An AI software supply chain-oriented security evidence fusion analysis method and system

Abstract

The application provides an AI software supply chain-oriented security evidence fusion analysis method and system, and relates to the technical field of AI software supply chain security. The method comprises receiving an AI software supply chain and extracting AI software components, and constructing an AI software supply chain topology network; performing security detection on the supply chain to generate a security scanning report, performing semantic field mapping and spatial coordinate marker alignment on the report based on the topology network, outputting security evidence data and performing high-dimensional space vector projection, calculating the semantic similarity of the security evidence data and historical empirical data, and screening risk evidence data with a semantic similarity higher than a threshold; extracting risk evidence data of the same component and importing a hybrid judgment engine to perform vulnerability existence judgment and vulnerability exploitation accessibility verification, prioritizing the risk evidence data based on vulnerability risk values, and outputting a risk priority evaluation report, thereby improving the security audit accuracy and disposal efficiency of the AI software supply chain.

Description

Technical Field

[0001] This invention relates to the field of AI software supply chain security technology, and in particular to a method and system for security evidence fusion analysis of AI software supply chains. Background Technology

[0002] With the widespread application of AI technologies such as large models and intelligent algorithms, the complexity of the AI ​​software supply chain has increased significantly, encompassing multiple components such as code, model files, and third-party dependency images. These components are intertwined with multi-level dependencies, leading to a continuous increase in the risk of supply chain attacks.

[0003] In existing AI software supply chain security testing technologies, the report fields output by various scanning tools are inconsistent, vulnerability evidence is fragmented, and there is a lack of unified data fusion standards. Conventional testing only lists vulnerability entries for individual components, without outlining direct and indirect propagation dependencies between components, making it impossible to build a reliable supply chain relationship view and pinpoint the propagation path of vulnerabilities in the business chain. Furthermore, existing technologies rely solely on fixed rules to match vulnerabilities, without combining historical vulnerability evidence for semantic verification, easily leading to numerous false positives and false negatives. Moreover, risk assessment based solely on basic vulnerability scores ignores component asset value, business scenario constraints, and attack path reachability, resulting in a lack of objective quantitative basis for risk classification. Traditional testing models lack evidence fusion and reasoning capabilities, failing to comprehensively deduce the true probability of vulnerability existence from multi-source scanning evidence, and also failing to verify whether vulnerabilities can be exploited externally. Operations personnel struggle to distinguish between urgent remediation items and low-risk vulnerabilities, leading to low efficiency in auditing and remediation work and failing to meet the actual security audit needs of the AI ​​software supply chain.

[0004] Therefore, it is necessary to provide a secure evidence fusion analysis method and system for the AI ​​software supply chain to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a security evidence fusion analysis method and system for the AI ​​software supply chain. This system solves the problems of inconsistent data formats, inability to fuse and verify evidence, difficulty in determining the actual existence of vulnerabilities and attack paths, one-sided risk assessment, and inability to accurately prioritize risk handling in existing security detection tools for the AI ​​software supply chain.

[0006] This invention provides a method for security evidence fusion analysis in the AI ​​software supply chain, the method comprising: Receive the AI ​​software supply chain to be audited and extract the AI ​​software components corresponding to the AI ​​software supply chain. Construct an AI software supply chain topology network with the AI ​​software components as nodes and the dependencies between the AI ​​software components as edges. The security scanning tool is scheduled to perform concurrent security detection on the AI ​​software supply chain, generate a security scanning report, and perform semantic field mapping and spatial coordinate label alignment on the vulnerability records in the security scanning report according to the topology network of the AI ​​software supply chain, and output security evidence data. Load historical empirical data from the historical vulnerability database, perform high-dimensional space vector projection on the security evidence data, calculate the semantic similarity between the security evidence data and the historical empirical data, and filter the security evidence data with semantic similarity higher than a preset similarity threshold as risk evidence data; Extract the risk evidence data from the same AI software component, import it into a two-stage hybrid judgment engine to determine the existence of vulnerabilities and verify the reachability of vulnerability exploitation, calculate the vulnerability risk value of the risk evidence data, prioritize the risk evidence data based on the vulnerability risk value from largest to smallest, and output a security evidence risk priority assessment report.

[0007] Preferably, the step of receiving the AI ​​software supply chain to be audited and extracting the AI ​​software components corresponding to the AI ​​software supply chain, and constructing an AI software supply chain topology network with the AI ​​software components as nodes and the dependencies between the AI ​​software components as edges, includes: The inventory file, image configuration file, and model dependency description file of the AI ​​software supply chain to be audited are parsed to extract the name, version number, and deployment carrier information of the AI ​​software components and generate a set of component metadata. Traverse the component metadata set and use a directed dependency resolution algorithm to identify the direct dependencies, indirect transitive dependencies and optional conditional dependencies between the AI ​​software components, and generate dependency triples. Loop detection and redundant edge removal are performed on the dependency triples. The component metadata set and the dependency triples are digitally encoded using an adjacency matrix to generate an initial topological adjacency matrix. Based on the initial topological adjacency matrix, the AI ​​software components are used as nodes, the dependencies between the AI ​​software components are used as edges, and a corresponding spatial identifier code is assigned to each node. A corresponding dependency weight coefficient is assigned to each edge to construct the AI ​​software supply chain topological network.

