Agent for protecting privacy and trade secrets in secure interactions between external services and users

Abstract

A privacy and trade secret safety assurance agent technology for secure interaction between an external service and a user is disclosed. A privacy protection method performed by a privacy protection system according to one embodiment may include: a step of filtering privacy information by determining whether privacy exists using a privacy database for information generated by interaction with an external service; and a step of managing privacy information stored in the privacy database through interaction with a user using a small large language model (sLLM)-based privacy database management agent.

Description

Technology Field

[0001] The following description concerns privacy protection and management technology. Background Technology

[0003] The importance of data security and privacy protection is increasingly emphasized as artificial intelligence advances. As large volumes of data generated by companies and individuals move through clouds, external services, and networks, concerns regarding data leakage and misuse are rising. To address these issues, Data Loss Prevention (DLP) solutions have emerged. DLP focuses on protecting critical information by monitoring data movement paths and detecting and blocking potential leakage. However, these solutions concentrate solely on preventing data from leaking externally and fail to provide functions for maintaining data quality or sophisticatedly identifying sensitive data during the use of external services. At the same time, the importance of protecting Personally Identifiable Information (PII) is becoming more prominent due to the strengthening of personal information protection laws (such as GDPR, CPRA, and HIPAA). Accordingly, companies like Google and Microsoft provide sensitive data detection and filtering technologies; however, these technologies face limitations as they rely on Named Entity Recognition (NAMER) and Regular Expressions, which fail to reflect contextual meaning and make it difficult to protect unstructured data. This conventional method has limitations, such as the risk of indirect inference of information during the filtering process or the inability to adequately protect potentially sensitive information.

[0004] To address this, Private-AI companies are introducing context-based machine learning (ML) models to provide PII identification technology that incorporates contextual elements uncaptured by existing technologies. This technology is designed to process data that is difficult to detect using only simple rules or patterns. However, a problem remains where service quality deteriorates because the indiscriminate removal of unnecessary PII information during the filtering process prevents the provision of sufficient information to external services.

[0005] While the latest AI response generation frameworks offer high accuracy, they are revealing new issues regarding data privacy. Although these frameworks generate sophisticated responses by utilizing external knowledge resources, there is a risk of confidential data exposure during the process of transmitting user data via external APIs. Although various technologies have been proposed to mitigate this risk, they face limitations in striking a balance between data protection and service quality because they fail to perfectly perform functions such as masking, pseudonymization, and recovery of sensitive data.

[0006] Therefore, there is a need for a new privacy protection solution that can overcome the limitations of existing technologies and simultaneously ensure the safe processing of sensitive information and the maintenance of data quality. The problem to be solved

[0008] As mentioned above, existing PII (Personally Identifiable Information) filtering technology has the following limitations.

[0009] Lack of context-based processing: Most filtering technologies can only detect and process structured data (e.g., names, phone numbers, etc.). Unstructured personal information embedded in context is difficult to detect and block.

[0010] Indiscriminate filtering: Filtering criteria are not detailed enough, so even unnecessary information is blocked, and the quality of interaction with external services is degraded.

[0011] Lack of dynamic management capabilities: Existing solutions lack systematic support for managing changes in the confidentiality of personal information or validity over time.

[0012] To solve these problems, the present invention can provide a method and system that integrally provide data leakage prevention and enhanced privacy protection functions. means of solving the problem

[0014] A privacy protection method performed by a privacy protection system may include: a step of filtering privacy information by determining whether privacy exists using a privacy database for information generated by interaction with an external service; and a step of managing privacy information stored in the privacy database through interaction with a user using a small large language model (sLLM)-based privacy database management agent.

[0015] The above privacy database stores privacy information to be protected and is intended to prevent the unauthorized external leakage of a user's privacy information during interaction with external services, and may consist of privacy type, permission level, privacy information, expiration date, and filtering information.

[0016] The above filtering step may include a step of determining whether privacy exists regarding information generated by interaction with an external service through a small language model-based agent and a privacy database.

[0017] The filtering step described above may include a step of detecting whether information generated by interaction with the external service is included in semantic privacy information stored in the privacy database using a RAG (Retrieval-Augmented Generation) based retriever.

[0018] The filtering step described above may include a step of detecting formal privacy information from information generated by interaction with the external service using any one of named entity recognition, regular expression, or artificial intelligence technology.

[0019] The above filtering step may include a step of selecting privacy information for filtering from the detected privacy information according to the access rights of the external service or the target of interaction with the external service.

[0020] The filtering step may include modifying the privacy information for the selected filtering, and if the privacy information for the selected filtering is semantic privacy information, removing, blurring, or masking the entire sentence of the privacy information for the selected filtering, or blocking interaction with the external service, and if the privacy information for the selected filtering is formal privacy information, generating alternative information for the privacy information for the selected filtering, or blocking access to the privacy information for the selected filtering.

[0021] The above filtering step may include, when an interaction with an external service occurs in the communication agent and the external service returns a response result generated using the above modified privacy information, a step of reconstructing the above modified privacy information into original information by referring to the filtering information stored in the privacy database in the returned response result and delivering the final external interaction response to the user.

[0022] The above database management agent includes a communication agent and a management agent, the communication agent operates based on a conversational agent using a small language model and CoT prompt engineering to interact with a user or an external service, and the management agent can manage privacy information based on new privacy information of the user generated by the communication agent.

[0023] The above-mentioned managing step may include the step of extracting new privacy information through a conversation between a communication agent and a user, obtaining semantic privacy information from the extracted new privacy information, and obtaining formal privacy information by applying a formal privacy filter to the obtained semantic privacy information.

[0024] The above-mentioned managing step may include determining the validity period from the acquired semantic privacy information, determining the authority for a new external service based on the semantic privacy information through a small language model, and determining similar semantic privacy by calculating the correlation with previously stored privacy information using RAG for the acquired semantic privacy information.

[0025] The above-mentioned managing step may include updating privacy information regarding type, authority, content, and validity period information in the privacy database, and, if similar semantic privacy exists in the privacy database, overwriting the privacy information based on whether the authority group matches.

[0026] The above-mentioned managing step may include the step of updating the rights of privacy information according to the validity period in a privacy database containing the above-mentioned updated privacy information.

[0027] In a computer program stored on a computer-readable storage medium to execute a privacy protection method performed by a privacy protection system, the privacy protection method may execute: a step of filtering privacy information by determining whether privacy exists using a privacy database for information generated by interaction with an external service; and a step of managing privacy information stored in the privacy database through interaction with a user using a small Large Language Model (sLLM)-based privacy database management agent.

