Dynamic desensitization digital avatar interaction method and system based on layered architecture

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a layered architecture-based dynamic desensitization digital avatar interaction method, which comprises the following steps that: a base brain layer generates general dialogue content, and generates dialogue content containing general information based on a large-scale pre-training language model; the user brain layer processes personalized data, stores identity information, emotional intimacy and sensitive topic tags of dialogue objects, and marks dialogue contents according to user-defined rules; and the agent decision-making layer executes dynamic desensitization, performs dynamic desensitization processing on the dialogue content generated by the base brain layer according to the mark and the rule of the user brain layer, and outputs the desensitized dialogue content. Through the layered architecture design, the general dialogue generation capability and the personalized privacy protection function are decoupled, and the balance between personalization and privacy protection is realized. The base brain layer provides powerful universal dialogue generation ability, and the user brain layer focuses on personalized data processing and privacy protection.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and specifically provides a dynamic desensitization digital avatar interaction method and system based on a hierarchical architecture. Background Art

[0002] With the rapid development of artificial intelligence and natural language processing technologies, digital avatar interaction systems have gradually become important tools for information exchange and customer service. However, existing technologies face many challenges and limitations while achieving general dialogue generation and personalized interaction.

[0003] On the one hand, although existing general large models have powerful language generation capabilities and can handle a wide range of general dialogue tasks, they have obvious deficiencies in terms of personalization and privacy protection. Directly using these models will lead to the leakage of user privacy information, while fully local training will sacrifice the generality and performance of the models. For example, although patent document CN115374129A involves the privacy protection of digital avatars, it fails to effectively solve the problem of sensitive information leakage.

[0004] On the other hand, traditional desensitization technologies can usually only perform static replacement on fixed fields (such as ID numbers and mobile phone numbers), and cannot dynamically adjust the desensitization strategy according to the identity of the dialogue partner, emotional intimacy, and topic sensitivity. This static desensitization method is particularly inefficient when dealing with complex social scenarios and cannot meet the privacy protection needs of users in different social relationships. For example, the desensitization technology described in patent document CN113139080A is only applicable to the desensitization processing of fixed fields and cannot achieve dynamic desensitization.

[0005] In addition, most existing digital human systems adopt enterprise-preset models, and users cannot independently train the personality characteristics and social memories of the avatars through daily chat behaviors. Such systems lacking user-independent training capabilities are difficult to meet the needs of users for personalized interaction and cannot dynamically adjust interaction strategies according to user preferences and social relationships. For example, although patent document CN115830680B involves the interaction method of a digital human system, it does not provide a user-independent training function.

[0006] Therefore, there are obvious deficiencies in the balance between personalization and privacy protection, the flexibility and accuracy of dynamic desensitization, and user-independent training capabilities in existing technologies. For this reason, technical personnel in this field have proposed a dynamic desensitization digital avatar interaction method and system based on a hierarchical architecture to solve the above problems. Summary of the Invention

[0007] Aiming at the deficiencies of the existing technology, the present invention provides a dynamic desensitization digital avatar interaction method and system based on a hierarchical architecture, which solves the problems raised in the above background art.

[0008] To achieve the above objectives, the present invention is realized through the following technical solutions: A dynamic desensitization digital avatar interaction method based on a hierarchical architecture, comprising the following steps:

[0009] The base brain layer generates general conversation content, generating conversation content containing general information based on a large-scale pre-trained language model;

[0010] The user brain layer processes personalized data, stores the identity information, emotional intimacy, and sensitive topic labels of the conversation object, and marks the conversation content according to user-defined rules;

[0011] The proxy decision layer performs dynamic desensitization, dynamically desensitizing the conversation content generated by the base brain layer according to the marks and rules of the user brain layer, and outputting the desensitized conversation content;

[0012] Application of privacy-enhanced technology, using federated learning for local training to ensure user data privacy, and processing sensitive labels through homomorphic encryption technology to ensure data privacy during transmission and processing.

