APP sensitive feature detection and analysis method based on large language model

Abstract

The invention relates to the technical field of information security detection, in particular to an APP sensitive feature detection analysis method based on a large language model, which comprises the following steps: analyzing and extracting webpage HTML data corresponding to an APP sensitive resource file to establish a multi-dimensional sensitive feature data set; performing labeling processing on the basis of a sensitive feature constraint mode corpus by utilizing the multi-dimensional sensitive feature data set to obtain a multi-dimensional sensitive feature data labeling corpus; according to the multi-dimensional sensitive feature data annotation corpus, an APP sensitive feature detection analysis result is obtained based on a large language model, a multi-dimensional privacy policy sensitive feature constraint mode is innovatively constructed, theme information, laws and regulations and market requirements are fused, accurate privacy protection guidance is provided for developers, and a unified evaluation criterion is established.

Description

Technical Field

[0001] This invention relates to the field of information security detection technology, specifically to a method for detecting and analyzing sensitive features of an app based on a large language model. Background Technology

[0002] In recent years, emerging issues related to personal information protection have attracted widespread national attention. As sensitive information related to app permissions, privacy policies disclose how and why sensitive user data is collected, shared, used, and stored. However, due to the complexity of legal and technical terminology surrounding privacy policies, the documents are often overly verbose and have complex sources, making it difficult for users to quickly grasp and understand the privacy information involved and collected during app operation. This "unnoticed" authorization behavior during rapid operation seriously threatens privacy information security. Therefore, the ability to automatically analyze and detect the compliance of privacy policies and help users understand the privacy access behavior of applications is of great significance for protecting user information security. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method for detecting and analyzing sensitive features of apps based on a large language model. By constructing a training application of multi-dimensional privacy policy sensitive features based on different categories and a large language model, the efficiency and accuracy of app privacy policy compliance detection are improved.

[0004] To achieve the above objectives, this invention provides a method for detecting and analyzing sensitive features of an app based on a large language model, comprising:

[0005] S1. Use the HTML data of the webpage corresponding to the sensitive resource files of the APP to parse and extract data to establish a multi-dimensional sensitive feature dataset;

[0006] S2. Using the multidimensional sensitive feature dataset, an annotation process is performed based on the sensitive feature constraint pattern corpus to obtain a multidimensional sensitive feature data annotation corpus;

[0007] S3. Based on the multidimensional sensitive feature data-annotated corpus, obtain the APP sensitive feature detection and analysis results using a large language model.

[0008] Preferably, the step of using the HTML data of the webpage corresponding to the sensitive resource files of the APP to parse and extract data to establish a multidimensional sensitive feature dataset includes:

[0009] S1-1. Using the HTML data of the webpage corresponding to the sensitive resource files of the APP, obtain the multi-dimensional feature elements of the webpage HTML data based on the crawling algorithm;

[0010] S1-2. Use the multi-dimensional feature elements of the web page HTML data to parse and extract them to establish a multi-dimensional sensitive feature dataset;

[0011] The multidimensional feature elements include web addresses, images, text, audio, and video.

[0012] Furthermore, using the HTML data of the webpage corresponding to the sensitive resource files of the APP, the multi-dimensional feature elements of the webpage HTML data obtained based on the crawling algorithm include:

[0013] Use the sensitive resource files of the APP to obtain the corresponding webpage HTML data;

[0014] Using the HTML data of the webpage, a web crawler algorithm is used to obtain webpage address data, image information data, text information data, audio data information, and video data information respectively;

[0015] The webpage address data, image information data, text information data, audio data information, and video data information are used as multidimensional feature elements of the webpage HTML data.

[0016] Furthermore, the process of parsing and extracting multidimensional feature elements from the webpage HTML data to establish a multidimensional sensitive feature dataset includes:

[0017] S1-2-1. Establish document data volume thresholds based on privacy policies;

[0018] S1-2-2. Determine whether the data volume of the multidimensional feature elements of the web page HTML data is less than the document data volume threshold. If so, delete the multidimensional feature elements that are less than the document data volume threshold. Otherwise, use the multidimensional feature elements of the web page HTML data to establish a multidimensional sensitive feature dataset.

[0019] The document data size threshold is 2kb.

[0020] Furthermore, the multidimensional sensitive feature data annotation corpus is obtained by annotating the multidimensional sensitive feature dataset based on the sensitive feature constraint pattern corpus, including:

[0021] S2-1. Preprocess the multidimensional sensitive feature dataset to establish a sensitive information permission mapping.

[0022] S2-2. Based on the sensitive information permission mapping, a multi-dimensional sensitive feature data annotation corpus is obtained by annotating the sensitive feature constraint pattern corpus.

[0023] Furthermore, the preprocessing of the multidimensional sensitive feature dataset to establish a sensitive information permission mapping includes:

[0024] S2-1-1. Using the multidimensional sensitive feature dataset, perform punctuation removal, stop word bag construction, language judgment and filtering, web page bullet point removal and title bar noise removal in sequence to obtain a multidimensional sensitive feature preprocessing dataset.

[0025] S2-1-2. Using the multi-dimensional sensitive features, the dataset is preprocessed and dangerous permission mapping is performed according to the corresponding subset type to establish the initial sensitive information permission mapping.

[0026] S2-1-3. Determine whether the initial sensitive information permission mapping has overlapping mappings. If so, obtain the data volume of the initial sensitive information permission mapping with overlapping mappings and execute S2-1-4. Otherwise, retain the initial sensitive information permission mapping as the sensitive information permission mapping.

[0027] S2-1-4. Based on the amount of data in the overlapping initial sensitive information permission mappings, obtain a subset of the initial sensitive information permission mappings with a relatively smaller amount of data as the abnormal initial sensitive information permission mappings.

