Interaction system based on user digital avatar and privacy protection method

Abstract

The invention relates to the technical field of artificial intelligence and social computing, and discloses a user-based digital duplicate interaction system and privacy protection method.The user-based digital duplicate interaction system comprises a dynamic personality modeling engine used for extracting behavior features, emotion features and personality features of a user through a multi-source heterogeneous data fusion framework and generating a five-dimensional personality feature matrix; the emotion migration training module adopts a double-loop adversarial training framework, comprises a main loop network and an adversarial network, and is used for generating an emotion response mode; and the multi-modal interaction control module can realize a three-layer decision architecture, including semantic understanding, emotion calculation and decision generation, and is used for processing interaction between the user and the digital clones. Through adoption of a differential privacy technology, privacy protection is carried out on interaction data between a user and a digital avatar. The differential privacy ensures that the influence of single user data on a statistical result is limited within a certain range by adding a proper amount of noise in the data, so that the privacy of the user is protected.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and social computing, and specifically provides an interactive system based on user digital avatars and a privacy protection method. Background Art

[0002] In the current social network and digital interaction environment, the interaction between users and digital avatars is becoming increasingly frequent. Especially in multi-user scenarios, users' demand for privacy protection is becoming more urgent. However, there are many deficiencies in the existing technology in terms of privacy protection, making it difficult to meet users' high requirements for security and privacy.

[0003] Firstly, traditional identity authentication methods rely on combinations of usernames and passwords, or two-factor authentication mechanisms. Although these methods can verify user identities to a certain extent, they cannot effectively hide users' real identity information, making it easy for identity information to be leaked. For example, usernames and passwords can be stolen through cyberattacks or data leaks, and although two-factor authentication increases security, it still cannot completely avoid the risk of identity information being maliciously exploited. In addition, these methods also cannot support the secure binding between users and digital avatars, and cannot ensure the legality and credibility of the interaction process.

[0004] Secondly, in terms of data privacy protection, the existing technology uses data encryption methods. Although encryption technology can protect the confidentiality of data to a certain extent, after the data is decrypted, users' privacy information may still be exposed. Especially in multi-user interaction scenarios, the sharing and analysis of data will lead to an increased risk of privacy leakage. In addition, the existing technology lacks a real-time monitoring mechanism for users' behavior patterns and cannot detect and prevent abnormal behaviors in a timely manner, thus unable to effectively prevent privacy leakage.

[0005] Therefore, those skilled in the art have proposed an interactive system based on user digital avatars and a privacy protection method to solve the above problems. Summary of the Invention

[0006] Aiming at the deficiencies of the existing technology, the present invention provides an interactive system based on user digital avatars and a privacy protection method, which solves the problems raised in the above background art.

[0007] To achieve the above objectives, the present invention is realized through the following technical solutions: An interactive system based on user digital avatars, comprising:

[0008] A dynamic personality modeling engine, used to extract users' behavioral features, emotional features, and personality features through a multi-source heterogeneous data fusion architecture and generate a five-dimensional personality feature matrix;

[0009] The emotional transfer training module, adopting a dual-loop adversarial training framework, includes a main loop network and an adversarial network, and is used to generate emotional response patterns;

[0010] The multimodal interaction control module can implement a three-layer decision-making architecture, including semantic understanding, emotional computing, and decision-making generation, and is used to process the interaction between the user and the digital avatar;

[0011] The intelligent expression generation module generates emoji symbols that match the user's emotional state through a symbolic semantic mapping algorithm;

[0012] The cross-platform interaction protocol supports end-to-end encryption and multiple data formats, and is used to achieve cross-platform interaction between the user and the digital avatar;

[0013] The self-evolution management module deploys an online learning mechanism and realizes the self-evolution of the system through a dual-version model pool and model update conditions.

[0014] Preferably, the dynamic personality modeling engine includes:

[0015] The text feature extraction module adopts a BERT-LSTM network, where the hidden layer dimension ≥ 768;

[0016] The social graph analysis module adopts a graph attention network, where the number of heads ≥ 4;

[0017] The temporal behavior modeling module adopts a temporal convolutional network, where the number of channels ≥ 64.