[0013] Preferably, the dynamic desensitization processing includes:

[0014] Adjustment of the desensitization strategy based on identity labels, dynamically adjusting the desensitization intensity according to the identity of the conversation object;

[0015] Adjustment of the desensitization strategy based on sensitive values, dynamically adjusting the desensitization method according to the sensitive value of the sensitive topic, including entity replacement, content deletion, or obfuscation processing;

[0016] Adjustment of the desensitization strategy based on emotional intimacy, used to dynamically adjust the desensitization strategy according to the emotional intimacy between the conversation object and the user to adapt to different social relationships.

[0017] Preferably, the user brain layer conducts local training through federated learning, and the training data includes:

[0018] Local text data of the user's historical conversations;

[0019] Pairs of labels and rules annotated by the user through a visual interface;

[0020] Conversation correction data feedback by the user in real time, used to continuously optimize the model of the user brain layer.

[0021] Preferably, the proxy decision layer uses homomorphic encryption technology to process sensitive labels to ensure data privacy during transmission and processing, specifically including:

[0022] Using the Paillier encryption algorithm to encrypt sensitive labels;

[0023] Decrypt and process the encrypted sensitive tags at the proxy decision layer to ensure that the data is available but invisible.

[0024] Preferably, the dynamic desensitization policy is optimized through equations and matrix operations to achieve a more flexible desensitization decision logic, specifically including:

[0025] Define a desensitization decision function to dynamically adjust the desensitization policy according to the identity, topic type, and emotional intimacy of the conversation object;

[0026] Use matrix operations to optimize the calculation process of the desensitization decision function and improve the response speed and processing efficiency of the system.

[0027] Preferably, the user brain layer supports the user to dynamically adjust tags and rules through a visual interface to adapt to different conversation scenarios and user needs, specifically including:

[0028] Provide a user-friendly visual interface that allows users to customize tags and rules;

[0029] Support users to update tags and rules in real time, and the system responds immediately and adjusts the desensitization policy.

[0030] Preferably, the base brain layer and the user brain layer perform data interaction through the API interface to ensure the modularity and scalability of the system, specifically including:

[0031] The base brain layer transmits the generated general conversation content to the proxy decision layer through the API interface;

[0032] The user brain layer transmits the tags and rules to the proxy decision layer through the API interface;

[0033] The proxy decision layer returns the desensitized conversation content to the user side through the API interface.

[0034] Preferably, the proxy decision layer fuses the original content of the base brain layer and the tags and rules of the user brain layer through the multi-head attention mechanism to output the desensitized content, specifically including:

[0035] Use the multi-head attention mechanism to perform weighted fusion on the original content and tags;

[0036] Dynamically adjust the desensitization policy according to the fusion result to ensure the accuracy and naturalness of the desensitized content.

[0037] Preferably, it also includes supporting multi-language conversations and dynamically adjusting the desensitization policy according to language characteristics to meet the privacy protection needs in different language environments, specifically including:

[0038] Support the generation of conversations and desensitization processing in multiple languages;

[0039] Dynamically adjust the desensitization strategy according to the grammar and cultural characteristics of different languages to ensure the effectiveness of privacy protection.

[0040] A dynamic desensitization digital avatar interaction system based on a hierarchical architecture, comprising:

[0041] The base brain layer, which contains a large-scale pre-trained language model for generating general conversation content;

[0042] The user brain layer, which contains a three-dimensional tag library and a rule engine for storing the identity information, emotional intimacy, and sensitive topic tags of the conversation object, and marking the conversation content according to user-defined rules;

[0043] The proxy decision layer is used to perform dynamic desensitization processing on the conversation content generated by the base brain layer according to the marks and rules of the user brain layer, and output the desensitized conversation content;

[0044] The privacy enhancement module uses federated learning for local training to ensure user data privacy, and processes sensitive tags through homomorphic encryption technology to ensure data privacy during transmission and processing;

[0045] The multilingual support module supports multilingual conversations and dynamically adjusts the desensitization strategy according to language characteristics to meet the privacy protection requirements in different language environments.