[0008] Preferably, the scheduling security scanning tool performs concurrent security checks on the AI ​​software supply chain, generates a security scan report, and performs semantic field mapping and spatial coordinate label alignment on the vulnerability records in the security scan report according to the AI ​​software supply chain topology network, outputting security evidence data, including: The security scanning tools, including static code scanning tools, AI model vulnerability detection tools, and third-party component compliance scanning tools, are scheduled to perform concurrent security detection on the AI ​​software supply chain, generating multiple security scanning reports in heterogeneous formats. Field normalization and cleaning are performed on the multiple heterogeneous security scanning reports to obtain the original vulnerability records. Construct a cross-tool semantic mapping dictionary, establish mapping rules between tool-private fields and evidence unification fields, convert the tool-private fields in the original vulnerability record into the evidence unification fields, and generate an intermediate vulnerability record containing the evidence unification fields; Read the spatial identifier code corresponding to each AI software component in the AI ​​software supply chain, match it with the affected component information in the intermediate vulnerability record, embed the spatial identifier code into the intermediate vulnerability record, and output the security evidence data.

[0009] Preferably, the process of loading historical evidence data from the historical vulnerability database, performing high-dimensional space vector projection on the security evidence data, calculating the semantic similarity between the security evidence data and the historical evidence data, and selecting security evidence data with semantic similarity higher than a preset similarity threshold as risk evidence data includes: The vulnerability CVE number, vulnerability triggering conditions, affected component version, and supply chain attack path description text are read from the security evidence data, concatenated to form a security evidence text sequence, and the security evidence text sequence is segmented and filtered for stop words using a segmented sliding window strategy. The pre-trained domain vector encoder is invoked to map the security evidence text sequence after word segmentation and stop word filtering to a high-dimensional continuous vector space to generate the corresponding security evidence vector. At the same time, the historical empirical data is read from the historical vulnerability database to generate the corresponding historical empirical vector. Using a weighted improved RBF kernel distance, the vector distance between the secure evidence vector x and the historical empirical vector y is calculated and mapped to the semantic similarity D(x,y) as follows: ; In the formula, exp() is an exponential function with base e; I is the kernel width adjustment coefficient; I is the total dimension of the high-order continuous vector space; Let x be the i-th component of the security evidence vector. The corresponding dimension weight coefficients; Let be the i-th component of the historical empirical vector y; Security evidence data with semantic similarity higher than the preset similarity threshold are selected as risk evidence data.

[0010] Preferably, the vulnerability determination includes: The two-stage hybrid decision engine includes a rule-based decision sub-engine and a probabilistic graph reasoning sub-engine. The rule-determination sub-engine loads the vulnerability determination rule set of the AI ​​software supply chain. Based on the vulnerability determination rule set, it reads the affected component version range, vulnerability CVE number, and component runtime configuration constraint information from the risk evidence data and performs hard rule matching, outputting the vulnerability existence confidence level. ; Collect multiple risk evidence data corresponding to each of the AI ​​software components, and establish a confidence level for the vulnerabilities corresponding to the multiple risk evidence data. The basic fusion confidence level is obtained by performing evidence fusion weighted calculation. for: ; In the formula, Let K be the credibility coefficient of the risk evidence data corresponding to the kth risk evidence data; K is the total number of risk evidence data corresponding to the AI ​​software component. Using the probabilistic graphical inference sub-engine, the AI ​​software components are used as inference variables. A Bayesian inference network is constructed by combining the dependencies between the AI ​​software components, and the basic fusion confidence is then used. As the prior probability input, the probability of the posterior vulnerability actually existing is calculated. The probability of the actual existence of the posterior vulnerability is compared with a preset existence probability threshold. If the posterior probability is greater than the preset existence probability threshold, the AI ​​software component is determined to have a vulnerability, and the vulnerability determination result is output.

[0011] Preferably, the vulnerability exploit reachability verification includes: Obtain the vulnerability assessment results, extract the vulnerability attack prerequisites, component access permission requirements and component interface constraints, and construct a set of vulnerability exploitation prerequisite constraints. Traverse the AI ​​software supply chain topology network, taking the entry node of the AI ​​software supply chain topology network as the target starting point and the AI ​​software component with the vulnerability as the target ending point, and use a directed depth-first path search algorithm to search for accessible dependency paths from the target starting point to the target ending point. Each AI software component on the accessible dependency path is verified to ensure that it meets the preconditions for exploiting the vulnerability. If each AI software component on the accessible dependency path meets the preconditions for exploiting the vulnerability, the accessible dependency path is marked as a reachable attack path. If each AI software component on the accessible dependency path does not meet the preconditions for exploiting the vulnerability, the accessible dependency path is marked as an unreachable attack path. The number of reachable attack paths and the length of the shortest reachable attack path are counted to generate a reachability attack quantification index, which is then linked to the risk evidence data.

[0012] Preferably, calculating the vulnerability risk value of the risk evidence data includes: Parse the vulnerability CVE number carried in the risk evidence data, search the pre-built vulnerability information database based on the vulnerability CVE number, and obtain the vulnerability CVSS score. Simultaneously extract the probability of the actual existence of the posterior vulnerability. The number of reachable attack paths The shortest attack path length ; Introducing attenuation coefficient Calculate the initial vulnerability risk value of the risk evidence data. for: ; Read the component asset importance weights configured in the system. AI business scenario risk correction coefficient Regarding the initial vulnerability risk value The vulnerability risk value was obtained by making corrections. for: .

