Chrome extension permission risk assessment method based on dimensionality reduction driven belief rule base

Abstract

The invention discloses a Chrome extension permission risk assessment method based on a dimensionality reduction driven confidence rule base, and is suitable for the field of browser extension security. The invention provides a PermDR-BRB framework based on DR-BRB, the framework introduces multi-strategy dimensionality reduction including feature selection, weighted combination, PCA and Fast t-SNE into a confidence rule base model, not only can the Chrome extension permission risk be effectively assessed, but also the problem of combination rule explosion in a confidence rule base is solved. Experimental results show that the PermDR-BRB achieves the highest classification accuracy of 99.00% and the lowest MSE of 0.000555, and the effectiveness and the accuracy of the PermDR-BRB framework are proved.

Description

Technical Field

[0001] This invention relates to the field of data processing, specifically to browser extension security, and is a method for assessing Chrome extension permission risks based on a dimensionality-reduced trust rule base. Background Technology

[0002] Browser extensions offer users highly customized features. In actual Chrome extension development scenarios, some attackers request all necessary permissions at once during the initial extension development process. This is done to avoid triggering Chrome's warning mechanism during subsequent extension updates. After obtaining these permissions, attackers use these high-privilege calls to the extension API to develop malicious browser extensions. These malicious extensions are often used to carry out various attacks, causing serious harm to users. Attackers use malicious extensions to steal sensitive user information, such as modifying page content and forging authentication requests, to obtain illegal economic benefits. For example, in actual security analysis, a malicious extension called Routecue (aeebcpcmmalnaaddhbahbjepgoehlbla.1.1) was discovered. It steals OAuth tokens and then obtains sensitive information such as user documents, forms, and channels from third-party services. This attack pattern clearly shows that extension permissions are a key prerequisite for malicious behavior. Effective supervision of extension permission requests and usage will greatly enhance user information security.

[0003] However, the current Chrome extension security mechanism has significant limitations in permission assessment and management, and its permission management mechanism is rather crude. Specifically, the existing Chrome extension permission management mechanism cannot allow developers to precisely control the permissions of individual extensions. This deficiency often forces developers to request permissions that exceed their actual usage needs, thus violating the principle of least privilege. For example, the ad-blocking extension Adblock Plus explicitly declares global host permissions in its manifest file.<all_urls> This means the extension can access the content and requests of all websites. Adblock Plus's core function is ad filtering, which primarily relies on the `chrome.webRequest` and `chrome.declarativeNetRequestWithHostAccess` APIs to intercept, modify, or block web page requests. Due to Chrome's permission system, extensions must declare relevant host permissions when using these APIs. However, ads may appear on various websites, and Adblock Plus cannot declare permissions only for the ad domain. To ensure the comprehensiveness and integrity of its ad-blocking functionality, developers can only request global host permissions. But the result is that the extension effectively gains access to and modification rights for all website requests, including even sites involving sensitive user information such as banks and email providers. This undoubtedly increases the risk of user information leakage and security, highlighting the urgent need to improve the current Chrome extension permission management mechanism.

[0004] As can be seen from the above, extended access assessment is of paramount importance. It clearly demonstrates the actual scope of extended access and helps the invention accurately determine the risk level of extended permissions. Currently, there are three main research methods in the field of risk assessment. The first is a knowledge-driven approach. This method relies on the accumulated knowledge of experts, focusing on building models with expert experience and interpretability. Examples include the Analytic Hierarchy Process (AHP), Fault Tree Analysis (FTA), and Decision Trees. The second is a data-driven approach. Data-driven methods primarily utilize historical data, employing statistical analysis, machine learning, and other techniques for prediction. Support Vector Machines (SVMs) are an example of this approach. The third is a hybrid approach. This method combines the advantages of knowledge-driven and data-driven approaches, typically integrating expert knowledge with data analysis to improve model accuracy through multiple methods. Evidence-based reasoning and confidence rule bases are common hybrid approaches.

[0005] Belief Rule Base (BRB) is a hybrid-driven approach. It's a rule-based reasoning method used to build and reason about rule bases for expert systems under conditions of uncertainty. It integrates traditional IF-THEN rules, Dempster-Shafer evidence theory, decision theory, and fuzzy theory, enabling it to handle incomplete, imprecise, and uncertain information. BRB effectively balances the accuracy of data-driven approaches with the interpretability of knowledge-driven approaches. This characteristic lays a preliminary foundation for conducting permission risk assessments for Chrome extensions.

