Personal information protection compliance audit detection method and system based on AI

Abstract

The invention provides an AI-based personal information protection compliance audit detection method and system, and the method comprises the steps: collecting personal information data from a plurality of heterogeneous data sources, and generating a high-quality data set through data cleaning and redundancy elimination processing; personal sensitive information is automatically identified and marked by using natural language processing and a deep learning model; establishing a data classification system, performing static or dynamic classification and tagging on the marked personal sensitive information, and performing authority control according to a classification result; a risk assessment model based on time sequence analysis is adopted to monitor the data access frequency, the storage position and the encryption state in real time, and the potential leakage risk is predicted; multi-party cooperative training is realized through a federated learning technology, and a global risk assessment model is generated, thereby ensuring that original data is not shared. According to the scheme, the backstepping of the individual identity is effectively prevented, the timeliness of risk prediction is improved, and the accuracy of predicting the potential leakage risk is improved on the premise of protecting privacy.

Description

Technical Field

[0001] The present invention relates to, and specifically to, an AI-based personal information protection compliance audit detection method and system. Background Art

[0002] With the rapid development of big data and artificial intelligence technologies, the collection, storage, and processing of personal information have been widely used in various fields, such as finance, healthcare, and e-commerce. However, the centralized management and cross-institutional sharing of large amounts of personal information data also pose significant risks of privacy breaches. Existing technologies primarily rely on traditional encryption methods and access control strategies, such as static encryption of data or fixed rule-based permission allocation. These methods are somewhat effective for static data or single scenarios, but their limitations are becoming increasingly apparent with the surge in data volume, the heterogeneity of data sources, and the dynamic changes in data usage scenarios.

[0003] Specifically, existing technologies have the following problems: First, traditional methods often lack efficient pre-processing and sensitive information identification capabilities when dealing with large-scale heterogeneous data, resulting in uneven data quality and affecting the accuracy of subsequent protection. Secondly, static classification and permission control mechanisms cannot adapt to changes in data access frequency, storage environment or regulatory requirements, making it difficult to achieve real-time risk monitoring and dynamic adjustment. In addition, in multi-party collaboration scenarios, existing technologies usually require sharing of original data for joint analysis, which increases the risk of data leakage. Even if technologies such as federated learning are used, there is a lack of effective means to quantify global risks. Existing risk assessment models are mostly based on a single time point or local data, and fail to fully combine the advantages of time series analysis and multi-party collaboration, limiting the accuracy and comprehensiveness of predicting potential leakage risks.

[0004] Therefore, there is an urgent need for a personal information protection method that can integrate big data processing, artificial intelligence recognition and privacy protection technologies to solve the privacy protection issues in heterogeneous data management, dynamic classification management and multi-party collaboration, while improving the real-time and accuracy of risk assessment. Summary of the Invention

[0005] In view of this, the present invention provides an AI-based personal information protection compliance audit detection method, which includes the following steps:

[0006] Collect personal information data from multiple heterogeneous data sources and generate high-quality data sets through data cleaning and de-redundancy processing;

[0007] Utilize natural language processing and deep learning models to automatically identify and label sensitive personal information in the high-quality datasets, including but not limited to national identification numbers, phone numbers, and biometric data;

[0008] Establish a data classification system to statically or dynamically classify and label marked personal sensitive information, and implement permission control based on the classification results;

[0009] Adopt a risk assessment model based on time series analysis to monitor data access frequency, storage location, and encryption status in real time to predict potential leakage risks;

[0010] Federated learning technology is used to achieve multi-party collaborative training and generate a global risk assessment model. This global risk assessment model ensures that the original data is not shared.

[0011] In particular, the step of generating a high-quality data set by data cleaning and redundancy removal includes: generating a high-quality data set by the following expression:

[0012]

[0013] D clean represents the high-quality dataset generated after cleaning and de-redundancy processing; S represents the total number of heterogeneous data sources; N s represents the number of data entries in the sth heterogeneous data source; d s,i represents the i-th original data in the s-th data source; f clean (d s,i ,θ s ) represents the cleaning function, which is affected by the parameter θ s Control, used to remove noise and outliers, where the parameter θ s represents the cleaning threshold of the sth data source; g dedup (d s,i ,H s ) represents the deduplication function, based on historical data H s Determine the historical data set of the s-th data source for deduplication; U represents the data set merging operation to ensure the integration of all data sources.

[0014] In particular, the step of automatically identifying the high-quality data set and marking personal sensitive information using natural language processing and deep learning models specifically includes: for each record in the high-quality data set, first using natural language processing technology to perform a preliminary analysis of the record content, and extracting a data set with sensitive information according to predefined regular expression rules as the result of preliminary identification; inputting the result of the preliminary identification into a deep learning model, further analyzing the sensitivity of each paragraph of text, and forming a data set marked with sensitive information.