[0046] The present invention provides a dynamic desensitization digital avatar interaction method and system based on a hierarchical architecture.

[0047] It has the following beneficial effects:

[0048] 1. Through the hierarchical architecture design, the present invention decouples the general conversation generation ability and the personalized privacy protection function, achieving a balance between personalization and privacy protection. The base brain layer provides a powerful general conversation generation ability, while the user brain layer focuses on the processing of personalized data and privacy protection. This architecture design enables the system to provide a highly personalized interaction experience while ensuring user privacy. Users can customize tags and desensitization rules through the user brain layer according to their own needs and preferences, so as to flexibly adjust the privacy protection strategy in different social scenarios.

[0049] 2. Through the dynamic desensitization model and optimization algorithm, the present invention significantly improves the flexibility and accuracy of desensitization processing. The system dynamically adjusts the desensitization strategy according to the identity, emotional intimacy of the conversation object, and the sensitivity of the topic. This desensitization method based on multi-dimensional decision logic can more accurately identify and process sensitive information, avoiding the limitations of traditional static desensitization technologies.

[0050] 3. The present invention adopts privacy-enhancing technologies such as federated learning and homomorphic encryption, significantly enhancing the data security and privacy protection capabilities of the system. Among them, federated learning allows the user's brain layer to perform small-sample training on local devices without uploading the original data to the base brain layer, thus maximizing the protection of user data privacy. At the same time, homomorphic encryption technology ensures the non-decryptability of sensitive tags during transmission and processing, further guaranteeing data security. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is the overall interaction flowchart of the present invention;

[0052] Figure 2 is the dynamic desensitization processing flowchart of the present invention;

[0053] Figure 3 is the training flowchart of the user's brain layer of the present invention;

[0054] Figure 4 is the desensitization flowchart of the proxy decision layer of the present invention;

[0055] Figure 5 is the hierarchical architecture design flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0057] Embodiment 1:

[0058] Please refer to the attached Figure 1 - attached Figure 5 , the embodiment of the present invention provides a dynamic desensitization digital avatar interaction method based on a hierarchical architecture, including the following steps:

[0059] The base brain layer generates general conversation content, generating conversation content containing general information based on a large-scale pre-trained language model;

[0060] Specifically, the base brain layer is the core of the entire system and is responsible for generating general conversation content. It generates conversation content containing general information based on a large-scale pre-trained language model.

[0061] Using large-scale pre-trained language models (such as DeepSeek-R1) as the basis, these models are trained on a large amount of text data and can generate high-quality dialogue content. It supports 1024-token context to ensure the coherence and integrity of the dialogue. The generated dialogue content contains general information, such as answering common questions like salary and address.

[0062] The user brain layer processes personalized data, stores the identity information of the dialogue object, emotional intimacy, sensitive topic tags, and marks the dialogue content according to user-defined rules.

[0063] The user brain layer supports the user to dynamically adjust tags and rules through a visual interface to adapt to different dialogue scenarios and user needs, specifically including:

[0064] Provide a user-friendly visual interface that allows users to customize tags and rules.

[0065] Support users to update tags and rules in real time, and the system responds immediately and adjusts the desensitization strategy.

[0066] The user brain layer conducts local training through federated learning, and the training data includes:

[0067] Local text data of the user's historical conversations.

[0068] Pairs of tags and rules annotated by the user through the visual interface.

[0069] Conversation correction data feedback by the user in real time, which is used to continuously optimize the model of the user brain layer.

[0070] The base brain layer and the user brain layer conduct data interaction through the API interface to ensure the modularity and scalability of the system, specifically including:

[0071] The base brain layer transmits the generated general dialogue content to the proxy decision layer through the API interface.