[0006] However, in practical extended permission risk assessment systems, traditional confidence rule bases cannot meet actual needs. This is mainly because BRBs combine rules in the form of Cartesian products. As the number of assessment indicators and reference points increases, the number of BRB rules grows exponentially, easily leading to a rule explosion problem. To alleviate this problem, the academic community has proposed several feasible solutions, including building hierarchical confidence rule bases based on fault trees, reducing rules based on decision trees, and using interval confidence rule bases. Summary of the Invention

[0007] To alleviate the problem of rule explosion, this invention proposes for the first time the most comprehensive confidence rule base method based on multi-model dimensionality reduction, namely the Dimensionality Reduction–Driven Belief RuleBase (DRBRB). This invention employs multiple methods for dimensionality reduction, including feature selection, weighted combination, average combination, principal component analysis, and fast t-distribution random neighborhood embedding, and combines these with the confidence rule base to conduct Chrome extension permission risk assessment. Specifically, the main contributions of this invention are as follows:

[0008] (1) This invention proposes for the first time a novel extended permission risk assessment framework based on a dimensionality reduction-driven confidence rule base, namely the Permission Assessment Framework based on DRBRB (PAF-DRBRB). This framework is based on DRBRB theory and incorporates multi-dimensional features such as over-authorization detection, single-risk permission detection, and combined permission risk detection, and is used to conduct risk assessment of Chrome extended permissions.

[0009] (2) This invention proposes for the first time a comprehensive confidence rule base method based on multi-model dimensionality reduction. This method can effectively reduce the dimensionality of evaluation indicators in the confidence rule base and alleviate the problem of combinatorial rule explosion caused by too many evaluation indicators in BRB.

[0010] (3) Given the current limitations of the Chrome Web Store (CWS) in terms of the fine-grained management of permission evaluation, this invention proposes a new single-permission evaluation mechanism. This mechanism evaluates the risk level of a single permission from aspects such as the danger of the permission granting behavior, whether the permission triggers a warning, and whether the permission is directly related to user privacy, providing users with a valuable reference basis.

[0011] (4) This invention proposes a method for detecting abuse of permissions. This method can detect whether there is excessive authorization behavior in the extension and identify the specific excessive authorization situation.

[0012] The technical solution of this invention is:

[0013] The Chrome extension permission risk assessment method based on a dimensionality reduction-driven confidence rule base includes the following steps:

[0014] Step 1: Obtain the source data of the Chrome extension, including the CRX extension package and metadata;

[0015] Step 2: Extract risk characteristic indicators of extended permissions to accurately capture the risks hidden in extended permissions;

[0016] The extraction of extended permission risk characteristic indicators includes the following steps:

[0017] S21. Extract over-authorization features to reflect the mismatch between extended functionality and permission requests;

[0018] S22. Extract high-risk combined permission features to demonstrate the risks arising from interactions between different permissions;

[0019] S23. Design and extract single-permission risk features to focus on the sensitive permissions themselves;

[0020] Step 3: Quantify the extracted permission features and assign them realistic labels; by quantifying permission features from different dimensions, an accurate assessment of Chrome extension risks can be achieved, including the following steps:

[0021] S31, Single-permission score generation;

[0022] S32. Generation of a rating for abuse of privileges;

[0023] S33, Generation of combined permission scores;

[0024] S34. Generation of overall risk score;

[0025] Step 4: Establish an extended permission risk assessment model based on a dimensionality reduction-driven confidence rule base;

[0026] The modeling process of the extended permission risk assessment model is as follows:

[0027] S41. Determine the feature set of extended permission evaluation indicators;

[0028] S42. Perform dimensionality reduction on the feature set of extended permissions to obtain the dimensionality-reduced permission indicators.

[0029] S43. Construct confidence rules for the dimensionality-reduced permission indicators and assign initial confidence levels based on expert knowledge;

[0030] Step 5: Reason about the model based on evidence;

[0031] The DRBRB model in the PAF-DRBRB framework is adopted, and the evidence-based reasoning module is used in the reasoning stage to realize rule fusion and decision inference.

[0032] Step 6: Optimize the parameters of the extended permission risk assessment model based on the dimensionality reduction-driven confidence rule base;

[0033] Step 7: Real-world experiment assessment of the risks associated with extended permissions;

[0034] The effectiveness of PAF-DRBRB, a Chrome extension risk assessment framework based on DRBRB, is demonstrated. DRBRB implements five BRB variants: Feature Selection-Confidence Rule Base, Weighted Combination-Confidence Rule Base, Average Combination-Confidence Rule Base, Principal Component Analysis-Confidence Rule Base, and Fast t-SNE-Confidence Rule Base. Furthermore, the permission risk level is defined as high, medium, and low.

[0035] Furthermore, step S21 includes:

[0036] S211. To extract over-granted permissions, you must first extract the permissions declared in the extended declaration. ; This indicates a specific permission in the extended declaration;

[0037] S212, Next, based on the abstract syntax tree analysis, extract the permissions actually used for the extension. ; This indicates an extension of a specific permission that is actually being used;

[0038] S213, then based on permissions and extended API benchmarks A comparison is made between the permissions declared in the extended statement and the permissions actually used. This is used to identify whether extensions have excessive or redundant permissions. Excessively authorized extensions are represented as follows:

[0039]

[0040] in and These represent the extended declaration and the actual set of permissions used, respectively. This indicates a mapping from extended APIs to permissions; This represents the difference between the set of permissions declared in the extension declaration and the set of permissions actually used by the extension; This indicates a set of permissions that have been excessively granted. This indicates that the extension has excessive permissions; otherwise, it does not.

