but the license traceability record is still retained. For example, the first threshold can be set to 0.8, and the second threshold can be set to 0.4; if a percentage system is used, it can be set to 80 and 40 points respectively. This threshold is used to distinguish between three situations: obvious similarity, partial similarity, and low similarity. When multiple open-source models have high similarity to the target model, the system selects the one with the highest risk level as the primary conclusion and lists other related models as secondary criteria; if multiple medium-risk conclusions exceed the cumulative threshold, the overall result is upgraded by one level. To improve the interpretability of the conclusion, the system synchronously writes the trigger criteria when generating the compliance assessment results, including the number of the hit open-source model, the corresponding license number, the details of similarity across dimensions, the total similarity score, the license constraint level, the hit rule number, and the suggested action. The suggested action may be to supplement the source proof, replace the relevant model, complete the signature document, adjust the deployment method, suspend commercial deployment, or submit for manual legal review. Based on the above analysis, this step integrates the similarity relationships at the technical level of the model with the constraints at the legal level of the license into the same judgment logic. This allows the output conclusions to directly correspond to the company's R&D, release, and commercialization decisions, thereby resolving the inconsistency caused by the separation of model analysis and license review. It should be understood that the above example is merely illustrative and not limiting.

[0070] S107. Output the compliance assessment results.

[0071] In this application, outputting compliance assessment results refers to providing the generated compliance conclusions and their associated evidence to external systems, user interfaces, or subsequent processing modules in a readable, storable, and traceable format. The output content should include at least the target model identifier, overall compliance status, a list of matched open-source models, similarity scores for each open-source model, license constraint levels, trigger rule identifiers, and suggested actions. In specific implementation, the system can write the compliance assessment results into a structured result table. The result table fields should include at least the task number, target model number, assessment time, final risk level, highest risk source model number, license number, conclusion summary, and audit log address. Simultaneously, it should generate a user-facing assessment report file, which should retain dimensional similarity descriptions, license clause summaries, and rectification suggestions. If the enterprise R&D platform supports workflow linkage, the system can synchronously push corresponding status codes to the model release approval module, legal review module, and R&D rectification module after outputting the results. For example, when the conclusion is high-risk, it can automatically block the launch application; when the conclusion is that rectification is required, it can automatically create a rectification work order with attached license obligations to be processed. If the result requires manual review, the system packages the target model, relevant open-source models, open-source license text, and similarity evidence and pushes them to the manual review queue. To ensure traceability, the output stage also writes the model version, license version, feature extraction algorithm version, scoring weight configuration, and threshold configuration used in this assessment into the audit log, enabling the assessment process to be reproduced during subsequent reviews. For front-end display, the system can present the results using text conclusions, tabular views, or risk matrix views; for external interfaces, the system can return data using JSON (JavaScript Object Notation), database records, or message queue events. If the compliance assessment results are updated, such as due to revisions to license library rules or changes in the target model version, the system archives historical results and outputs the difference fields between the latest and historical results. Based on the above analysis, the output is not only a result display action but also forms a closed-loop connection with the R&D process, legal process, and audit process, enabling the compliance assessment results to directly enter the subsequent rectification, release control, and accountability tracking stages. It should be understood that the above example is for demonstration purposes only and is not a limitation.

[0072] Based on the above analysis, this application provides a compliance assessment method for a model, including acquiring a target model to be tested, multiple open-source models, and multiple open-source licenses; extracting first multi-dimensional feature data from the target model; extracting second multi-dimensional feature data from the multiple open-source models; determining the similarity score between the target model and the multiple open-source models based on the first and second multi-dimensional feature data; determining the license constraint level based on the multiple open-source licenses; determining the compliance assessment result based on the similarity score and the license constraint level; and outputting the compliance assessment result. In this application, by connecting model access, unified feature extraction, multi-dimensional similarity analysis, structured license grading, and joint judgment rules into a single processing flow, the reuse relationship between the target model and multiple open-source models, and the constraint relationship corresponding to the multiple open-source licenses, are correlated and calculated within the same assessment framework, thereby outputting a compliance conclusion with quantitative and auditable basis. Based on this technical approach, even with significant code differences, the degree of similarity between the target model and the open-source models can still be identified through the joint representation of the first and second multi-dimensional feature data. Furthermore, by combining the license constraint level with a unified mapping of relevant restrictions, a consistent risk grading result can be formed. It should be understood that the above example is merely illustrative and not limiting.

[0073] In one possible implementation, the first multidimensional feature data includes first structural feature data, which includes at least one of a structured topological graph, a layered structural feature vector, and a set of structural labels.