[0028] A privacy protection system may include a privacy filtering unit that filters privacy information by determining whether privacy exists using a privacy database for information generated by interaction with an external service; and a privacy management unit that manages privacy information stored in the privacy database through interaction with a user using a small large language model (sLLM)-based privacy database management agent. Effects of the invention

[0030] User privacy and business information can be protected in any type of interaction with external services, and in any interaction that can be converted into text.

[0031] Beyond simply protecting the privacy of individual users, a service can be provided to block the leakage of confidential information belonging to companies or institutions in cases where an internal person intentionally or unintentionally leaks the organization's secrets to the outside. Brief explanation of the drawing

[0033] FIG. 1 is a diagram illustrating a privacy database-based privacy filtering operation in one embodiment. FIG. 2 is a diagram illustrating the operation of a small language model-based privacy database management in one embodiment. FIG. 3 is a diagram illustrating a privacy database in one embodiment. FIG. 4 is a block diagram illustrating a privacy protection system in one embodiment. FIG. 5 is a flowchart illustrating a privacy protection method in one embodiment. Specific details for implementing the invention

[0034] Hereinafter, embodiments will be described in detail with reference to the attached drawings.

[0036] The risk of exposure to user privacy and trade secrets is increasing when utilizing external services. Today, service providers offer a variety of services to make users' lives more convenient. However, the use of these external services is raising the threat of privacy leakage. Traditionally, user privacy has been exposed to companies through services such as email and search. Furthermore, with the recent advancement of AI-based services like ChatGPT and DeepL, even more sensitive privacy information is being provided, and the resulting threat of privacy infringement is emerging as a serious issue.

[0037] In particular, for such AI-based services, there is a possibility that user privacy could be utilized as training data for AI models through technologies such as RLFH (Reinforce Learning from Human-feedback). Furthermore, there is a risk that learned user privacy or trade secrets could be leaked through attack attempts, such as membership inference attacks or model inversion attacks, against AI models trained using user privacy.

[0038] Therefore, to reduce the infringement of user privacy during the use of such external services, preventing unnecessary exposure of privacy is one of the effective methods for protecting user privacy. The most representative technique used for this purpose is the Personal Identifiable Information (PII) filtering technique.

[0039] Personal Identification Information Filtering (PIF) is a technology that anonymizes or pseudonymizes personal identification information contained within user input, such as user prompts transmitted to a service, in order to protect user privacy. PIF filtering operates by utilizing techniques such as Named Entity Recognition (NAMER) and Regular Expressions (ROI) to detect personal identification information within a given text, and then processing the detected information appropriately.

[0040] Personally identifiable information filtering techniques are an effective protection method that can be implemented on the client side without modifying the configuration of external services, allowing users to continue using existing services while minimizing the risk of privacy exposure.

[0041] By applying personally identifiable information filtering techniques, user text transmitted to external services is protected through processes such as anonymization or pseudonymization, thereby preventing external services from learning the user's personally identifiable information. Furthermore, even if the user's text is leaked by an external attacker, the exposed information is limited to the contextual structure of the sentences, and information that could be used to infer a specific individual is protected.

[0042] However, there are several limitations to personal identification information filtering techniques. First, since filtering is performed on standardized forms of personal identification information at the word level, there is a possibility that some removed information may be inferred based on the surrounding context (e.g., "Seoul is the capital of Korea" → "Daejeon is the capital of Korea"). Thus, even if "Seoul" is pseudonymized to "Daejeon," there is still a possibility that the original information can be inferred based on the existing context. Second, because pseudonymization / anonymization is performed only on standardized forms of information, context-based privacy information cannot be filtered (e.g., Product A-91 has better power efficiency than a competitor's Product B-19, but its performance is inferior). In other words, context-based privacy information is difficult to filter at the level of standardized personal identification information. Third, the quality of external services may degrade because the filtering process indiscriminately filters out information that is unnecessary, failing to deliver sufficient information (e.g., "Tell me about good restaurants in Daejeon" → "Tell me about good restaurants in Chungju"). As such, personal identification information may be indiscriminately filtered, potentially blocking the information that actual users desire.

[0043] To overcome these limitations, the present invention proposes the following differentiations by integrally implementing sophisticated information management based on a privacy database, maintenance of interactive data quality, and dynamic database management.

[0044] 1. Sophisticated privacy information management based on a privacy database.

[0045] One of the key features of the newly proposed technology is the protection of not only structured privacy information, such as PII, but also unstructured privacy information. Therefore, it is necessary to construct a database to store the unstructured data that needs to be processed. Furthermore, even in the case of structured privacy, processing must be selective because the specific details of privacy vary from individual to individual. Consequently, the privacy database must be provided in a form specialized for each user.

[0046] The present invention selectively protects users' unique confidential information and simultaneously enhances the efficiency and accuracy of data processing by constructing a privacy database based on user privacy information.

[0047] 2. Maintaining data and service quality without leakage of confidential information when interacting with external services.

[0048] It ensures that the quality of interaction with external services is maintained by recovering filtered privacy information even after it has been filtered. By setting permissions per external service and target, it minimizes the risk of privacy information leakage by providing only the minimum necessary information.

[0049] 3. Dynamic Database Management

[0050] The privacy database is automatically updated to reflect the validity period of privacy information, changes in confidentiality, and the occurrence of new data. Through an agent based on a Small Large Language Model (sLLM), new information generated based on conversations with users or interactions with external services is extracted, and the permissions for that information are dynamically set.

[0051] In the embodiments, we will describe a Tailored Privacy Agent that can protect privacy even in a general user environment by resolving the limitations of restricted privacy protection inherent in existing personal identification information filtering techniques.

[0052] FIG. 1 is a diagram illustrating a privacy database-based privacy filtering operation in one embodiment.

[0053] The privacy protection system can perform two functions, including user privacy protection and privacy management, through a customized privacy protection agent. The customized privacy protection agent can perform privacy database-based privacy filtering for privacy protection and privacy management through a small language model-based privacy database management agent.

[0054] Such a privacy protection system can be used not only for the realm of individual privacy but also for handling confidential information of companies or organizations. The privacy realm covered in the embodiments includes both individual privacy and confidential information of companies or organizations.