[0041] Furthermore, step S23 includes:

[0042] Based on the degree of danger of granting permissions, whether permissions will trigger warnings, and whether permissions directly involve user privacy, a preliminary risk assessment of single permissions is conducted, and a preliminary risk level for each single permission is assigned. Then, the knowledge accumulated from the single permission assessment is integrated into an expert knowledge base. This process is described as follows:

[0043]

[0044]

[0045]

[0046] in: Indicates permissions The corresponding characteristics of the degree of danger of the authorized behavior; Indicates permissions Whether the alarm characteristics trigger a system security warning; Indicates permissions Does it directly involve privacy-related features of user privacy data? Indicates a single-authority evaluation set; Indicates the single-access assessment level; An evaluation function representing a single-privilege evaluation level; This represents a single-access risk fusion assessment function. For risk level mapping function; For expert knowledge base.

[0047] Furthermore, in step S31:

[0048] The single-permission score generator performs a security risk assessment on permissions with extended declarations, dividing permissions into four predefined levels: critical, high, medium, and low, with each level corresponding to a specific weight. By calculating a weighted average of all permissions, a standardized score reflecting the overall risk contribution of a single permission is derived. The calculation formula is as follows:

[0049]

[0050] in, Representing the The weight of each risk level, It refers to the number of permissions at that level. This indicates a single-access risk score.

[0051] Furthermore, in step S32:

[0052] The permission abuse rating generator identifies requests for extended permissions that exceed functional requirements. The evaluation results are presented in binary form: a score of 1 indicates unnecessary permissions, suggesting potential over-authorization or security risks; a score of 0 indicates that the extended permission request is legitimate. Mathematically, this can be expressed as:

[0053]

[0054] in, This represents the privilege abuse score or over-granting score. A value of 1 indicates over-granting or privilege abuse, while a value of 0 indicates no abuse.

[0055] Furthermore, in step S33:

[0056] The combined permission scoring generator focuses on assessing the potential risks of permission combinations, as these combinations could be maliciously extended and abused. For each predefined rule, a corresponding empirical weight is assigned. Then, the rules are matched against extended permissions, and the generated combined permission risk score reflects the weighted matching degree between the extended permissions and all rules. The calculation formula is as follows:

[0057]

[0058] in, This represents the total number of rules for combining sensitive permissions. It is the first The weight of each rule, This indicates the degree to which the extension matches the rule. This indicates the risk score for combined permissions.

[0059] Furthermore, in step S34:

[0060] After obtaining the scores from steps S1-S3, the total risk score generator integrates the single-permission score, permission abuse score, and combined permission score to calculate the final risk score for each extension. The total risk score is denoted as... The calculation formula is:

[0061]

[0062] Finally, based on the 33rd and 67th percentiles of the final risk score, the risk level is further divided into three categories: low, medium, and high.

[0063] Furthermore, in step S42, the dimensionality-reduced set of indicator features is: Five joint dimensionality reduction strategies were employed to reduce the dimensionality of the extended permission metrics. These strategies included:

[0064]

[0065]

[0066]

[0067]

[0068]

[0069] in, This represents the set of index features after processing by their respective dimensionality reduction methods. , , , and These represent five dimensionality reduction functions: feature selection, weighted combination, average combination, principal component analysis, and fast t-SNE. Indicates the original dimensions of the evaluation indicators; The set of parameters representing the dimensionality reduction process.

[0070] Furthermore, in step S5:

[0071] set up Indicates the first input sample The dimensionality reduction attribute and the first The degree of matching between the reference values ​​of the rule premises Indicates the first The first feature These are reference points, taking into account the importance of the rules. and the contribution of each attribute Based on this, the first Activation weight of the rule Calculate using the following formula:

[0072]

[0073] Indicates the first Each sample, matching degree This can be derived from the following formula:

[0074]

[0075] In the inference phase of DRBRB, an evidence-based reasoning method is used to fuse the activated rules, thereby obtaining the confidence levels for each output level. The number of rules involved in the reasoning. Indicates the total number of possible output levels, the first... The overall belief level corresponding to each output Calculated using the following formula:

[0076]

[0077]

[0078] Wherein, the normalization coefficient is ;

[0079] Finally, the overall output of the system is obtained by weighted summing of the belief level of each output level and its corresponding utility value, as shown in the following formula:

[0080]

[0081]

[0082] in, This represents the expected utility value; This represents the confidence distribution of the output results; Indicates the reference level of the result; Indicates the confidence level of a reference grade for a specific result; Indicates the number of reference levels for the result.

[0083] Compared to existing technologies, this invention has the following advantages: Compared to six mainstream machine learning models—random forest, extreme gradient boosting, gradient boosting machine, multilayer perceptron, support vector regression, and linear regression—this invention can achieve similar accuracy. This indicates that these models can effectively perform extended permission risk assessment. However, although random forest and gradient boosting models perform well in overall dimensions, their internal decision-making processes are relatively opaque, making it difficult to provide clear reasoning. In contrast, this invention is based on a confidence rule base. Its reasoning process consists of explicit rules and confidence distributions, possessing good interpretability and traceability.