[0018] Preferably, the main loop network of the emotional transfer training module includes 3 layers of Bi-LSTM, and the adversarial network adopts a WassersteinGAN architecture.

[0019] Preferably, the three-layer decision-making architecture of the multimodal interaction control system includes:

[0020] The semantic understanding layer, by adopting a BERT+CRF joint model;

[0021] The emotional computing layer, by adopting a 32-dimensional emotional state transition matrix;

[0022] The decision-making generation layer, by adopting Monte Carlo tree search.

[0023] Preferably, the intelligent expression generation module selects the most matching emoji symbols from the database through a symbolic semantic mapping algorithm, combining emotional intensity, semantic relevance, and context weight.

[0024] Preferably, the self-evolution management module realizes online learning through a dual-version model pool and model update conditions. When the amount of new data exceeds 1000 and the test accuracy rate increases by more than 2%, a model hot update is triggered.

[0025] A privacy protection method based on a user digital avatar interaction system, comprising:

[0026] An identity binding layer, which adopts an account association proof based on zk-SNARKs to ensure the anonymity and non-forgeability of the user identity;

[0027] A data isolation layer, which adopts differential privacy technology to protect the privacy of interaction data;

[0028] A behavior verification layer, which adopts a dynamic time warping algorithm to detect abnormal behaviors.

[0029] Preferably, the identity binding layer uses zero-knowledge proof technology to ensure the anonymity and non-forgeability of the user identity during the interaction process, and at the same time supports the secure binding between the user and the digital avatar.

[0030] Preferably, the behavior verification layer uses the DTW algorithm to monitor the user's behavior pattern in real time. When the similarity between the behavior pattern and the user's normal behavior pattern is lower than the threshold, the privacy protection mechanism is triggered to prevent privacy leakage.

[0031] The present invention provides a user digital avatar interaction system and a privacy protection method. It has the following beneficial effects:

[0032] 1. The present invention protects the privacy of the interaction data between the user and the digital avatar by adopting differential privacy technology. Differential privacy ensures that the impact of a single user's data on the statistical results is limited within a certain range by adding an appropriate amount of noise to the data, thereby protecting the user's privacy. Compared with traditional data encryption methods, differential privacy technology can continuously protect the user's privacy during the process of data analysis and use, without exposing the user information due to data decryption.

[0033] 2. The present invention realizes the anonymity and non-forgeability of the user identity by adopting an account association proof technology based on zk-SNARKs. During the interaction process between the user and the digital avatar, the user's real identity information is hidden, and only zero-knowledge proof is used for identity verification. This method not only protects the user's privacy but also effectively prevents the risk of malicious use of identity information. Compared with traditional identity verification methods, the verification method of the present invention can quickly and efficiently complete identity verification without revealing any user identity information, greatly enhancing the security of the system.

[0034] 3. The present invention uses the dynamic time warping algorithm to monitor the user's behavior pattern in real time, and can quickly detect abnormal situations in the user's behavior. When the similarity between the user's behavior pattern and the normal behavior pattern is lower than the set threshold, the system will immediately trigger the privacy protection mechanism to timely prevent potential privacy leakage risks. Compared with traditional static privacy protection methods, this method can dynamically adapt to changes in user behavior and protect the user's privacy in real time. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flowchart of the dynamic personality modeling engine of the present invention;

[0036] Figure 2 It is a flowchart of the emotion transfer training module of the present invention;

[0037] Figure 3 It is a flowchart of the privacy protection method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] 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 of 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.