[0015] In particular, the step of statically classifying and labeling the marked personal sensitive information specifically includes: statically classifying and labeling the sensitive data using the following expression:

[0016]

[0017] d r Indicates the rth data record; w NLP represents the weight of the natural language processing model, indicating its contribution to recognition; P NLP (d r ,λ) represents the probability of sensitive information in the natural language processing model, which is controlled by the parameter λ, w DL represents the weight of the deep learning model, indicating its contribution to recognition; P DL (d r ,φ) represents the probability of sensitive information of the deep learning model, which is controlled by the parameter φ and includes the internal parameter set of the deep learning model; T(d r ,ψ) represents the sensitive type classification function, which is controlled by the parameter ψ and includes the trained classifier parameters, d r Indicates the rth data record, S labeled Represents a dataset after sensitive information is marked.

[0018] In particular, the steps of dynamically classifying and labeling the marked personal sensitive information and performing permission control based on the classification results specifically include: for each record marked with sensitive information, the system first calculates the sensitivity score based on the content of its sensitive information and the characteristics at the current moment, and calculates the total score in combination with the compliance score; based on the total score, the system assigns the record to different risk categories and sets corresponding permission control for each record.

[0019] In particular, the step of performing permission control according to the classification result is implemented by the following expression:

[0020] Among them, C t represents the classification data set at time t, including high, medium, and low risk categories and permissions; R represents the total number of data records, s r Indicates the sensitive information marking result of the rth record, Q sens (s r ,t) represents the sensitivity scoring function based on the labeling result of sensitive information changing with time t; κ represents the weight of the sensitivity score; R comp (s r ,ζ t ) represents the dynamic classification compliance function, which is subject to the time-dependent parameter ξ t control, τ represents the weight of dynamic classification compliance; V(t,ξ t ) represents the dynamic update function, according to time t and parameter ξ t Adjust classification rules; t Represents the classification threshold or rule adjustment parameter at time t; through integral calculation, the classification is dynamically updated over time.

[0021] In particular, the risk assessment model based on time series analysis adds noise to data analysis through a differential privacy mechanism to prevent individual identity inference. The risk value is calculated using the following expression:

[0022]

[0023] Among them, Y t is the risk assessment value at time t, reflecting the prediction of sensitive data leakage at time t;

[0024] μ is the risk baseline, β i Y is the influence coefficient of historical data on current risk, which controls the influence of the data at the past i-th moment; t-i represents the risk value at time ti, reflecting the historical trend of the data; σ is the volatility coefficient, which controls the degree of volatility of the risk value; B(H,t) is the time-varying external influencing factor, and H is the external historical influencing data;

[0025] γ j is the periodic weight, δ k is the coefficient of the impact of external features on risk assessment, k represents the identifier of the external feature, and the external features include data access frequency, storage location, and encryption status; sin(2πjt / T) is a periodic function that reflects the periodic fluctuation of data leakage risk;

[0026] X k,t The external feature value represents the value of the kth external feature at time t; ε t Noise that represents differential privacy.

[0027] In particular, the steps for generating a global risk assessment model include: achieving multi-party collaborative training through federated learning, where each data party trains the model locally and generates encrypted model update parameters; and aggregating all encrypted parameters through a weighted aggregation algorithm to generate a global risk assessment model.

[0028] In particular, the global risk assessment model is implemented by the following expression:

[0029]

[0030] Among them, M 2 Represents the global risk value, obtained through multi-party collaborative training, w ai represents the data weight of the ai-th data owner, Δm ai represents the local model update amount of the ai-th data owner; n1 represents the total number of data parties participating in federated learning;

[0031] α ajThe weighted coefficient of the encryption model parameter of the ajth aggregation server owner reflects the sensitivity; Enc(m aj ) represents the encrypted model parameters of the ajth aggregation server owner; m1 represents

[0032] The total number of data parties involved in the encryption operation;

[0033] γ ak represents the weight of the akth third-party regulator in the communication process, Com(m ak ) represents the communication aggregation result, reflecting the model update integration; p1 represents the total number of data parties participating in the communication aggregation model update; ∈ represents the random error term, which is used to capture the unmodeled disturbance or information loss in the federated learning process.

[0034] In particular, the present invention also discloses an AI-based personal information protection compliance audit detection system, which includes:

[0035] The data collection and preprocessing module is used to collect personal information data from multiple heterogeneous data sources and generate high-quality data sets through data cleaning and redundancy removal;

[0036] A multi-dimensional sensitive information identification module, which uses natural language processing and deep learning models to automatically identify the high-quality dataset and label personal sensitive information, including but not limited to ID numbers, phone numbers, and biometric data;

[0037] The classification management module is used to establish a data classification system, statically or dynamically classify and label marked personal sensitive information, and perform permission control based on the classification results;

[0038] A local intelligent risk assessment module uses a risk assessment model based on time series analysis to monitor data access frequency, storage location, and encryption status in real time to predict potential leakage risks;

[0039] The global intelligent risk assessment module is used to achieve multi-party collaborative training through federated learning technology and generate a global risk assessment model. Through this global risk assessment model, it is ensured that the original data is not shared.

[0040] Beneficial effects:

[0041] Compared with the existing technology, the AI-based personal information protection compliance audit detection method of the present invention has achieved the following significant technical effects:

[0042] Efficient heterogeneous data processing capabilities

[0043] Through data collection and redundancy removal, this invention can extract personal information from multiple heterogeneous data sources and generate high-quality datasets through cleansing and redundancy removal. This process significantly improves data quality, reduces the impact of noise and duplicate data on subsequent sensitive information identification and classification, and lays a solid foundation for the entire protection and detection process.