[0084] This invention also has a natural advantage in handling uncertainty and incomplete information. By employing an evidence-based reasoning mechanism to assign confidence to input features, it effectively characterizes this uncertainty. Furthermore, this invention maintains good performance even with small sample sizes. Unlike machine learning models that rely on large-scale data-driven learning, this invention expresses the relationship between input and output through rule structures and confidence parameters, thus having a lower dependence on the size of the training samples. Attached Figure Description

[0085] Figure 1 This is a diagram of the overall structure of the PAF-DRBRB frame;

[0086] Figure 2 A graph showing the fitting results of PAF-DRBRB predicted and actual values;

[0087] Figure 3 for Figure 2 The magnified view of the enlarged area shows the region with the most significant differences among samples 23-43;

[0088] Figure 4 A graph showing the predictive indicators for PAF-DRBRB results;

[0089] Figure 5 This is a graph showing the error analysis of the PAF-DRBRB results. Detailed Implementation

[0090] like Figure 1 As shown, the specific steps of the Chrome extension permission risk assessment method based on a dimensionality-reduced confidence rule base of the present invention can be described as follows:

[0091] Step 1: Obtain the source data for the Chrome extension, including the CRX extension package and metadata.

[0092] This invention uses a data acquisition module to collect extension data from Chrome Web Store (CWS). The collected data mainly consists of two categories: firstly, CRX packages, which aid in structured analysis and static feature extraction of extensions; and secondly, metadata information, covering permission statements, developer information, user reviews, installation counts, etc., which can be used to analyze the permission risk characteristics of extensions. However, in actual Chrome extension permission risk assessment, simply collecting extension data from CWS is insufficient to fully depict the high-risk behavior patterns of extensions. The characteristics of removed extensions can help identify abnormal permission requests or malicious functional patterns, and can also provide negative samples for the model. Based on this, this invention further filters extensions that have been removed by Google from the dataset. In the PAF-DRBRB framework designed in this invention, the Filtering Unit is responsible for filtering samples of extensions removed by Google.

[0093] Step 2: Extract risk characteristic indicators for extended permissions.

[0094] In research on Chrome extension permission risk assessment, feature selection is a crucial step in constructing the assessment framework. To accurately capture the risks hidden in extension permissions, this invention selects three core features: high-risk combined permission features, over-granting features, and single-permission risk features. High-risk combined permission features reveal the risks arising from interactions between different permissions. Over-granting features reflect mismatches between extension functionality and permission requests, often indicating malicious intent. Single-permission risk features focus on the sensitive permissions themselves, such as location access or microphone access. In considering single-permission risk features, this invention also integrates other risk characteristics of the extension, such as whether the permission is directly related to user privacy, whether it will trigger warnings from Google, and the danger of the permission granting behavior.

[0095] The extraction of these three types of permission risk features is undertaken by the over-authorization detector, combined permission detector, and single permission detector designed in this invention, respectively. This provides basic feature support for subsequent risk assessment. The extraction process of these three types of features is described in detail below.

[0096] S21. Extract over-authorization features.

[0097] To extract over-granted permissions, you must first extract the permissions declared in the extended declaration. , This represents a specific permission declared in the extension; in the Chrome extension permission risk assessment system of this invention, the over-authorization detector plays a crucial role in detecting over-authorization. The over-authorization detector primarily constructs the set of over-authorization permissions from two different sources:

[0098] (1) Extend the permission list declared in the configuration file. When obtaining this source information, the version number of the Manifest should be given special consideration. Chrome extensions mainly use Manifest V2 and Manifest V3. For Manifest V2, the permission list can be obtained directly from the permissions field. However, for Manifest V3, the permission classification method has changed. In this case, both the permissions and host_permissions fields need to be checked. The permissions field currently mainly covers API permissions, while the host_permissions field contains the website addresses that the extension can access.

[0099] (2) API call set extracted from the source code AST. This process first requires decompressing the CRX file to obtain all its internal source files. Then, all JavaScript files are selected and analyzed in detail. For each JavaScript file, a parser is used to convert it into an abstract syntax tree. Based on the abstract syntax tree, the permissions actually used for extensions are extracted. , This indicates a specific permission that is actually used in the extension. Then, based on the permission and the extended API baseline... A comparison is made between the permissions declared in the extended statement and the permissions actually used. This is used to identify whether extensions have excessive or redundant permissions. Excessively authorized extensions are represented as follows:

[0100]

[0101] in, and These represent the extended declaration and the actual set of permissions used, respectively. This indicates a mapping from extended APIs to permissions; This represents the difference between the set of permissions declared in the extension declaration and the set of permissions actually used by the extension; This indicates a set of permissions that have been excessively granted. This indicates that the extension has redundant or excessive permissions. Using this tree structure, all nodes can be traversed, ultimately aggregating the discovered extension APIs into a complete set and storing them in a unified database.