[0072] Specifically, the base brain layer is responsible for generating general dialogue content, which is generated based on large-scale pre-trained language models and contains general information. Through the API interface, the base brain layer transmits this generated content to the proxy decision layer, ensuring the standardization and efficiency of data transmission. This modular design enables the base brain layer to focus on the generation of dialogue content without having to handle subsequent desensitization logic, thereby improving the overall efficiency and maintainability of the system.

[0073] The user brain layer transmits tags and rules to the proxy decision layer through the API interface.

[0074] Specifically, the user's brain layer stores the user's personalized data, including the identity information of the conversation partner, emotional intimacy, sensitive topic tags, and the desensitization rules defined by the user. Through the API interface, the user's brain layer transmits these tags and rules to the proxy decision-making layer. This design not only ensures the privacy of user data but also enables the user to flexibly adjust the desensitization strategy to adapt to different conversation scenarios and requirements. The use of the API interface ensures the security and reliability of data transmission while maintaining the modularity and scalability of the system.

[0075] The proxy decision-making layer returns the desensitized conversation content to the user side through the API interface.

[0076] Specifically, the proxy decision-making layer receives the general conversation content from the base brain layer and the tags and rules from the user's brain layer, and performs dynamic desensitization processing based on this information. After the processing is completed, the proxy decision-making layer returns the desensitized conversation content to the user side through the API interface. This design ensures the flexibility and accuracy of the desensitization process, and at the same time, through the standardized transmission of the API interface, it guarantees the modularity and scalability of the system. Users can obtain the conversation content that has been privacy-protected immediately, thus enjoying a high-quality interaction experience while ensuring privacy.

[0077] The proxy decision-making layer performs dynamic desensitization, dynamically desensitizes the conversation content generated by the base brain layer according to the tags and rules of the user's brain layer, and outputs the desensitized conversation content;

[0078] The proxy decision-making layer uses homomorphic encryption technology to process sensitive tags to ensure the privacy of data during transmission and processing, specifically including:

[0079] Using the Paillier encryption algorithm to encrypt sensitive tags;

[0080] Decrypting and processing the encrypted sensitive tags in the proxy decision-making layer to ensure that the data is available but invisible.

[0081] Specifically, the proxy decision-making layer uses homomorphic encryption technology to process sensitive tags to ensure the privacy of data during transmission and processing. This technology allows data to be processed in an encrypted state, ensuring the availability of the data while keeping it invisible, thus effectively protecting user privacy.

[0082] Using the Paillier encryption algorithm to encrypt sensitive tags. The Paillier encryption algorithm is an additive homomorphic encryption algorithm that supports addition operations and scalar multiplication operations on ciphertexts. Specifically, the Paillier algorithm has the following properties:

[0083] Additive homomorphism: If E(x) and E(y) are two ciphertexts, then E(x)·E(y) = E(x + y).

[0084] Scalar multiplication homomorphism: If E(x) is a ciphertext and k is a plaintext constant, then E(x) k = E(k·x).

[0085] Before transmitting the sensitive label to the proxy decision layer, the user's brain layer encrypts the sensitive label using the Paillier encryption algorithm. The specific steps are as follows:

[0086] Key generation: The user's brain layer generates a pair of public and private keys. The public key is used for encryption, and the private key is used for decryption.

[0087] Encryption operation: Encrypt the sensitive label m to generate the ciphertext c = E(m).

[0088] Transmit the ciphertext: Transmit the encrypted ciphertext c to the proxy decision layer through the API interface.

[0089] Example: Assume the sensitive label m = 0.7 (representing the sensitive value of the salary topic). After encryption using the Paillier encryption algorithm, the ciphertext c is obtained. The user's brain layer transmits c to the proxy decision layer.

[0090] The proxy decision layer fuses the original content of the base brain layer and the labels and rules of the user's brain layer through the multi-head attention mechanism, and outputs the desensitized content, specifically including:

[0091] Use the multi-head attention mechanism to perform weighted fusion of the original content and the markings;

[0092] Dynamically adjust the desensitization strategy according to the fusion result to ensure the accuracy and naturalness of the desensitized content.

