Software supply chain-oriented high-risk component security knowledge base construction method
Through the joint extraction model of network crawler and entity-relationship, a high-risk component security knowledge base for IoT devices is built, which solves the problems of inaccurate resolution of dependencies between components and incomplete vulnerability information, and realizes efficient security knowledge base management and patching position.
Patent Information
- Application Number
- CN202410879987.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-02
- Publication Date
- 2024-11-29
AI Technical Summary
When building a security knowledge base for high-risk components of IoT devices, there are problems such as inaccurate resolution of dependencies between components, fuzzy description of vulnerabilities, and insufficient collection of vulnerabilities, resulting in low management efficiency and inaccurate information of security knowledge bases. comprehensive.
Using a network crawler strategy combining deep traversal and breadth traversal to capture component information, parse configuration files to build a dependency data set, combine entity-relationship joint extraction model to identify high-risk components and version ranges from public vulnerability databases, reversely extract vulnerability functions, and Based on semantic correlation search security patches, build a comprehensive security knowledge base for high-risk components.
It realizes accurate analysis of deep dependencies between components, eliminates the ambiguity of security knowledge, accurately identifys and builds a comprehensive high-risk component security knowledge base, and improves the clarity of description of vulnerability information and patch positioning efficiency.
Smart Images

Figure CN119046945A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of Internet of Things security technology, and specifically is a method for constructing a high-risk component security knowledge base for a software supply chain. Background Art
[0002] The components contained in IoT firmware play a decisive role in the performance, functionality and security of the device. Given the importance of IoT device security, in-depth research on these components has become particularly critical. In this context, some researchers have conducted in-depth analysis of third-party software libraries in different software ecosystems. Abdalkareem et al. studied the impact of lightweight libraries in the NPM ecosystem (Abdalkareem R, Nourry O, Wehaibi S. Why do developers use trivial packages? an empirical case study on npm, Proceedings of the 201711th joint meeting on foundations of software engineering, 2017), and analyzed 230,000 software packages to reveal the security vulnerabilities and performance issues that these libraries may cause. Ma et al. built a GitHub-based framework (Ma Y, Bogart C, Amreen S. World of code: an infrastructure formining the universe of open source VCS data, IEEE, 2019), covering more than 1.2 million git objects, providing a rich data foundation for metadata and dependency analysis of software libraries. Decan et al. pointed out the impact of ecosystem management tools and strategies on dependencies from the perspective of dependency management in different ecosystems (Decan A, Mens T, Constantinou E. On the evolution of technical lag in the npm package dependency network, IEEE, 2018). In particular, their study in the NPM ecosystem showed that a large number of dependencies had version lags, highlighting the importance of dependency management. For the collection framework and analysis model of different software ecosystems, existing research provides technical and data support for metadata and dependency analysis of software libraries. However, there is a lack of a comprehensive security knowledge base of "firmware-component-vulnerability-patch". Research on dependency management and other aspects also needs to be strengthened. It is necessary to further design automated tools to improve the efficiency of component collection and dependency management, so as to achieve comprehensive association and efficient management of IoT component security knowledge base.