[0044] This invention leverages multi-dimensional recognition technology that combines natural language processing and deep learning models to automatically and accurately identify and label various types of sensitive personal information, including but not limited to ID numbers, phone numbers, and biometric data. Compared to traditional single-rule matching methods, this invention demonstrates higher precision and recall when identifying complex patterns and context-sensitive sensitive information.

[0045] This invention establishes a data classification system that supports both static and dynamic classification of sensitive data. Static classification is suitable for initial processing or stabilization scenarios, providing efficient and consistent classification results. Dynamic classification, on the other hand, dynamically adjusts classification and access control by monitoring sensitivity changes and regulatory requirements in real time. This dual-mode classification mechanism significantly enhances the system's flexibility and adaptability, enabling it to cope with diverse application scenarios and changing external conditions.

[0046] In the local intelligent risk assessment step, this paper uses a risk assessment model based on time series analysis to monitor data access frequency, storage location, and encryption status in real time. It also incorporates differential privacy mechanisms to add noise, effectively preventing the reverse inference of individual identities. This approach not only improves the timeliness of risk prediction but also enhances the accuracy of predicting potential leaks while protecting privacy.

[0047] Through a global intelligent risk assessment process, this paper leverages federated learning technology to enable multi-party collaborative training, generating a global risk assessment model that quantifies the overall risk of collaboration without sharing raw data. The global risk value is calculated using a weighted sum of squares, innovatively reflecting the consistency of model updates, significantly reducing the risk of data leakage while improving the comprehensiveness and reliability of risk assessments.

[0048] This invention incorporates time-dependent parameters and a dynamic update mechanism into dynamic classification and risk assessment, enabling real-time adjustments to classification rules and compliance scores based on regulatory changes. This adaptive capability ensures the system consistently complies with the latest personal information protection regulations. It also enhances system robustness through error terms and noise mechanisms, mitigating the impact of external interference on protection effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is the AI-based personal information protection compliance audit detection method proposed in this invention;

[0050] Figure 2 This is the AI-based personal information protection compliance audit detection system proposed in this invention. DETAILED DESCRIPTION

[0051] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0052] The present invention provides an AI-based personal information protection compliance audit detection method, the method, such as Figure 1 As shown, the following steps are included:

[0053] Step S101: Collect personal information data from multiple heterogeneous data sources and generate a high-quality data set through data cleaning and redundancy removal. The data collection and preprocessing steps generate a high-quality data set using the following formula:

[0054]

[0055] D clean represents the high-quality dataset generated after cleaning and de-redundancy processing; S represents the total number of heterogeneous data sources; N s represents the number of data entries in the sth heterogeneous data source; d s,i represents the i-th original data in the s-th data source; f clean (d s,i ,θ s ) represents the cleaning function, which is affected by the parameter θ s Control, used to remove noise and outliers, where the parameter θ s represents the cleaning threshold of the sth data source; g dedup (d s,i ,H s ) represents the deduplication function, based on historical data H s Determine the historical data set of the sth data source for deduplication; U represents the data set merging operation to ensure the integration of all data sources. In this embodiment, f clean (d s,i ,θ s ) can be specifically:

[0056]

[0057] In this embodiment, based on the linear function of the threshold, if the data value d s,i and mean μ s The deviation is less than the threshold θ s , then retain it, otherwise remove it.

[0058] or f clean (d s,i ,θ s ) can be specifically a statistical function based on the Z score, which is in the form of:

[0059]

[0060] Use standard deviation σ s Determine outliers, θ s Typically set to 2 or 3.

[0061] g dedup (d s,i ,H s ) represents the deduplication function, which can be in the form of:

[0062]

[0063] Step S102: Utilize natural language processing and deep learning models to automatically identify and label personal sensitive information, including but not limited to ID number, phone number, and biometric data.

[0064] In this embodiment, for each record in the data set, natural language processing technology is first used to perform a preliminary analysis of the record content, and a data set containing sensitive information is extracted according to predefined regular expression rules, so as to determine which parts may be sensitive information and use them as the results of preliminary identification; the results of the preliminary identification are input into a deep learning model to further analyze the sensitivity of each paragraph of text to form a data set marked with sensitive information.

[0065] Step S103: Establish a data classification system, statically or dynamically classify and label sensitive data, and perform permission control based on the classification results.

[0066] Reasons and scenarios for choosing static classification: Static classification is suitable for situations where data characteristics are relatively stable or during the initial processing phase. For example, when data first enters the system, it is necessary to quickly assign sensitivity levels (such as high, medium, and low risk) without considering changes over time or real-time updates.

[0067] Technical Advantages: Computational simplicity: One-time classification based on fixed rules or models, resulting in low resource consumption. Strong consistency: Results remain stable over time, making them suitable for long-term storage or auditing. Historical data can be archived, or fixed permissions (such as read-only) can be assigned to sensitive information, provided regulatory requirements remain unchanged.

[0068] Reasons and scenarios for choosing dynamic classification: Dynamic classification is suitable for scenarios where data risks change over time or due to external conditions (such as regulations and access frequency). For example, real-time monitoring of sensitive data status ensures that classification reflects the latest risks.