[0074] In its implementation, after receiving the configuration file, structure description file, or framework definition file of the target model, the system parses the layer definitions, connection definitions, and module nesting relationships to generate at least one of the following: a structured topology graph, layer structure feature vectors, and a structure label set. If the target model adopts a Transformer, CNN (Convolutional Neural Network), or hybrid architecture, the system can extract structure labels based on the corresponding layer names, parameter fields, and module relationships, and encode the number of layers, the number of repeating units, and cross-layer connection information into layer structure feature vectors. The structured topology graph can be stored as an adjacency list, adjacency matrix, or attributed graph structure to facilitate alignment and comparison with similar structured data of open-source models; the layer structure feature vectors can be represented using fixed-length floating-point vectors to adapt to subsequent distance calculations; the structure label set can be stored using discrete label encoding to support direct comparison of architecture categories and structural patterns. In practical applications, other model parsing formats can also be used for this feature representation, and this application does not limit this.

[0075] In one possible implementation, structural features such as neural network layers, network topology, activation function type, attention mechanism design, inter-layer connection method, and model parameter magnitude can be extracted from the target model by parsing at least one of the model configuration file, network structure visualization tool, and custom parsing script, and then integrated to obtain the first structural feature data.

[0076] During operation, the system incorporates the aforementioned first structural feature data into the first multi-dimensional feature data and matches it with the corresponding structural features extracted from multiple open-source models, thereby uniformly measuring the similarity between the target model and the open-source models at the structural level. Since the structured topology graph, layer structure feature vectors, and structural label set can respectively reflect the model's connection relationships, hierarchical distribution, and architectural attributes, they can collectively constitute a complete representation of the model's ontological structure.

[0077] After adopting this implementation method, the structural information of the target model can be stably expressed in at least one of graph structure, vector structure and label structure, so that the similarity score can simultaneously cover hierarchical relationship, connection pattern and architectural attribute, thereby improving the completeness and consistency of structural identification in model compliance assessment, and enhancing the ability to compare the structure of models that have been fine-tuned, pruned or partially modified.

[0078] Based on the foregoing embodiments, the first multi-dimensional feature data further includes first parameter feature data, which includes at least one of parameter distribution feature vector, key layer parameter tensor, and parameter statistical index set.

[0079] In its implementation, after obtaining the parameter file of the target model, the system performs parsing and normalization on the parameter matrix and extracts the corresponding parameter information according to the hierarchy or module division. When the first parameter feature data uses a parameter distribution feature vector, the system can statistically analyze the mean, variance, proportion of non-zero elements, and amplitude distribution of the weights of each layer, and concatenate the statistical results into a vector in a preset order. When the first parameter feature data uses a key layer parameter tensor, the system reads the weight tensor from the preset key layer and retains its dimensional structure for dimension-by-dimensional comparison with the corresponding tensor of the open-source model. When the first parameter feature data uses a set of parameter statistical indicators, the system calculates statistical indicators for the global parameters of the model or the parameters of a specified layer, and maps the indicator values ​​to a unified data format. The above three types of data can be used individually or in combination to form the parameter class representation in the first multi-dimensional feature data.

[0080] In one possible implementation, features such as the numerical distribution of weights / biases of each layer, the dimension and sparsity of the weight matrix, the parameter similarity of key layers (such as attention layers and fully connected layers), and the parameter initialization method can be extracted from the target model through tensor feature extraction algorithms, parameter distribution statistics (mean / variance / standard deviation), and cosine distance preprocessing, and then integrated to obtain the first parameter feature data.

[0081] Under this structure, the first parameter feature data can describe the parameter state of the target model from three levels: overall distribution, local key layers, and statistical indicators, and provide a unified input for subsequent similarity calculations with open-source models. Based on the combined expression of parameter distribution feature vectors, key layer parameter tensors, and parameter statistical indicator sets, the system can output more complete parameter feature results, thereby improving the representability and comparability of the target model parameter features.

[0082] In one possible implementation, the first multi-dimensional feature data includes first data feature data, which includes at least one of data fingerprint hash value, data distribution feature vector, and training set overlap value.

[0083] In practical implementation, the training data of the target model can first undergo cleaning, deduplication, and normalization. Then, the hash module generates data fingerprint hash values, and the statistics module extracts data distribution feature vectors. Subsequently, the overlap calculation module outputs the training set overlap value based on the fingerprint matching results or the intersection-union relationship of the sample sets. The above-mentioned data features can be used individually for similarity scoring or combined to form more complete data-side evidence. When only partial training data is obtained, any corresponding feature value can still be generated based on the collected samples and used in the evaluation. The hash algorithm can be SHA (SecureHash Algorithm) series, MD5 (Message Digest Algorithm 5), or other algorithms suitable for data identification generation. In practical applications, other models of the algorithm can also be selected, and this application does not limit this.

[0084] In one possible implementation, features such as training data fingerprints, data distribution features, training set overlap, and data preprocessing methods can be extracted from the training dataset of the target model through methods such as training data hash fingerprint generation, data distribution histogram comparison, text feature extraction (for text-based models) / image feature extraction (for visual-based models), and integrated into the first data feature data.