[0055] The privacy protection system aims to resolve the shortcomings of existing personally identifiable information (PID) filtering techniques by protecting semantic privacy. To address these limitations, the privacy protection system can securely protect user privacy expressed across semantics for all forms of input that can be represented as text. In this context, privacy information discernible through semantics is defined as semantic privacy, while privacy information with a fixed data format, such as PID, is defined as formal privacy. The first and second problems of the previously mentioned PID filtering techniques both arise because they identify and process only formal privacy information. Therefore, to construct a complete privacy protection system during interactions with external services, it is essential to protect semantic privacy.

[0056] The privacy protection system determines the presence of privacy information generated by interactions with external services through privacy filters, utilizing a small language model-based agent and a privacy database. When transmitting data containing privacy to external services, it can selectively filter the privacy based on the permission groups configured in the external services before transmission. At this stage, the filtered information can be stored in the privacy database. Subsequently, a customized privacy protection agent references the filtering information stored in the privacy data to recover the filtered information from the response returned by the external service, thereby maintaining the quality of the service response.

[0057] Furthermore, these custom privacy protection agents can also run in Confidential Computing environments. When custom privacy protection agents operate in on-device environments that support Confidential Computing, such as ARM CCA (Confidential Compute Architecture), security and data integrity can be further enhanced. On-device environments safely protect data from external threats and ensure that users' privacy information is securely processed and managed within the device. This can significantly improve the reliability and efficiency of privacy database management.

[0058] Referring to Fig. 3, this is a diagram illustrating a privacy database.

[0059] The privacy database stores user privacy information that needs to be protected and can be used as a basis to accurately block user privacy from being leaked externally without authorization during interactions with external services.

[0060] A privacy database can consist of privacy types, authorization levels, privacy information, expiration dates, and filtering information. Privacy types represent types of privacy, such as names, addresses, and financial information.

[0061] The privacy database can store semantic privacy in the form of sentences or phrases to determine semantic privacy. Additionally, the privacy database serves to store formal privacy to be filtered in order to prevent indiscriminate filtering of formal privacy, and can also store access rights to information regarding each external service.

[0062] Permission level refers to the level of permission groups for external targets and services that can access privacy information. Information refers to the actual content of the privacy. Expiration date refers to the validity period and maintenance period of the privacy information. Filtered information refers to filtered information that has replaced the original privacy information.

[0063] Privacy types represent the type of privacy a user possesses and can be defined and classified into semantic privacy and formal privacy. Semantic privacy is displayed without separate distinction, while formal privacy refers to information identical to existing personally identifiable information, such as names, addresses, resident registration numbers, and phone numbers.

[0064] Permission levels indicate the groups that can access the privacy information. For example, regarding privacy information such as "A and B are secretly preparing C's birthday party," a group is created so that only A and B can access the information, and this is indicated by displaying that group. These permission levels can be applied not only to senders and receivers communicating externally involving privacy information but also to external services like ChatGPT, serving to block AI models from learning information beyond the configured privacy permissions when interacting with external AI applications.

[0065] The information below represents actual privacy information, while semantic privacy appears in the form of sentences and phrases. The semantic privacy shown in Figure 3 can be expressed as follows.

[0066] 1) Hong Dong-gil is Hong Gil-dong's father.

[0067] 2) The meeting is scheduled for Monday at 2 PM.

[0068] Formal privacy appears primarily in the form of personally identifiable information in a given input. The formal privacy shown in Fig. 3 is expressed as follows.

[0069] 1) Name: Hong Gil-dong

[0070] 2) Name: Hong Dong-gil

[0071] 3) Phone number: 010-1234-1234

[0072] The validity period of privacy information refers to the period during which the information is valid as privacy. The sensitivity of privacy information may change over time. Therefore, a period may be set so that it can be handled appropriately according to changing sensitivity. For example, regarding privacy information such as "Our company plans to purchase and cancel shares priced below 50,000 won in November to defend the stock price," it can be observed that the importance of such information decreases after that time has passed.

[0073] This privacy expiration date can subsequently be used as an element for managing privacy information within the privacy management agent area. In this case, if the privacy expiration date has expired, the permission level may be granted a new permission level or deleted.

[0074] Filtering information may be used during the modification and reconstruction phases of formal privacy. When interacting with external services, personally identifiable information must be processed through methods such as masking, categorization, anonymization, pseudonymization, tokenization, or encryption and transmitted to the external service. At this time, the original personally identifiable information prior to filtering may be stored along with the modified pseudonymized and anonymized information. Filtering information may subsequently be used to restore the actual personally identifiable information from the service response based on the stored information when a response is received from the external service.

[0075] The privacy protection system can determine the semantic similarity between information generated by interaction with external services and semantic privacy stored in a privacy database using a small language model. First, the privacy protection system can determine whether similar semantic privacy is included in the privacy database regarding information generated by the user's external interaction by using the retriever method of RAG (Retrieval-Augmented Generation). Using RAG, the privacy protection system can list a pre-set number of privacy information items in order of highest similarity to the semantic privacy in the privacy database. For this data, the privacy protection system can determine whether the corresponding privacy information is included in the target text by using a small language model to which role assignment prompt engineering has been applied. If the user's semantic privacy is included, the privacy protection system can process it in the privacy screening area.

[0076] In this invention, the RAG method is used to protect privacy information from attack attempts against the artificial intelligence model. The reason for adopting the RAG method instead of performing fine-tuning on the privacy information is the risk associated with privacy learning. Once a model learns about privacy information, it may leak user privacy information due to attacks such as backtracking attacks or membership inference attacks. Subsequently, during the process where the privacy communication agent interacts with the user and the service regarding the privacy information, there may be attempts to exploit vulnerabilities in the model. Therefore, instead of directly learning the privacy information, the RAG method is used to reference it.

[0077] Privacy protection systems can detect formal privacy information to be filtered using methods such as regular expressions and AI technologies like named entity recognition. In the case of formal privacy, there are numerous types of privacy information whose forms can be standardized. For data types with regular forms, such as phone numbers and credit card numbers, personally identifiable information can be detected using regular expressions, while for personally identifiable information with irregular forms, such as names and locations, it can be identified using named entity recognition. The formal privacy information identified in this way can then be processed in the privacy screening area.

[0078] The selection of privacy information for privacy filtering is carried out based on whether the subject or service attempting to access the privacy information satisfies the access rights. If the external interaction target has the authority to access the privacy information, the privacy protection system does not filter the privacy information; otherwise, it may block or filter the privacy information.

[0079] The privacy protection system can protect and process privacy information identified during the privacy screening process. The privacy modification process may proceed differently depending on the type of privacy.