[0039] Embodiment:

[0040] Please refer to the attached Figure 1 - attached Figure 3 , the embodiment of the present invention provides a user digital avatar interaction system, including:

[0041] A dynamic personality modeling engine for extracting the user's behavior features, emotional features, and personality features through a multi-source heterogeneous data fusion architecture and generating a five-dimensional personality feature matrix;

[0042] Specifically, the main function of the dynamic personality modeling engine is to extract the user's behavior features, emotional features, and personality features through a multi-source heterogeneous data fusion architecture and generate a five-dimensional personality feature matrix. This engine can comprehensively analyze various data sources of the user, thereby constructing a comprehensive and dynamic user personality model. This model provides a basis for the interaction of the digital avatar that is closer to the real user behavior and emotion, making the interaction more natural, personalized, and adaptable.

[0043] The dynamic personality modeling engine includes:

[0044] A text feature extraction module using a BERT-LSTM network, where the hidden layer dimension ≥ 768;

[0045] Specifically, BERT is a pre-trained language model based on the Transformer architecture, capable of learning the context representation of text. Through a bidirectional encoder structure, it generates context-related embedding vectors for each word or token in the input text. LSTM is a recurrent neural network structure that can effectively handle long-term dependencies in long sequence data. In this module, BERT is used to extract the context semantic information of the text, while LSTM is used to capture the temporal dependencies in the text sequence. A hidden layer dimension ≥ 768 enables the model to have sufficient capacity to learn complex text features.

[0046] The social graph analysis module uses a graph attention network, where the number of heads ≥ 4;

[0047] The social graph analysis module uses a graph attention network, where the number of heads ≥ 4. Its main function is to extract social graph features from the user's social relationship data and analyze the user's roles, influence, and relationship patterns in the social network.

[0048] GAT is a graph neural network based on the attention mechanism, capable of learning the representations of nodes in a graph. It assigns different weights to the neighbors of each node through the attention mechanism, thus better capturing the relationships between nodes. The formula of GAT is as follows:

[0049] e ij =LeakyReLU(a(Wh i ,Wh j ));

[0050]

[0051] Among them, e ij is the attention coefficient between node i and node j, a is the attention mechanism function, W is the learnable weight matrix, h i and h j are the feature vectors of node i and node j respectively, is the set of neighbor nodes of node i, α ij is the normalized attention weight, and σ is the activation function. In this module, the multi-head attention mechanism allows the model to learn the relationships between nodes from different perspectives, thus more comprehensively capturing the structural information of the social graph. The number of heads ≥ 4 means that the model can simultaneously learn the features of multiple attention heads, further improving the model's expressiveness.

[0052] The temporal behavior modeling module uses a temporal convolutional network, where the number of channels ≥ 64.

[0053] Specifically, the temporal behavior modeling module uses a temporal convolutional network. Its main function is to extract temporal behavior features from the user's behavior logs and analyze the user's behavior patterns and habits.

[0054] The dynamic personality modeling engine can extract user features from multiple dimensions and generate a five-dimensional personality feature matrix by integrating a text feature extraction module, a social graph analysis module, and a temporal behavior modeling module. This multi-source heterogeneous data fusion architecture enables the model to more comprehensively understand the user's behavior, emotions, and personality characteristics, providing a solid foundation for the interaction of digital avatars.

[0055] The emotion transfer training module, which adopts a dual-loop adversarial training framework including a main loop network and an adversarial network, is used to generate emotion response patterns.

[0056] The main loop network of the emotion transfer training module consists of 3 layers of Bi-LSTM, and the adversarial network adopts the WassersteinGAN architecture.

[0057] The dual-loop adversarial training framework is a training architecture that combines the idea of generative adversarial networks, aiming to improve the generation quality and authenticity of emotion response patterns through adversarial learning between the generator and the discriminator. This framework usually contains two main parts: the main loop network and the adversarial network.

[0058] The role of the main loop network is to generate emotion response patterns. In the present invention, the main loop network adopts 3 layers of Bi-LSTM (Bidirectional Long Short-Term Memory Network).

[0059] Bi-LSTM is a variant of LSTM that takes into account both the forward and backward information of the input sequence and can better capture the bidirectional dependencies in sequence data. Bi-LSTM is a variant of LSTM that takes into account both the forward and backward information of the input sequence and can better capture the bidirectional dependencies in sequence data.