[0085] The data fingerprint hash value is used to provide a fast and stable basis for data identity comparison, the data distribution feature vector is used to characterize the statistical attribute differences of the training data, and the training set overlap value is used to directly reflect the degree of reuse of training samples. When at least one of the three is included in the first multi-dimensional feature data, the similarity of the data source of the target model can be quantified and output under a unified feature framework.

[0086] By adopting this specific implementation method, the data-side features of the target model no longer rely solely on manual review, but can form calculable and traceable quantitative results using hash identifiers, distribution vectors, or overlap values. This improves the integrity and consistency of training data similarity identification and provides a more stable data basis for subsequent compliance assessments.

[0087] In one possible implementation, the first multi-dimensional feature data includes first behavioral feature data, which includes at least one of the following: inference result feature vector, behavioral index set, and output similarity value.

[0088] In practical implementation, standard test samples for behavior analysis can be input into the target model first, and their inference responses can be extracted. Then, the responses are converted into inference result feature vectors using preset encoding rules. For categorical outputs, the probabilities of each category can be concatenated into vectors along fixed dimensions. For generative outputs, the text representation can be mapped into semantic embedding vectors. Subsequently, a set of behavioral indicators is statistically analyzed based on multiple rounds of test results, and the similarity score is calculated after a sample-by-sample comparison with the corresponding output of the open-source model. The above features can be used individually or in combination to form the first behavioral feature data and participate in subsequent similarity scoring calculations. In practical applications, other models can be selected for the encoding model and similarity calculation model, which is not limited in this application.

[0089] In terms of working principle, the first behavioral feature data transforms the performance of the target model at the behavioral level into a computable data representation by characterizing the model's output content, output stability, and similarity relationships between outputs. It is then compared with the behavioral features corresponding to multiple open-source models, so that the similarity score not only reflects the degree of closeness at the structural or parameter level, but also reflects whether the reasoning behavior is consistent.

[0090] In one possible implementation, features such as model inference output, intermediate layer output features, prediction accuracy / recall rate on the test set, and inference logic consistency can be extracted from the target model through standardized test set (public general set + industry customized set) inference, intermediate layer feature extraction, and behavior consistency statistics, and then integrated into the first behavior feature data.

[0091] With this implementation method, the behavioral characteristics of the target model can be uniformly expressed in the form of vectors, indicators and numerical values, which facilitates automated comparison and quantitative judgment in compliance assessment. It can also improve the ability to identify the behavioral similarity of models that have been fine-tuned, distilled or locally modified, thereby making the similarity score results more complete and stable.

[0092] One possible implementation also includes: inputting the compliance assessment results into a preset visualization report generation model to obtain and output a compliance report, which includes an inspection overview, model similarity analysis, license analysis, compliance assessment, rectification suggestions, and data traceability.

[0093] In its implementation, after generating the compliance assessment results, the system structurally encapsulates the model identifier, similarity score, license constraint level, risk assessment label, and evidence chain information from the results into a unified data object, which is then input into the visual report generation model. The visual report generation model can use a combination of preset templates and rule-driven approaches to perform summary extraction, graphic rendering, and text generation on different fields, converting the similarity score into a comparison chart, the license constraint level into a risk identifier, and the evidence chain into traceability entries, thereby generating a compliance report that can be directly reviewed or archived. After output, the report can be exported as an electronic document, a webpage, or an embedded view within the system, allowing R&D, legal, and management personnel to view the complete assessment basis within a single interface. Regarding the specific model framework used, other model architectures can be selected in practical applications; this application does not impose any limitations on this.

[0094] The report includes the following information: Detection overview (including inspection time, target model information, open-source model version, and usage scenario configuration); Model similarity analysis (including similarity scores for each dimension, overall similarity score, similar region heatmap, and similar region location report); License analysis (including a matching table of the associated open-source model and license, a license constraint extraction table, and constraint application results); Compliance assessment (including infringement risk level (including score and judgment basis) and compliance conclusion (including core basis); Rectification suggestions (including personalized rectification measures, rectification completion verification standards, and rectification time limit suggestions); and Data traceability (including the source, algorithm version, and parameter configuration of all detection data).

[0095] By directly inputting compliance assessment results into the report generation model, model similarity, license constraints, and final conclusions can be consistently expressed within the same report. Assessment evidence is presented synchronously with the results, ensuring consistent data support for compliance reviews, archiving, and subsequent remediation. Because the report includes an overview of the inspection, analytical conclusions, remediation recommendations, and data traceability, the assessment results are more complete in terms of readability and traceability, and information omissions caused by manual compilation are reduced.

[0096] Based on the foregoing embodiments, it further includes:

[0097] Based on the compliance report, a risk visualization interface is generated, which is used to graphically display the compliance risks of the target model.

[0098] Displays a risk visualization interface.