[0028] S2-1-5. Use the initial sensitive information permission mapping to delete the abnormal initial sensitive information permission mapping to obtain the sensitive information permission mapping;

[0029] The overlapping of the mappings refers to the situation where there is a non-single correspondence between the mappings.

[0030] Furthermore, based on the sensitive information permission mapping and the sensitive feature constraint pattern corpus, a multi-dimensional sensitive feature data annotation corpus is obtained by annotation processing, including:

[0031] Using the aforementioned sensitive information permission mapping, a basic annotated corpus of multi-dimensional sensitive feature data is obtained by performing multi-feature annotation processing based on the sensitive feature constraint pattern corpus;

[0032] Using the corresponding action target annotations in the basic annotated corpus, fine-grained extraction processing is performed to obtain data collectors, data collection conditions, data collection types, and data-related actions to establish fine-grained action target extraction results;

[0033] The basic annotated corpus of the aforementioned multidimensional sensitive feature data and the fine-grained extraction results of action targets are used as the multidimensional sensitive feature data annotated corpus;

[0034] The multi-feature annotation includes annotations for privacy policy overview, privacy policy title, developer information, action target, processing technology, data sharing, data security, user rights, data retention, international data transmission, changes to privacy policy, legal basis, advertising use, child protection, and link validity.

[0035] Furthermore, based on the multidimensional sensitive feature data-annotated corpus and a large language model, the APP sensitive feature detection and analysis results include:

[0036] S3-1. Using the multidimensional sensitive feature data-annotated corpus, establish a sensitive feature classification and detection model based on a large language model;

[0037] S3-2. Obtain the APP sensitive feature detection and analysis results using the aforementioned sensitive feature classification and detection model.

[0038] Furthermore, the sensitive feature classification and detection model established based on a large language model using the multidimensional sensitive feature data-annotated corpus includes:

[0039] S3-1-1. Use the HTML data of the corresponding web pages in the basic annotated corpus of the multidimensional sensitive feature data annotation corpus as the first dataset, and use the basic annotated corpus of the multidimensional sensitive feature data annotation corpus as the second dataset.

[0040] S3-1-2. Based on the first dataset and the second dataset, train a sensitive feature extraction model using RoBERTa.

[0041] S3-1-3. Using the multidimensional sensitive feature data annotation corpus, obtain the three-dimensional tensor of sensitive features based on the sensitive feature extraction model;

[0042] S3-1-4. Based on the three-dimensional tensor of the sensitive features, a sensitive feature variable enhancement model is established by training on T5.

[0043] S3-1-5. Using the multidimensional sensitive feature data-annotated corpus, obtain sensitive feature text sequences based on the sensitive feature variable enhancement model;

[0044] S3-1-6. Using the aforementioned sensitive feature text sequence, a sensitive feature structure extraction model is established based on UIE for training.

[0045] S3-1-7. The aforementioned sensitive feature basic extraction model, sensitive feature variable enhancement model, and sensitive feature structured extraction model are used as sensitive feature classification and detection models.

[0046] Furthermore, the results of APP sensitive feature detection and analysis obtained using the aforementioned sensitive feature classification and detection model include:

[0047] S3-2-1. Obtain the three-dimensional tensors of webpage address data sensitive features, image information data sensitive features, text information data sensitive features, audio data information sensitive features, and video data information sensitive features corresponding to the multi-dimensional sensitive feature data annotation corpus.

[0048] S3-2-2, Calculate the similarity of the three-dimensional tensor of the web address sensitive features based on the channel dimension using the three-dimensional tensor of the web address sensitive features;

[0049] S3-2-3. Calculate the similarity of the three-dimensional tensor of the image information data sensitive features based on the spatial dimension using the three-dimensional tensor of the image information data sensitive features.

[0050] S3-2-4. Calculate the similarity of the three-dimensional tensor of the sensitive features of the text information data based on the length dimension using the three-dimensional tensor of the sensitive features of the text information data.

[0051] S3-2-5. Calculate the similarity of the three-dimensional tensor of the sensitive features of the audio data based on the time dimension using the three-dimensional tensor of the sensitive features of the audio data.

[0052] S3-2-6. Calculate the similarity of the three-dimensional tensors of the sensitive features of video data based on the time dimension using the three-dimensional tensors of the sensitive features of video data.

[0053] S3-2-7. Obtain the average three-dimensional tensor similarity of sensitive features for webpage address sensitive features, image information data sensitive features, text information data sensitive features, audio data sensitive features, and video data sensitive features, respectively.

[0054] S3-2-8. Using the average value of the three-dimensional tensor similarity of the sensitive features, obtain the absolute value of the similarity difference between the three-dimensional tensor similarity of the web address sensitive features, the image information data sensitive features, the text information data sensitive features, the audio data information sensitive features, and the video data information sensitive features.

[0055] S3-2-9. Obtain the corresponding three-dimensional tensor as the feature tensor of the tuning parameter based on the absolute value of the similarity difference;

[0056] S3-2-10. Standardize the tuning parameter feature tensor to obtain the standardized tuning parameter feature dimension.

[0057] S3-2-11. After optimizing the sensitive feature classification and detection model using the standardized tuning parameter feature dimension, obtain the APP sensitive feature detection and analysis results.

[0058] Compared with the closest existing technology, the present invention has the following advantages:

[0059] This project innovatively constructs a multi-dimensional privacy policy sensitive feature constraint model, integrating thematic information, laws and regulations, and market requirements to provide developers with precise privacy protection guidance and establish a unified evaluation benchmark. It utilizes deep learning technology to achieve sensitive feature classification and detection, and applies natural language processing to privacy policy compliance analysis, representing a technological breakthrough in the field of mobile application security. The project can efficiently detect the compliance of app privacy policies, helping companies avoid legal risks, enhance user privacy awareness, and reduce data misuse. Attached Figure Description

[0060] Figure 1 This is a flowchart of an APP sensitive feature detection and analysis method based on a large language model provided by the present invention. Detailed Implementation

[0061] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0062] 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, 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.