[0093] Dynamic desensitization processing includes:

[0094] Adjust the desensitization strategy based on the identity label, and dynamically adjust the desensitization intensity according to the identity of the conversation object;

[0095] Adjust the desensitization strategy based on the sensitive value, and dynamically adjust the desensitization method according to the sensitive value of the sensitive topic, including entity replacement, content deletion, or fuzzification processing;

[0096] The dynamic desensitization strategy is optimized through equations and matrix operations to achieve a more flexible desensitization decision logic, specifically including:

[0097] Define the desensitization decision function, and dynamically adjust the desensitization strategy according to the identity of the conversation object, the topic type, and the emotional intimacy;

[0098] Use matrix operations to optimize the calculation process of the desensitization decision function, and improve the response speed and processing efficiency of the system.

[0099] Specifically, according to the identity of the conversation partner (such as colleagues, family members, strangers, etc.), dynamically adjust the intensity of the desensitization strategy. Different identities of conversation partners require different levels of privacy protection, so the system can flexibly adjust the desensitization strategy according to the identity tags.

[0100] Dynamically adjust the desensitization method according to the sensitivity value of the sensitive topic, including entity replacement, content deletion or obfuscation. The sensitivity value reflects the privacy level of the topic, and the system selects the appropriate desensitization method according to the size of the sensitivity value.

[0101] By defining a desensitization decision function and using matrix operations to optimize the desensitization decision logic, a more flexible desensitization strategy can be achieved. This optimization method can improve the response speed and processing efficiency of the system. Use matrix operations to efficiently calculate the desensitization decision function and improve the response speed and processing efficiency of the system.

[0102] Specifically, the system processes sensitive topics by dynamically adjusting the desensitization method. The specific methods include entity replacement, content deletion or obfuscation. The selection of these desensitization methods is based on the size of the sensitivity value, which is a quantitative indicator used to reflect the privacy level of the topic. For example, when the sensitivity value is low, only slight obfuscation is required; while when the sensitivity value is high, more strict entity replacement or content deletion is needed. To achieve this flexible desensitization strategy, the system defines a desensitization decision function. This function dynamically adjusts the desensitization strategy according to factors such as the identity of the conversation partner, topic type, and emotional intimacy. In this way, the system can select the most appropriate desensitization method according to different situations and user needs. To further improve the response speed and processing efficiency of the system, the present invention uses matrix operations to optimize the desensitization decision logic. Matrix operations can efficiently process a large amount of data, making the calculation process of the desensitization decision function faster. This optimization method not only improves the performance of the system but also ensures the stability and reliability of the system under high load conditions.

[0103] Adjust the desensitization strategy based on emotional intimacy, which is used to dynamically adjust the desensitization strategy according to the emotional intimacy between the conversation partner and the user to adapt to different social relationships.

[0104] Application of privacy enhancement technology, using federated learning for local training to ensure user data privacy, and processing sensitive tags through homomorphic encryption technology to ensure data privacy during transmission and processing.

[0105] It also includes supporting multi-language conversations and dynamically adjusting the desensitization strategy according to language characteristics to meet the privacy protection needs in different language environments, specifically including:

[0106] Supporting the generation of conversations and desensitization processing in multiple languages;

[0107] Dynamically adjust the desensitization strategy according to the grammar and cultural characteristics of different languages to ensure the effectiveness of privacy protection.

[0108] A dynamic desensitization digital avatar interaction system based on a hierarchical architecture, comprising:

[0109] The base brain layer, which contains a large-scale pre-trained language model for generating general conversation content;

[0110] Specifically, the base brain layer is the underlying architecture of the system and contains a large-scale pre-trained language model. Its main function is to generate general conversation content and provide the basic language generation ability for the entire system.