[0099] In this application, the risk visualization interface analyzes and graphically arranges the risk data in the compliance report, transforming the originally scattered compliance conclusions into interface content that can be directly viewed interactively. After the risk visualization interface is displayed, legal personnel, R&D personnel, and management personnel can view the risk level of the target model and its basis within the same interface, and make review, confirmation, and rectification judgments accordingly.

[0100] In one possible implementation, the risk visualization interface includes core indicators, risk warnings, trend changes, and multi-model comparisons. The core indicators include the overall similarity score of the target model, the infringement risk score, and the risk level. The risk warnings include high-risk / extremely high-risk points marked with red pop-ups and low / medium-risk points marked with yellow prompts. The trend changes are displayed as a curve showing the trend change of the similarity score and risk level if the same model is detected multiple times. The multi-model comparison is displayed as a ranking of the risk levels and core risk points of each model if multiple models are detected.

[0101] After adopting the above implementation method, the risk information in the compliance report can be uniformly mapped into a visual result, the compliance risk performance of the target model is more intuitive, the readability and review efficiency of the report content are improved, and the correspondence between risk conclusions and model sources and license restrictions is enhanced.

[0102] Building upon the aforementioned embodiments, further, based on multiple open-source licenses, the license constraint level is determined, including: extracting license structured information and license association rules from multiple open-source licenses; and determining the license constraint level based on the license structured information and license association rules.

[0103] In implementation, after receiving multiple open-source license texts, the system first parses each license, converting the original terms into a unified data representation. Then, it normalizes the restrictive clauses within this data representation, forming comparable structured information. Subsequently, the system matches each structured field according to preset rules to identify whether it contains restrictions on copying and distribution, closed-source distribution restrictions, patent retaliation clauses, or obligations to open-source in the same manner. The system then determines the constraint strength based on the field combination results. For cases where multiple constraints exist within the same open-source license, the system cumulatively evaluates the constraints according to license association rules, using the most stringent constraint as the constraint level for that open-source license. For cases where multiple open-source licenses coexist, the system can also determine the overall applicable license constraint level based on merging logic in the rule base.

[0104] During operation, the open-source license text is structurally extracted to form a unified set of fields. License association rules then perform constraint mapping on this set of fields, ultimately outputting a license constraint level that can be directly used for compliance assessment. This level can be further linked with model similarity results to determine the scope of license restrictions imposed on the target model in reuse, fine-tuning, deployment, or commercialization scenarios.

[0105] In one possible implementation, the license structure information includes fields such as commercial permission / prohibition, attribution requirements, open source requirements, infectious clauses, and modified distribution requirements.

[0106] In one possible implementation, license association rules include license scope (e.g., model type, use case, and distribution form) and conflict rules (e.g., constraint priority after multiple licenses are stacked).

[0107] In one possible implementation, human experience information is obtained. This human experience information consists of standardized experience data and scenario judgment rules accumulated during long-term open-source compliance audits, open-source project implementation, and license conflict assessment. Based on this human experience information, the license constraint level is optimized.

[0108] Specifically, human experience information includes at least one of the following: experience with priority of special clauses, experience with license combination conflicts, experience with scenario adaptation constraints, experience with license version iteration, and experience with industry compliance precedents.

[0109] By combining the actual use cases of open-source code, license combination status, version characteristics, and industry compliance requirements, differentiated optimization and adjustments are performed on the initial level results. For high-risk scenarios such as those with strong constraint special clauses, license mutual exclusion conflicts, and commercial external release, the constraint level is increased. For low-risk scenarios such as those with legal exemption conditions, private isolated use, and non-profit implementation, redundant constraint levels are reasonably decreased. At the same time, accurate supplementary judgments are made for ambiguous boundary scenarios that machines cannot determine, and the level determination deviation of old version algorithms is forcibly fixed and corrected. After the level optimization output is completed, this solution will also continuously accumulate the scenario adaptation rules, level correction logic, and special judgment criteria into the human experience base to achieve dynamic iterative updates of experience data. This optimization method effectively makes up for the shortcomings of pure algorithm judgment, which relies heavily on structured data and neglects actual implementation compliance requirements. It significantly improves the accuracy, scenario adaptability, and industry authority of license constraint level classification in various complex scenarios, and can provide comprehensive and reliable technical support for open-source software compliance risk assessment, open-source license selection, project compliance management, and risk avoidance.

[0110] With this implementation, license terms can be used in an automatic determination in a structured form. The generation of license constraint levels no longer relies on manual reading of the text line by line. Furthermore, the restriction strength of multiple open source licenses can be merged into a consistent conclusion through unified rules, thereby improving the accuracy, stability, and traceability of license analysis.

[0111] In one possible implementation, the compliance assessment result is determined based on the similarity score and the license constraint level, including: determining the infringement risk level of the target model based on the similarity score and the license constraint level; and generating the compliance assessment result based on the infringement risk level, the similarity score, the license constraint level, and the use case information of the target model.