[0102] S22. Extract high-risk combination permission features.

[0103] In the Chrome extension security assessment system of this invention, the frequently used permission combinations by malicious extensions are considered as one of the key indicators for assessment. To extract these permission combination characteristics, a combined permission detector is designed. This module can effectively identify and extract frequently used permission combinations by malicious extensions, thereby providing strong data support for the security assessment of Chrome extensions.

[0104] S23. Design and extract single-access risk features.

[0105] Based on the degree of danger of granting permissions, whether permissions will trigger warnings, and whether permissions directly involve user privacy, a preliminary risk assessment of single permissions is conducted, and a preliminary risk level for each single permission is assigned. Then, the knowledge accumulated from the single permission assessment is integrated into an expert knowledge base. This process is described as follows:

[0106]

[0107]

[0108]

[0109] in: Indicates permissions The corresponding characteristics of the degree of danger of the authorized behavior; Indicates permissions Whether the alarm characteristics trigger a system security warning; Indicates permissions Does it directly involve privacy-related features of user privacy data? Indicates a single-authority evaluation set; Indicates the single-access assessment level; An evaluation function representing a single-privilege evaluation level; This represents a single-access risk fusion assessment function. For risk level mapping function; For expert knowledge base.

[0110] Specifically, in the Chrome extension security assessment system of this invention, based on past experience in extension permission security analysis, this invention conducts a preliminary risk assessment of single permissions. This invention classifies single permissions into risk levels according to three criteria: whether the permission is directly related to user privacy, whether the permission will trigger a Google warning, and the degree of danger of the permission granting behavior. Based on these criteria, this invention divides single permissions into four risk levels: Low, Medium, High, and Critical.

[0111] For permissions that demonstrate a high level of threat across all three assessment dimensions, this invention defines their security level as "Critical". If a permission presents potential risk in only two dimensions, it will be classified as "High". If the risk is only present in a single dimension, the corresponding permission will be assessed as "Medium". Permissions that perform well across all assessment metrics and demonstrate robust security are assigned a "Low" rating.

[0112] Furthermore, this invention establishes a clear rule: any permission that triggers a Chrome security warning is directly identified as Critical. It is particularly noteworthy that in this study, the webRequest permission was consistently classified as the highest risk level. When this permission is combined with...<all_urls> When all permissions are used simultaneously, attackers can monitor all network requests, potentially stealing sensitive information entered by users on login pages. For example, attackers have used a malicious extension called "CredentialsStealer" to illegally collect users' private data.

[0113] Step 3: Quantify the extracted permission features and assign them real labels.

[0114] In the Chrome extension risk assessment process of this invention, step 3 focuses on quantifying the extracted permission features and assigning them realistic labels, thereby achieving an accurate assessment of Chrome extension risks. To achieve this goal, this step utilizes multiple feature generators to quantify the permission features from different dimensions.

[0115] S31, Single-permission rating generation.

[0116] The single-permission score generator performs a security risk assessment on each declared permission. It divides permissions into four predefined levels: critical, high, medium, and low, with each level corresponding to a specific weight. By calculating a weighted average of all permissions, a normalized score reflecting the overall risk contribution of a single permission is derived. The calculation formula is as follows:

[0117]

[0118] in, Representing the The weight of each risk level, It refers to the number of permissions at that level. This indicates a single-access risk score.

[0119] S32, Generation of abuse of privileges score.

[0120] A permission abuse rating generator identifies requests for permissions that exceed functional requirements. The evaluation results are presented in binary form: a score of 1 indicates unnecessary permissions, suggesting potential over-authorization or security risks; a score of 0 indicates the request for extended permissions is justified. Mathematically, this can be expressed as:

[0121]

[0122] in, This represents the privilege abuse score or over-granting score. A value of 1 indicates over-granting or privilege abuse, while a value of 0 indicates no abuse.

[0123] S33, Generation of combined permission scores.

[0124] The combined permission score generator focuses on assessing the potential risks of permission combinations, as these combinations can be maliciously extended and abused. For each predefined rule, a corresponding empirical weight is assigned, and then the rule is matched against the extended permissions. The generated score reflects the weighted matching degree between the extended permissions and all rules, calculated using the following formula:

[0125]

[0126] in, This represents the total number of rules for combining sensitive permissions. It is the first The weight of each rule, This indicates the degree to which the extension matches the rule. This indicates the risk score for combined permissions.

[0127] S34. Generation of overall risk score.

[0128] After obtaining the scores from the above categories, the total risk score generator integrates the single-permission score, the permission abuse score, and the combined permission score to calculate the final risk score for each extension. The total risk score is then recorded as follows: The calculation formula is:

[0129]

[0130] Finally, based on the 33rd and 67th percentiles of the final risk score, the extended risk is divided into three levels: low, medium, and high. This classification method ensures category balance while preserving the natural clustering of permission risk characteristics, providing an interpretable and reliable labeling system for subsequent permission assessments.