[0063] Example 1:

[0064] This invention provides a method for detecting and analyzing sensitive features of apps based on a large language model, such as... Figure 1 As shown, it includes:

[0065] S1. Use the HTML data of the webpage corresponding to the sensitive resource files of the APP to parse and extract data to establish a multi-dimensional sensitive feature dataset;

[0066] S2. Using the multidimensional sensitive feature dataset, an annotation process is performed based on the sensitive feature constraint pattern corpus to obtain a multidimensional sensitive feature data annotation corpus;

[0067] S3. Based on the multidimensional sensitive feature data-annotated corpus, obtain the APP sensitive feature detection and analysis results using a large language model.

[0068] S1 specifically includes:

[0069] S1-1. Using the HTML data of the webpage corresponding to the sensitive resource files of the APP, obtain the multi-dimensional feature elements of the webpage HTML data based on the crawling algorithm;

[0070] S1-2. Use the multi-dimensional feature elements of the web page HTML data to parse and extract them to establish a multi-dimensional sensitive feature dataset;

[0071] The multidimensional feature elements include web addresses, images, text, audio, and video.

[0072] S1-1 specifically includes:

[0073] S1-1-1. Obtain the corresponding webpage HTML data using the sensitive resource files of the APP;

[0074] S1-1-2. Using the HTML data of the webpage, obtain webpage address data, image information data, text information data, audio data information and video data information respectively based on the crawling algorithm;

[0075] S1-1-3. The webpage address data, image information data, text information data, audio data information, and video data information are used as multi-dimensional feature elements of the webpage HTML data.

[0076] S1-2 specifically includes:

[0077] S1-2-1. Establish document data volume thresholds based on privacy policies;

[0078] S1-2-2. Determine whether the data volume of the multidimensional feature elements of the web page HTML data is less than the document data volume threshold. If so, delete the multidimensional feature elements that are less than the document data volume threshold. Otherwise, use the multidimensional feature elements of the web page HTML data to establish a multidimensional sensitive feature dataset.

[0079] The document data size threshold is 2kb.

[0080] S2 specifically includes:

[0081] S2-1. Preprocess the multidimensional sensitive feature dataset to establish a sensitive information permission mapping.

[0082] S2-2. Based on the sensitive information permission mapping, a multi-dimensional sensitive feature data annotation corpus is obtained by annotating the sensitive feature constraint pattern corpus.

[0083] S2-1 specifically includes:

[0084] S2-1-1. Using the multidimensional sensitive feature dataset, perform punctuation removal, stop word bag construction, language judgment and filtering, web page bullet point removal and title bar noise removal in sequence to obtain a multidimensional sensitive feature preprocessing dataset.

[0085] S2-1-2. Using the multi-dimensional sensitive features, the dataset is preprocessed and dangerous permission mapping is performed according to the corresponding subset type to establish the initial sensitive information permission mapping.

[0086] S2-1-3. Determine whether the initial sensitive information permission mapping has overlapping mappings. If so, obtain the data volume of the initial sensitive information permission mapping with overlapping mappings and execute S2-1-4. Otherwise, retain the initial sensitive information permission mapping as the sensitive information permission mapping.

[0087] S2-1-4. Based on the amount of data in the overlapping initial sensitive information permission mappings, obtain a subset of the initial sensitive information permission mappings with a relatively smaller amount of data as the abnormal initial sensitive information permission mappings.

[0088] S2-1-5. Use the initial sensitive information permission mapping to delete the abnormal initial sensitive information permission mapping to obtain the sensitive information permission mapping;

[0089] The overlapping of the mappings refers to the situation where there is a non-single correspondence between the mappings.

[0090] S2-2 specifically includes:

[0091] S2-2-1. Using the aforementioned sensitive information permission mapping, a basic annotated corpus of multi-dimensional sensitive feature data is obtained by performing multi-feature annotation processing based on the sensitive feature constraint pattern corpus.

[0092] S2-2-2, Using the corresponding action target annotations in the basic annotation corpus, perform fine-grained extraction processing to obtain data collectors, data collection conditions, data collection types, and data-related actions to establish fine-grained action target extraction results;

[0093] S2-2-3. Use the basic annotated corpus of the multidimensional sensitive feature data and the fine-grained extraction results of the action target as the multidimensional sensitive feature data annotated corpus;

[0094] The multi-feature annotation includes annotations for privacy policy overview, privacy policy title, developer information, action target, processing technology, data sharing, data security, user rights, data retention, international data transmission, changes to privacy policy, legal basis, advertising use, child protection, and link validity.

[0095] S3 specifically includes:

[0096] S3-1. Using the multidimensional sensitive feature data-annotated corpus, establish a sensitive feature classification and detection model based on a large language model;

[0097] S3-2. Obtain the APP sensitive feature detection and analysis results using the aforementioned sensitive feature classification and detection model.

[0098] S3-1 specifically includes:

[0099] S3-1-1. Use the HTML data of the corresponding web pages in the basic annotated corpus of the multidimensional sensitive feature data annotation corpus as the first dataset, and use the basic annotated corpus of the multidimensional sensitive feature data annotation corpus as the second dataset.

[0100] S3-1-2. Based on the first dataset and the second dataset, train a sensitive feature extraction model using RoBERTa.

[0101] S3-1-3. Using the multidimensional sensitive feature data annotation corpus, obtain the three-dimensional tensor of sensitive features based on the sensitive feature extraction model;

[0102] S3-1-4. Based on the three-dimensional tensor of the sensitive features, a sensitive feature variable enhancement model is established by training on T5.