[0013] A security evidence fusion analysis system for the AI ​​software supply chain, the system comprising: The topology network construction module is used to receive the AI ​​software supply chain to be audited and extract the AI ​​software components corresponding to the AI ​​software supply chain. The AI ​​software components are used as nodes and the dependencies between the AI ​​software components are used as edges to construct the AI ​​software supply chain topology network. The security evidence output module is used to schedule security scanning tools to perform concurrent security detection on the AI ​​software supply chain, generate a security scanning report, perform semantic field mapping and spatial coordinate label alignment on the vulnerability records in the security scanning report according to the topology network of the AI ​​software supply chain, and output security evidence data. The risk evidence screening module is used to load historical empirical data from the historical vulnerability database, perform high-dimensional space vector projection on the security evidence data, calculate the semantic similarity between the security evidence data and the historical empirical data, and screen the security evidence data with semantic similarity higher than a preset similarity threshold as risk evidence data. The assessment report output module is used to extract the risk evidence data of the same AI software component, import it into a two-stage hybrid judgment engine to determine the existence of vulnerabilities and verify the accessibility of vulnerability exploitation, calculate the vulnerability risk value of the risk evidence data, prioritize the risk evidence data based on the vulnerability risk value from largest to smallest, and output a security evidence risk priority assessment report.

[0014] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor performs the steps of a security evidence fusion analysis method for an AI software supply chain as described in any of the above.

[0015] A readable storage medium storing a computer program, which, when executed by a processor, is used to implement the steps of a secure evidence fusion analysis method for an AI software supply chain as described above.

[0016] Compared with related technologies, the security evidence fusion analysis method and system for AI software supply chain provided by this invention has the following beneficial effects: This invention receives the AI ​​software supply chain to be audited and extracts the corresponding AI software components. Using AI software components as nodes and dependencies between them as edges, it constructs an AI software supply chain topology network. It then schedules security scanning tools to perform concurrent security checks on the AI ​​software supply chain, generating a security scan report. Based on the AI ​​software supply chain topology network, it performs semantic field mapping and spatial coordinate alignment on the vulnerability records in the security scan report, outputting security evidence data. It loads historical empirical data from a historical vulnerability database, performs high-dimensional spatial vector projection on the security evidence data, calculates the semantic similarity between the security evidence data and the historical empirical data, and selects security evidence data with semantic similarity higher than a preset similarity threshold as risk evidence data. It extracts risk evidence data for the same AI software component and imports it into a two-stage hybrid judgment engine to determine vulnerability existence and verify vulnerability exploitability, calculating the vulnerability risk value of the risk evidence data. Based on the vulnerability risk value, it prioritizes the risk evidence data from largest to smallest, outputting a security evidence risk priority assessment report. This allows for accurate identification of vulnerabilities in the AI ​​software supply chain, improving the accuracy and efficiency of security audits of the AI ​​software supply chain.

[0017] This invention constructs a topological relational view of the AI ​​software supply chain, unifying vulnerability data standards across various scanning tools, eliminating data silos from heterogeneous reports, integrating multi-source security evidence, significantly reducing the workload of manually compiling vulnerability data, and filtering invalid or low-relevance alerts through high-dimensional semantic matching, thereby significantly reducing false alarms and improving the credibility of vulnerability evidence. This invention employs a hybrid reasoning model combining rules and probabilistic graphs, integrating multiple pieces of evidence to comprehensively deduce the true probability of vulnerability existence, avoiding the bias caused by single-rule judgments. It uses topological network retrieval to accessable attack paths, verifying the actual exploitation conditions of vulnerabilities, and eliminating invalid vulnerabilities that cannot be triggered, ensuring that security audit results closely reflect the real business operating environment. This invention constructs risk quantification indicators by comprehensively considering vulnerability severity, the true probability of vulnerability existence, the number of attack paths, component asset weights, and business scenario attributes. This enables vulnerability classification and ranking, intuitively distinguishing high-risk priority remediation items, thereby achieving automated vulnerability analysis of the AI ​​software supply chain, comprehensively improving the accuracy and efficiency of security audits of the AI ​​software supply chain, and effectively reducing the security risks of supply chain attacks on AI businesses. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a secure evidence fusion analysis method for the AI ​​software supply chain, provided as an embodiment of the present invention; Figure 2 A system block diagram of a secure evidence fusion analysis system for the AI ​​software supply chain provided in this embodiment of the invention; Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 The diagram shown is a flowchart of a secure evidence fusion analysis method for the AI ​​software supply chain provided by an embodiment of the present invention. Figure 1The execution subject of the method shown can be a software and / or hardware device. The execution subject of this invention can include, but is not limited to, at least one of the following: user equipment, network equipment, etc. User equipment can include, but is not limited to, computers, smartphones, personal digital assistants (PDAs), and the aforementioned electronic devices. Network equipment can include, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers. This embodiment does not limit this. Steps S1 to S4 are detailed as follows: S1. Receive the AI ​​software supply chain to be audited and extract the AI ​​software components corresponding to the AI ​​software supply chain. Construct an AI software supply chain topology network with the AI ​​software components as nodes and the dependencies between the AI ​​software components as edges.

[0021] The process of receiving the AI ​​software supply chain to be audited and extracting the corresponding AI software components, constructing an AI software supply chain topology network with the AI ​​software components as nodes and the dependencies between the AI ​​software components as edges, includes: The inventory file, image configuration file, and model dependency description file of the AI ​​software supply chain to be audited are parsed to extract the name, version number, and deployment carrier information of the AI ​​software components and generate a set of component metadata. Traverse the component metadata set and use a directed dependency resolution algorithm to identify the direct dependencies, indirect transitive dependencies and optional conditional dependencies between the AI ​​software components, and generate dependency triples. Loop detection and redundant edge removal are performed on the dependency triples. The component metadata set and the dependency triples are digitally encoded using an adjacency matrix to generate an initial topological adjacency matrix. Based on the initial topological adjacency matrix, the AI ​​software components are used as nodes, the dependencies between the AI ​​software components are used as edges, and a corresponding spatial identifier code is assigned to each node. A corresponding dependency weight coefficient is assigned to each edge to construct the AI ​​software supply chain topological network.

