Personal Data Field Detection System

Abstract

A personal data field detection system is applicable to a database server and includes: a storage unit and a processing unit electrically connected to the database server and the storage unit; the processing unit is configured to: receive multiple field data from the database server, each field data including a field name, a field annotation and multiple field values; for each field data, generate preliminary judgment data based on multiple hit keywords, multiple exclusion keywords, multiple format rules, and the field data; input the preliminary judgment data into a large language model to generate a preliminary judgment confidence level; select the field data with higher priority as multiple designated field data based on the preliminary judgment confidence levels; and for each designated field data, generate advanced judgment data based on the large language model.

Claims

1. A personal data field detection system, suitable for communicating with a database server, and comprising: a processing unit electrically connected to the database server; and a storage unit electrically connected to the processing unit, storing multiple sensitive data types, multiple matching keywords corresponding to each of the sensitive data types, multiple exclusion keywords corresponding to each of the sensitive data types, multiple format rules corresponding to the sensitive data types, and a large language model; the processing unit receives multiple field data from the database server, wherein, Each field of data includes a field name, a field annotation, and multiple field values. For each field of data, the processing unit performs a preliminary judgment procedure based on the hit keywords, the excluded keywords, the format rules, the field name, the field annotation, and the field values ​​to generate preliminary judgment data corresponding to that field of data. For each preliminary judgment data, the processing unit inputs the preliminary judgment data into the large language model to generate a preliminary judgment confidence level corresponding to the preliminary judgment data, wherein the preliminary judgment confidence level indicates the accuracy of the large language model in judging the preliminary judgment data. The processing unit sorts the field data corresponding to the preliminary judgment data with the respective preliminary judgment confidence levels according to the preliminary judgment confidence levels, and selects the field data with higher ranking as multiple designated field data. For each specified field of data, the processing unit inputs the hit keywords, the excluded keywords, the format rules, the field name, and the field value into the large language model to generate advanced judgment data corresponding to the specified field of data.

2. The personal data field detection system as described in claim 1, wherein, The storage unit also stores multiple sensitive asset scores corresponding to the sensitive asset types. The initial judgment data includes an initial judgment result, an initial type judgment result, keyword judgment data, and format rule judgment data. For each field of data, when the processing unit executes the initial judgment procedure, the processing unit determines whether the field name matches one of the excluded keywords. For each field of data, when the processing unit determines that the field name does not match any of the excluded keywords, it determines whether the field name matches one of the hit keywords. For each field of data, when the processing unit determines that the field name matches one of the hit keywords, it generates keyword judgment data including the hit keyword and uses the sensitive asset score corresponding to the sensitive asset type of the hit keyword as a field name score. For each field of data, the processing unit determines whether the field annotation contains one of the hit keywords and whether it contains a specific tag. For each field of data, when the processing unit determines that the field annotation contains one of the hit keywords and that the field annotation contains the specific mark, the sensitive resource score corresponding to the sensitive resource type of the hit keyword is taken as a field annotation score; the processing unit takes multiple field values ​​as multiple sample values, and for each sample value, when the processing unit determines that the sample value conforms to one of the format rules, the sensitive resource type corresponding to the conforming format rule is taken as a filtered sensitive resource type; the processing unit generates format rule judgment data according to the format rules corresponding to the filtered sensitive resource types; the processing unit generates a field value score according to the filtered sensitive resource types and the number of sample values; the processing unit calculates a field comprehensive score corresponding to the field data based on the field name score, the field annotation score, and the field value score; when the processing unit determines that the field comprehensive score is greater than or equal to a comprehensive score threshold, it generates an initial judgment result indicating that it is a sensitive resource. The processing unit determines whether the score of the field value is greater than or equal to a field value score threshold; when the processing unit determines that the score of the field value is greater than or equal to the field value score threshold, it generates the initial type judgment result based on the sensitive data type of the selected sensitive data; when the processing unit determines that the score of the field value is less than the field value score threshold, it generates the initial type judgment result based on the sensitive data type corresponding to the hit keyword.