[0003] According to the "2023 China Software Supply Chain Security Analysis Report", the number of new open source software vulnerabilities in 2022 is 7682. Given the continued use of open source and security research, the number of reported open source vulnerabilities will continue to increase. Using existing public vulnerability libraries to match component vulnerabilities is an important means of efficiently detecting vulnerabilities. However, the existing public vulnerability libraries have many vulnerability entries and complex descriptions. There are problems such as inconsistent description information and lack of vulnerability functions in different public vulnerability libraries. For this reason, existing research mainly focuses on vulnerability information identification and vulnerability function collection. Lv et al. proposed a tool Advance (Lv T, Li R, Yang Y. Rtfm! automatic assumption discovery and verification derivation from library document forapi misuse detection, Proceedings of the 2020 ACM SIGSAC conference oncomputer and communications security, 2020) that uses NLP technology to analyze C language library documents. Frequent subtree mining and identification of commonly used code descriptions are used, and finally lexical analysis and semantic analysis are used to parse parameters. Nikitopoulos et al. established a cross-language vulnerability dataset with commit data (Nikitopoulos G, Dritsa K, Louridas P. CrossVul: across-language vulnerability dataset with commit data, Proceedings of the29th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering, 2021). Each file stores information including its original repository, commit identifier, file name, associated vulnerability type, and CVE identifier.Bhandari et al. proposed an automatic vulnerability information collection method to automatically collect and organize a comprehensive vulnerability dataset CVEfixes from common vulnerabilities and exposure records in NVD (Bhandari G, Naseer A, Moonen L. CVEfixes: automated collection of vulnerabilities and their fixes from open-source software, Proceedings of the 17th International Conference on Predictive Models and Data Analytics in Software Engineering, 2021). Cheng Y et al. proposed a vulnerability detection system called VERI (Cheng Y, Yang S, Lang Z, Shi Z, Sun L. VERI: A Large-scale Open-Source Components Vulnerability Detection in IoT Firmware, Computers & Security, 2023). First, vulnerability descriptions are widely collected from NVD, and then an entity-relationship joint extraction model is trained to identify the version range of the vulnerability. Finally, the extracted information is converted into a unified version range, and a version-vulnerability relationship database is maintained. Existing research on the collection of high-risk component vulnerabilities in IoT firmware mostly builds a security vulnerability knowledge base based on the vulnerability information provided in official website documents. However, there are still problems such as vague descriptions of text information sentences and incomplete collection of vulnerability information. At the same time, existing research is not sufficient in collecting vulnerable functions. Summary of the invention
[0004] In order to accurately analyze the deep dependencies between components and effectively eliminate the ambiguity of existing security knowledge, the present invention provides a method for constructing a high-risk component security knowledge base for the software supply chain.
[0005] The present invention adopts the following technical solution: a method for constructing a high-risk component security knowledge base for a software supply chain, comprising: S100: Capture and collect component information in the field of Internet of Things, and collect basic information of binary components; S200: Parse the configuration files of the components, obtain the dependencies between the components, and construct three data sets, namely, a component set, a metadata set, and a dependency set, to represent the component supply chain knowledge base; S300: using the public vulnerability database and combining the entity-relationship joint extraction model, identifying the high-risk component names and version ranges from the public vulnerability database description information, and obtaining the vulnerability-version mapping relationship; S400: Collect vulnerability diff files from the open source vulnerability database, reverse extract the vulnerability function, add the function where the vulnerability trigger statement is located as the vulnerability function to the vulnerability knowledge base, and establish a vulnerability-version relationship knowledge base for high-risk components; S500: Based on the semantic relevance of vulnerability information and code submission information, the submitted codes in the high-risk component patch submission range are searched and ranked, and security patches are determined from the top-ranked positions, and a security patch knowledge base is constructed based on this.
[0006] In some embodiments, the basic information in step S100 includes: name, developer, version history, update log, description and code repository link.
[0007] In some embodiments, step S300 includes: S301: crawling data from a public vulnerability database, deduplicating the crawled results to obtain a complete set of vulnerability information, decomposing the information into character sequences according to the vulnerability description, and converting the sequences into word vectors through word embedding technology; S302: extract each vulnerability number, related component name and version range by combining entity-relationship joint extraction model, and output the information in the form of tuples; S303: Perform standardization processing to uniformly represent vulnerability numbers, component names, and version ranges.
[0008] In some embodiments, the entity-relationship joint extraction model in step S302 includes: Bi-LSTM layer, inputting the embedding vector to the Bi-LSTM layer, the Bi-LSTM layer obtains the forward and backward hidden state sequences; A self-attention layer, which is used to calculate the attention weight of each word and perform weighted summation on the output of the Bi-LSTM layer to obtain a global representation of the sequence; A fully connected layer processes the outputs of the Bi-LSTM and self-attention layers to extract vulnerability numbers, component names, and version ranges.
[0009] In some embodiments, step S400 includes: S401: performing preprocessing, deleting comments and blank codes in the patch statements of the collected vulnerability diff files, and identifying key variables of each patch statement according to the type of the patch statement; S402: Parse the source code to obtain the control flow graph and data flow information of the program; then identify the data dependency and determine the definition location and usage location of each variable; S403: Identify the trigger statement according to the feature of the vulnerability trigger statement, and add the function where the vulnerability trigger statement is located as a vulnerability function to the vulnerability knowledge base.