[0111] The user brain layer, which contains a three-dimensional tag library and a rule engine for storing the identity information, emotional intimacy, and sensitive topic tags of the conversation object, and marking the conversation content according to user-defined rules;

[0112] Specifically, the user brain layer is the personalized module of the system and contains a three-dimensional tag library and a rule engine. Its main function is to store the identity information, emotional intimacy, and sensitive topic tags of the conversation object, and mark the conversation content according to user-defined rules.

[0113] The proxy decision layer is used to perform dynamic desensitization processing on the conversation content generated by the base brain layer according to the marks and rules of the user brain layer, and output the desensitized conversation content;

[0114] The privacy enhancement module uses federated learning for local training to ensure user data privacy, and processes sensitive tags through homomorphic encryption technology to ensure the privacy of data during transmission and processing;

[0115] The multilingual support module supports multilingual conversations and dynamically adjusts the desensitization strategy according to language characteristics to meet the privacy protection requirements in different language environments.

[0116] Through the hierarchical architecture design, the present invention realizes an efficient, flexible, and privacy-protected digital avatar interaction system. The base brain layer provides the general conversation generation ability, the user brain layer processes personalized data, the proxy decision layer performs dynamic desensitization, the privacy enhancement module ensures data privacy, and the multilingual support module adapts to different language environments. This hierarchical architecture not only improves the modularity and scalability of the system, but also significantly enhances the user experience and privacy protection level.

[0117] To better demonstrate the practical application effects of the dynamic desensitization digital avatar interaction method based on a hierarchical architecture, the following uses specific embodiments to elaborate in detail on the operation mechanism and advantages of the present invention in different scenarios. These embodiments cover common social communication scenarios, including salary inquiries among colleagues, address inquiries between strangers, and health status exchanges among family members. Each embodiment follows the same processing flow, including the generation of general conversation content in the base brain layer, the processing of personalized data in the user brain layer, the execution of dynamic desensitization in the proxy decision layer, and the application of privacy enhancement technologies, and finally outputs conversation content that meets the privacy protection requirements.

[0118] Embodiment 2: Scenario of a colleague asking about salary

[0119] Scenario description: When the user is communicating with a colleague, they are asked about their salary last month.

[0120] Input dialogue: "How much was your salary last month?"

[0121] The processing flow is as follows:

[0122] The base brain layer generates general conversation content:

[0123] The base brain layer generates general conversation content based on a large-scale pre-trained language model (such as DeepSeek-R1): "My salary last month was 20,000 yuan."

[0124] The user brain layer processes personalized data:

[0125] Identity label recognition: Identify the conversation object as "colleague" through the contact list API.

[0126] Sensitive value calculation: According to the topic type "salary", calculate the sensitive value S = 0.7.

[0127] Rule matching: According to the user-defined rule (colleague + salary → partial desensitization), mark that this conversation content needs to be partially desensitized.

[0128] The proxy decision layer executes dynamic desensitization:

[0129] Entity replacement: Replace "20,000 yuan" with "XX yuan".

[0130] Supplementary note: Add "at the industry average level".

[0131] Application of privacy enhancement technologies:

[0132] Use federated learning for local training to ensure user data privacy.

[0133] Encrypt sensitive labels through the Paillier encryption algorithm to ensure data privacy during transmission and processing.

[0134] Output dialogue: "My salary last month was XX yuan, which is at the industry average level."

[0135] Example 2: Scenario of a stranger asking for the address

[0136] Scenario description: When the user is communicating with a stranger, they are asked about their residential address.

[0137] Input dialogue: "Where do you live?"

[0138] The processing flow is as follows:

[0139] The base brain layer generates general dialogue content:

[0140] The base brain layer generates general dialogue content based on a large-scale pre-trained language model: "I live in Chaoyang District, Beijing."

[0141] The user brain layer processes personalized data:

[0142] Identity label recognition: Recognize that the dialogue object is a "stranger".

[0143] Sensitivity value calculation: According to the topic type "address", calculate the sensitivity value S = 0.95.