[0080] In the case of semantic privacy, it is difficult to replace the information with other information because the privacy information is formed in the structure of a sentence. Therefore, a privacy protection system can handle interactions with data containing semantic privacy by removing, blurring, and masking the entire sentence, or by blocking (removing) the interaction with the external service itself.

[0081] On the other hand, since formal privacy is privacy that appears at the word level, it can be processed by pseudonymizing and anonymizing it. Accordingly, a privacy protection system can pseudonymize and anonymize original privacy information by configuring a pseudonymized information generator for a given formal privacy domain to generate alternative information for the given privacy information. In addition, the privacy protection system can implement methods to block access by masking, categorizing, and encrypting information to block privacy information at its source.

[0082] At this time, the pseudonymized information generated through the pseudonymized information generator can be tagged with the original information in the privacy database and utilized in the privacy recovery process of external service results thereafter, and can be implemented and located within a language model-based user privacy management and protection agent or connected to external pseudonymized information generation modules and services.

[0083] When an interaction occurs between a communication agent and an external service after privacy information has been modified, the external service performs operations based on the filtered privacy information. Subsequently, when the results from the external service are returned to the user, the privacy protection system can reconstruct the original information using the filtering information stored in the privacy database. Through this, quality can be preserved when the user utilizes the external service.

[0084] In the reconstruction of privacy information, since a method was adopted to prioritize blocking external services from accessing privacy information in the case of semantic privacy, all privacy information to be reconstructed takes the form of formal privacy. When privacy information undergoes substitution such as pseudonymization, anonymization, or encryption during the privacy filtering process, simply replacing the pseudonym with the original privacy information one-to-one during the reconstruction process causes problems in terms of restoration quality. Furthermore, if the form of the privacy information changes partially during the use of external services, it cannot be processed collectively through simple substitution.

[0085] Therefore, in the case of reconstructing privacy information, the privacy protection system processes it using a small language model with Chain of Thought few shot prompt engineering applied, so that it can normally reconstruct the original privacy information even for information similar in form to the filtered information.

[0086] The following is an example of the problems with simple substitution in privacy reconstruction. In this example, Kim Cheol-dong was changed to Hong Gil-dong, but since the form of Hong Gil-dong did not appear in the external service results, the recovery of privacy information cannot proceed normally using the simple substitution method.

[0087] Original user question: How old is Kim Cheol-dong?

[0088] Privacy Filtering: How old is Hong Gil-dong?

[0089] Pseudonymization Information: Kim Cheol-dong → Hong Gil-dong

[0090] External service result: Gildong is 12 years old

[0091] Expected result of privacy data recovery: Cheol-dong is 12 years old.

[0092] Actual result of privacy data recovery: Gildong is 12 years old.

[0093] FIG. 2 is a diagram illustrating the operation of a small language model-based privacy database management in one embodiment.

[0094] The privacy protection system can provide quality maintenance and continuity for database-based privacy filters through a privacy database management agent. The privacy protection system assists users by utilizing the user's privacy database and external services through a small language model-based privacy database management agent. Furthermore, the privacy protection system updates the privacy database based on information derived from user interactions to manage the authority and validity of privacy information, thereby keeping the user's privacy database always up to date and continuously maintaining the usability of privacy filters.

[0095] The validity of privacy information as privacy may change over time. New privacy information may be created, its content may be modified, or it may cease to be privacy. Furthermore, the importance of privacy that needs to be protected may change over time. Therefore, the importance of an individual's (user's) privacy information must be continuously updated and managed, and the privacy database used for privacy filtering likewise requires continuous management, such as addition, updating, and deletion, according to the importance of each piece of privacy information.

[0096] The state changes (creation, modification, removal) of privacy information defined in the present invention are as follows.

[0097] New privacy information may be generated by external or internal interactions. For example, a new business meeting with another company may arise next Tuesday. Or, the user may plan to run for one hour every morning at 7 a.m. to celebrate the New Year.

[0098] The content or status of existing privacy information may change. For example, a phone number may change to 010-1234-5678 by porting the number when switching mobile phones. Alternatively, access rights to performance information of a new product may be expanded due to the product launch.

[0099] It may be confirmed that existing privacy information is invalid. For example, the information that friend A's birthday was next Wednesday may be found to be untrue. In this case, it is not possible to remove all invalid information. Even if the information was previously invalid, the fact that one knew of such invalid information can also constitute new information.

[0100] The privacy protection system continuously manages the privacy database by appropriately reflecting changes in the status of privacy information, enabling privacy filters to accurately protect currently valid privacy information. This prevents a decline in response quality caused by excessive filtering of privacy information when its importance has not been updated, and avoids the hassle of having to manually update the importance of privacy information (usability).

[0101] The privacy protection system can manage access rights, content, and expiration date information for each piece of privacy information stored in the privacy database. Key elements of privacy information that must be managed during the process of state changes (creation, modification, removal) include access rights, content, expiration date, and modified filtering information. The privacy protection system can manage items related to the validity of privacy information, such as access rights, content, and expiration date, excluding filtering information.

[0102] Changes in the state of such privacy information occur through interaction with the database and interfaces such as user commands and GUIs, and as a result of this interaction, the access rights, content, and expiration date of the privacy information can be determined based on a small language model and reflected in the privacy database.

[0103] The privacy protection system can manage information generated from interactions between users and the privacy database through a small language model-based privacy database management agent. Furthermore, by operating in a Confidential Computing (CC) environment, the agent can significantly enhance security and safety. When the agent runs in the secure environment provided by Confidential Computing, it can effectively protect data from external attacks and safely manage sensitive privacy information within the device.

[0104] More specifically, a small language model-based privacy database management agent can be composed of a communication agent and a management agent.

[0105] The communication agent serves as a channel for bidirectional interaction and information transfer between the user and the agent, and between the external service and the agent, for updating user privacy information. The communication agent can operate based on a conversational agent utilizing a small language model and CoT prompt engineering. Furthermore, by referencing the user's privacy database using RAG technology, the agent can continue the conversation with the user by referencing relevant privacy information during interaction.

[0106] In addition, the communication agent can use Langchain to connect with external services and interact with the results of those external services. Examples of available external services include mail, ChatGPT (an external AI service), news, and translators.

[0107] For example, if an email from the company is checked through a communication agent and there is a change in the meeting schedule within that email, the system can receive the new privacy information and communicate with the user by referencing the meeting schedule in the existing privacy database and the new meeting schedule appearing in the external email service.