[0103] S3-1-5. Using the multidimensional sensitive feature data-annotated corpus, obtain sensitive feature text sequences based on the sensitive feature variable enhancement model;

[0104] S3-1-6. Using the aforementioned sensitive feature text sequence, a sensitive feature structure extraction model is established based on UIE for training.

[0105] S3-1-7. The aforementioned sensitive feature basic extraction model, sensitive feature variable enhancement model, and sensitive feature structured extraction model are used as sensitive feature classification and detection models.

[0106] S3-2 specifically includes:

[0107] S3-2-1. Obtain the three-dimensional tensors of webpage address data sensitive features, image information data sensitive features, text information data sensitive features, audio data information sensitive features, and video data information sensitive features corresponding to the multi-dimensional sensitive feature data annotation corpus.

[0108] S3-2-2, Calculate the similarity of the three-dimensional tensor of the web address sensitive features based on the channel dimension using the three-dimensional tensor of the web address sensitive features;

[0109] S3-2-3. Calculate the similarity of the three-dimensional tensor of the image information data sensitive features based on the spatial dimension using the three-dimensional tensor of the image information data sensitive features.

[0110] S3-2-4. Calculate the similarity of the three-dimensional tensor of the sensitive features of the text information data based on the length dimension using the three-dimensional tensor of the sensitive features of the text information data.

[0111] S3-2-5. Calculate the similarity of the three-dimensional tensor of the sensitive features of the audio data based on the time dimension using the three-dimensional tensor of the sensitive features of the audio data.

[0112] S3-2-6. Calculate the similarity of the three-dimensional tensors of the sensitive features of video data based on the time dimension using the three-dimensional tensors of the sensitive features of video data.

[0113] S3-2-7. Obtain the average three-dimensional tensor similarity of sensitive features for webpage address sensitive features, image information data sensitive features, text information data sensitive features, audio data sensitive features, and video data sensitive features, respectively.

[0114] S3-2-8. Using the average value of the three-dimensional tensor similarity of the sensitive features, obtain the absolute value of the similarity difference between the three-dimensional tensor similarity of the web address sensitive features, the image information data sensitive features, the text information data sensitive features, the audio data information sensitive features, and the video data information sensitive features.

[0115] S3-2-9. Obtain the corresponding three-dimensional tensor as the feature tensor of the tuning parameter based on the absolute value of the similarity difference;

[0116] S3-2-10. Standardize the tuning parameter feature tensor to obtain the standardized tuning parameter feature dimension.

[0117] S3-2-11. After optimizing the sensitive feature classification and detection model using the standardized tuning parameter feature dimension, obtain the APP sensitive feature detection and analysis results.

[0118] In this embodiment, a method for detecting and analyzing sensitive features of an app based on a large language model is implemented as follows:

[0119] I. Data collectors targeting multidimensional sensitive features:

[0120] The first part involves data collection and preprocessing. The input consists of multi-dimensional sensitive data features, such as privacy policies, descriptive text, titles, and categories. To achieve this, we first identify document pages related to the application's sensitive resource files. Using web scraping algorithms, we employ the Python third-party library BeautifulSoup to parse and extract information from these web pages. BeautifulSoup supports various parsers, such as HTML, XML, and HTML5. Its core function is to parse and extract various useful information from the HTML data of web pages, such as URLs (Url, Href), images, text, audio (MP3), and video (MP4, AVI). These elements are embedded in the HTML structure of the web page and stored in corresponding elements according to different levels.

[0121] Furthermore, since an app's popularity is determined by its ranking and download volume, and different categories have different requirements for accessing users' sensitive information (e.g., communication, social, and maps), multiple strategies are developed to collect seed links based on app ranking, download volume, and category coverage. Secondly, after obtaining seed Android app links, different filtering criteria and conditions are created to improve the quality of collected sensitive resource documents, including but not limited to: text resource size, language, and content. Therefore, based on previous research and observations of the average word count of privacy policies, a minimum size of 2KB is set for privacy policy documents, and they are limited to privacy policy text described in English. By constructing a data collector oriented towards multidimensional sensitive features, a dataset is built, thereby providing high-quality data support for the implementation of [the project / project].

[0122] II. NLP Preprocessing Stage:

[0123] Through the data collection process, various sensitive data features were gathered, such as privacy policies, descriptive text, titles, and categories. Since this sensitive information is described using natural language, NLP techniques were used to preprocess different types of data, applying different preprocessing methods to different sensitive data types. For example, privacy policies and long text documents written in natural language involved preprocessing techniques including, but not limited to, removing punctuation, creating stop word bags, language judgment and filtering, removing bullet points and noise from title bars on web pages, and normalizing links and emails. Secondly, based on the characteristics of the AOFU permission mechanism, dangerous permissions related to sensitive information were mapped to corresponding permission groups, such as location, contacts, and calls, laying the foundation for subsequent model detection.

[0124] III. Creating a Multi-Dimensional Privacy Policy Sensitive Feature Constraint Pattern:

[0125] Based on thematic information related to APP privacy policies, relevant laws and regulations, and application market requirements, a multi-dimensional privacy policy sensitive feature model is constructed. Regarding thematic information extraction, based on previous research and by observing APP privacy policies in different application areas and industries, privacy-related thematic information is collected, such as personal information collection, data usage, and sharing methods. These themes are categorized and extracted to lay the foundation for the construction of multi-granularity standards. In terms of legal and regulatory analysis, privacy regulations in different countries and regions are analyzed, comprehensively considering the European General Data Protection Regulation (GDPR) and privacy regulations in other countries and regions, such as the national standard GB / T 35273-2020 "Personal Information Security Specification" (PISS) released in September 2020, and the "Network Security Standard Practice Guide—Security Guidelines for Mobile Internet Applications (Apps) Using Software Development Kits (SDKs)" (NOSSTC) released in November 2020. The requirements and regulations for accessing sensitive data are determined, and key legal clauses and norms are identified for inclusion in the testing standards. Regarding app store requirements, we comprehensively consider the official app release guidelines published by Android developers. For example, the Google Play app store requires all developers of apps released on Google Play to declare how they collect and process user data, and to detail how they protect this data through security measures such as encryption, thus defining their requirements and regulations for accessing sensitive data. Based on this, and taking into account the theme information of privacy policies, laws and regulations, and app store requirements, we have developed a multi-layered privacy policy sensitive feature detection model.