[0112] In its implementation, the system receives the calculated similarity score and license constraint level, maps them to a preset risk matrix, and outputs the infringement risk level of the target model. The risk matrix can be set by combining high and low similarity scores with strong license constraints; a higher similarity score and stronger license constraint level correspond to a higher infringement risk level, while a lower similarity score and weaker license constraint level correspond to a lower infringement risk level. Subsequently, the system inputs the infringement risk level, similarity score, license constraint level, and the target model's usage scenario information into compliance judgment rules. Combining factors such as whether there is commercial deployment, whether interfaces are provided to third parties, whether copyright statements need to be retained, and whether additional license obligations are triggered in the scenario, the system generates a corresponding compliance assessment result. This result may include judgments such as "Allowed to go online," "Restricted use," "Requires rectification before use," or "Prohibited use," and can simultaneously output risk descriptions related to the judgment.

[0113] During operation, the system first aggregates risks based on similarity scores and license constraint levels. Then, it uses usage scenario information to modify the risk aggregation results in a contextualized manner. This allows the compliance assessment results to simultaneously reflect the coupling relationship between model similarity, license constraint strength, and the actual deployment environment. Since the infringement risk level serves as an intermediate layer output, subsequent compliance judgments can directly reference the unified risk classification results, thereby forming traceable and verifiable compliance conclusions.

[0114] With this implementation method, the risk assessment of the target model no longer relies solely on a single similarity or a single license clause. Instead, it integrates both types of information and usage scenarios into the same assessment chain, ensuring that the compliance assessment results are consistent with the actual application environment and improving the stability of risk classification and the interpretability of the output conclusions.

[0115] One possible implementation also includes: determining a corresponding early warning instruction based on the infringement risk level, wherein the early warning instruction includes at least one of an early warning information push instruction and a model locking instruction; and executing the early warning instruction.

[0116] In this embodiment, after determining the infringement risk level of the target model based on similarity score, license constraint level, and target model usage scenario information, the system can input the infringement risk level into a preset instruction mapping rule or risk handling strategy table to match the warning instruction corresponding to the infringement risk level. For cases where the infringement risk level is low but has reached the alert threshold, the warning instruction can be configured as a warning information push instruction. The system encapsulates the infringement risk level, target model identifier, associated open-source model identifier, and corresponding license constraint information into a warning message and sends it through the enterprise's internal management platform, email service, instant messaging interface, or compliance review system so that compliance personnel can intervene and verify it in a timely manner. For cases where the infringement risk level reaches the freezing threshold, the warning instruction can be configured as a model locking instruction. The system will restrict the target model's online call permissions, inference service permissions, or training iteration permissions through the permission control interface, model registry status update interface, or deployment management interface, placing the target model in a controlled state where it cannot be further published, called, or disseminated. In practical applications, the warning information push instruction and the model locking instruction can also be generated simultaneously and executed in parallel based on the same infringement risk level; this application does not limit this.

[0117] When executing an early warning command, the system first calls the corresponding execution module based on the command type. For example, the early warning information push command corresponds to the message sending module, and the model locking command corresponds to the access control module. The message sending module can write the risk source, infringement risk level, associated license terms, similar model names, and suggested actions into the early warning content and complete the sending. The access control module locks the target model's status, writes control flags indicating that it cannot be deployed, downloaded, or invoked, and simultaneously updates the log records. Through this process, the system can directly link the infringement risk level to subsequent actions, ensuring that compliance assessment results not only remain at the output level but are also transformed into actionable risk control results.

[0118] By adopting this implementation method, a correspondence is established between the infringement risk level and the early warning and handling actions, enabling high-risk models to be promptly alerted or immediately restricted, reducing compliance risks caused by continued research and development, deployment and commercial use, while ensuring the traceability and consistency of the risk handling process.

[0119] In one possible implementation, the similarity score between the target model and multiple open-source models is determined based on the first multi-dimensional feature data and the second multi-dimensional feature data, including: determining the first score for each dimension based on the first multi-dimensional feature data and the second multi-dimensional feature data, wherein the first score is the similarity score of the same dimension feature data in the first multi-dimensional feature data and the second multi-dimensional feature data; obtaining the first weight for each dimension; and determining the similarity score between the target model and multiple open-source models based on the first score for each dimension and the corresponding first weight.

[0120] In its implementation, the system first performs multi-dimensional feature alignment between the target model and each open-source model, mapping feature data of the same dimension to a unified feature space, and then obtains the corresponding first score according to a preset similarity function. Subsequently, the system reads the first weight of each dimension and weights and sums the first scores of each dimension with their corresponding first weights to obtain the similarity score between the target model and a single open-source model. For multiple open-source models, the system can perform the above calculations separately to form a set of similarity scores between the target model and multiple open-source models, and output this set to the subsequent license constraint analysis module.