[0022] The inventory files, image configuration files, and model dependency description files of the AI ​​software supply chain to be audited are structured and parsed to extract the corresponding identifiers, version identifiers, and deployment carrier attributes of each AI software component, and then aggregated into a set of component metadata.

[0023] A directed dependency resolution algorithm is used to traverse the entire component metadata set to identify the relationships between components, including direct dependencies that are directly bound to components, indirect transitive dependencies formed through multiple layers of components, and optional conditional dependencies that are only effective due to runtime environment constraints. These component relationships are then encapsulated into dependency triples.

[0024] For dependency triples, a loop detection operation is first performed to identify and remove cyclic dependency links, followed by a redundant edge removal operation to delete duplicate and invalid dependency associations, thereby simplifying the topology association data. An adjacency matrix encoding mechanism is used to simultaneously digitize the component metadata set and the cleaned dependency triples, outputting an initial topology adjacency matrix.

[0025] Based on the initial topology adjacency matrix, AI software components are defined as network nodes, and various dependencies between components are defined as directed edges. Each network node is assigned a unique spatial identifier code for accurate matching of subsequent security evidence data with topology nodes; each edge representing a dependency relationship is configured with a corresponding dependency weight coefficient to distinguish the importance of business associations in different dependency links.

[0026] S2, schedule the security scanning tool to perform concurrent security detection on the AI ​​software supply chain, generate a security scanning report, perform semantic field mapping and spatial coordinate label alignment on the vulnerability records in the security scanning report according to the topology network of the AI ​​software supply chain, and output security evidence data.

[0027] The scheduling security scanning tool performs concurrent security checks on the AI ​​software supply chain, generates a security scan report, and performs semantic field mapping and spatial coordinate label alignment on the vulnerability records in the security scan report based on the AI ​​software supply chain topology network, outputting security evidence data, including: The security scanning tools, including static code scanning tools, AI model vulnerability detection tools, and third-party component compliance scanning tools, are scheduled to perform concurrent security detection on the AI ​​software supply chain, generating multiple security scanning reports in heterogeneous formats. Field normalization and cleaning are performed on the multiple heterogeneous security scanning reports to obtain the original vulnerability records. Construct a cross-tool semantic mapping dictionary, establish mapping rules between tool-private fields and evidence unification fields, convert the tool-private fields in the original vulnerability record into the evidence unification fields, and generate an intermediate vulnerability record containing the evidence unification fields; Read the spatial identifier code corresponding to each AI software component in the AI ​​software supply chain, match it with the affected component information in the intermediate vulnerability record, embed the spatial identifier code into the intermediate vulnerability record, and output the security evidence data.

[0028] Static code scanning tools, AI model vulnerability detection tools, and third-party component compliance scanning tools are used in parallel to perform security testing on the AI ​​software supply chain. These tools can independently output heterogeneous security scan reports with incompatible formats. Due to the inconsistent data field specifications of various scanning tools, direct merging creates data silos. Therefore, all heterogeneous security scan reports are uniformly normalized and cleaned, redundant and invalid fields are removed, and missing fields are filled in to obtain the original vulnerability records.

[0029] A cross-tool semantic mapping dictionary is constructed to be compatible with all security scanning tools. This dictionary contains mapping conversion rules that correspond one-to-one between tool-specific fields and evidence uniformity fields. Based on the mapping conversion rules, the original vulnerability records are traversed, and all tool-specific fields with inconsistent definitions are converted into globally universal evidence uniformity fields. After the conversion, an intermediate vulnerability record containing the evidence uniformity fields is generated.

[0030] Retrieve the spatial identifier codes assigned to each AI software component within the AI ​​software supply chain, extract the affected component information recorded in the intermediate vulnerability records, establish a matching association between the component entity and the corresponding spatial identifier code, embed the matched spatial identifier code into the corresponding intermediate vulnerability record, realize the spatial coordinate binding and alignment between the vulnerability record and the supply chain topology node, and output security evidence data.

[0031] S3, Load historical empirical data from the historical vulnerability database, perform high-dimensional space vector projection on the security evidence data, calculate the semantic similarity between the security evidence data and the historical empirical data, and select the security evidence data with semantic similarity higher than a preset similarity threshold as risk evidence data.

[0032] The process involves loading historical empirical data from the historical vulnerability database, performing high-dimensional space vector projection on the security evidence data, calculating the semantic similarity between the security evidence data and the historical empirical data, and selecting security evidence data with semantic similarity higher than a preset similarity threshold as risk evidence data, including: The vulnerability CVE number, vulnerability triggering conditions, affected component version, and supply chain attack path description text are read from the security evidence data, concatenated to form a security evidence text sequence, and the security evidence text sequence is segmented and filtered for stop words using a segmented sliding window strategy. The pre-trained domain vector encoder is invoked to map the security evidence text sequence after word segmentation and stop word filtering to a high-dimensional continuous vector space to generate the corresponding security evidence vector. At the same time, the historical empirical data is read from the historical vulnerability database to generate the corresponding historical empirical vector. Using a weighted improved RBF kernel distance, the vector distance between the secure evidence vector x and the historical empirical vector y is calculated and mapped to the semantic similarity D(x,y) as follows: ; In the formula, exp() is an exponential function with base e; I is the kernel width adjustment coefficient; I is the total dimension of the high-order continuous vector space; Let x be the i-th component of the security evidence vector. The corresponding dimension weight coefficients; Let be the i-th component of the historical empirical vector y; Security evidence data with semantic similarity higher than the preset similarity threshold are selected as risk evidence data.