[0126] LEVEL1-TFP: mainly includes 15 Topic Feature Patterns (TFPs), as shown below:

[0127] FTP-PI: This stands for Privacy Policy Summary, a description of the privacy policy document, including the personal pronouns used in the definition document.

[0128] FTP-TI: This refers to the title of the privacy policy. App stores such as Google Play require that the words "Privacy Policy" be clearly listed in the title of the privacy policy document.

[0129] FTP-DI: This stands for Developer Information. Entities mentioned in the app's Google Play listing (such as the developer or company) must appear in the privacy policy, or the app name must appear in the privacy policy. It also includes contact information or a mechanism for contacting the app developer regarding privacy issues.

[0130] FTP-AP: Actions and Purposes: This specifies what types of data will be collected, accessed, and used, including but not limited to personal information and device information. It clearly lists the purpose of using the data, such as providing specific services, improving user experience, or providing personalized recommendations.

[0131] FTP-PT: This refers to processing technology. GDPR and CCPA regulations require that the processing of personal data and sensitive user data must follow strict secure data processing procedures. Data collectors must clearly explain the processing procedures for sensitive user data to ensure data privacy and security, such as how cookies and other similar technologies are collected and used.

[0132] FTP-DS: which stands for Data Sharing, refers to how developers share and disclose information with third parties, including company affiliates, service providers, or advertising partners.

[0133] FTP-DE: This stands for Data Security. GDPR employs a range of security measures and methods to protect user information, aiming to ensure the security and privacy of personal data.

[0134] FTP-UR: This stands for User Rights. GDPR requires data subjects to have the right to access, correct, and erase data, providing users with ways to exercise their data privacy rights and control, such as the right to opt out of data sharing, the right to correct or delete personal data, and settings for user privacy and security.

[0135] FTP-DR stands for Data Retention, which refers to the developer's data retention and deletion policies, and the requirements for data storage and duration.

[0136] FTP-IDT: International Data Transfer. Privacy policies should clearly describe the circumstances surrounding international data transfers, especially when cross-border data transfers are involved.

[0137] FTP-PC: This refers to changes in privacy policies. App stores require developers to inform users and ensure they are aware of these changes whenever their privacy policies are updated or changed.

[0138] FTP-LAW: This refers to the legal basis. The privacy policy should clearly provide relevant legal grounds to demonstrate that the processing of personal data is legal. Privacy laws and regulations cited in the privacy policy, such as the General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA), are examples.

[0139] FTP-AD: This refers to advertising use. If an app uses personal and sensitive user data for ad delivery, it must comply with Google Play's advertising policies. For example, if an app uses or collects device location data based on permissions for advertising purposes, it must clearly inform users and explicitly state this in its mandatory app privacy policy, which should include links to all relevant advertising network privacy policies involving the use of location data.

[0140] FTP-CH: This stands for Child Protection. Google Play prioritizes children's privacy, especially when it comes to personal data, to help users ensure that apps are safe for all users. For example, its privacy policy clearly defines the age range for children and the conditions for data access.

[0141] FTP-LV: This refers to the validity of the link. Before releasing an application, developers need to upload the correct link related to the privacy policy. The privacy policy URL should be accessible, meaning users can easily find and access it. It's crucial to ensure the privacy policy URL is stable and doesn't change frequently. Therefore, URL validity (FTP-LV) is an important prerequisite for verifying the compliance of the privacy policy.

[0142] LEVEL2-WACP: Based on the FTP-AP topic, i.e. actions and purposes, further fine-grained pattern extraction is performed. The privacy policy should clearly state under what circumstances user data will be accessed and for what purpose this data will be used.

[0143] 1. Data Collectors: The privacy policy should clearly state who collects the data. Typically, the data collector refers to the developer or operator, but it can also explicitly indicate that third-party services or organizations they work with (such as advertising partners, analytics tool providers, etc.) may be involved in the data collection process.

[0144] 2. Data collection conditions: Explain under what circumstances user data will be collected, such as registration, login, use of specific functions, etc.

[0145] 3. Data types to be collected: List in detail what types of data will be collected, including but not limited to personal information, device information, etc.

[0146] 4. Data-related actions: An app's access to, collection of, and use of data refers to three basic behaviors involved in the app's operation:

[0147] (1) Access Data: This refers to the information that an app obtains from users through technical means. This includes information voluntarily provided by users (such as forms filled out during registration), information obtained through device sensors (such as geographical location), and information automatically collected by the app (such as IP address, device information, etc.). Keywords include, but are not limited to, provide, acquire, read, collect, input, access, etc.

[0148] (2) Data collection: This refers to the purposeful storage of user information for later use. This may include storing user-provided information in a database or storing information on the user's device using technologies such as cookies. Keywords include, but are not limited to, process, display, store, pay, analyze, call, save, show, etc.

[0149] (3) Using data: Using data refers to applying data to a specific purpose, such as analysis, processing, storage, sharing, or display, after obtaining access rights and legally collecting the data. Keywords include, but are not limited to, Use, Utilize, Employ, Apply, etc.

[0150] 5. Purpose of Data Acquisition: Developers should clearly state the purpose of data usage, such as providing specific services, improving user experience, or personalized recommendations. When conducting these activities, developers must comply with applicable privacy regulations and clearly inform users in their privacy policies to build trust and protect user privacy rights.