[0144] Rule matching: According to the user-defined rule (stranger + address → complete desensitization), mark that this dialogue content needs to be completely desensitized.

[0145] The proxy decision layer performs dynamic desensitization:

[0146] Entity deletion: Remove "Chaoyang District, Beijing".

[0147] Fuzzy reply: Add "Live in the city".

[0148] Application of privacy-enhancing technologies:

[0149] Use federated learning for local training to ensure user data privacy.

[0150] Encrypt sensitive labels through the Paillier encryption algorithm to ensure data privacy during transmission and processing.

[0151] Output dialogue: "I live in the city."

[0152] Example 3: Scenario of a family member asking about health status

[0153] Scenario description: When the user is communicating with a family member, they are asked about their health status.

[0154] Input dialogue: "How are you feeling lately?"

[0155] The processing flow is as follows:

[0156] The base brain layer generates general conversation content:

[0157] The base brain layer generates general conversation content based on a large-scale pre-trained language model: "I haven't been feeling well lately. I have a bit of a cold."

[0158] The user brain layer processes personalized data:

[0159] Identity label recognition: Identifies the conversation object as "family member" through the contact list API.

[0160] Sensitivity value calculation: Based on the topic type "health", calculates the sensitivity value S = 0.2.

[0161] Rule matching: According to the user-defined rule (family member + health → no desensitization), marks that the conversation content does not need to be desensitized.

[0162] The proxy decision layer performs dynamic desensitization:

[0163] Since the sensitivity value is low and the conversation object is a family member, directly outputs the original content.

[0164] Application of privacy-enhancing technology: Uses federated learning for local training to ensure user data privacy.

[0165] Encrypts sensitive labels through the Paillier encryption algorithm to ensure data privacy during transmission and processing.

[0166] Outputs the conversation: "I haven't been feeling well lately. I have a bit of a cold."

[0167] To more intuitively demonstrate the actual effects and advantages of the dynamic desensitization digital avatar interaction method based on a hierarchical architecture, the following uses two tables to conduct a detailed comparison of the desensitization effects in different scenarios and the improvement of privacy protection and user experience. These data not only verify the significant improvement of the present invention in terms of privacy protection and user experience, but also demonstrate its superiority in multi-language support and system response speed.

[0168]

[0169]

[0170] Table 1 (Comparison of desensitization effects in different scenarios)

[0171]

[0172] Table 2 (Data on privacy protection and user experience improvement)

[0173] Conclusion: As can be seen from the data in Table 1 and Table 2, the present invention is superior to traditional methods in terms of privacy protection, user experience, system response speed, and multilingual support.

[0174] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic desensitization digital twin interaction method based on a hierarchical architecture, characterized in that It includes the following steps: The base brain layer generates general conversation content, generating conversation content containing general information based on a large-scale pre-trained language model; The user brain layer processes personalized data, stores the identity information, emotional intimacy, and sensitive topic tags of the conversation object, and marks the conversation content according to user-defined rules; The proxy decision layer performs dynamic desensitization, dynamically desensitizing the conversation content generated by the base brain layer according to the marks and rules of the user brain layer, and outputting the desensitized conversation content; Application of privacy-enhancing technologies, using federated learning for local training to ensure user data privacy, and processing sensitive tags through homomorphic encryption technology to ensure data privacy during transmission and processing.

2. The dynamic desensitization digital avatar interaction method based on a hierarchical architecture according to claim 1, characterized in that The dynamic desensitization processing includes: Desensitization strategy adjustment based on identity tags, dynamically adjusting the desensitization intensity according to the identity of the conversation object; Desensitization strategy adjustment based on sensitive values, dynamically adjusting the desensitization method according to the sensitive values of sensitive topics, including entity replacement, content deletion, or obfuscation processing; Desensitization strategy adjustment based on emotional intimacy, used to dynamically adjust the desensitization strategy according to the emotional intimacy between the conversation object and the user to adapt to different social relationships.