[0033] The vulnerability CVE numbers, vulnerability triggering conditions, affected component versions, and supply chain attack path descriptions are extracted from the security evidence data and then concatenated to obtain a security evidence text sequence. A segmented sliding window strategy is used to clean the security evidence text sequence, performing word segmentation and stop word filtering sequentially to remove redundant words with no actual semantic value, retaining semantic text that represents the vulnerability.

[0034] A pre-trained domain vector encoder adapted for AI supply chain security scenarios is invoked to project the security evidence text sequence, after word segmentation and stop word filtering, into a high-dimensional continuous vector space, generating a security evidence vector representing the semantic features of this vulnerability. Simultaneously, historical empirical data stored in the historical vulnerability database is retrieved, and vector transformation is performed using source code logic to generate corresponding historical empirical vectors. This ensures that the security evidence vector and the historical empirical vector reside in the same feature space, guaranteeing semantic comparability for subsequent distance calculations.

[0035] The vector space distance between the security evidence vector and the historical empirical vector is calculated by weighted improving the RBF kernel. Then, an exponential function is used to realize the mapping transformation from vector space distance to semantic similarity. In the formula, the kernel width adjustment coefficient is used to control the similarity decay rate, the weight coefficients of each dimension are used to distinguish the contribution of different vector dimensions to the vulnerability semantics, and the total dimension of the high-dimensional space uniformly constrains the feature dimensions of the security evidence vector and the historical empirical vector.

[0036] By comparing the semantic similarity of each security evidence data with the system's preset similarity threshold, security evidence data whose semantic similarity reaches the threshold is marked and retained, and uniformly classified as risk evidence data. This can filter out invalid alarms with low relevance and retain only valid risk evidence that highly matches the semantics of historical vulnerability evidence.

[0037] S4. Extract the risk evidence data of the same AI software component, import it into a two-stage hybrid judgment engine to determine the existence of vulnerabilities and verify the reachability of vulnerability exploitation, calculate the vulnerability risk value of the risk evidence data, prioritize the risk evidence data based on the vulnerability risk value from largest to smallest, and output a security evidence risk priority assessment report.

[0038] The determination of the existence of the vulnerability includes: The two-stage hybrid decision engine includes a rule-based decision sub-engine and a probabilistic graph reasoning sub-engine. The rule-determination sub-engine loads the vulnerability determination rule set of the AI ​​software supply chain. Based on the vulnerability determination rule set, it reads the affected component version range, vulnerability CVE number, and component runtime configuration constraint information from the risk evidence data and performs hard rule matching, outputting the vulnerability existence confidence level. ; Collect multiple risk evidence data corresponding to each of the AI ​​software components, and establish a confidence level for the vulnerabilities corresponding to the multiple risk evidence data. The basic fusion confidence level is obtained by performing evidence fusion weighted calculation. for: ; In the formula, Let K be the credibility coefficient of the risk evidence data corresponding to the kth risk evidence data; K is the total number of risk evidence data corresponding to the AI ​​software component. Using the probabilistic graphical inference sub-engine, the AI ​​software components are used as inference variables. A Bayesian inference network is constructed by combining the dependencies between the AI ​​software components, and the basic fusion confidence is then used. As the prior probability input, the probability of the posterior vulnerability actually existing is calculated. The probability of the actual existence of the posterior vulnerability is compared with a preset existence probability threshold. If the posterior probability is greater than the preset existence probability threshold, the AI ​​software component is determined to have a vulnerability, and the vulnerability determination result is output.

[0039] The rule-based judgment sub-engine loads a set of vulnerability judgment rules adapted to the AI ​​software supply chain scenario, extracts the affected component version range, vulnerability CVE number, and component runtime configuration constraint information contained in a single piece of risk evidence data, and performs hard rule matching based on the set of vulnerability judgment rules to output the vulnerability existence confidence of the corresponding single piece of risk evidence.

[0040] In scenarios where a single AI software component is associated with multiple pieces of risk evidence data, the confidence level of vulnerability existence corresponding to all risk evidence of that component is unified and evidence fusion weighted calculation is performed to obtain the basic fusion confidence level. The evidence credibility coefficient in the formula is used to differentiate the credibility of different risk evidence, thereby achieving the unification of multiple pieces of risk evidence under the same component and eliminating the one-sidedness of the judgment result of a single piece of evidence.

[0041] The probabilistic graphical inference sub-engine is launched, using AI software components as inference variables. A Bayesian inference network is built by combining the component dependencies recorded in the AI ​​software supply chain topology network. The basic fusion confidence is imported into the Bayesian inference network as the network prior probability. Based on the Bayesian probabilistic inference mechanism, the probability of the existence of posterior vulnerabilities that can reflect real business scenarios is solved.

[0042] The probability of a vulnerability actually existing is compared with the system's preset probability threshold. When the probability of a vulnerability actually existing is greater than the threshold, the corresponding AI software component is identified as having a vulnerability, and the vulnerability determination result is output.