[0121] In one possible implementation, the first multi-dimensional feature data includes at least one of first structural feature data, first parameter feature data, first data feature data, and first behavioral feature data. The second multi-dimensional feature data includes at least one of second structural feature data, second parameter feature data, second data feature data, and second behavioral feature data. Based on the first and second multi-dimensional feature data, a first score for each dimension is determined. The first score includes at least one of the following: structural dimension score, parameter dimension score, data dimension score, and behavioral dimension score. Specifically, the structural dimension score is calculated based on the first and second structural feature data. Both the first and second structural feature data can contain at least one of the following: a structured topology graph, layer structure feature vectors, and a set of structure labels. In actual calculation, the consistency of core architectural information such as the neural network layer arrangement, topological branch structure, and inter-layer connection logic is verified by accurately traversing and aligning the nodes and connecting edges of the network topology graphs corresponding to the two sets of models. The feature matching accuracy of the layer structure feature vectors and the overlap coverage of the structure label set are combined to comprehensively calculate the overall structural matching degree of the model, which is used as the first score of the structural dimension. The parameter dimension score is determined based on the first and second parameter feature data. Both types of parameter feature data cover at least one of the following: parameter distribution feature vector, key layer parameter tensor, and parameter statistical index set. This dimension uses a cosine similarity algorithm to perform a global similarity comparison of the parameter distribution feature vectors of the two sets of data. At the same time, it performs a refined difference comparison on the parameter tensors of the core key layers of the model, and combines the deviation of statistical indicators such as parameter mean, variance, parameter magnitude, and parameter sparsity to integrate and quantify the parameter distribution similarity result, which is the first score of the parameter dimension. The data dimension score is calculated based on the first and second data feature data. The corresponding first and second data feature data include at least one of the following: data fingerprint hash value, data distribution feature vector, and training set overlap value. The uniqueness of the training data is compared by using a hash fingerprint precise matching method, the overlap value of the training sets of the two models is accurately calculated, and the similarity of the data distribution feature vectors is matched. The homogeneity and distribution consistency of the model training data are comprehensively evaluated to obtain the first score of the data dimension. The behavioral dimension score is generated based on the first behavioral feature data and the second behavioral feature data. The two types of behavioral feature data cover at least one of the following: the feature vector of the reasoning result, the behavioral index set, and the output similarity value. A composite algorithm combining cosine similarity and accuracy consistency is adopted. The final reasoning results and the output features of the intermediate layers of the network are compared in a comprehensive manner based on the standard test set to verify the consistency of the model's reasoning logic, output accuracy, and behavioral performance. The output similarity and the degree of behavioral matching are quantitatively calculated to finally obtain the first score of the behavioral dimension.

[0122] In terms of working principle, this scheme calculates local similarity across different dimensions in a hierarchical manner, and then summarizes them uniformly based on a first weight. This allows the similarity relationship between the target model and multiple open-source models to simultaneously reflect differences at the structural, parameter, data, and behavioral levels. The introduction of the first weight allows the influence of different dimensions on the overall score to be explicitly controlled, thereby ensuring that the scoring results are consistent with the requirements of model compliance review.

[0123] With this implementation method, the similarity score between the target model and multiple open-source models no longer depends on a single feature, but is determined comprehensively based on multi-dimensional features and their weights. This can improve the completeness and consistency of the scoring results and provide more stable input for subsequent license constraint level judgment and compliance assessment results generation.

[0124] Based on the aforementioned embodiments, further, the first weights of each dimension are obtained, including: modifying the preset second weights based on the model type of the target model to obtain the first weights of each dimension.

[0125] During operation, the system first obtains the target model's type information, then calls the preset weight library to read the second weight, and reallocates the weights of each dimension according to the type information to obtain the first weight, which participates in the subsequent weighted calculation of similarity scores. Since the first weight is generated driven by the target model type, the same set of basic features can present different importance distributions under different model types, thus making the similarity score more in line with the actual judgment requirements of the target model.

[0126] For example, if the target model is a vehicle-side intelligent driving model, the weight of the data dimension can be increased to 0.3 (the default is 0.2), and the weight of the other dimensions can be appropriately reduced.

[0127] By adopting the above method, the weight configuration can be automatically adjusted according to the target model type, avoiding the scoring deviation caused by using the same weight for different types of models, making the similarity analysis results of each dimension more consistent with the actual feature distribution of the model, and improving the consistency and adaptability of subsequent compliance assessments.

[0128] In this application, the target model to be detected, multiple open-source models, and multiple open-source licenses corresponding to each open-source model are first obtained, and the target model and each open-source model are uniformly parsed. First multi-dimensional feature data is extracted from the target model, and second multi-dimensional feature data is extracted from the open-source models. The multi-dimensional feature data can simultaneously cover multiple dimensions such as structure, parameters, data, and behavior. Specifically, structural feature data includes at least one of the following: a structured topology graph, layer structure feature vectors, and a set of structure labels; parameter feature data includes at least one of the following: parameter distribution feature vectors, key layer parameter tensors, and a set of parameter statistical indicators; data feature data includes at least one of the following: data fingerprint hash value, data distribution feature vectors, and training set overlap values; and behavioral feature data includes at least one of the following: inference result feature vectors, behavioral indicator sets, and output similarity values.