3. The dynamic desensitization digital twin interaction method based on a hierarchical architecture according to claim 1, wherein The user brain layer conducts local training through federated learning, and the training data includes: Local text data of the user's historical conversations; Label and rule pairs annotated by the user through the visualization interface; Conversation correction data feedback by the user in real time, used to continuously optimize the model of the user brain layer.

4. The dynamic desensitization digital avatar interaction method based on a hierarchical architecture according to claim 1, wherein, The proxy decision layer uses homomorphic encryption technology to process sensitive tags to ensure data privacy during transmission and processing, specifically including: Using the Paillier encryption algorithm to encrypt sensitive tags; Decrypting and processing the encrypted sensitive tags at the proxy decision layer to ensure that the data is available but invisible.

5. The dynamic desensitization digital avatar interaction method based on a hierarchical architecture according to claim 1, characterized in that, The dynamic desensitization strategy is optimized through equations and matrix operations to achieve a more flexible desensitization decision logic, specifically including: Defining a desensitization decision function to dynamically adjust the desensitization strategy according to the identity, topic type, and emotional intimacy of the conversation object; Using matrix operations to optimize the calculation process of the desensitization decision function to improve the response speed and processing efficiency of the system.

6. The dynamic desensitization digital avatar interaction method based on a hierarchical architecture according to claim 1, characterized in that, The user brain layer supports the user to dynamically adjust tags and rules through the visualization interface to adapt to different conversation scenarios and user needs, specifically including: Providing a user-friendly visualization interface that allows users to customize tags and rules; Supporting users to update tags and rules in real time, and the system responds immediately and adjusts the desensitization strategy.

7. The dynamic desensitization digital avatar interaction method based on a hierarchical architecture according to claim 1, wherein The base brain layer and the user brain layer conduct data interaction through the API interface to ensure the modularity and scalability of the system, specifically including: The base brain layer transmits the generated general conversation content to the proxy decision layer through the API interface; The user brain layer transmits the marks and rules to the proxy decision layer through the API interface; The proxy decision layer returns the desensitized conversation content to the user side through the API interface.

8. The dynamic desensitization digital avatar interaction method based on a hierarchical architecture according to claim 1, wherein The proxy decision layer fuses the original content of the base brain layer and the tags and rules of the user brain layer through the multi-head attention mechanism, and outputs the desensitized content, specifically including: Using the multi-head attention mechanism to perform weighted fusion of the original content and the marks; Dynamically adjust the desensitization strategy according to the fusion result to ensure the accuracy and naturalness of the desensitized content.

9. The method for dynamic desensitization digital avatar interaction based on a hierarchical architecture according to claim 1, characterized in that, It also includes supporting multi-language conversations and dynamically adjusting the desensitization strategy according to language characteristics to meet the privacy protection requirements in different language environments, specifically including: Supporting the generation and desensitization processing of conversations in multiple languages; Dynamically adjusting the desensitization strategy according to the grammar and cultural characteristics of different languages to ensure the effectiveness of privacy protection.

10. A dynamic desensitization digital avatar interaction system based on a hierarchical architecture, which is applied to the dynamic desensitization digital avatar interaction method based on the hierarchical architecture according to any one of claims 1-9, and is characterized in that, Including: The base brain layer, which contains a large-scale pre-trained language model for generating general conversation content; The user brain layer, which contains a three-dimensional tag library and a rule engine for storing the identity information, emotional intimacy, and sensitive topic tags of the conversation object, and marking the conversation content according to user-defined rules; The proxy decision layer for dynamically desensitizing the conversation content generated by the base brain layer according to the marks and rules of the user brain layer and outputting the desensitized conversation content; The privacy enhancement module, which uses federated learning for local training to ensure user data privacy, and processes sensitive tags through homomorphic encryption technology to ensure the privacy of data during transmission and processing; The multi-language support module, which supports multi-language conversations and dynamically adjusts the desensitization strategy according to language characteristics to meet the privacy protection requirements in different language environments.

Images