[0043] The vulnerability exploit reachability verification includes: Obtain the vulnerability assessment results, extract the vulnerability attack prerequisites, component access permission requirements and component interface constraints, and construct a set of vulnerability exploitation prerequisite constraints. Traverse the AI ​​software supply chain topology network, taking the entry node of the AI ​​software supply chain topology network as the target starting point and the AI ​​software component with the vulnerability as the target ending point, and use a directed depth-first path search algorithm to search for accessible dependency paths from the target starting point to the target ending point. Each AI software component on the accessible dependency path is verified to ensure that it meets the preconditions for exploiting the vulnerability. If each AI software component on the accessible dependency path meets the preconditions for exploiting the vulnerability, the accessible dependency path is marked as a reachable attack path. If each AI software component on the accessible dependency path does not meet the preconditions for exploiting the vulnerability, the accessible dependency path is marked as an unreachable attack path. The number of reachable attack paths and the length of the shortest reachable attack path are counted to generate a reachability attack quantification index, which is then linked to the risk evidence data.

[0044] Extract the vulnerability attack prerequisites, component access permission requirements, and component interface constraints from the vulnerability assessment results, and construct a set of pre-exploitation constraints. This set covers all the prerequisites that allow external attackers to successfully trigger the vulnerability, avoiding distortions in reachability judgments caused by fragmented or missing verification standards.

[0045] Traverse the AI ​​software supply chain topology network, set the entry node exposed to the outside of the topology network as the search starting point, set the AI ​​software components with confirmed vulnerabilities as the search endpoint, and use the directed depth-first path search algorithm to traverse and search along the directed dependency edges of the topology network. During the traversal, capture all accessible dependency paths that can connect the starting point and the endpoint.

[0046] Each accessible dependency path retrieved is subjected to constraint validation. Each AI software component along the path is checked to ensure it matches all pre-existing constraints within the vulnerability exploitation pre-constraint set. If all components along a single path satisfy all pre-existing constraints, it indicates an attacker can exploit the vulnerability through this link, and the path is marked as a reachable attack path. If any component within a single path fails to satisfy the pre-existing constraints, the attacker cannot trigger the vulnerability along this link, and the corresponding path is marked as an unreachable attack path. This distinguishes between effective attack paths with actual threats and theoretically invalid paths.

[0047] After the path marking is completed, the number of reachable attack paths and the length of the shortest reachable attack path are counted to generate a reachability attack quantification index, which is then linked to the corresponding risk evidence data so that each piece of risk evidence carries its own corresponding attack reachability characteristics.

[0048] The calculation of the vulnerability risk value of the risk evidence data includes: Parse the vulnerability CVE number carried in the risk evidence data, search the pre-built vulnerability information database based on the vulnerability CVE number, and obtain the vulnerability CVSS score. Simultaneously extract the probability of the actual existence of the posterior vulnerability. The number of reachable attack paths The shortest attack path length ; Introducing attenuation coefficient Calculate the initial vulnerability risk value of the risk evidence data. for: ; Read the component asset importance weights configured in the system. AI business scenario risk correction coefficient Regarding the initial vulnerability risk value The vulnerability risk value was obtained by making corrections. for: .

[0049] The vulnerability CVE number in a single piece of risk evidence data is analyzed, and the corresponding vulnerability CVSS score is retrieved from the system's pre-built vulnerability information database based on the number. The probability of the vulnerability's actual existence, the number of reachable attack paths, and the length of the shortest attack path are extracted simultaneously to avoid the bias caused by a single indicator assessment. An attenuation coefficient is introduced to calculate the initial vulnerability risk value to intuitively represent the original security threat level of the vulnerability. The attenuation coefficient is used to represent the weakening effect of the attack path length on the actual threat level of the vulnerability.

[0050] The system retrieves the pre-configured component asset importance weights and AI business scenario risk correction coefficients, which correspond to the correction dimensions of the component's own business value and the inherent risk level of the AI ​​business scenario, respectively. The component asset importance weights, AI business scenario risk correction coefficients, and initial vulnerability risk values ​​are coupled and calculated to output the vulnerability risk value.

[0051] like Figure 2 The diagram shown is a system block diagram of a secure evidence fusion analysis system for the AI ​​software supply chain provided by an embodiment of the present invention. The system includes: The topology network construction module is used to receive the AI ​​software supply chain to be audited and extract the AI ​​software components corresponding to the AI ​​software supply chain. The AI ​​software components are used as nodes and the dependencies between the AI ​​software components are used as edges to construct the AI ​​software supply chain topology network. The security evidence output module is used to schedule security scanning tools to perform concurrent security detection on the AI ​​software supply chain, generate a security scanning report, perform semantic field mapping and spatial coordinate label alignment on the vulnerability records in the security scanning report according to the topology network of the AI ​​software supply chain, and output security evidence data. The risk evidence screening module is used to load historical empirical data from the historical vulnerability database, perform high-dimensional space vector projection on the security evidence data, calculate the semantic similarity between the security evidence data and the historical empirical data, and screen the security evidence data with semantic similarity higher than a preset similarity threshold as risk evidence data. The assessment report output module is used to extract the risk evidence data of the same AI software component, import it into a two-stage hybrid judgment engine to determine the existence of vulnerabilities and verify the accessibility of vulnerability exploitation, calculate the vulnerability risk value of the risk evidence data, prioritize the risk evidence data based on the vulnerability risk value from largest to smallest, and output a security evidence risk priority assessment report.

[0052] Figure 2 The apparatus of the illustrated embodiment can be used to perform corresponding actions. Figure 1 The steps in the method embodiments shown are implemented in a similar manner and have similar technical effects, and will not be repeated here.

[0053] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor performs the steps of a security evidence fusion analysis method for an AI software supply chain as described in any of the above.

[0054] like Figure 3 The diagram shown is a hardware structure schematic of an electronic device according to an embodiment of the present invention. The electronic device 30 includes: a processor 31, a memory 32, and a computer program; wherein... The memory 32 is used to store the computer program, and the memory may also be flash memory. The computer program is, for example, an application program or functional module that implements the above method.