[0151] In summary, the LEVEL2-WACP model mainly includes the subject (Who), action (Action), sensitive information (What), condition (Condition), and purpose (Purpose) related to the collection and use of sensitive information, and generates a fine-grained detection model for accessing sensitive features based on this.

[0152] Sensitive information includes not only users' personal information, such as general personal information like name, phone number, email address, home address, health data, and biometric data, but also dangerous permissions related to calendar, camera, contacts, and location, as shown in the table below:

[0153]

[0154] Android 6.0 introduced the concept of "permission groups," further dividing dangerous permissions into nine groups containing 24 permissions. These permissions cover data most relevant to user privacy, such as camera, phone, and contacts. When a user grants a permission, other permissions in that group are also granted by default. For example, the PHONE permission group includes phone-related permissions such as CALL_PHONE, READ_CALL_LOG, and WRITE_CALL_LOG. When a user grants an app the CALL_PHONE permission to read the phone, other permissions like READ_CALL_LOG and WRITE_CALL_LOG are also granted by default.

[0155] IV. Constructing an annotated corpus based on a sensitive feature-constrained pattern corpus:

[0156] In the process of labeling datasets, privacy policies are often lengthy texts. Processing large amounts of content is time-consuming and labor-intensive. It requires not only mapping multi-granularity sensitive feature constraint patterns to the text in the corpus, but also ensuring that the corpus contains samples of different categories and features to meet diversity requirements. The challenges of labeling privacy policies stem from their complexity, diversity, and the ambiguity of legal provisions. The complex language and structure of privacy policies, often using legal and technical terminology and complex sentence structures, make reading and comprehension difficult. Different applications may have different privacy policies, covering varying topics and content, thus requiring labeling for different policies. Privacy policies may contain repetitive and redundant content, requiring the identification of their actual meaning to avoid duplicate labeling. Furthermore, privacy policies may sometimes rely on subjective interpretation; labelers may have different understandings of the meaning of certain terms. Therefore, it is necessary to evaluate and validate the constructed multi-granularity privacy policy labeled corpus, continuously optimizing the labeled corpus and improving the quality of the labeled dataset.

[0157] Inspired by previous research, this study aims to improve annotation quality by focusing on three aspects: dataset selection, preprocessing, and annotation personnel, in order to properly control the annotation process and the quality of the annotated data corpus. The dataset primarily focuses on top-ranking apps in app stores, and its coverage is enhanced by categorical features. In the annotation phase, natural language processing (NLP) techniques are first used to assist in the analysis and filtering of privacy policy documents. Possible paragraphs are selected through keyword bag-of-words analysis. Based on this, a combination of manual and automatic annotation is employed to further analyze the document structure of privacy policies. Relevant paragraphs and sentences are selected by establishing a tag library and a common keyword library. In the implementation phase, professional annotators, based on open-source tagging tools, first discuss the initial tagging scheme and refine the tagging scheme and process according to existing research findings. Unlike traditional annotation work, this study annotates sensitive features of the privacy policy at different granularities during the dataset annotation process, including a topic feature annotation set (LEVEL1-TFP) and a fine-grained sensitive resource access annotation set (LEVEL2-WACP), and maps them to the constructed sensitive feature constraint patterns. Using the Docano open-source text annotation tool, we completed text classification and sequence labeling tasks. The Fleiss' Kappa algorithm was used to evaluate the labeled data, ensuring the accuracy and compliance of the annotations. By constructing constraint patterns for different categories of apps and a manually annotated corpus, we provided a high-quality training dataset for building sensitive resource analysis and detection models.

[0158] V. Constructing a sensitive feature classification and detection model based on a large-scale language model:

[0159] This work proposes a novel task: analyzing sensitive resource documents based on long texts, automatically providing topic and semantic tags for paragraphs and sentences, and generating sensitive feature patterns. Therefore, based on multi-dimensional privacy policy-constrained sensitive feature patterns and manually labeled datasets, this work utilizes NLP techniques, the Universal Information Extraction Framework (UIE), and large language models such as RoBERTa to establish a sensitive feature classification, detection, and generation model based on a large-scale language model. First, labeled data is used to fine-tune the NER model and the dialogue language model to complete named entity recognition, summary extraction, and named entity extraction tasks. To achieve this goal, pre-trained models based on UIE, T5, and RoBERTa are constructed. The Universal Information Extraction Framework (UIE) enables unified modeling of tasks such as entity extraction, relation extraction, event extraction, and sentiment analysis, and provides good transfer and generalization capabilities across different tasks. T5 (Text-to-Text Transfer Transformer) is an open-source large-scale language model proposed by Google, a Transformer-based pre-trained language model. The T5 model has 1.1 billion parameters and can perform a variety of natural language processing tasks, such as text classification, text generation, and question answering systems, while also having high accuracy and versatility.

[0160] Based on this, information is extracted for different categories using a pre-trained model. Different detection techniques are employed for different sensitive feature patterns.

[0161] In this embodiment, a method for detecting and analyzing sensitive features of an app based on a large language model, the actual processing steps for the three-dimensional tensor similarity of sensitive features of web page addresses, image information data, text information data, audio data, and video data are as follows:

[0162] 1. Three-dimensional tensor structure of webpage address sensitivity features:

[0163] Assume the shape of the three-dimensional tensor is (C, H, W), where:

[0164] Channel dimension (C): may represent different categories of sensitive features (such as URL path, parameters, domain name, encoding method, etc.).

[0165] Height (H) and width (W): may represent time series (such as request frequency), character-level encoding (such as one-hot vectors of letters), or spatial structure (such as matrix expansion of session logs).

[0166] 2. Three-dimensional tensor structure of image address-sensitive features:

[0167] Assume the tensor shape is (C, H, W), where:

[0168] Channel dimension (C): can represent:

[0169] Different sensitive attributes (such as URL hash, domain name encoding, parameter type, etc.).