[0129] After feature extraction, the first multi-dimensional feature data of the target model and the second multi-dimensional feature data of each open-source model are compared along the same dimensions to determine the first score for each dimension. The first score is the similarity score between feature data of the same dimension. Then, the first weight of each dimension is obtained. The first weight is obtained by modifying the preset second weight according to the model type of the target model. Finally, the first score of each dimension is weighted and calculated with the corresponding first weight to determine the similarity score between the target model and each of the multiple open-source models.

[0130] Simultaneously, text and rule parsing is performed on multiple open-source licenses to extract structured license information and license association rules. Based on this information, the license constraint level is determined. The structured license information includes the license's scope of authorization, distribution conditions, modification requirements, attribution requirements, open-source contagion requirements, commercial restrictions, patent clauses, and liability limitations. License association rules characterize the compatibility, inheritance, conflict, or overlapping relationships between different licenses, thereby forming the license constraint level corresponding to each open-source model.

[0131] After obtaining the similarity score and license constraint level, the infringement risk level of the target model is determined by combining the two. The infringement risk level, similarity score, license constraint level, and the target model's use case information are then input into the evaluation logic to generate a compliance assessment result. Use case information may include whether the target model is used for commercial release, internal R&D, online services, secondary distribution, or embedded deployment. If the infringement risk level reaches a preset trigger condition, a corresponding warning instruction is determined based on that infringement risk level. The warning instruction includes at least one of a warning information push instruction and a model locking instruction, and the corresponding warning information push or model locking control is executed. The compliance assessment result is then output.

[0132] In one possible implementation, the compliance assessment results are input into a pre-defined visualization report generation model, which generates and outputs a compliance report. The compliance report includes an inspection overview, model similarity analysis, license analysis, compliance assessment, remediation recommendations, and data traceability. Based on this compliance report, a risk visualization interface is generated to graphically display the compliance risks of the target model and show the risk visualization interface.

[0133] Figure 2 A schematic diagram of the structure of the compliance assessment device for the model provided in this application embodiment is shown below. Figure 2 As shown, the compliance assessment device 20 for the model provided in this embodiment includes an acquisition module 201, a processing module 202, and an output module 203.

[0134] Module 201 is used to acquire the target model to be detected, multiple open-source models, and multiple open-source licenses.

[0135] Processing module 202 is used to extract first multi-dimensional feature data from the target model;

[0136] Processing module 202 is also used to extract second multi-dimensional feature data from multiple open-source models;

[0137] The processing module 202 is also used to determine the similarity score between the target model and multiple open-source models based on the first multi-dimensional feature data and the second multi-dimensional feature data.

[0138] Processing module 202 is also used to determine the license constraint level based on multiple open source licenses;

[0139] Processing module 202 is also used to determine the compliance assessment results based on similarity scores and license constraint levels;

[0140] Output module 203 is used to output the compliance assessment results.

[0141] In one possible embodiment, the first multidimensional feature data includes first structural feature data, which includes at least one of a structured topology graph, a layered structure feature vector, and a set of structural labels.

[0142] In one possible embodiment, the first multidimensional feature data includes first parameter feature data, which includes at least one of the following: parameter distribution feature vector, key layer parameter tensor, and parameter statistical index set.

[0143] In one possible embodiment, the first multi-dimensional feature data includes first data feature data, which includes at least one of data fingerprint hash value, data distribution feature vector, and training set overlap value.

[0144] In one possible embodiment, the first multi-dimensional feature data includes first behavioral feature data, which includes at least one of the following: inference result feature vector, behavioral index set, and output similarity value.

[0145] In one possible embodiment, it also includes:

[0146] The compliance assessment results are input into a pre-defined visualization report generation model to obtain and output a compliance report. The compliance report includes an inspection overview, model similarity analysis, license analysis, compliance assessment, rectification suggestions, and data traceability.

[0147] In one possible embodiment, it also includes:

[0148] Based on the compliance report, a risk visualization interface is generated, which is used to graphically display the compliance risks of the target model.

[0149] Displays a risk visualization interface.

[0150] In one possible implementation, the license constraint level is determined based on multiple open-source licenses, including:

[0151] Extract structured license information and license association rules from multiple open-source licenses;

[0152] Based on the structured information of the license and the rules associated with the license, the level of license constraint is determined.

[0153] In one possible embodiment, the compliance assessment result is determined based on the similarity score and the license constraint level, including:

[0154] Based on similarity scores and license constraint levels, the infringement risk level of the target model is determined;

[0155] Based on the infringement risk level, similarity score, license constraint level, and the use case information of the target model, a compliance assessment result is generated.