[0055] The processor 31 is configured to execute the computer program stored in the memory to implement the various steps performed by the device in the above method. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0056] Alternatively, the memory 32 can be either standalone or integrated with the processor 31.

[0057] When the memory 32 is a device independent of the processor 31, the device may further include: Bus 33 is used to connect the memory 32 and the processor 31.

[0058] A readable storage medium storing a computer program, which, when executed by a processor, is used to implement the steps of a secure evidence fusion analysis method for an AI software supply chain as described above.

[0059] The readable storage medium can be a computer storage medium or a communication medium. A communication medium includes any medium that facilitates the transfer of computer programs from one location to another. A computer storage medium can be any available medium accessible to a general-purpose or special-purpose computer. For example, a 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 ASIC can be located in a user equipment. Of course, the processor and the readable storage medium can also exist as discrete components in a communication device. The readable storage medium can be a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0060] The present invention also provides a program product including executable instructions stored in a readable storage medium. At least one processor of the device can read the executable instructions from the readable storage medium, and the at least one processor executes the executable instructions to cause the device to implement the methods provided in the various embodiments described above.

[0061] In the embodiments of the above-described device, 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 manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for security evidence fusion analysis in the AI ​​software supply chain, characterized in that, The method includes: Receive the AI ​​software supply chain to be audited and extract the AI ​​software components corresponding to the AI ​​software supply chain. Construct an AI software supply chain topology network with the AI ​​software components as nodes and the dependencies between the AI ​​software components as edges. The security scanning tool is scheduled to perform concurrent security detection on the AI ​​software supply chain, generate a security scanning report, and perform semantic field mapping and spatial coordinate label alignment on the vulnerability records in the security scanning report according to the topology network of the AI ​​software supply chain, and output security evidence data. Load historical empirical data from the historical vulnerability database, perform high-dimensional space vector projection on the security evidence data, calculate the semantic similarity between the security evidence data and the historical empirical data, and filter the security evidence data with semantic similarity higher than a preset similarity threshold as risk evidence data; Extract the risk evidence data from the same AI software component, import it into a two-stage hybrid judgment engine to determine the existence of vulnerabilities and verify the reachability of vulnerability exploitation, calculate the vulnerability risk value of the risk evidence data, prioritize the risk evidence data based on the vulnerability risk value from largest to smallest, and output a security evidence risk priority assessment report.

2. The method for security evidence fusion analysis of AI software supply chain according to claim 1, characterized in that, The process of receiving the AI ​​software supply chain to be audited and extracting the corresponding AI software components, constructing an AI software supply chain topology network with the AI ​​software components as nodes and the dependencies between the AI ​​software components as edges, includes: The inventory file, image configuration file, and model dependency description file of the AI ​​software supply chain to be audited are parsed to extract the name, version number, and deployment carrier information of the AI ​​software components and generate a set of component metadata. Traverse the component metadata set and use a directed dependency resolution algorithm to identify the direct dependencies, indirect transitive dependencies and optional conditional dependencies between the AI ​​software components, and generate dependency triples. Loop detection and redundant edge removal are performed on the dependency triples. The component metadata set and the dependency triples are digitally encoded using an adjacency matrix to generate an initial topological adjacency matrix. Based on the initial topological adjacency matrix, the AI ​​software components are used as nodes, the dependencies between the AI ​​software components are used as edges, and a corresponding spatial identifier code is assigned to each node. A corresponding dependency weight coefficient is assigned to each edge to construct the AI ​​software supply chain topological network.

3. The method for security evidence fusion analysis of AI software supply chain according to claim 1, characterized in that, The scheduling security scanning tool performs concurrent security checks on the AI ​​software supply chain, generates a security scan report, and performs semantic field mapping and spatial coordinate label alignment on the vulnerability records in the security scan report based on the AI ​​software supply chain topology network, outputting security evidence data, including: The security scanning tools, including static code scanning tools, AI model vulnerability detection tools, and third-party component compliance scanning tools, are scheduled to perform concurrent security detection on the AI ​​software supply chain, generating multiple security scanning reports in heterogeneous formats. Field normalization and cleaning are performed on the multiple heterogeneous security scanning reports to obtain the original vulnerability records. Construct a cross-tool semantic mapping dictionary, establish mapping rules between tool-private fields and evidence unification fields, convert the tool-private fields in the original vulnerability record into the evidence unification fields, and generate an intermediate vulnerability record containing the evidence unification fields; Read the spatial identifier code corresponding to each AI software component in the AI ​​software supply chain, match it with the affected component information in the intermediate vulnerability record, embed the spatial identifier code into the intermediate vulnerability record, and output the security evidence data.

4. The method for security evidence fusion analysis of AI software supply chain according to claim 1, characterized in that, The process involves loading historical empirical data from the historical vulnerability database, performing high-dimensional space vector projection on the security evidence data, calculating the semantic similarity between the security evidence data and the historical empirical data, and selecting security evidence data with semantic similarity higher than a preset similarity threshold as risk evidence data, including: The vulnerability CVE number, vulnerability triggering conditions, affected component version, and supply chain attack path description text are read from the security evidence data, concatenated to form a security evidence text sequence, and the security evidence text sequence is segmented and filtered for stop words using a segmented sliding window strategy. The pre-trained domain vector encoder is invoked to map the security evidence text sequence after word segmentation and stop word filtering to a high-dimensional continuous vector space to generate the corresponding security evidence vector. At the same time, the historical empirical data is read from the historical vulnerability database to generate the corresponding historical empirical vector. Using a weighted improved RBF kernel distance, the vector distance between the secure evidence vector x and the historical empirical vector y is calculated and mapped to the semantic similarity D(x,y) as follows: ； In the formula, exp() is an exponential function with base e; I is the kernel width adjustment coefficient; I is the total dimension of the high-order continuous vector space; Let x be the i-th component of the security evidence vector. The corresponding dimension weight coefficients; Let be the i-th component of the historical empirical vector y; Security evidence data with semantic similarity higher than the preset similarity threshold are selected as risk evidence data.