[0069] Technical Advantages: Real-time: Ability to quickly respond to environmental changes, such as the implementation of new regulations or changes in data access patterns. Adaptability: Dynamically adjust classification rules to keep the system aligned with actual needs. Application Example: Dynamically adjust the risk category and permissions of sensitive data (e.g., restrict high-frequency access to data) when data is frequently accessed or regulations are updated.

[0070] The static classification process specifically includes:

[0071] Establish a data classification system to statically classify and label marked personal sensitive information, and perform permission control based on the classification results. Specifically, static classification and labeling of sensitive data can be achieved through the following expressions:

[0072]

[0073] d r Indicates the rth data record; w NLP represents the weight of the natural language processing model, indicating its contribution to recognition; P NLP (d r ,λ) represents the probability of sensitive information in the natural language processing model, which is controlled by the parameter λ, w DL represents the weight of the deep learning model, indicating its contribution to recognition; P DL (d r ,φ) represents the probability of sensitive information of the deep learning model, which is controlled by the parameter φ, including the set of internal parameters of the deep learning model, such as the weights, biases or other hyperparameters of the neural network (such as learning rate, number of layers).

[0074] T(d r ,ψ) represents the sensitive type classification function, which is controlled by the parameter ψ and includes the trained classifier parameters, such as the branch conditions of the decision tree. In static classification, the parameter ψ is fixed to ensure consistent classification results.

[0075] d r Indicates the rth data record, S labeled Represents a dataset after sensitive information is marked;

[0076] P NLP (d r ,λ) represents the natural language processing model, which can be a model based on conditional random field (CRF)

[0077]

[0078] T(d r ,ψ) represents the sensitive type classification function, which is controlled by the parameter ψ. The sensitive type classification function,

[0079]

[0080] ψ is the rule set, score measures the matching degree, T(d r ,ψ) returns the type with the highest score. Further, permission control can be performed based on the classification results.

[0081] In this embodiment, sensitive data is dynamically classified and labeled, and permission control is performed based on the classification results. Specifically, for each record marked with sensitive information, the system first calculates a sensitivity score based on the content of its sensitive information and the characteristics at the current moment, and calculates a total score in combination with the compliance score; based on the total score, the system assigns the record to different risk categories and sets corresponding permission controls for each record. In order to ensure the real-time and adaptability of the classification, the system dynamically adjusts the classification rules. This adjustment is based on the data characteristics at the current time point and external inputs (such as regulatory changes), thereby updating the classification standards or thresholds. Over time, the system continuously repeats the above process, cumulatively updates the classification and permission status of each record, and ultimately forms a dynamically evolving classification data set that reflects the latest risk status and permission configuration of the data.

[0082] The real-time dynamic classification management step implements classification and authority control through the following formula:

[0083]

[0084] Among them, C t represents the classification data set at time t, including high, medium, and low risk categories and permissions; R represents the total number of data records, s r Indicates the sensitive information marking result of the rth record, Q sens (s r ,t) represents the sensitivity scoring function based on the change of the labeling result of sensitive information over time t. In this embodiment, the sensitive information types include ID card number, phone number, biometric data (such as "fingerprint data"), or other sensitive information (such as address, email address, etc.); κ represents the weight of the sensitivity score; in this embodiment,

[0085]

[0086] Among them, η represents the global adjustment coefficient of the sensitivity score, which controls the overall score amplitude, and v field The dynamic weight of the sensitive field field indicates its relative importance. ρ indicates the growth rate of the sensitivity index, which controls the speed of the sensitivity value over time. sens(field) indicates the inherent sensitivity value of the field field, which is predetermined based on the type. This allows the system to evaluate the sensitivity of different sensitive information in the record separately. fieldand sens(field) assign weights and sensitivity values to each field, refining the granularity of dynamic classification. ficld represents the time offset factor of the field field, represents the initial delay of sensitivity, ω represents the influence coefficient of access frequency, freq(s r ,t) Adjust the contribution of frequency to sensitivity and record s r The access frequency at time t, dynamic statistics, σ t Represents the time decay adjustment parameter, which controls the speed at which the frequency effect changes over time. comp (s r ,ζ t ) represents the dynamic classification compliance function, which is subject to the time-dependent parameter ξ t control, τ represents the weight of dynamic classification compliance.

[0087] In this embodiment,

[0088]

[0089] θ c represents the global scaling factor of the compliance score, which is used to adjust the score range, κ r Indicates the sensitivity coefficient of the rule-th regulation, which controls the steepness of compliance changes. r ,rule) records r The compliance score of rule rule, π rule Indicates the time impact factor of the rule, indicating the change in the importance of the rule over time t. c Represents the weight coefficient of the penalty item, which adjusts the contribution of non-compliance penalty to the score. r ,t,ζ t ) indicates record s r At time t based on ζ t Non-compliance penalty value.

[0090] V(t,ξ t ) represents the dynamic update function, according to time t and parameter ξ t Adjust classification rules; t Represents the classification threshold or rule adjustment parameter at time t; through integral calculation, the classification is dynamically updated over time.