[0060] The formula for Bi-LSTM is as follows:

[0061] Where, is the hidden state of the forward LSTM at time step t, is the hidden state of the backward LSTM at time step t, h t is the merged hidden state, and [;] represents the vector concatenation operation.

[0062] In the emotion transfer training module, the design of 3 layers of Bi-LSTM means that the input data will undergo 3 layers of such bidirectional processing. The output of each layer will be used as the input of the next layer, enabling a deeper extraction of emotion features and context information. This multi-layer structure helps to capture more complex patterns and relationships.

[0063] Specifically, the working process of the emotion transfer training module is as follows:

[0064] Input data: Input the user's emotional input data into the main recurrent network.

[0065] Emotional feature extraction: Extract emotional features through a 3-layer Bi-LSTM network to capture the emotional information and context relationships in the input data.

[0066] Generate emotional response patterns: According to the extracted emotional features, the generator generates emotional response patterns.

[0067] Discriminator evaluation: Input the generated emotional response patterns into the adversarial network, and the discriminator evaluates the quality of the generated emotional response patterns.

[0068] Feedback and optimization: According to the evaluation results of the discriminator, the generator adjusts its own parameters to generate higher-quality emotional response patterns. At the same time, the discriminator also adjusts its own parameters according to the generation results of the generator to better distinguish between real data and generated data.

[0069] Iterative training: Repeat the above steps until the emotional response patterns generated by the generator are close enough to the real emotional response patterns and the discriminator cannot distinguish between real data and generated data.

[0070] A multi-modal interaction control module that can implement a three-layer decision-making architecture, including semantic understanding, emotional computing, and decision-making generation, for handling interactions between users and digital avatars;

[0071] The three-layer decision-making architecture of the multi-modal interaction control system includes:

[0072] Semantic understanding layer, by adopting a BERT+CRF joint model;

[0073] Emotional computing layer, by adopting a 32-dimensional emotional state transition matrix;

[0074] Decision-making generation layer, by adopting Monte Carlo tree search.

[0075] Specifically, the multi-modal interaction control system is a core module in the user digital avatar interaction system, responsible for handling the complex interaction process between users and digital avatars. The system adopts a three-layer decision-making architecture, which is responsible for semantic understanding, emotional computing, and decision-making generation respectively. This hierarchical architecture can effectively process multi-modal inputs and generate reasonable interaction responses.

[0076] The semantic understanding layer is the first layer of the multimodal interaction control system. Its main function is to extract semantic information from the user's input and understand the user's intention and content. In the present invention, the semantic understanding layer adopts a combined BERT+CRF model. Among them, BERT is used to extract the context semantic information of the input text and generate high-quality semantic embedding vectors. CRF then performs sequence labeling based on the output of BERT, further refining the result of semantic understanding to ensure accurate parsing of the user's intention and content.

[0077] The emotion computing layer is the second layer of the multimodal interaction control system. Its main function is to analyze the user's emotional state based on the output of the semantic understanding layer and calculate the emotion transfer pattern. In the present invention, the emotion computing layer adopts a 32-dimensional emotion state transition matrix.

[0078] The emotion state transition matrix is a tool for modeling the dynamic changes of emotions. It represents the transition probabilities between different emotion states through a matrix. Suppose the set of emotion states is {E1, E2, …, E 32}, and each element E i of the emotion state transition matrix A represents the probability of transitioning from emotion state E1 to emotion state E j . The matrix A is represented by the following formula:

[0079]

[0080] where a ij ≥ 0 and The intelligent emoji generation module generates emoji symbols that match the user's emotional state through a symbolic semantic mapping algorithm. In the emotion computing layer, the system first determines the user's current emotional state based on the output of the semantic understanding layer, and then calculates the emotion transfer pattern through the emotion state transition matrix. This transfer pattern can help the system predict the dynamic changes of the user's emotions and generate corresponding emotional responses.