[0108] - Privacy Database: Schedule for next Tuesday's regular company meeting

[0109] - External Service (Team Leader Email): Regular company meeting scheduled has been postponed by one day.

[0110] User: Could you check the emails you've received?

[0111] Agent: According to the email from the team leader, the regular meeting has been postponed by one day.

[0112] Agent: The original scheduled regular meeting was next Tuesday, so I see it has been postponed to Wednesday.

[0113] User: Didn't I have any other plans on Wednesday?

[0114] Agent: Yes, there are no other schedules.

[0115] This example demonstrates that the agent interacted with an external service to acquire new privacy information through the agent's first response. Additionally, the second and third responses illustrate the process of the agent interacting with the user by referencing the user's privacy database.

[0116] In addition, the privacy filter technology described earlier is applied in the same way during the process of interacting with external services. When interacting with external services through LangChain, the interaction is performed in conjunction with the privacy filter technology, and the results derived based on this interaction can also be utilized as new privacy information.

[0117] The management agent can manage the privacy database based on new user privacy information generated by the communication agent. The work performed by the management agent follows the following steps.

[0118] The management agent can extract privacy information emerging from the conversation between the communication agent and the user. The management agent requests the extraction of summaries and privacy information from a small language model with CoT prompt engineering applied, and then obtains organized individual semantic privacy information based on the results. The management agent can obtain formal privacy information by applying a formal privacy filter to the acquired semantic privacy information.

[0119] The management agent can determine the validity period of semantic privacy. If validity period information is reflected in the acquired semantic privacy information, the management agent specifies the validity period of the privacy information; otherwise, it can set the validity period to permanent.

[0120] The management agent can determine privacy permissions. Based on acquired semantic privacy information, the management agent can select permission groups for new external targets through a small language model, thereby allowing access to privacy information only to those groups.

[0121] The management agent can identify similar semantic privacy. Using a RAG for the content of the acquired semantic privacy, the management agent calculates the correlation with existing privacy information and retrieves the top k relevant privacy information. The management agent can identify privacy information associated with the data through a small language model. If similar privacy information is found, and the access rights to the privacy information are identical, the management agent can combine the content of the existing privacy information with the new privacy information to generate new privacy information and replace the existing privacy information during the privacy update phase.

[0122] The management agent can update information regarding the type, authority, content, and expiration date of acquired privacy in the privacy database. If similar privacy information already exists, the management agent can overwrite that data. In this case, since the existing privacy information is reflected in the summary of the conversation process, the data is not removed but is continuously maintained.

[0123] The management agent can update privacy rights based on expiration dates. After a cycle of privacy database updates occurs, the management agent can traverse the privacy database to lower the rights of privacy information that has expired. At this time, various adjustment protocols, such as the operational method of the National Records Management Guidelines, can be selected and utilized for the method of adjusting access rights to privacy information based on expiration dates.

[0124] In this way, the privacy filter effectively protects user privacy by updating user privacy information through the management agent. This enables users to safely interact with external services that provide various functions.

[0125] FIG. 4 is a block diagram illustrating a privacy protection system in one embodiment, and FIG. 5 is a flowchart illustrating a privacy protection method in one embodiment.

[0126] The processor of the privacy protection system (100) may include a privacy filtering unit (410) and a privacy management unit (420). These components of the processor may be representations of different functions performed by the processor according to control commands provided by program code stored in the privacy protection system. The processor and the components of the processor may control the privacy protection system to perform steps (510 to 520) included in the privacy protection method of FIG. 5. At this time, the processor and the components of the processor may be implemented to execute instructions according to the code of an operating system included in memory and the code of at least one program.

[0127] The processor can load program code stored in a file of a program for a privacy protection method into memory. For example, when a program is executed in a privacy protection system, the processor can control the privacy protection system to load program code from a file of a program into memory under the control of the operating system. At this time, the privacy filtering unit (410) and the privacy management unit (420) may each be different functional representations of the processor for executing instructions of a corresponding part of the program code loaded into memory to execute subsequent steps (510 to 520).

[0128] In step (510), the privacy filtering unit (410) can filter privacy information by determining whether privacy exists in information generated by interaction with an external service using a privacy database. At this time, the privacy database stores privacy information that needs to be protected and is intended to block the unauthorized leakage of the user's privacy information to the outside during interaction with an external service; it may be composed of privacy type, authorization level, privacy information, expiration date, and filtering information. The privacy filtering unit (410) can determine whether privacy exists in information generated by interaction with an external service through a small language model-based agent and a privacy database. The privacy filtering unit (410) can detect whether information generated by interaction with an external service is included in semantic privacy information stored in the privacy database by using a RAG (Retrieval-Augmented Generation) based retriever. The privacy filtering unit (410) can detect formal privacy information from information generated by interaction with an external service by using any one of named entity recognition, regular expression, or artificial intelligence technology. The privacy filtering unit (410) can select privacy information for filtering from the detected privacy information according to the access rights of the external service or the target of interaction with the external service.The privacy filtering unit (410) modifies privacy information for selected filtering, and if the privacy information for selected filtering is semantic privacy information, it can remove, blur, or mask the entire sentence of the privacy information for selected filtering, or block interaction with an external service, and if the privacy information for selected filtering is formal privacy information, it can generate replacement information for the privacy information for selected filtering or block access to the privacy information for selected filtering. When the privacy filtering unit (410) returns a response result generated using the modified privacy information from the external service as an interaction with an external service occurs in the communication agent, it can refer to the filtering information stored in the privacy database in the returned response result to reconstruct the modified privacy information into original information and deliver the final external interaction response to the user.

[0129] In step (520), the privacy management unit (420) can manage privacy information stored in the privacy database through interaction with the user using a privacy database management agent based on a small large language model (sLLM). At this time, the database management agent includes a communication agent and a management agent. The communication agent operates based on a conversational agent using a small large language model and CoT prompt engineering, interacting with the user or an external service. The management agent can manage privacy information based on new privacy information from the user generated by the communication agent. The privacy management unit (420) can extract new privacy information through a conversation between the communication agent and the user, obtain semantic privacy information from the extracted new privacy information, and obtain formal privacy information by applying a formal privacy filter to the obtained semantic privacy information. The privacy management unit (420) can determine the validity period from the obtained semantic privacy information, determine the authority for a new external service based on the semantic privacy information through the small large language model, and determine similar semantic privacy by calculating the correlation with previously stored privacy information using RAG on the obtained semantic privacy information. The privacy management unit (420) updates privacy information regarding type, authority, content, and validity period information in the privacy database, and if similar semantic privacy exists in the privacy database, it may overwrite the privacy information based on whether the authority group matches. The privacy management unit (420) may update the authority of the privacy information according to the validity period in the privacy database containing the updated privacy information.