[0091] Step S104: Using a risk assessment model based on time series analysis, real-time monitoring of data access frequency, storage location, and encryption status is performed to predict potential leakage risks. The risk assessment model based on time series analysis in the local intelligent risk assessment step uses a differential privacy mechanism to add noise to data analysis to prevent individual identity inference. Specifically, the risk value is calculated as follows:

[0092]

[0093] Among them, Y t is the risk assessment value at time point t, reflecting the prediction of sensitive data leakage at time t.

[0094] μ is the risk baseline, β i Y is the influence coefficient of historical data on current risk, which controls the influence of the data at the past i-th moment. t-i The risk value at time point ti reflects the historical trend of the data, σ is the volatility coefficient, which controls the degree of volatility of the risk value; B(H,t) is the time-varying external influencing factor, H is the external historical influencing data,

[0095] γ j is the periodic weight, δ k is the coefficient of the impact of external features on risk assessment, k represents the identifier of the external features, which include data access frequency, storage location, and encryption status; sin(2πjt / T) is a periodic function that reflects the periodic fluctuation of data leakage risk.

[0096] X k,t The external feature value represents the value of the kth external feature at time point t; ε t Noise that represents differential privacy.

[0097] Step S105: Federated learning technology is used to implement multi-party collaborative training to generate a global risk assessment model. This global risk assessment model ensures that the original data is not shared. Federated learning enables multi-party collaborative training. Each data source trains the model locally and generates encrypted model update parameters. A weighted aggregation algorithm aggregates all encrypted parameters to generate a global risk assessment model.

[0098] The global risk assessment model specifically includes:

[0099]

[0100] Among them, M 2 Represents the global risk value, obtained through multi-party collaborative training, w ai represents the data weight of the ai-th data owner, Δm ai represents the local model update amount of the ai-th data owner; n1 represents the total number of data parties participating in federated learning;

[0101] α aj The weighted coefficient of the encryption model parameter of the ajth aggregation server owner reflects the sensitivity; Enc(m aj ) represents the model parameters encrypted by the owner of the ajth aggregation server; m1 represents the total number of data parties participating in the encryption operation;

[0102] γ ak represents the weight of the akth third-party regulator in the communication process, Com(m ak ) represents the communication aggregation result, reflecting the model update integration; p1 represents the total number of data parties participating in the communication aggregation model update; ∈ represents the random error term, which is used to capture the unmodeled disturbance or information loss in the federated learning process.

[0103] The present invention also discloses an AI-based personal information protection compliance audit detection system, which includes:

[0104] The data collection and preprocessing module is used to collect personal information data from multiple heterogeneous data sources and generate high-quality data sets through data cleaning and redundancy removal;

[0105] Personal information data is collected from multiple heterogeneous data sources, and a high-quality data set is generated through data cleaning and de-redundancy processing. The data cleaning and de-redundancy processing steps generate a high-quality data set using the following formula:

[0106]

[0107] D clean represents the high-quality dataset generated after cleaning and de-redundancy processing; S represents the total number of heterogeneous data sources; N s represents the number of data entries in the sth heterogeneous data source; d s,i represents the i-th original data in the s-th data source; f clean (d s,i ,θ s ) represents the cleaning function, which is affected by the parameter θ s Control, used to remove noise and outliers, where the parameter θ s represents the cleaning threshold of the sth data source; g dedup (d s,i ,H s ) represents the deduplication function, based on historical data H s Determine the historical data set of the sth data source for deduplication; U represents the data set merging operation to ensure the integration of all data sources. In this embodiment, f clean (d s,i ,θ s ) can be specifically:

[0108]

[0109] A linear function based on a threshold, where if the data value d s,i and mean μ s The deviation is less than the threshold θ s , then retain it, otherwise remove it.

[0110] or f clean (d s,i ,θ s ) can be specifically a statistical function based on the Z score, which is in the form of:

[0111]

[0112] Use standard deviation σ s Determine outliers, θ s Usually set to 2 or 3

[0113] g dedup (d s,i ,H s ) represents the deduplication function, which can be in the form of:

[0114]

[0115] Multi-dimensional sensitive information identification module: Utilizes natural language processing and deep learning models to automatically identify and label personal sensitive information, including but not limited to ID numbers, phone numbers, and biometric data; for each record in the dataset, natural language processing technology is first used to perform a preliminary analysis of the record content, and predefined regular expression rules are used to determine which parts may be sensitive information as the result of preliminary identification; the results of the preliminary identification are input into a deep learning model to further analyze the sensitivity of each paragraph of text to form a dataset marked with sensitive information.

[0116] Classification management module: Establish a data classification system, statically or dynamically classify and label sensitive data, and perform permission control based on the classification results.

[0117] Reasons and scenarios for choosing static classification: Static classification is suitable for situations where data characteristics are relatively stable or during the initial processing phase. For example, when data first enters the system, it is necessary to quickly assign sensitivity levels (such as high, medium, and low risk) without considering changes over time or real-time updates.

[0118] Technical Advantages: Computational simplicity: One-time classification based on fixed rules or models, resulting in low resource consumption. Strong consistency: Results remain stable over time, making them suitable for long-term storage or auditing. Historical data can be archived, or fixed permissions (such as read-only) can be assigned to sensitive information, provided regulatory requirements remain unchanged.