[0081] The decision-making generation layer is the third layer of the multimodal interaction control system. Its main function is to generate the final interaction decision based on the outputs of the semantic understanding layer and the emotion computing layer. In the present invention, the decision-making generation layer adopts Monte Carlo tree search. Monte Carlo tree search is a decision tree search algorithm based on the Monte Carlo method. By simulating a large number of interaction scenarios, it evaluates the advantages and disadvantages of different decision paths and selects the optimal interaction decision. This algorithm can effectively handle complex interaction environments and generate high-quality interaction responses.

[0082] The three - layer decision - making architecture of the multimodal interaction control system can effectively handle the interaction between users and digital avatars through the collaborative work of the semantic understanding layer, the emotion computing layer, and the decision - making generation layer. The semantic understanding layer extracts semantic information through the BERT+CRF joint model, the emotion computing layer analyzes the emotional dynamics through a 32 - dimensional emotion state transition matrix, and the decision - making generation layer generates the optimal interaction decision through Monte Carlo tree search. This hierarchical architecture can not only process multimodal inputs but also generate natural, accurate, and emotion - rich interaction responses, enhancing the user experience.

[0083] The intelligent expression generation module selects the most matching emojis from the database through the symbolic semantic mapping algorithm, combining emotional intensity, semantic relevance, and context weights.

[0084] The cross - platform interaction protocol, which supports end - to - end encryption and multiple data formats, is used to achieve cross - platform interaction between users and digital avatars.

[0085] The self - evolution management module deploys an online learning mechanism and realizes the self - evolution of the system through a dual - version model pool and model update conditions.

[0086] The self - evolution management module realizes online learning through a dual - version model pool and model update conditions. When the amount of new data exceeds 1000 and the test accuracy rate increases by more than 2%, it triggers a model hot update.

[0087] A privacy protection method for a user digital avatar interaction system includes:

[0088] The identity binding layer adopts the account association proof based on zk - SNARKs to ensure the anonymity and non - forgeability of user identities.

[0089] The identity binding layer ensures the anonymity and non - forgeability of user identities during the interaction process through zero - knowledge proof technology, and at the same time supports the secure binding between users and digital avatars.

[0090] The data isolation layer adopts differential privacy technology to protect the privacy of interaction data.

[0091] Differential privacy is a data privacy protection technology that ensures the impact of individual user data on statistical results is limited within a certain range by adding an appropriate amount of noise to the data. In this way, even if the data is analyzed and used, individual user privacy information cannot be inferred from the statistical results.

[0092] The behavior verification layer adopts the dynamic time warping algorithm to detect abnormal behaviors.

[0093] The behavior verification layer uses the DTW algorithm to monitor the user's behavior pattern in real time. When the similarity between the behavior pattern and the user's normal behavior pattern is lower than the threshold, the privacy protection mechanism is triggered to prevent privacy leakage.

[0094] DTW is an algorithm used to measure the similarity between two time series. It aligns the two time series through dynamic programming techniques to calculate the similarity between them. The system monitors the user's behavior pattern in real time through the DTW algorithm. It compares the user's current behavior pattern with the normal behavior pattern to determine whether there is an abnormality. If the similarity between the user's behavior pattern and the normal pattern is lower than the set threshold, the system will trigger the privacy protection mechanism. This mechanism can detect and prevent potential privacy leakage risks in a timely manner. For example, when it detects that the user's account has been illegally accessed or there are abnormal operations, the system can immediately take measures to protect the user's privacy.

[0095] To verify the actual application effect of the present invention in the user digital twin interaction system, we designed and implemented experiments in multiple scenarios. These experiments cover typical scenarios such as social networks and multi-user online collaboration platforms, aiming to verify the significant advantages of the present invention in aspects such as privacy protection, identity verification efficiency, abnormal behavior detection, and user experience by comparing the performance indicators of traditional methods and the methods of the present invention. The following are specific embodiments and experimental results.

[0096] Embodiment 2: Interaction between users and digital twins in social networks

[0097] In social networks, users interact with other users through digital twins. This embodiment verifies the effect of the present invention in protecting user privacy and improving the interaction experience.