[0131] Furthermore, considering the numerous limitations of relying solely on expert manual annotation to expand permission risks during actual evaluation, on the one hand, the dataset is massive, and expert review would consume a significant amount of time and manpower; on the other hand, manual annotation is easily influenced by expert subjective experience, leading to inconsistent labels and low repeatability. Therefore, this invention employs multiple generators to initially score permissions, followed by expert sampling verification to ensure label quality.

[0132] Step 4: Establish an extended permission risk assessment model based on a dimensionality reduction-driven confidence rule base.

[0133] This invention employs the DRBRB model within the PAF-DRBRB framework. This model has broad applicability and can be used for risk assessment of most complex systems, not just Chrome extension permission assessment. However, the assessment metrics for complex systems are typically complex and contain numerous reference points. When using BRB combination rules for risk assessment, the number of rules grows exponentially, easily leading to rule explosion and consequently degrading model performance. To address this issue, this invention uses various dimensionality reduction techniques to process the multidimensional attributes of complex systems. These techniques include feature selection, weighted combination, average combination, principal component analysis, and fast t-distributed random neighborhood embedding. By using these techniques, the multidimensional attributes of complex systems are mapped to a low-dimensional space, effectively solving the rule combination explosion problem.

[0134] Specifically, this invention designs a dimensionality reduction processor to mitigate the impact of the curse of dimensionality. This framework offers flexible processing methods: on the one hand, feature selection methods can be used to remove variables that contribute no substantial information to the evaluation results; on the other hand, a combination of weighted and averaged methods can be used to integrate multidimensional inputs into a unified comprehensive index based on structural similarity. Furthermore, principal component analysis and fast t-distributed random neighborhood embedding techniques can be used to map the massive original feature space into a compact low-dimensional representation. By integrating principal component analysis and fast t-distributed random neighborhood embedding techniques, the DRBRB model can control the growth of the confidence rule base within a manageable range, ensuring computational feasibility while preserving the high-fidelity causal relationships required for accurate system evaluation. After dimensionality reduction, an initial extended permission evaluation model is constructed based on the confidence rule base, establishing confidence rules.

[0135] The modeling process in this step is as follows:

[0136] S41. Determine the feature set of extended permission evaluation indicators. ;

[0137] S42. Set of indicator features for extended permissions Dimensionality reduction is performed, and the resulting set of indicator features is: In this step, five joint dimensionality reduction strategies are used to reduce the dimensionality of the extended permissions metrics. These strategies include:

[0138]

[0139]

[0140]

[0141]

[0142]

[0143] in, This represents the set of index features after processing by their respective dimensionality reduction methods. , , , and These represent five dimensionality reduction functions: feature selection, weighted combination, average combination, principal component analysis, and fast t-SNE. Indicates the original dimensions of the evaluation indicators; The set of parameters representing the dimensionality reduction process.

[0144] S43. Construct confidence rules for the dimensionality-reduced permission indicators and assign initial confidence levels based on expert knowledge.

[0145] Step 5: Reason about the model based on evidence.

[0146] In the extended permission risk assessment model based on a dimensionality reduction-driven confidence rule base constructed in this invention patent, step 5 focuses on using evidence reasoning methods to reason about the model in order to obtain accurate risk assessment output results.

[0147] The DRBRB model within the PAF-DRBRB framework employed in this invention utilizes an evidence-based reasoning module to achieve rule fusion and decision inference during the reasoning phase. The core task of this module is to evaluate the consistency between the input sample and each confidence rule, thereby calculating the activation strength of the corresponding rule and laying the foundation for subsequent reasoning.

[0148] set up Indicates the first input sample The dimensionality reduction attribute and the first The degree of matching between the reference values ​​of the rule premises Indicates the first The first feature One reference point. The importance of rules should be considered comprehensively. and the contribution of each attribute Based on this, the first Activation weight of the rule Calculate using the following formula:

[0149]

[0150] Indicates the first The matching degree between input attributes and rule reference values ​​for each sample. This can be derived from the following formula:

[0151]

[0152] In the reasoning phase of DRBRB, an evidence-based reasoning method is used to fuse the activated rules, thereby obtaining the confidence level of each output level. Each output level represents the final overall risk level of the extended permissions. Let... The number of rules involved in the reasoning. Indicates the total number of possible output levels. The overall belief level corresponding to each output Calculated using the following formula:

[0153]

[0154]

[0155] Wherein, the normalization coefficient is .

[0156] Finally, the overall output of the system is obtained by weighted summing of the belief level of each output level and its corresponding utility value, as shown in the following formula:

[0157]

[0158]

[0159] in, This represents the expected utility value; This represents the confidence distribution of the output results; Indicates the reference level of the result; Indicates the confidence level of a reference grade for a specific result; Indicates the number of reference levels for the result.