[0119] Reasons and scenarios for choosing dynamic classification: Dynamic classification is suitable for scenarios where data risks change over time or due to external conditions (such as regulations and access frequency). For example, real-time monitoring of sensitive data status ensures that classification reflects the latest risks.

[0120] Technical Advantages: Real-time: Ability to quickly respond to environmental changes, such as the implementation of new regulations or changes in data access patterns. Adaptability: Dynamically adjust classification rules to keep the system aligned with actual needs. Application Example: Dynamically adjust the risk category and permissions of sensitive data (e.g., restrict high-frequency access to data) when data is frequently accessed or regulations are updated.

[0121] The static classification management process specifically includes:

[0122] Establish a data classification system, statically classify and label sensitive data, and implement permission control based on the classification results. Specifically, static classification and labeling of sensitive data can be achieved through the following expressions:

[0123]

[0124] d r Indicates the rth data record; w NLP represents the weight of the natural language processing model, indicating its contribution to recognition; P NLP (d r ,λ) represents the probability of sensitive information in the natural language processing model, which is controlled by the parameter λ, w DL represents the weight of the deep learning model, indicating its contribution to recognition; P DL (d r ,φ) represents the probability of sensitive information of the deep learning model, which is controlled by the parameter φ, including the set of internal parameters of the deep learning model, such as the weights, biases or other hyperparameters of the neural network (such as learning rate, number of layers).

[0125] T(d r ,ψ) represents the sensitive type classification function, which is controlled by the parameter ψ and includes the trained classifier parameters, such as the branch conditions of the decision tree. In static classification, the parameter ψ is fixed to ensure consistent classification results.

[0126] d r Indicates the rth data record, S labeled Represents a dataset after sensitive information is marked;

[0127] P NLP (d r ,λ) represents the natural language processing model, which can be a model based on conditional random fields (CRF):

[0128]

[0129] T(d r ,ψ) represents the sensitive type classification function, which is controlled by the parameter ψ. The sensitive type classification function,

[0130]

[0131] ψ is the rule set, score measures the matching degree, T(d r ,ψ) returns the type with the highest score.

[0132] Furthermore, permission control can be performed based on the classification results.

[0133] In this embodiment, sensitive data is dynamically classified and labeled, and permission control is performed based on the classification results. Specifically, for each record marked with sensitive information, the system first calculates a sensitivity score based on the content of its sensitive information and the characteristics at the current moment, and calculates a total score in combination with the compliance score; based on the total score, the system assigns the record to different risk categories and sets corresponding permission controls for each record. In order to ensure the real-time and adaptability of the classification, the system dynamically adjusts the classification rules. This adjustment is based on the data characteristics at the current time point and external inputs (such as regulatory changes), thereby updating the classification standards or thresholds. Over time, the system continuously repeats the above process, cumulatively updates the classification and permission status of each record, and ultimately forms a dynamically evolving classification data set that reflects the latest risk status and permission configuration of the data.

[0134] The real-time dynamic classification management implements classification and authority control through the following formula:

[0135]

[0136] Among them, C t represents the classification data set at time t, including high, medium, and low risk categories and permissions; R represents the total number of data records, s r Indicates the sensitive information marking result of the rth record, Q sens (s r ,t) represents the sensitivity scoring function based on the change of the labeling result of sensitive information over time t. In this embodiment, the sensitive information types include ID card number, phone number, biometric data (such as "fingerprint data"), or other sensitive information (such as address, email address, etc.); κ represents the weight of the sensitivity score; in this embodiment,

[0137]

[0138] Among them, η represents the global adjustment coefficient of the sensitivity score, which controls the overall score amplitude, and ν field Indicates the dynamic weight of the sensitive field field, indicating its relative importance, ρ represents the sensitivity exponential growth rate, which controls the speed of the sensitivity value over time, and sens(field) represents the inherent sensitivity value of the field field, which is predetermined based on the type. This allows the system to evaluate the sensitivity of different sensitive information in the record separately, through ν fieldand sens(field) assign weights and sensitivity values to each field, refining the granularity of dynamic classification. ficld represents the time offset factor of the field field, represents the initial delay of sensitivity, ω represents the influence coefficient of access frequency, freq(s r ,t) Adjust the contribution of frequency to sensitivity and record s r Access frequency at time t, dynamic statistics, σ t Represents the time decay adjustment parameter, which controls the speed at which the frequency effect changes over time. comp (s r ,ζ t ) represents the dynamic classification compliance function, which is subject to the time-dependent parameter ξ t Control, τ represents the weight of dynamic classification compliance. In this embodiment,

[0139]

[0140] θ c represents the global scaling factor of the compliance score, which is used to adjust the score range, κ r Indicates the sensitivity coefficient of the rule-th regulation, which controls the steepness of compliance changes. r ,rule) records r The compliance score of rule rule, π rule Indicates the time impact factor of the rule, indicating the change in the importance of the rule over time t. c Represents the weight coefficient of the penalty item, which adjusts the contribution of non-compliance penalty to the score. r ,t,ζ t ) indicates record s r At time t based on ζ t The non-compliance penalty value,

[0141] V(t,ξ t ) represents the dynamic update function, according to time t and parameter ξ t Adjust classification rules; t Represents the classification threshold or rule adjustment parameter at time t; through integral calculation, the classification is dynamically updated over time.