[0098] Among them, the number of users is 1000 active users, and each user conducts an average of 50 interactions per day

[0099] Test period: 4 weeks.

[0100] The test metrics are as follows:

[0101] Privacy leakage risk: Detect privacy leakage through simulated attacks.

[0102] Identity verification efficiency: The time to verify the user's identity.

[0103] Abnormal behavior detection rate: The proportion of detected abnormal behaviors.

[0104] User satisfaction: Obtain the user's satisfaction with the interaction experience through a questionnaire survey.

[0105] The experimental results are shown in Table 1:

[0106] Indicator Traditional method Method of the present invention Improvement percentage Risk of privacy leakage 15％ 5％ 66.67％ Authentication efficiency (seconds) 3.5 1.0 71.43％ Abnormal behavior detection rate 45％ 80％ 77.78％ User satisfaction (out of 10) 6.5 8.0 23.08％

[0107] Table 1

[0108] Conclusion: As can be seen from Table 1, the privacy leakage risk of the traditional method is 15%, while the method of the present invention reduces this risk to 5%, with an improvement rate of 66.67%. Thus, through the differential privacy technology and the zk-SNARKs authentication mechanism, the user privacy can be effectively protected, the leakage of sensitive information can be prevented, and a more secure interaction environment can be provided for users. The average time taken for authentication by the traditional method is 3.5 seconds, while the method of the present invention only requires 1.0 second, and the authentication efficiency is increased by 71.43%. Thus, the authentication mechanism of the present invention is not only more secure, but also significantly improved in terms of efficiency, can quickly complete the user authentication, reduce the waiting time of users, and improve the user experience. The abnormal behavior detection rate of the traditional method is 45%, while the method of the present invention, through the dynamic time warping (DTW) algorithm, increases the detection rate to 80%, with an improvement rate of 77.78%. Thus, the present invention can more accurately identify and prevent abnormal behaviors, timely discover potential privacy leakage risks, and further ensure user privacy and system security.

[0109] Embodiment 2: Multi-user Online Collaboration Platform

[0110] Scenario description: In a multi-user online collaboration platform, users conduct team collaboration and project management through digital avatars. This embodiment verifies the privacy protection and interaction efficiency of the present invention in complex interaction scenarios. 2000 users conduct an average of 80 interactions per day.

[0111] Test period: 8 weeks

[0112] The test metrics are as follows:

[0113] Privacy leakage risk: Detect the privacy leakage situation through simulated attacks.

[0114] Authentication efficiency: The time for verifying the user identity.

[0115] Abnormal behavior detection rate: The proportion of detected abnormal behaviors.

[0116] Team collaboration efficiency: Evaluate the team collaboration efficiency through the project completion time and quality.

[0117] The experimental results are shown in Table 2 below:

[0118]

[0119] Table 2:

[0120] Conclusion: The privacy leakage risk of the traditional method is 18%, while the method of the present invention reduces this risk to 3%, with an improvement rate of 83.33%. Thus, through differential privacy technology and the zk-SNARKs authentication mechanism, it is possible to effectively protect user privacy, prevent the leakage of sensitive information, and provide a more secure interaction environment for users. The average time-consuming for authentication of the traditional method is 4.0 seconds, while the method of the present invention only requires 1.2 seconds, and the authentication efficiency is increased by 70%. This shows that the authentication mechanism is not only more secure but also significantly improved in terms of efficiency, capable of quickly completing user authentication, reducing the waiting time of users, and enhancing the user experience. The abnormal behavior detection rate of the traditional method is 50%, and the present invention increases the detection rate to 90% through the dynamic time warping algorithm, with an improvement rate of 80%. It can be inferred from this that the dynamic time warping algorithm can more accurately identify and prevent abnormal behaviors, timely discover potential privacy leakage risks, and further ensure user privacy and system security. The improvement in the team collaboration efficiency of the traditional method is 10%, while the method of the present invention increases this efficiency to 20%, with an improvement rate of 100%. Through the optimized multi-modal interaction control module and intelligent expression generation module of the present invention, the team collaboration efficiency is significantly improved, helping the team to complete projects faster and improving the overall work efficiency.