[0010] In some embodiments, step S500 includes: S501: selecting three groups of features with different weights from among a plurality of patch-related semantic information, extracting the three groups of features with different weights and converting them into feature vectors by weight; S502: Use historical vulnerabilities and corresponding security patches as training data, mark the relevance score of each patch, and assign the score based on the matching degree between the patch and the vulnerability. Use the LambdaMART machine ranking model for training. The training process optimizes the model parameters so that patches with high relevance to the training data are ranked higher. S503: In actual applications, new vulnerability information is input, and a candidate set of patches to be sorted is obtained through feature extraction and vectorization steps. The trained LambdaMART model is used to predict the relevance of these candidate patches and generate a sorted list. Based on the sorting results, the top-ranked patches are selected as security patches, and the security patch knowledge base is updated and constructed.
[0011] In some embodiments, in step S501, the three groups of features with different weights include: Vulnerability characteristics, specifically the unique identifier CVE-ID that describes the vulnerability and the Bug-ID that describes the specific vulnerability in the software, as well as the specific link to it; Code features, specifically the variables, functions, file names, relative and absolute paths, code modification lines and locations in the source code involved in the description information; Semantic features: specifically, the semantic correlation between vulnerability information and code submission information. Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention focuses on how to accurately parse the deep dependencies between components, effectively eliminate the ambiguity of existing security knowledge, accurately identify the complex dependencies between components, and automatically build a comprehensive high-risk component security knowledge base.
[0012] 2. In order to solve the problems of non-standard description of vulnerability information and lack of vulnerability functions, the present invention uses natural language processing methods to construct a mapping relationship between vulnerabilities and versions, reversely extracts vulnerability functions based on vulnerability diff files, and constructs a high-risk component vulnerability knowledge base.
[0013] 3. The present invention transforms the location problem of security patches into a screening and searching problem in code submission in source code repositories, designs a weighted patch searching method that can integrate vulnerability information and code submission information, and constructs a security patch knowledge base.
[0014] 4. The present invention effectively integrates component data from different software ecosystems, public vulnerability databases, and patch source code repositories to build a high-risk component security knowledge base for the software supply chain. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 A schematic diagram of the component supply chain knowledge base construction process of the method of the present invention; Figure 2 A schematic diagram of component dependency of the method of the present invention; Figure 3 A schematic diagram of the vulnerability knowledge base construction process of the method of the present invention; Figure 4 The present invention is a schematic diagram of the security patch knowledge base construction process of the method of the present invention. DETAILED DESCRIPTION
[0016] The present invention is described in detail below through specific embodiments in conjunction with the accompanying drawings, but does not constitute a limitation of the present invention.
[0017] This embodiment provides an example of building a high-risk component security knowledge base using the method of the present invention.
[0018] S100: A web crawler strategy combining deep traversal and breadth traversal is used to crawl and collect component information, and on this basis, basic metadata of the components is collected.
[0019] Specifically, in this embodiment, for different open source software ecosystems, a web crawler strategy combining deep traversal and breadth traversal is used to crawl and collect component information. On this basis, basic metadata of the component is collected, such as name, developer, version history, update log, description, and code repository link.
[0020] The depth traversal starts from the root node (such as the main component page), visits the current page and crawls information, and recursively visits all linked pages until the preset depth is reached or all links have been visited. The breadth traversal starts from the root node. Visit the current page and crawl information, add all links of the current page to the queue, and visit each page in the queue in turn until the queue is empty.
[0021] S200: Parse the configuration files of the components, obtain the dependencies between the components, and construct three data sets, namely, a component set, a metadata set, and a dependency set, to represent the component supply chain knowledge base; Specifically, in this embodiment, combined with the inherent characteristics of the Internet of Things software supply chain, the configuration file of the component is parsed to obtain the dependencies between the components, that is, a certain component uses other components or third-party libraries. The configuration file is responsible for managing third-party dependency sources, compilation environments, and related files during component development. The dependency relationship of the component supply chain is established through the name and version information of the third-party component. The 2.7.13 version of the third-party library spring-boot depends on the 5.3.28 version of spring-context and spring-core, and spring-context also uses the 5.3.28 version of spring-core. While spring-context depends on spring-beans, spring-aop, and spring-expression, all three also use the spring-core library.