[0142] Local Intelligent Risk Assessment Module: This module uses a risk assessment model based on time series analysis to monitor data access frequency, storage location, and encryption status in real time, predicting potential leakage risks. In this module, the time series analysis-based risk assessment model uses a differential privacy mechanism to add noise to data analysis to prevent individual identity inference. Specifically, the risk value is calculated as follows:

[0143]

[0144] Among them, Y t is the risk assessment value at time point t, reflecting the prediction of sensitive data leakage at time t.

[0145] μ is the risk baseline, β i Y is the influence coefficient of historical data on current risk, which controls the influence of the data at the past i-th moment. t-i The risk value at time point ti reflects the historical trend of the data, σ is the volatility coefficient, which controls the degree of volatility of the risk value; B(H,t) is the time-varying external influencing factor, H is the external historical influencing data,

[0146] γ j is the periodic weight, δ k is the coefficient of the impact of external features on risk assessment, k represents the identifier of the external features, which include data access frequency, storage location, and encryption status; sin(2πjt / T) is a periodic function that reflects the periodic fluctuation of data leakage risk.

[0147] X k,t The external feature value represents the value of the kth external feature at time point t; ε t Noise that represents differential privacy.

[0148] The global intelligent risk assessment module uses federated learning technology to achieve multi-party collaborative training and generate a global risk assessment model. This global risk assessment model ensures that the original data is not shared. Federated learning enables multi-party collaborative training. Each data provider trains the model locally and generates encrypted model update parameters. A weighted aggregation algorithm aggregates all encrypted parameters to generate a global risk assessment model.

[0149] The global risk assessment model specifically includes:

[0150]

[0151] Among them, M 2 Represents the global risk value, obtained through multi-party collaborative training, w ai represents the data weight of the ai-th data owner, Δm ai represents the local model update amount of the ai-th data owner; n1 represents the total number of data parties participating in federated learning;

[0152] α aj The weighted coefficient of the encryption model parameter of the ajth aggregation server owner reflects the sensitivity, Enc(m aj ) represents the encrypted model parameters of the ajth aggregation server owner; m1 represents

[0153] The total number of data parties involved in the encryption operation;

[0154] γ ak represents the weight of the akth third-party regulator in the communication process, Com(m ak ) represents the communication aggregation result, reflecting the model update integration, p1 represents the total number of data parties participating in the communication aggregation model update; ∈ represents the random error term, which is used to capture the unmodeled disturbance or information loss in the federated learning process.

[0155] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0156] It is obvious to those skilled in the art that the embodiments of the present invention are not limited to the details of the above-mentioned exemplary embodiments, and that the embodiments of the present invention can be implemented in other specific forms without departing from the spirit or essential features of the embodiments of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the embodiments of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes that fall within the meaning and scope of the equivalent elements of the claims are included in the embodiments of the present invention. Any figure marks in the claims should not be regarded as limiting the claims involved. In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units, modules or devices stated in the system, device or terminal claims may also be implemented by the same unit, module or device through software or hardware. Words such as first and second are used to indicate names and do not indicate any particular order.

[0157] Finally, it should be noted that the above implementation methods are only used to illustrate the technical solutions of the embodiments of the present invention and are not limiting. Although the embodiments of the present invention are described in detail with reference to the above preferred implementation methods, ordinary technicians in this field should understand that the technical solutions of the embodiments of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based personal information protection compliance audit detection method, characterized in that: The method comprises the following steps: Collect personal information data from multiple heterogeneous data sources and generate high-quality data sets through data cleaning and de-redundancy processing; Utilize natural language processing and deep learning models to automatically identify the high-quality datasets and label sensitive personal information, including but not limited to national identification numbers, phone numbers, and biometric data; Establish a data classification system to statically or dynamically classify and label marked personal sensitive information, and implement permission control based on the classification results; Adopt a risk assessment model based on time series analysis to monitor data access frequency, storage location, and encryption status in real time to predict potential leakage risks; Federated learning technology is used to achieve multi-party collaborative training and generate a global risk assessment model. This global risk assessment model ensures that the original data is not shared.

2. The method according to claim 1, characterized in that The step of generating a high-quality data set by data cleaning and redundancy removal includes generating a high-quality data set by the following expression: D clean represents the high-quality dataset generated after cleaning and de-redundancy processing; S represents the total number of heterogeneous data sources; N s represents the number of data entries in the sth heterogeneous data source; d s,i represents the i-th original data in the s-th data source; f clean (d s,i ,θ s ) represents the cleaning function, which is affected by the parameter θ s Control, used to remove noise and outliers, where the parameter θ s represents the cleaning threshold of the sth data source; g dedup (d s,i ,H s ) represents the deduplication function, based on historical data H s Determine the historical data set of the s-th data source for deduplication; U represents the data set merging operation to ensure the integration of all data sources.

3. The method according to any one of claims 1-2, characterized in that The steps of automatically identifying the high-quality data set and marking personal sensitive information using natural language processing and a deep learning model specifically include: for each record in the high-quality data set, first using natural language processing technology to perform a preliminary analysis of the record content, and extracting a data set containing sensitive information according to predefined regular expression rules as the result of preliminary identification; inputting the result of the preliminary identification into a deep learning model, further analyzing the sensitivity of each paragraph of text, and forming a data set marked with sensitive information.