[0121] Through innovative technologies such as differential privacy technology, zk-SNARKs authentication, and dynamic time warping algorithm, the present invention effectively addresses the deficiencies in the prior art and provides a more secure, efficient, and reliable solution for the interaction between users and digital avatars.

[0122] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art 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 user digital avatar-based interaction system, characterized in that, Including: A dynamic personality modeling engine, which is used to extract the user's behavioral characteristics, emotional characteristics and personality characteristics through a multi-source heterogeneous data fusion architecture, and generate a five-dimensional personality characteristic matrix; An emotional transfer training module, which adopts a dual-loop adversarial training framework, including a main loop network and an adversarial network, and is used to generate an emotional response pattern; A multi-modal interaction control module, which can implement a three-layer decision-making architecture, including semantic understanding, emotional computing and decision-making generation, and is used to process the interaction between the user and the digital avatar; An intelligent expression generation module, which generates emoji symbols that match the user's emotional state through a symbolic semantic mapping algorithm; A cross-platform interaction protocol, which supports end-to-end encryption and multiple data formats, and is used to achieve cross-platform interaction between the user and the digital avatar; A self-evolution management module, which deploys an online learning mechanism and realizes the self-evolution of the system through a dual-version model pool and model update conditions.

2. The interactive system based on the user digital avatar according to claim 1, characterized in that, The dynamic personality modeling engine includes: A text feature extraction module, which adopts a BERT-LSTM network, where the hidden layer dimension ≥ 768; A social graph analysis module, which adopts a graph attention network, where the number of heads ≥ 4; A temporal behavior modeling module, which adopts a temporal convolutional network, where the number of channels ≥ 64.

3. The interactive system based on user digital avatars according to claim 1, characterized in that, The main loop network of the emotional transfer training module includes 3 layers of Bi-LSTM, and the adversarial network adopts a WassersteinGAN architecture.

4. The interactive system based on user digital avatar according to claim 1, wherein The three-layer decision-making architecture of the multi-modal interaction control system includes: A semantic understanding layer, which adopts a BERT+CRF joint model; An emotional computing layer, which adopts a 32-dimensional emotional state transition matrix; A decision-making generation layer, which adopts Monte Carlo tree search.

5. The interactive system based on user digital avatars according to claim 1, characterized in that, The intelligent expression generation module selects the most matching emoji symbols from the database through a symbolic semantic mapping algorithm, combining emotional intensity, semantic relevance and context weight.

6. The interactive system based on user digital avatar according to claim 1, wherein The self-evolution management module realizes online learning through a dual-version model pool and model update conditions. When the amount of new data exceeds 1000 and the test accuracy rate increases by more than 2%, it triggers a model hot update.

7. A privacy protection method for a user digital avatar interaction system, based on the user digital avatar interaction system according to any one of claims 1-6, characterized in that, Including: An identity binding layer, which adopts an account association proof based on zk-SNARKs to ensure the anonymity and non-forgeability of the user's identity; A data isolation layer, which adopts differential privacy technology to protect the privacy of interaction data; A behavior verification layer, which adopts a dynamic time warping algorithm to detect abnormal behaviors.

8. A privacy protection method for a user digital avatar interaction system according to claim 7, characterized in that, The identity binding layer ensures the anonymity and non-forgeability of the user's identity during the interaction process through zero-knowledge proof technology, and at the same time supports the secure binding between the user and the digital avatar.

9. A privacy protection method for a user digital twin interaction system according to claim 7, characterized in that The behavior verification layer monitors the user's behavior pattern in real time through the DTW algorithm. When the similarity between the behavior pattern and the user's normal behavior pattern is lower than the threshold, it triggers a privacy protection mechanism to prevent privacy leakage.

Images