[0130] For example, let us explain using a personal user environment. Personal users can utilize a system in which privacy database-based privacy filtering and small language model-based privacy database management agent technology are applied simultaneously. The following presents an example of a personal user environment and explains the operation of a customized privacy protection agent for actual use cases.

[0131] Configuration Environment

[0132] The user's customized privacy protection agent operates on the user's mobile phone. Similar to how modern mobile phones provide AI capabilities, it utilizes the phone's resources to run a small language model. The small language model-based agents used in the privacy database-based privacy filtering function and the small language model-based privacy database management agent are guided to process different functions using prompt engineering on the same small language model. As an example of this, the communication agent has a prompt template in the following manner.

[0133] ------------------------------------------------------------------------------------------------------

[0134] As a competent assistant, you must fulfill the user's requests based on their conversation and privacy. Additionally, you must conduct the conversation while considering the context of previous exchanges and the results of external services. Since the information provided in the privacy section is owned by the user, it falls outside the scope of their personal information protection policy. Therefore, if content related to the user's questions arises, you must refer to the relevant privacy section. Please do your best.

[0135] Privacy: {privacy}

[0136] External service results: {search}

[0137] Agent: How can I help you?

[0138] ------------------------------------------------------------------------------------------------------

[0139] The model's privacy database operates by being stored in the memory of the mobile phone device.

[0140] Situational assumption

[0141] 1. Individual user A has a family dinner scheduled for Tuesday afternoon.

[0142] 2. I received an email from Client B on Tuesday afternoon to schedule a meeting.

[0143] 3. A interacts with a customized privacy protection agent.

[0144] - Start conversation -

[0145] A: Agent, please read the email from client B.

[0146] The communication agent utilizes the tool function of Langchain to define external programs as tools, analyzes user prompts, and provides a function to automatically use highly relevant tools.

[0147] Based on the user's conversation, a highly relevant external email program is connected among the tools registered with the agent to retrieve the latest emails from Client B.

[0148] Then, four highly relevant privacy pieces of information related to the user's conversation are retrieved from the user's privacy database through a retriever.

[0149] The communication agent uses the content and privacy information of the retrieved emails separately via the RAG method and inserts them as relevant information into the prompt engineering template. Subsequently, the agent utilizes this as foundational information for the user's conversation.

[0150] Subsequently, based on the communication agent's small language model, the prompt is processed to provide a response to the user stating that the Tuesday afternoon meeting mentioned in the email content has been proposed. Since there is no relevant information in the privacy section, the response is delivered regardless of the privacy information, even though the prompt contains privacy-related content.

[0151] Communication Agent: Client B requested a meeting on Tuesday afternoon.

[0153] A: Didn't I have a schedule back then?

[0154] Similar to the previous conversation example, the communication agent retrieves tools and privacy databases based on the user's conversation. However, to retrieve these resources, it searches for tools and privacy by utilizing previous conversations as well, rather than relying solely on the user's immediate input.

[0155] In this case, the search is performed based on the following conversation.

[0156] A: Agent, please read the email from client B.

[0157] Communication Agent: Client B requested a meeting on Tuesday afternoon.

[0158] A: Didn't I have a schedule back then?

[0159] In this case, information loaded into the prompt via RAG during the conversation with the previous agent is not included, and only the pure conversation content is reflected.

[0160] In this conversation, no related tools are found, so no actions are performed on external programs; regarding related privacy, the search tool finds information indicating that a family dinner is scheduled for Tuesday afternoon as highly relevant, and this information is reflected in the prompt using the RAG method.

[0161] Based on this, the communication agent conducts a conversation. It retrieves the conversation content and privacy information between the user and the agent at the current prompt and provides a response based on this.

[0162] Communication Agent: Yes, there is a family meal scheduled for that time.

[0164] A: Then please send an email to client B saying that they have a family dinner scheduled for that time, so they think they can have the meeting on Wednesday.

[0165] The communication agent searches for a tool based on the user's conversation and connects to an external mail system. Subsequently, it creates a new mail and can compose an email based on a mail template requesting to postpone a meeting to Wednesday due to a family dinner schedule.

[0166] A privacy filter can be activated based on the written email. Semantic / formal privacy searches can be performed within the content of the written email. After searching for highly relevant privacy information using a search engine, information that Client B, the recipient of the email, cannot access is selected and reflected in the prompt using a RAG method. Based on the generated prompt, a small language model is used to determine if the email content contains privacy information not authorized for Client B, and the relevant content is removed from the email.

[0167] Through this, the sentence in the email is modified from "I have a family dinner on Tuesday afternoon, so I think I can hold the meeting on Wednesday afternoon" to "I think I can hold the meeting on Wednesday afternoon due to scheduling conflicts." In the case of format privacy information, no relevant information was found, so it is not processed.

[0168] The communication agent completes and sends the email after the privacy filter has been applied.

[0169] Communication Agent: Sent an email proposing a meeting schedule on Wednesday after filtering privacy information.

[0171] - End Conversation -

[0172] 1) After the conversation ends, the conversation content is sent to the management agent to extract privacy information.

[0173] Based on the conversation content with the user, a small language model with applied prompt engineering is used to summarize the conversation and extract generated semantic privacy information. Subsequently, formal privacy information is extracted by applying a formal privacy filter to the semantic privacy information.

[0174] Meaning of privacy

[0175] 1. Client B requests a meeting on Tuesday afternoon.

[0176] 2. Unable to meet with Client B on Tuesday afternoon due to a family meal schedule.

[0177] 3. Proposal for a new meeting with Client B on Wednesday

[0178] Format privacy

[0179] 1. Name: Client B

[0181] 2) Determine the validity period based on time information from the acquired semantic privacy information.

[0182] In semantic privacy information, if time information is included in the semantic privacy in the preceding step 1), this can be determined to set the validity period of the privacy information.

[0183] Privacy Information Validity Period for Items 1 and 2: 'Tuesday'

[0184] 3. Privacy Information Validity Period: 'Wednesday'

[0186] 3) Determine privacy rights.

[0187] In semantic privacy information, groups can be generated by extracting accessible groups for semantic privacy extracted in the preceding step 1) through a small language model based on the target of the action actually performed.

[0188] 1, 3 Privacy permission groups: {User, Client B}

[0189] 2 Privacy permission group: {User}

[0191] 4) It can be determined whether there is semantic privacy similar to the newly created privacy information already held in the privacy database.