4. The method according to any one of claims 1 to 3, characterized in that The steps for statically classifying and labeling the marked personal sensitive information specifically include: statically classifying and labeling the sensitive data using the following expressions: d r Indicates the rth data record; w NLP represents the weight of the natural language processing model, indicating its contribution to recognition; P NLP (d r ,λ) represents the probability of sensitive information in the natural language processing model, which is controlled by the parameter λ, w DL represents the weight of the deep learning model, indicating its contribution to recognition; P DL (d r ,φ) represents the probability of sensitive information of the deep learning model, which is controlled by the parameter φ and includes the internal parameter set of the deep learning model; T(d r ,ψ) represents the sensitive type classification function, which is controlled by the parameter ψ and includes the trained classifier parameters, d r Indicates the rth data record, S labeled Represents a dataset after sensitive information is marked.

5. The method according to any one of claims 1 to 3, characterized in that The steps of dynamically classifying and labeling marked personal sensitive information and performing permission control based on the classification results specifically include: for each record marked with sensitive information, the system first calculates a sensitivity score based on the content of its sensitive information and the characteristics at the current moment, and calculates a total score in combination with the compliance score; based on the total score, the system assigns the record to different risk categories and sets corresponding permission controls for each record.

6. The method according to claim 5, characterized in that The step of performing permission control based on the classification results is implemented by the following expression: Among them, C t represents the classification data set at time t, including high, medium, and low risk categories and permissions; R represents the total number of data records, s r Indicates the sensitive information marking result of the rth record, Q sens (s r ,t) represents the sensitivity scoring function based on the labeling result of sensitive information changing with time t; κ represents the weight of the sensitivity score; R comp (s r ,ζ t ) represents the dynamic classification compliance function, which is subject to the time-dependent parameter ξ t control, τ represents the weight of dynamic classification compliance; V(t,ξ t ) represents the dynamic update function, according to time t and parameter ξ t Adjust classification rules; t Represents the classification threshold or rule adjustment parameter at time t; through integral calculation, the classification is dynamically updated over time.

7. The method according to claim 1, characterized in that The risk assessment model based on time series analysis adds noise to data analysis through the differential privacy mechanism to prevent the reverse inference of individual identities. The risk value is calculated using the following expression: Among them, Y t is the risk assessment value at time t, reflecting the prediction of sensitive data leakage at time t; μ is the risk baseline, β i Y is the influence coefficient of historical data on current risk, which controls the influence of the data at the past i-th moment; t-i represents the risk value at time ti, reflecting the historical trend of the data; σ is the volatility coefficient, which controls the degree of volatility of the risk value; B(H,t) is the time-varying external influencing factor, and H is the external historical influencing data; γ j is the periodic weight, δ k is the coefficient of the impact of external features on risk assessment, k represents the identifier of the external feature, and the external features include data access frequency, storage location, and encryption status; sin(2πjt / T) is a periodic function that reflects the periodic fluctuation of data leakage risk; X k,t The external feature value represents the value of the kth external feature at time t; ε t Noise that represents differential privacy.

8. The method according to any one of claims 1 to 7, characterized in that The steps for generating a global risk assessment model include: achieving multi-party collaborative training through federated learning, where each data party trains the model locally and generates encrypted model update parameters; and aggregating all encrypted parameters through a weighted aggregation algorithm to generate a global risk assessment model.

9. The method according to claim 8, characterized in that The global risk assessment model is implemented by the following expression: Among them, M 2 Represents the global risk value, obtained through multi-party collaborative training, w ai represents the data weight of the ai-th data owner, Δm ai represents the local model update amount of the ai-th data owner; n1 represents the total number of data parties participating in federated learning; α aj The weighted coefficient of the encryption model parameter of the ajth aggregation server owner reflects the sensitivity; Enc(m aj ) represents the model parameters encrypted by the owner of the ajth aggregation server; m1 represents the total number of data parties participating in the encryption operation; γ ak represents the weight of the akth third-party regulator in the communication process, Com(m ak ) represents the communication aggregation result, reflecting the model update integration; p1 represents the total number of data parties participating in the communication aggregation model update; ∈ represents the random error term, which is used to capture the unmodeled disturbance or information loss in the federated learning process.

10. An AI-based personal information protection compliance audit detection system, characterized by: The system includes: The data collection and preprocessing module is used to collect personal information data from multiple heterogeneous data sources and generate high-quality data sets through data cleaning and redundancy removal; A multi-dimensional sensitive information identification module, which uses natural language processing and deep learning models to automatically identify the high-quality dataset and label personal sensitive information, including but not limited to ID numbers, phone numbers, and biometric data; The classification management module is used to establish a data classification system, statically or dynamically classify and label marked personal sensitive information, and perform permission control based on the classification results; Local intelligent risk assessment module, which is used to adopt risk assessment model based on time series analysis, Monitor data access frequency, storage location, and encryption status in real time to predict potential leakage risks; The global intelligent risk assessment module is used to achieve multi-party collaborative training through federated learning technology and generate a global risk assessment model. Through this global risk assessment model, it is ensured that the original data is not shared.

Images