[0156] In one possible embodiment, it also includes:

[0157] Based on the level of infringement risk, a corresponding early warning instruction is determined. The early warning instruction includes at least one of an early warning information push instruction and a model locking instruction.

[0158] Execute the warning command.

[0159] In one possible embodiment, based on first multi-dimensional feature data and second multi-dimensional feature data, a similarity score between the target model and multiple open-source models is determined, including:

[0160] Based on the first multi-dimensional feature data and the second multi-dimensional feature data, a first score for each dimension is determined. The first score is the similarity score of the same dimension feature data in the first multi-dimensional feature data and the second multi-dimensional feature data.

[0161] Obtain the first weight for each dimension;

[0162] Based on the first score and corresponding first weight of each dimension, the similarity score between the target model and multiple open source models is determined.

[0163] In one possible implementation, obtaining the first weights for each dimension includes:

[0164] Based on the model type of the target model, the preset second weights are modified to obtain the first weights for each dimension.

[0165] The compliance assessment device for the model provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0166] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 3 As shown, the electronic device 30 provided in this embodiment includes at least one processor 301 and a memory 302. Optionally, the device 30 further includes a communication component 303. The processor 301, memory 302, and communication component 303 are connected via a bus.

[0167] In a specific implementation, at least one processor 301 executes computer execution instructions stored in memory 302, causing at least one processor 301 to perform the above-described method.

[0168] The specific implementation process of processor 301 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0169] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0170] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0171] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0172] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0173] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0174] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0175] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0176] The division of units is merely a logical functional division; 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 indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0177] 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; that is, 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 according to actual needs.

[0178] In addition, the functional units in the various embodiments of the present invention 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.

[0179] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part 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 of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0180] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0181] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A compliance assessment method for a model, characterized in that, include: Obtain the target model to be detected, multiple open-source models, and multiple open-source licenses; Extract the first multi-dimensional feature data from the target model; Extract second multi-dimensional feature data from the multiple open-source models; Based on the first multi-dimensional feature data and the second multi-dimensional feature data, a similarity score between the target model and the multiple open-source models is determined; Based on the aforementioned multiple open-source licenses, the license constraint level is determined; The compliance assessment result is determined based on the similarity score and the license constraint level; Output the compliance assessment results.

2. The method according to claim 1, characterized in that, The first multi-dimensional feature data includes first structural feature data, which includes at least one of a structured topology graph, a layer structure feature vector, and a set of structural labels.

3. The method according to claim 1, characterized in that, The first multi-dimensional feature data includes first parameter feature data, which includes at least one of the following: parameter distribution feature vector, key layer parameter tensor, and parameter statistical index set.

4. The method according to claim 1, characterized in that, The first multi-dimensional feature data includes first data feature data, which includes at least one of data fingerprint hash value, data distribution feature vector, and training set overlap value.

5. The method according to claim 1, characterized in that, The first multi-dimensional feature data includes first behavioral feature data, which includes at least one of the following: inference result feature vector, behavioral index set, and output similarity value.

6. The method according to any one of claims 1-5, characterized in that, Also includes: The compliance assessment results are input into a preset visualization report generation model to obtain and output a compliance report, which includes an inspection overview, model similarity analysis, license analysis, compliance assessment, rectification suggestions, and data traceability.

7. The method according to claim 6, characterized in that, Also includes: Based on the compliance report, a risk visualization interface is generated, which is used to graphically display the compliance risks of the target model. The risk visualization interface is displayed.

8. The method according to claim 1, characterized in that, The determination of the license constraint level based on the multiple open-source licenses includes: Extract the structured license information and license association rules from the multiple open-source licenses; Based on the structured information of the license and the association rules of the license, the license constraint level is determined.

9. The method according to claim 1, characterized in that, The determination of the compliance assessment result based on the similarity score and the license constraint level includes: Based on the similarity score and the license constraint level, the infringement risk level of the target model is determined; Based on the infringement risk level, the similarity score, the license constraint level, and the use case information of the target model, a compliance assessment result is generated.

10. The method according to claim 9, characterized in that, Also includes: Based on the infringement risk level, a corresponding early warning instruction is determined, and the early warning instruction includes at least one of an early warning information push instruction and a model locking instruction; Execute the aforementioned warning command.

11. The method according to claim 1, characterized in that, The step of determining the similarity score between the target model and the multiple open-source models based on the first multi-dimensional feature data and the second multi-dimensional feature data includes: Based on the first multi-dimensional feature data and the second multi-dimensional feature data, a first score for each dimension is determined. The first score is the similarity score of the same dimension feature data in the first multi-dimensional feature data and the second multi-dimensional feature data. Obtain the first weight for each dimension; Based on the first score and corresponding first weight of each dimension, the similarity score between the target model and the multiple open-source models is determined.

12. The method according to claim 1, characterized in that, The process of obtaining the first weight for each dimension includes: Based on the model type of the target model, the preset second weights are modified to obtain the first weights for each dimension.

13. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-12.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-12.

Images