[0170] Multimodal features (such as the fusion of text descriptions and image visual features).

[0171] Comparison before and after desensitization (e.g., original features vs. encrypted features).

[0172] Spatial dimension (H, W): can represent:

[0173] Character-level encoding of address strings (such as a one-hot matrix).

[0174] The spatial distribution after image processing (such as the pixel matrix of a QR code or barcode).

[0175] 3. Common structures of three-dimensional tensors for text data:

[0176] Assume the shape of the 3D tensor is (L, D, C) (common in NLP tasks):

[0177] L (length dimension): sequence length (such as the number of words or characters in a sentence).

[0178] D (Feature Dimension): Embedding dimension of words / characters (such as BERT's 768-dimensional vector).

[0179] C (channel dimension): Multiple feature sources (such as different embedding layers, multilingual encoding, time steps).

[0180] 4. Typical structure of audio three-dimensional tensor

[0181] Assume the tensor shape is (T, F, C):

[0182] T (Time dimension): Number of audio frames or time steps (e.g., one frame every 10ms).

[0183] F (Frequency Dimension): Number of frequency bands or spectral characteristics (e.g., 80 bands in the Mel spectrum).

[0184] C (Channel Dimension): Multi-source features (such as left and right channels, multi-microphone signals, and combinations of different acoustic features).

[0185] 5. Typical structure of a video 3D tensor:

[0186] Assuming the tensor shape is (T,H,W,C) (time × height × width × channels) or (C,T,H,W) (as in PyTorch format):

[0187] Time dimension (T): Video frame sequence (dynamically changing).

[0188] Spatial dimension (H,W): pixels or regions in each frame of an image.

[0189] Channel dimension (C): Multimodal features (such as RGB, optical flow, depth map) or multi-task features (such as sensitive information such as face, license plate, etc.).

[0190] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0191] This invention is illustrated by referring to flowchart illustrations of methods, apparatus (systems), and computer program products according to embodiments of this application.

[0192] The flowchart and / or block diagram are used for description. It should be understood that each step in the flowchart and / or block diagram can be implemented by computer program instructions.

[0193] / or blocks, and combinations of flowcharts and / or blocks in block diagrams. These computer program instructions may be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, produce a machine for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0194] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0195] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0196] 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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting and analyzing sensitive features of an app based on a large language model, characterized in that, include: S1. Use the HTML data of the webpage corresponding to the sensitive resource files of the APP to parse and extract data to establish a multi-dimensional sensitive feature dataset; S2. Using the multidimensional sensitive feature dataset, an annotation process is performed based on the sensitive feature constraint pattern corpus to obtain a multidimensional sensitive feature data annotation corpus; S3. Based on the multidimensional sensitive feature data-annotated corpus, obtain the APP sensitive feature detection and analysis results using a large language model.

2. The method for detecting and analyzing sensitive features of an app based on a large language model as described in claim 1, characterized in that, The step of using the HTML data of the webpage corresponding to the sensitive resource files of the APP to parse and extract data to establish a multidimensional sensitive feature dataset includes: S1-1. Using the HTML data of the webpage corresponding to the sensitive resource files of the APP, obtain the multi-dimensional feature elements of the webpage HTML data based on the crawling algorithm; S1-2. Use the multi-dimensional feature elements of the web page HTML data to parse and extract them to establish a multi-dimensional sensitive feature dataset; The multidimensional feature elements include web addresses, images, text, audio, and video.

3. The method for detecting and analyzing sensitive features of an app based on a large language model as described in claim 2, characterized in that, Using the HTML data of the webpage corresponding to the sensitive resource files of the APP, the multi-dimensional feature elements of the webpage HTML data obtained by the web crawler algorithm include: Use the sensitive resource files of the APP to obtain the corresponding webpage HTML data; Using the HTML data of the webpage, a web crawler algorithm is used to obtain webpage address data, image information data, text information data, audio data information, and video data information respectively; The webpage address data, image information data, text information data, audio data information, and video data information are used as multidimensional feature elements of the webpage HTML data.

4. The APP sensitive feature detection and analysis method based on a large language model as described in claim 3, characterized in that, The process of parsing and extracting multidimensional feature elements from the webpage HTML data to establish a multidimensional sensitive feature dataset includes: S1-2-1. Establish document data volume thresholds based on privacy policies; S1-2-2. Determine whether the data volume of the multidimensional feature elements of the web page HTML data is less than the document data volume threshold. If so, delete the multidimensional feature elements that are less than the document data volume threshold. Otherwise, use the multidimensional feature elements of the web page HTML data to establish a multidimensional sensitive feature dataset. The document data size threshold is 2kb.

5. The method for detecting and analyzing sensitive features of an app based on a large language model as described in claim 2, characterized in that, The multidimensional sensitive feature data annotation corpus obtained by annotating the aforementioned multidimensional sensitive feature dataset based on the sensitive feature constraint pattern corpus includes: S2-1. Preprocess the multidimensional sensitive feature dataset to establish a sensitive information permission mapping. S2-2. Based on the sensitive information permission mapping, a multi-dimensional sensitive feature data annotation corpus is obtained by annotating the sensitive feature constraint pattern corpus.