[0160] Step 6: Optimize the parameters of the extended permission risk assessment model based on the dimensionality reduction-driven confidence rule base.

[0161] To ensure model accuracy, this invention incorporates a dedicated optimization module within the DRBRB framework. After comprehensive consideration, the Projection Covariance Matrix Adaptation Evolutionary Strategies (P-CMA-ES) algorithm was selected for parameter optimization. This choice is based on the superior performance of the P-CMA-ES algorithm in handling the non-convex and constrained parameter space of DRBRB. Unlike traditional gradient-based optimization algorithms, the P-CMA-ES algorithm integrates projection operators, strictly adhering to the inherent "sum of one" and non-negativity constraints of confidence levels, thus guaranteeing the validity of the inference results at the physical level. Therefore, this step utilizes P-CMA-ES to optimize the DRBRB model within the PAF-DRBRB framework to ensure model accuracy.

[0162] Step 7: Real-world experimental assessment of the risks of extended permissions.

[0163] As described in the six steps above, the PAF-DRBRB framework is established. Finally, the Chrome extension permission risk is assessed. This invention demonstrates the effectiveness of the DRBRB-based Chrome extension risk assessment framework PAF-DRBRB. DRBRB implements five BRB variants: feature selection-confidence rule base, weighted combination-confidence rule base, average combination-confidence rule base, principal component analysis-confidence rule base, and fast t-SNE-confidence rule base. Simultaneously, the permission risk level is defined as high, medium, and low.

[0164] Figure 2 and Figure 3 The figure shows the fitting results of the predicted and actual values ​​of the PAF-DRBRB method of the present invention. The figure compares the predicted and actual values ​​of the five BRB methods. Figure 4 The diagram shows the prediction index of the PAF-DRBRB results of this invention. This diagram is a performance ranking chart of the BRB method. Figure 5The error analysis results of the PAF-DRBRB method of this invention are shown in the figure, which is a histogram of the error distribution of five BRB methods. (a) shows the error distribution of the feature selection-confidence rule base; (b) shows the error distribution of the weighted combination-confidence rule base; (c) shows the error distribution of the average combination-confidence rule base; (d) shows the error distribution of the principal component analysis-confidence rule base; and (e) shows the error distribution of the fast t-SNE-confidence rule base. Experimental results show that the classification accuracy of all models exceeds 96%, with the feature selection-confidence rule base and the average combination-confidence rule base reaching the highest of 99.00%. Notably, the weighted combination-confidence rule base performs best in numerical accuracy, with a final MSE of only 0.000555, followed closely by the average combination method and the fast t-distribution random neighborhood embedding method. These results indicate that although different dimensionality reduction and combination strategies have slight differences in error convergence, they all maintain robust diagnostic capabilities in identifying Chrome extension permission risks.

Claims

1. A method for assessing Chrome extension permission risks based on a dimensionality-reduced confidence rule base, characterized in that... Includes the following steps: Step 1: Obtain the source data of the Chrome extension, including the CRX extension package and metadata; Step 2: Extract risk characteristic indicators of extended permissions to accurately capture the risks hidden in extended permissions; The extraction of extended permission risk characteristic indicators includes the following steps: S21. Extract over-authorization features to reflect the mismatch between extended functionality and permission requests; S22. Extract high-risk combined permission features to demonstrate the risks arising from interactions between different permissions; S23. Design and extract single-permission risk features to focus on the sensitive permissions themselves; Step 3: Quantify the extracted permission features and assign them realistic labels; by quantifying permission features from different dimensions, an accurate assessment of Chrome extension risks can be achieved, including the following steps: S31, Single-permission score generation; S32. Generation of a rating for abuse of privileges; S33, Generation of combined permission scores; S34. Generation of overall risk score; Step 4: Establish an extended permission risk assessment model based on a dimensionality reduction-driven confidence rule base; The modeling process of the extended permission risk assessment model is as follows: S41. Determine the feature set of extended permission evaluation indicators; S42. Perform dimensionality reduction on the feature set of extended permissions to obtain the dimensionality-reduced permission indicators. S43. Construct confidence rules for the dimensionality-reduced permission indicators and assign initial confidence levels based on expert knowledge; Step 5: Reason about the model based on evidence; The DRBRB model in the PAF-DRBRB framework is adopted, and the evidence-based reasoning module is used in the reasoning stage to realize rule fusion and decision inference. Step 6: Optimize the parameters of the extended permission risk assessment model based on the dimensionality reduction-driven confidence rule base; Step 7: Real-world experiment assessment of the risks associated with extended permissions; The effectiveness of PAF-DRBRB, a Chrome extension risk assessment framework based on DRBRB, is demonstrated. DRBRB implements five confidence rule base variants: feature selection-confidence rule base, weighted combination-confidence rule base, average combination-confidence rule base, principal component analysis-confidence rule base, and fast t-SNE-confidence rule base. At the same time, the permission risk level is set to three levels: high, medium, and low.