[0022] S300: Using the public vulnerability database and combining the entity-relationship joint extraction model, identify the high-risk component names and version ranges from the public vulnerability database description information to obtain the vulnerability-version mapping relationship; In this embodiment, a public vulnerability library is used, combined with an entity-relationship joint extraction model, to identify high-risk component names and version ranges from the public vulnerability database description information, and obtain a vulnerability-version mapping relationship. First, crawl data from public vulnerability databases such as NVD, CVE, and CNNVD, perform deduplication processing on the crawled results, obtain a complete set of vulnerability information, decompose the information into character sequences according to the vulnerability description, and convert the sequence into a word vector through word embedding technology; then use Bi-LSTM combined with self-attention to extract each CVE vulnerability number, related component name and version range, and output the information in the form of a tuple; finally, perform normalization processing to uniformly represent the CVE vulnerability number, component name and version range. In the description "OpenSSL before 1.0.1u", the relationship tuple of "OpenSSL" can be obtained as (1.27.2, before), which means that the CVE vulnerability exists in versions 1.27.2 and below, and then the tuple is converted into a unified version range form (oldest version, 1.27.2), and whether the boundary version 1.27.2 and the oldest version are included are recorded at the same time.
[0023] Specifically, the process is as follows: S301: crawling data from a public vulnerability database, deduplicating the crawled results to obtain a complete set of vulnerability information, decomposing the information into character sequences according to the vulnerability description, and converting the sequences into word vectors through word embedding technology; Crawl data from public vulnerability databases, deduplicate the crawled results, obtain a complete set of vulnerability information, and clean the vulnerability description text, including removing special characters, punctuation, and converting all characters to lowercase. Next, use word segmentation technology to decompose the cleaned text into separate word sequences. Then, remove common stop words (such as "and", "the", etc.) to reduce noise data. After completing text preprocessing, use word embedding technology to convert the segmented words into corresponding word vectors, and map each word to a high-dimensional vector space to capture the semantic similarity between words. Finally, the processed vulnerability description can be represented by these word vectors for further analysis and application.
[0024] S302: extract each vulnerability number, related component name and version range by combining entity-relationship joint extraction model, and output the information in the form of tuples; The entity-relationship joint extraction model in step S302 includes: Bi-LSTM layer, inputting the embedding vector to the Bi-LSTM layer, the Bi-LSTM layer obtains the forward and backward hidden state sequences; A self-attention layer, which is used to calculate the attention weight of each word and perform weighted summation on the output of the Bi-LSTM layer to obtain a global representation of the sequence; A fully connected layer processes the outputs of the Bi-LSTM and self-attention layers to extract vulnerability numbers, component names, and version ranges.
[0025] like Figure 3 As shown in the figure, data is crawled from the public vulnerability database, and the crawled results are deduplicated to obtain a complete set of vulnerability information. The information is decomposed into character sequences according to the vulnerability description, and the sequences are converted into word vectors through word embedding technology; then Bi-LSTM is combined with self-attention to first clean the text, segment the words, and convert them into word embedding vectors. The embedded vector is input to the Bi-LSTM layer to obtain the forward and backward hidden state sequences. Then the attention weight of each word is calculated, and the Bi-LSTM output is weighted summed to obtain the global representation of the sequence. The outputs of the Bi-LSTM and self-attention layers are processed by the fully connected layer to extract the vulnerability number, component name, and version range.
[0026] S303: Perform standardization processing to uniformly represent vulnerability numbers, component names, and version ranges.
[0027] The extracted information is output in the form of tuples, and the component names are standardized and formatted. Through the above steps, the extracted vulnerability numbers, component names, and version ranges are standardized, and a mapping relationship between vulnerabilities and component versions is established to form structured vulnerability-version mapping data.
[0028] S400: Collect vulnerability diff files from NVD and SARD libraries, extract vulnerability functions in reverse, add the functions where the vulnerability trigger statements are located as vulnerability functions to the vulnerability knowledge base, and establish a vulnerability-version relationship knowledge base for high-risk components.