5. The method for security evidence fusion analysis of AI software supply chain according to claim 1, characterized in that, The determination of the existence of the vulnerability includes: The two-stage hybrid decision engine includes a rule-based decision sub-engine and a probabilistic graph reasoning sub-engine. The rule-determination sub-engine loads the vulnerability determination rule set of the AI ​​software supply chain. Based on the vulnerability determination rule set, it reads the affected component version range, vulnerability CVE number, and component runtime configuration constraint information from the risk evidence data and performs hard rule matching, outputting the vulnerability existence confidence level. ; Collect multiple risk evidence data corresponding to each of the AI ​​software components, and establish a confidence level for the vulnerabilities corresponding to the multiple risk evidence data. The basic fusion confidence level is obtained by performing evidence fusion weighted calculation. for: ； In the formula, Let K be the credibility coefficient of the risk evidence data corresponding to the kth risk evidence data; K is the total number of risk evidence data corresponding to the AI ​​software component. Using the probabilistic graphical inference sub-engine, the AI ​​software components are used as inference variables. A Bayesian inference network is constructed by combining the dependencies between the AI ​​software components, and the basic fusion confidence is then used. As the prior probability input, the probability of the posterior vulnerability actually existing is calculated. The probability of the actual existence of the posterior vulnerability is compared with a preset existence probability threshold. If the posterior probability is greater than the preset existence probability threshold, the AI ​​software component is determined to have a vulnerability, and the vulnerability determination result is output.

6. The method for security evidence fusion analysis of AI software supply chain according to claim 5, characterized in that, The vulnerability exploit reachability verification includes: Obtain the vulnerability assessment results, extract the vulnerability attack prerequisites, component access permission requirements and component interface constraints, and construct a set of vulnerability exploitation prerequisite constraints. Traverse the AI ​​software supply chain topology network, taking the entry node of the AI ​​software supply chain topology network as the target starting point and the AI ​​software component with the vulnerability as the target ending point, and use a directed depth-first path search algorithm to search for accessible dependency paths from the target starting point to the target ending point. Each AI software component on the accessible dependency path is verified to ensure that it meets the preconditions for exploiting the vulnerability. If each AI software component on the accessible dependency path meets the preconditions for exploiting the vulnerability, the accessible dependency path is marked as a reachable attack path. If each AI software component on the accessible dependency path does not meet the preconditions for exploiting the vulnerability, the accessible dependency path is marked as an unreachable attack path. The number of reachable attack paths and the length of the shortest reachable attack path are counted to generate a reachability attack quantification index, which is then linked to the risk evidence data.

7. The method for security evidence fusion analysis of AI software supply chain according to claim 6, characterized in that, The calculation of the vulnerability risk value of the risk evidence data includes: Parse the vulnerability CVE number carried in the risk evidence data, search the pre-built vulnerability information database based on the vulnerability CVE number, and obtain the vulnerability CVSS score. Simultaneously extract the probability of the actual existence of the posterior vulnerability. The number of reachable attack paths The shortest attack path length ; Introducing attenuation coefficient Calculate the initial vulnerability risk value of the risk evidence data. for: ； Read the component asset importance weights configured in the system. AI business scenario risk correction coefficient Regarding the initial vulnerability risk value The vulnerability risk value was obtained by making corrections. for: 。 8. A secure evidence fusion and analysis system for the AI ​​software supply chain, characterized in that, The system, which is applied to a security evidence fusion analysis method for AI software supply chain as described in any one of claims 1-7, comprises: The topology network construction module is used to receive the AI ​​software supply chain to be audited and extract the AI ​​software components corresponding to the AI ​​software supply chain. The AI ​​software components are used as nodes and the dependencies between the AI ​​software components are used as edges to construct the AI ​​software supply chain topology network. The security evidence output module is used to schedule security scanning tools to perform concurrent security detection on the AI ​​software supply chain, generate a security scanning report, perform semantic field mapping and spatial coordinate label alignment on the vulnerability records in the security scanning report according to the topology network of the AI ​​software supply chain, and output security evidence data. The risk evidence screening module is used to load historical empirical data from the historical vulnerability database, perform high-dimensional space vector projection on the security evidence data, calculate the semantic similarity between the security evidence data and the historical empirical data, and screen the security evidence data with semantic similarity higher than a preset similarity threshold as risk evidence data. The assessment report output module is used to extract the risk evidence data of the same AI software component, import it into a two-stage hybrid judgment engine to determine the existence of vulnerabilities and verify the accessibility of vulnerability exploitation, calculate the vulnerability risk value of the risk evidence data, prioritize the risk evidence data based on the vulnerability risk value from largest to smallest, and output a security evidence risk priority assessment report.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor runs the computer program stored in the memory, the processor performs the steps of a secure evidence fusion analysis method for the AI ​​software supply chain as described in any one of claims 1-7.

10. A readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it is used to implement the steps of a secure evidence fusion analysis method for an AI software supply chain as described in any one of claims 1-7.

Images