2. The method according to claim 1, characterized in that, Step S21 includes: S211. To extract over-granted permissions, you must first extract the permissions declared in the extended declaration. ; This indicates a specific permission in the extended declaration; S212, Next, based on the abstract syntax tree analysis, extract the permissions actually used for the extension. ; This indicates an extension of a specific permission that is actually being used; S213, then based on permissions and extended API benchmarks A comparison is made between the permissions declared in the extended statement and the permissions actually used. This is used to identify whether extensions have excessive or redundant permissions. Excessively authorized extensions are represented as follows: in, and These represent the extended declaration and the actual set of permissions used, respectively. This indicates a mapping from extended APIs to permissions; This represents the difference between the set of permissions declared in the extension declaration and the set of permissions actually used by the extension; This indicates a set of permissions that have been excessively granted. This indicates that the extension has excessive permissions; otherwise, it does not.

3. The method according to claim 1, characterized in that, Step S23 includes: conducting a preliminary assessment of the risk level of a single permission based on dimensions such as the degree of danger of the permission granting behavior, whether the permission will trigger a warning, and whether the permission directly involves user privacy, and initially assigning a risk level to the single permission; then, integrating the knowledge accumulated from the single permission assessment into the expert knowledge base; this process is described as follows: in: Indicates permissions The corresponding characteristics of the degree of danger of the authorized behavior; Indicates permissions Whether the alarm characteristics trigger a system security warning; Indicates permissions Does it directly involve privacy-related features of user privacy data? Indicates a single-authority evaluation set; Indicates the single-access assessment level; An evaluation function representing a single-privilege evaluation level; This represents a single-access risk fusion assessment function. For risk level mapping function; For expert knowledge base.

4. The method according to claim 1, characterized in that, Step S31, Single-permission score generation in progress: The single-permission scoring generator performs a security risk assessment on the permissions of the extended declaration described in claim 2, dividing the permissions into four predefined levels: critical, high, medium, and low, with each level corresponding to a specific weight. By performing a weighted average calculation on all permissions, a standardized score reflecting the overall risk contribution of a single permission is obtained, calculated using the following formula: in, Representing the The weight of each risk level, It refers to the number of permissions at that level. This indicates a single-access risk score.

5. The method according to claim 1, characterized in that, Step S32, Generating a score for abuse of privileges: The permission abuse rating generator identifies requests for extended permissions that exceed functional requirements. The evaluation results are presented in binary form: a score of 1 indicates unnecessary permissions, suggesting potential over-authorization or security risks; a score of 0 indicates that the extended permission request is legitimate. Mathematically, this can be expressed as: in, This represents the privilege abuse score or over-granting score. A value of 1 indicates over-granting or privilege abuse, while a value of 0 indicates no abuse.

6. The method according to claim 1, characterized in that, Step S33: Generating combined permission scores: The combined permission scoring generator focuses on assessing the potential risks associated with permission combinations. The formula for calculating the combined permission risk score is as follows: in, This represents the total number of rules for combining sensitive permissions. It is the first The weight of each rule, This indicates the degree to which the extension matches the rule. This indicates the risk score for combined permissions.

7. The method according to claim 1, characterized in that, Step S34, Overall risk score generation in progress: After obtaining the scores from steps S1-S3, the total risk score generator integrates the single-permission score, permission abuse score, and combined permission score to calculate the final risk score for each extension, and records the total risk score as follows: The calculation formula is: Finally, based on the 33rd and 67th percentiles of the final risk score, the risk level is further divided into three categories: low, medium, and high.

8. The method according to claim 1, characterized in that, In step S42: the set of indicator features after dimensionality reduction is Five joint dimensionality reduction strategies were employed to reduce the dimensionality of the extended permissions metrics. These strategies included: in, This represents the set of index features after processing by their respective dimensionality reduction methods. , , , and These represent five dimensionality reduction functions: feature selection, weighted combination, average combination, principal component analysis, and fast t-SNE. Indicates the original dimensions of the evaluation indicators; The set of parameters representing the dimensionality reduction process.

9. The method according to claim 1, characterized in that, In step S5: set up Indicates the first input sample The dimensionality reduction attribute and the first The degree of matching between the reference values ​​of the rule premises Indicates the first The first feature These are reference points, taking into account the importance of the rules. and the contribution of each attribute Based on this, the first Activation weight of the rule Calculate using the following formula: Indicates the first Each sample, matching degree This can be derived from the following formula: In the inference phase of DRBRB, an evidence-based reasoning method is used to fuse the activated rules, thereby obtaining the confidence levels for each output level. The number of rules involved in the reasoning. Indicates the total number of possible output levels, the first... The overall belief level corresponding to each output Calculated using the following formula: Wherein, the normalization coefficient is ; Finally, the overall output of the system is obtained by weighted summing of the belief level of each output level and its corresponding utility value, as shown in the following formula: ； in, This represents the expected utility value; This represents the confidence distribution of the output results; Indicates the reference level of the result; Indicates the confidence level of a reference grade for a specific result; Indicates the number of reference levels for the result.

Images