[0029] Step S400 includes: S401: performing preprocessing, deleting comments and blank codes in the patch statements of the collected vulnerability diff files, and identifying key variables of each patch statement according to the type of the patch statement; S402: parse the source code through a static code analysis tool to obtain the control flow graph and data flow information of the program; then identify the data dependency and determine the definition location and usage location of each variable; For example: if statement A defines a variable and statement B uses this variable, a data dependency edge is drawn between statement A and statement B. Then analyze conditional statements, loop statements, etc. to determine which statements depend on the judgment of a certain condition. Finally, combine data dependency and control dependency, add each statement or variable in the program as a node to the graph, add data dependency edges and control dependency edges, and build a relationship dependency graph. According to the analysis goal, determine the key variables that need to be sliced, start from the statement where the key variable is located, trace all data dependencies and control dependencies related to it, extract related code snippets, and generate program slices.
[0030] S403: Identify the trigger statement according to the feature of the vulnerability trigger statement, and add the function where the vulnerability trigger statement is located as a vulnerability function to the vulnerability knowledge base.
[0031] Identify common vulnerability-related function calls, such as strcpy, malloc, free, etc. Capture exception handling code, such as try-catch blocks, error return code checks, etc. Then use static code analysis tools to scan the source code and match predefined features. Mark statements that match features as potential vulnerability trigger statements. Use static code analysis tools to build an AST of the source code. In the AST, trace upward from the identified vulnerability trigger statement node until you find the function node where the statement is located. Finally, store information such as function name, file path, line number, trigger statement features, including vulnerability number, vulnerability description, etc. Insert the marked vulnerability function and its detailed information into the vulnerability knowledge base according to the defined data structure.
[0032] S500: Based on the semantic relevance of vulnerability information and code submission information, the submitted codes in the high-risk component patch submission range are searched and ranked, and security patches are determined from the top-ranked positions, and a security patch knowledge base is constructed based on this.
[0033] Step S500 includes: S501: selecting three groups of features with different weights from among a plurality of patch-related semantic information, extracting the three groups of features with different weights and weightedly converting them into feature vectors; S502: Use historical vulnerabilities and corresponding security patches as training data, mark the relevance score of each patch, and assign the score based on the matching degree between the patch and the vulnerability. Use the LambdaMART machine ranking model for training. The training process optimizes the model parameters so that patches with high relevance to the training data are ranked higher. S503: In actual applications, new vulnerability information is input, and a candidate set of patches to be sorted is obtained through feature extraction and vectorization steps. The trained LambdaMART model is used to predict the relevance of these candidate patches and generate a sorted list. Based on the sorting results, the top-ranked patches are selected as security patches, and the security patch knowledge base is updated and constructed.
[0034] Specifically, in this embodiment, we focus on mainstream component code repositories, such as GitHub, GitLab, SourceForge, etc., and search and rank all submitted codes in the scope of IoT high-risk component patch submissions based on the semantic relevance of vulnerability information and code submission information. Among the many semantic information related to patches, three groups of features with different weights are selected: 1) Vulnerability features: Vulnerability information and code submission information may contain a unique identifier CVE-ID that describes the vulnerability and a Bug-ID that describes a specific vulnerability in the software, as well as a specific link. This type of information directly points to a specific vulnerability or code submission, and has a clear directionality, so this feature has the highest weight; 2) Code features: Code features refer to the variables, functions, file names, relative and absolute paths, code modification lines and locations in the source code involved in the description information. This type of information does not exclude the possibility that different code developers may have similar development habits, so its weight is lower than that of vulnerability features; 3) Semantic features: refers to the semantic correlation between vulnerability information and code submission information, such as vulnerability time, vulnerability type and impact, etc. Semantic feature information is presented in the form of natural language text, but due to the ambiguity of natural language, its directionality is the least clear compared to the above features, so it has the lowest weight.
[0035] After extracting three groups of features with different weights and converting them into feature vectors, the semantic relevance ranking problem of all code submission information and vulnerability information in the source code submission range obtained by dividing the source code library is converted into a ListWise problem in machine sorting learning. Finally, a LambdaMART machine sorting model is trained to rank all code submission information, determine the security patches from the top ranking positions, and use this to build a security patch knowledge base.