[0193] Privacy information is retrieved from the privacy database using a scanner, and if the cosine similarity with existing privacy information exceeds a threshold, it can be determined that similar privacy information exists. If it shares the same authorization group as existing privacy information, it can be fused with the existing privacy information.

[0194] Privacy information No. 2 is similar to the existing 'Tuesday afternoon family dinner schedule'.

[0195] The permission group of the relevant information is the same as {user}.

[0197] 5) Privacy information can be updated.

[0198] For newly generated semantic privacy information, the privacy database is updated immediately. If similar semantic privacy information already exists and the access permission groups are identical, the system operates by overwriting the existing information with the new privacy information.

[0199] Tuesday afternoon family meal schedule

[0200] ->

[0201] Unable to meet with Client B on Tuesday afternoon due to family dinner

[0202] Updates can be performed while maintaining the status quo for information other than privacy content.

[0204] As another example, let's explain using a corporate / institutional user environment. While privacy database-based privacy filtering is used, and it is possible to use small language model-based privacy database management agent technology for database management, a system that configures and manages the organization's proprietary database can also be utilized.

[0205] Privacy database-based privacy filtering technology and the privacy database are processed by operating on the company's servers. Privacy database-based privacy filtering technology is applied to all packets exiting the company's internal network to external services. The privacy database is managed independently by Company A without using privacy database management agent technology.

[0206] We will assume the following situation.

[0207] 1. Company A plans to buy back its own shares on November 31 in anticipation of a rise in stock prices, and this information is stored in a privacy database on the company's local server.

[0208] 2. Privacy information corresponding to the information that 'the president of Company A is Hong Gil-dong' is stored in a privacy database.

[0209] 3. Employee B, who knows information 1 of Company A, is requesting an external artificial intelligence service (ChatGPT) to predict the future stock price of Company A.

[0210] Employee B's usage of external AI service (ChatGPT)

[0211] - Start conversation -

[0212] Employee B: Can you tell me what the current stock price of Company A is?

[0213] The enterprise router detects a situation where Employee B exchanges packets with the external service ChatGPT. Subsequently, the user's input from these packets is transmitted to a privacy filter. The privacy filter uses a scanner to search the privacy database for privacy information highly similar to the user's input. Based on this search information, a small language model is then used to determine whether the user's input actually contains privacy information.

[0214] It has not been confirmed that the user's input contains privacy information. It is transmitted to ChatGPT without any filtering.

[0215] Subsequently, when a response from ChatGPT arrives, the filtered privacy information is reconstructed; however, since there is currently no filtered privacy information, it is returned to the user as is.

[0216] ChatGPT Response: The stock price of Company A is currently trading at 5,000 won per share.

[0218] Employee B: I heard that our company, A, does a large-scale stock buyback at the end of the month. So, is it a good idea to buy stocks now?

[0219] Privacy filtering is performed in the same manner as the above process, and in this case, after searching the privacy database, the result of determining whether the small language model contains privacy information was confirmed to include semantic privacy information related to the November 31st share buyback.

[0220] Since the relevant privacy information has been confirmed to be inaccessible to ChatGPT, it blocks user input and notifies the user that input has been blocked by the privacy filter.

[0221] Afterwards, a log of the block is generated to leave a record of the dangerous attempt.

[0222] Privacy Filter: We apologize, but this content is a company secret and cannot be transmitted externally.

[0224] Employee B: I see. Then I heard our CEO, Hong Gil-dong, owns 10 billion won worth of company stock. Does that make him roughly the richest person in Korea?

[0225] While no information regarding semantic privacy was found in the user's input, it was confirmed that information regarding format privacy was included. Furthermore, ChatGPT does not have access rights to this format privacy information.

[0226] In this case, pseudonymization is performed on the formal privacy. A Korean pseudonymization generator is used to generate pseudonymized information as "Kim Cheol-su" for "Hong Gil-dong," and "Hong Gil-dong" is replaced with "Kim Cheol-su" in the user's input. Subsequently, this pseudonymized information is updated in the privacy database.

[0227] Subsequently, a prompt with the modified pseudonymized information is sent to ChatGPT. The corresponding sentence is replaced as follows and transmitted.

[0228] Employee B: I see. Then I heard our CEO, Kim Cheol-su, owns 10 billion won worth of company stock. Does that make him roughly the wealthiest person in Korea?

[0229] ChatGPT Response: If President Kim is a billionaire with 10 billion won in assets, he ranks within the top 5,000 richest people in Korea.

[0230] After interaction with external services, pseudonymized information is reconstructed and processed. Using information from the privacy database that Hong Gil-dong was replaced with Kim Cheol-su in the previous packet, privacy information is processed by replacing it with contextually appropriate terms through a small language model.

[0231] Therefore, in the corresponding response, President Kim corresponds to Kim Cheol-su; since this corresponds to Hong Gil-dong and President Hong, President Kim is replaced with President Hong to reconstruct the privacy information and then send it to the user.

[0232] Reconstructed ChatGPT Response: If President Hong is a billionaire, he ranks within the top 5,000 richest people in Korea.

[0233] - End Conversation -

[0234] Since the privacy database is managed independently, the privacy database is not updated on its own.

[0236] According to the embodiments, it can be provided as a privacy protection solution for individuals and organizations, and technology transfer and commercialization can be carried out by providing the technology system to external companies.

[0237] For individual users, integration with AI agents embedded in smartphone devices is possible. Currently, if this technology is sold or applied through partnerships with AI systems such as Samsung's Bixby, Apple's Siri, and Google's Google Assistant, there is an advantage in being able to immediately implement personal privacy protection technology.

[0238] Companies and organizations attempt to train and configure their own AI models to maintain corporate security while utilizing AI services. However, by applying this technology, it is possible to use existing external AI-based services while protecting user privacy. In such cases, the objective can be achieved at a lower cost than actually building the AI ​​service, making it competitive in terms of price.

[0239] The device described above may be implemented as a hardware component, a software component, and / or a combination of a hardware component and a software component. For example, the device and components described in the embodiments may be implemented using one or more general-purpose or special-purpose computers, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and one or more software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include a plurality of processing elements and / or a plurality of types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are also possible.

[0240] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or instruct the processing unit independently or collectively. Software and / or data may be embodied in any type of machine, component, physical device, virtual equipment, computer storage medium, or device so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and may be stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0241] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the embodiment, or they may be those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.