6. The method for detecting and analyzing sensitive features of an app based on a large language model as described in claim 5, characterized in that, The process of preprocessing the multidimensional sensitive feature dataset to establish a sensitive information permission mapping includes: S2-1-1. Using the multidimensional sensitive feature dataset, perform punctuation removal, stop word bag construction, language judgment and filtering, web page bullet point removal and title bar noise removal in sequence to obtain a multidimensional sensitive feature preprocessing dataset. S2-1-2. Using the multi-dimensional sensitive features, the dataset is preprocessed and dangerous permission mapping is performed according to the corresponding subset type to establish the initial sensitive information permission mapping. S2-1-3. Determine whether the initial sensitive information permission mapping has overlapping mappings. If so, obtain the data volume of the initial sensitive information permission mapping with overlapping mappings and execute S2-1-4. Otherwise, retain the initial sensitive information permission mapping as the sensitive information permission mapping. S2-1-4. Based on the amount of data in the overlapping initial sensitive information permission mappings, obtain a subset of the initial sensitive information permission mappings with a relatively smaller amount of data as the abnormal initial sensitive information permission mappings. S2-1-5. Use the initial sensitive information permission mapping to delete the abnormal initial sensitive information permission mapping to obtain the sensitive information permission mapping; The overlapping of the mappings refers to the situation where there is a non-single correspondence between the mappings.

7. The method for detecting and analyzing sensitive features of an app based on a large language model as described in claim 6, characterized in that, Based on the sensitive information permission mapping and the sensitive feature constraint pattern corpus, a multi-dimensional sensitive feature data annotation corpus is obtained by annotation processing, including: Using the aforementioned sensitive information permission mapping, a basic annotated corpus of multi-dimensional sensitive feature data is obtained by performing multi-feature annotation processing based on the sensitive feature constraint pattern corpus; Using the corresponding action target annotations in the basic annotated corpus, fine-grained extraction processing is performed to obtain data collectors, data collection conditions, data collection types, and data-related actions to establish fine-grained action target extraction results; The basic annotated corpus of the aforementioned multidimensional sensitive feature data and the fine-grained extraction results of action targets are used as the multidimensional sensitive feature data annotated corpus; The multi-feature annotation includes annotations for privacy policy overview, privacy policy title, developer information, action target, processing technology, data sharing, data security, user rights, data retention, international data transmission, changes to privacy policy, legal basis, advertising use, child protection, and link validity.

8. A method for detecting and analyzing sensitive features of an app based on a large language model, as described in claim 3, characterized in that: Based on the multidimensional sensitive feature data-annotated corpus and a large language model, the APP sensitive feature detection and analysis results include: S3-1. Using the multidimensional sensitive feature data-annotated corpus, establish a sensitive feature classification and detection model based on a large language model; S3-2. Obtain the APP sensitive feature detection and analysis results using the aforementioned sensitive feature classification and detection model.

9. A method for detecting and analyzing sensitive features of an app based on a large language model, as described in claim 8, characterized in that: The sensitive feature classification and detection model based on a large language model, using the multidimensional sensitive feature data-annotated corpus, includes: S3-1-1. Use the HTML data of the corresponding web pages in the basic annotated corpus of the multidimensional sensitive feature data annotation corpus as the first dataset, and use the basic annotated corpus of the multidimensional sensitive feature data annotation corpus as the second dataset. S3-1-2. Based on the first dataset and the second dataset, train a sensitive feature extraction model using RoBERTa. S3-1-3. Using the multidimensional sensitive feature data annotation corpus, obtain the three-dimensional tensor of sensitive features based on the sensitive feature extraction model; S3-1-4. Based on the three-dimensional tensor of the sensitive features, a sensitive feature variable enhancement model is established by training on T5. S3-1-5. Using the multidimensional sensitive feature data-annotated corpus, obtain sensitive feature text sequences based on the sensitive feature variable enhancement model; S3-1-6. Using the aforementioned sensitive feature text sequence, a sensitive feature structure extraction model is established based on UIE for training. S3-1-7. The aforementioned sensitive feature basic extraction model, sensitive feature variable enhancement model, and sensitive feature structured extraction model are used as sensitive feature classification and detection models.

10. The method for detecting and analyzing sensitive features of an app based on a large language model as described in claim 9, characterized in that, The results of sensitive feature detection and analysis of the APP obtained using the aforementioned sensitive feature classification and detection model include: S3-2-1. Obtain the three-dimensional tensors of webpage address data sensitive features, image information data sensitive features, text information data sensitive features, audio data information sensitive features, and video data information sensitive features corresponding to the multi-dimensional sensitive feature data annotation corpus. S3-2-2, Calculate the similarity of the three-dimensional tensor of the web address sensitive features based on the channel dimension using the three-dimensional tensor of the web address sensitive features; S3-2-3. Calculate the similarity of the three-dimensional tensor of the image information data sensitive features based on the spatial dimension using the three-dimensional tensor of the image information data sensitive features. S3-2-4. Calculate the similarity of the three-dimensional tensor of the sensitive features of the text information data based on the length dimension using the three-dimensional tensor of the sensitive features of the text information data. S3-2-5. Calculate the similarity of the three-dimensional tensor of the sensitive features of the audio data based on the time dimension using the three-dimensional tensor of the sensitive features of the audio data. S3-2-6. Calculate the similarity of the three-dimensional tensors of the sensitive features of video data based on the time dimension using the three-dimensional tensors of the sensitive features of video data. S3-2-7. Obtain the average three-dimensional tensor similarity of sensitive features for webpage address sensitive features, image information data sensitive features, text information data sensitive features, audio data sensitive features, and video data sensitive features, respectively. S3-2-8. Using the average value of the three-dimensional tensor similarity of the sensitive features, obtain the absolute value of the similarity difference between the three-dimensional tensor similarity of the web address sensitive features, the image information data sensitive features, the text information data sensitive features, the audio data information sensitive features, and the video data information sensitive features. S3-2-9. Obtain the corresponding three-dimensional tensor as the feature tensor of the tuning parameter based on the absolute value of the similarity difference; S3-2-10. Standardize the tuning parameter feature tensor to obtain the standardized tuning parameter feature dimension. S3-2-11. After optimizing the sensitive feature classification and detection model using the standardized tuning parameter feature dimension, obtain the APP sensitive feature detection and analysis results.

Images