Claims
1. A method for constructing a high-risk component security knowledge base for a software supply chain, characterized in that: include: S100: Capture and collect component information in the field of Internet of Things, and collect basic information of binary components; S200: Parse the configuration files of the components, obtain the dependencies between the components, and construct three data sets, namely, a component set, a metadata set, and a dependency set, to represent the component supply chain knowledge base; S300: using the public vulnerability database and combining the entity-relationship joint extraction model, identifying the high-risk component names and version ranges from the public vulnerability database description information, and obtaining the vulnerability-version mapping relationship; S400: Collect vulnerability diff files from the open source vulnerability database, reverse extract the vulnerability function, add the function where the vulnerability trigger statement is located as the vulnerability function to the vulnerability knowledge base, and establish a vulnerability-version relationship knowledge base for high-risk components; S5 00: Based on the semantic relevance of vulnerability information and code submission information, the submitted codes in the high-risk component patch submission range are searched and ranked, and the security patches are determined from the top-ranked positions, and a security patch knowledge base is constructed based on this.
2. The method for constructing a high-risk component security knowledge base for a software supply chain according to claim 1, characterized in that: The basic information in step S100 includes: name, developer, version history, update log, description and code repository link.
3. The method for constructing a high-risk component security knowledge base for a software supply chain according to claim 1, characterized in that: The step S300 includes: S301: crawling data from a public vulnerability database, deduplicating the crawled results to obtain a complete set of vulnerability information, decomposing the information into character sequences according to the vulnerability description, and converting the sequences into word vectors through word embedding technology; S302: extract each vulnerability number, related component name and version range by combining entity-relationship joint extraction model, and output the information in the form of tuples; S303: Perform standardization processing to uniformly represent vulnerability numbers, component names, and version ranges.
4. The method for constructing a high-risk component security knowledge base for a software supply chain according to claim 1, characterized in that: The entity-relationship joint extraction model in step S302 includes: Bi-LSTM layer, inputting the embedding vector to the Bi-LSTM layer, the Bi-LSTM layer obtains the forward and backward hidden state sequences; A self-attention layer, which is used to calculate the attention weight of each word and perform weighted summation on the output of the Bi-LSTM layer to obtain a global representation of the sequence; A fully connected layer processes the outputs of the Bi-LSTM and self-attention layers to extract vulnerability numbers, component names, and version ranges.
5. The method for constructing a high-risk component security knowledge base for a software supply chain according to claim 1, characterized in that: The step S400 includes: S401: performing preprocessing, deleting comments and blank codes in the patch statements of the collected vulnerability diff files, and identifying key variables of each patch statement according to the type of the patch statement; S402: Parse the source code to obtain the control flow graph and data flow information of the program; then identify the data dependency and determine the definition location and usage location of each variable; S403: Identify the trigger statement according to the characteristics of the vulnerability trigger statement, and add the function where the vulnerability trigger statement is located as a vulnerability function to the vulnerability knowledge base.
6. The method for constructing a high-risk component security knowledge base for a software supply chain according to claim 1, characterized in that: The step S500 includes: S501: selecting three groups of features with different weights from among a plurality of patch-related semantic information, extracting the three groups of features with different weights and weightedly converting them into feature vectors; S502: Use historical vulnerabilities and corresponding security patches as training data, mark the relevance score of each patch, and assign the score based on the matching degree between the patch and the vulnerability. Use the LambdaMART machine ranking model for training. The training process optimizes the model parameters so that patches with high relevance to the training data are ranked higher. S503: In actual applications, new vulnerability information is input, and a candidate set of patches to be sorted is obtained through feature extraction and vectorization steps. The trained LambdaMART model is used to predict the relevance of these candidate patches and generate a sorted list. Based on the sorting results, the top-ranked patches are selected as security patches, and the security patch knowledge base is updated and constructed.
7. The method for constructing a high-risk component security knowledge base for a software supply chain according to claim 6, characterized in that: In step S501, the three groups of features with different weights include: Vulnerability characteristics, specifically the unique identifier CVE-ID that describes the vulnerability and the Bug-ID that describes the specific vulnerability in the software, as well as the specific link to it; Code features, specifically the variables, functions, file names, relative and absolute paths, code modification lines and locations in the source code involved in the description information; Semantic features: specifically, the semantic correlation between vulnerability information and code submission information.
Citation Information
Cited By
Multi-source software supply chain intelligent analysis method and system
CN119720225A