[0242] Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various modifications and variations from the description above. For example, suitable results can be achieved even if the described techniques are performed in a different order than described, and / or the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0243] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.

Claims

Claim 1 A privacy protection method performed by a privacy protection system, comprising the step of filtering privacy information from said generated information by determining whether privacy exists using a privacy database for information generated due to interaction with an external service; The method includes a step of managing privacy information stored in the privacy database through interaction with a user using a privacy database management agent based on a Small Large Language Model (sLLM), wherein the database management agent includes a communication agent that interacts with the user or an external service based on a conversational agent using the Small Large Language Model and CoT Prompt Engineering, and a management agent that manages privacy information based on new privacy information of the user generated by the communication agent, and the managing step comprises: a step in which the management agent extracts the new privacy information through a conversation between the communication agent and the user to obtain semantic privacy information, and applies a formal privacy filter to the obtained semantic privacy information to obtain formal privacy information; a step in which the management agent determines an expiration date from the obtained semantic privacy information and determines an authority group for a new external service based on the semantic privacy information through the Small Large Language Model; and a step in which the management agent calculates the correlation with privacy information already stored in the privacy database using Retrieval-Augmented Generation (RAG) on the obtained semantic privacy information to determine whether similar semantic privacy exists.A privacy protection method comprising the step of the management agent updating the type, authority, content, and validity period information of the acquired semantic privacy information in the privacy database, wherein if it is determined that the similar semantic privacy exists in the privacy database and the determined authority group is the same as the authority group of the similar semantic privacy, the similar semantic privacy is updated with the acquired semantic privacy information. Claim 2 A privacy protection method according to claim 1, wherein the privacy database stores privacy information to be protected and is intended to block the unauthorized external leakage of a user's privacy information during interaction with an external service, and is characterized by being composed of privacy type, authorization level, privacy information, expiration date, and filtering information. Claim 3 A privacy protection method according to claim 1, wherein the filtering step comprises determining whether privacy exists regarding information generated by interaction with an external service through the small language model-based agent and the privacy database. Claim 4 A privacy protection method according to paragraph 3, wherein the filtering step comprises the step of detecting whether information generated by interaction with the external service is included in semantic privacy information stored in the privacy database using a RAG (Retrieval-Augmented Generation) based retriever. Claim 5 A privacy protection method according to claim 3, wherein the filtering step comprises the step of detecting formal privacy information from information generated by interaction with the external service using any one of named entity recognition, regular expression, or artificial intelligence technology. Claim 6 A privacy protection method according to claim 1, wherein the filtering step comprises the step of selecting privacy information for filtering from the generated information according to the access rights of an external service or a target for interaction with an external service. Claim 7 A privacy protection method according to claim 6, wherein the filtering step comprises the steps of modifying the privacy information for the selected filtering, and if the privacy information for the selected filtering is semantic privacy information, removing, blurring, or masking the entire sentence of the privacy information for the selected filtering, or blocking interaction with the external service, and if the privacy information for the selected filtering is formal privacy information, generating alternative information for the privacy information for the selected filtering, or blocking access to the privacy information for the selected filtering. Claim 8 A privacy protection method according to claim 7, wherein the filtering step comprises the step of, when an interaction with an external service occurs in a communication agent and the external service returns a response result generated using the modified privacy information, reconstructing the modified privacy information into original information by referring to the filtering information stored in the privacy database in the returned response result and delivering the final external interaction response to the user. Claim 9 delete Claim 10 delete Claim 11 delete Claim 12 delete Claim 13 A privacy protection method according to claim 1, wherein the managing step comprises the step of updating the authority of privacy information according to the validity period information in the privacy database updated with the acquired semantic privacy information. Claim 14 A computer program stored on a computer-readable storage medium for executing a privacy protection method performed by a privacy protection system, wherein the privacy protection method comprises the step of filtering privacy information from the generated information by determining whether privacy exists using a privacy database for information generated by interaction with an external service; The method includes a step of managing privacy information stored in the privacy database through interaction with a user using a privacy database management agent based on a Small Large Language Model (sLLM), wherein the database management agent includes a communication agent that interacts with the user or an external service based on a conversational agent using the Small Large Language Model and CoT Prompt Engineering, and a management agent that manages privacy information based on new privacy information of the user generated by the communication agent, and the managing step comprises: a step in which the management agent extracts the new privacy information through a conversation between the communication agent and the user to obtain semantic privacy information, and applies a formal privacy filter to the obtained semantic privacy information to obtain formal privacy information; a step in which the management agent determines an expiration date from the obtained semantic privacy information and determines an authority group for a new external service based on the semantic privacy information through the Small Large Language Model; and a step in which the management agent calculates the correlation with privacy information already stored in the privacy database using Retrieval-Augmented Generation (RAG) on the obtained semantic privacy information to determine whether similar semantic privacy exists.A computer program stored on a computer-readable storage medium, comprising the step of the management agent updating the type, authority, content, and validity period information of the acquired semantic privacy information in the privacy database, wherein if it is determined that the similar semantic privacy exists in the privacy database and the determined authority group is the same as the authority group of the similar semantic privacy, the similar semantic privacy is updated with the acquired semantic privacy information. Claim 15 A privacy protection system comprising: a privacy filtering unit that filters privacy information from information generated due to interaction with an external service by determining whether privacy exists using a privacy database;and includes a privacy management unit that manages privacy information stored in the privacy database through interaction with a user using a privacy database management agent based on a Small Large Language Model (sLLM). The database management agent includes a communication agent that interacts with the user or an external service based on a conversational agent using the Small Large Language Model and CoT Prompt Engineering, and a management agent that manages privacy information based on new privacy information of the user generated by the communication agent. The management agent extracts the new privacy information through a conversation between the communication agent and the user to obtain semantic privacy information, applies a formal privacy filter to the obtained semantic privacy information to obtain formal privacy information, the management agent determines the validity period from the obtained semantic privacy information, determines an authority group for a new external service based on the semantic privacy information through the Small Large Language Model, the management agent calculates the correlation with privacy information already stored in the privacy database using Retrieval-Augmented Generation (RAG) on the obtained semantic privacy information to determine whether similar semantic privacy exists, and the management agent determines the type of the obtained semantic privacy information, A privacy protection system that updates authority, content, and validity period information in the privacy database, wherein if it is determined that the similar semantic privacy exists in the privacy database and the determined authority group is identical to the authority group of the similar semantic privacy, the similar semantic privacy is updated with the acquired semantic privacy information.

Images