3. The personal data field detection system as described in claim 2, wherein, For each initial judgment data, when the processing unit inputs the initial judgment data into the large language model, the large language model determines whether the keyword judgment data contains one of the hit keywords and whether the format rule judgment data contains one of the format rules. When the processing unit causes the large language model to determine that the keyword judgment data contains one of the hit keywords, or determines that the format rule judgment data contains one of the format rules, an initial judgment confidence level indicating a medium level of confidence is generated. When the processing unit causes the large language model to determine that the keyword judgment data does not contain one of the hit keywords and determines that the format rule judgment data does not contain one of the format rules, an initial judgment confidence level indicating a low level of confidence is generated. When the processing unit causes the large language model to determine that the keyword judgment data contains one of the hit keywords, and determines that the format rule judgment data contains one of the format rules, the processing unit causes the large language model to determine whether the sensitive data type corresponding to the hit keyword in the keyword judgment data matches the selected sensitive data type; when the processing unit causes the large language model to determine that the sensitive data type corresponding to the hit keyword in the keyword judgment data matches the selected sensitive data type, the initial judgment confidence level indicating a high confidence level is generated.

4. The personal data field detection system as described in claim 2, wherein, When the processing unit generates the field value score based on the sensitive data types and the number of sampled values, it counts the number of sensitive data types; the processing unit divides the number of sensitive data types by the number of sampled values ​​to calculate the field value score.

5. The personal data field detection system as described in claim 1 is also suitable for communicating with a terminal device, wherein, The processing unit is electrically connected to the terminal device; the processing unit transmits the advanced judgment data corresponding to the specified field data to the terminal device; for the specified field data, the processing unit receives from the terminal device a plurality of correct judgment counts corresponding to the sensitive data types, a plurality of incorrect judgment counts corresponding to the sensitive data types, a plurality of missed judgment counts corresponding to the sensitive data types, a plurality of correct judgment rates corresponding to a high confidence interval and corresponding to the sensitive data types, and a plurality of field numbers corresponding to the high confidence interval and corresponding to the sensitive data types; for each sensitive data type, when the processing unit determines that the number of incorrect judgments is greater than or equal to a misjudgment threshold, it transmits a correction suggestion notification corresponding to the sensitive data type to the terminal device, wherein the correction suggestion notification indicates that a new exclusion keyword is suggested; for each sensitive data type, the processing unit calculates a rule health score corresponding to the sensitive data type based on the correct judgment count, the incorrect judgment count, and the missed judgment count. For each sensitive asset type, when the processing unit determines that the rule health score is lower than the rule health threshold, it sends a correction suggestion notification corresponding to that sensitive asset type to the terminal device, wherein the correction suggestion notification indicates a suggestion to add or modify format rules; For each sensitive asset type, the processing unit calculates a correct judgment rate, a false judgment rate, and a false judgment rate corresponding to that sensitive asset type based on the number of correct judgments, the number of false judgments, and the number of false judgments; For each sensitive asset type, when the processing unit determines that the correct judgment rate is less than a correct judgment rate threshold and that the false judgment rate is greater than or equal to a false judgment rate threshold, it sends a correction suggestion notification corresponding to that sensitive asset type to the terminal device, wherein the correction suggestion notification indicates a suggestion to add excluded keywords; For each sensitive asset type, when the processing unit determines that the correct judgment rate is less than the correct judgment rate threshold and that the false judgment rate is greater than or equal to a false judgment rate threshold, it sends a correction suggestion notification corresponding to that sensitive asset type to the terminal device, wherein the correction suggestion notification indicates a suggestion to add hit keywords; For each sensitive data type, when the processing unit determines that the correct judgment rate corresponding to the high confidence interval is less than another correct judgment rate threshold and the number of columns corresponding to the high confidence interval is greater than or equal to a column number threshold, it transmits the correction suggestion notification corresponding to the sensitive data type to the terminal device.

Images