An intelligent conversation privacy protection method, device, equipment and storage medium

Abstract

The application discloses a kind of intelligent conversation privacy protection method, device, equipment and storage medium, it is related to the field of intelligent conversation, including: determining the user authority of target user, and according to user authority, using target large language model determines the context authority vector corresponding to target user;Obtain the conversation that target user inputs through pre-set user interface, and respond to conversation using target large language model according to context authority vector, to generate conversation response;Generate the evaluation data corresponding to conversation response, and according to target data set, combine evaluation data to optimize target large language model.The online intelligent conversation system is designed using pre-training large language model, and by introducing context authority vector, the correspondence of user authority and enterprise data is realized, so as to better protect enterprise data in model, realize authorized access, reduce the risk of model data leakage, and combine user evaluation data, continuously optimize the performance and privacy protection ability of model.

Description

Technical Field

[0001] This invention relates to the field of intelligent dialogue, and in particular to a method, apparatus, device, and storage medium for protecting privacy in intelligent dialogue. Background Technology

[0002] With the rapid development and widespread application of large language models such as BERT (Bidirectional Encoder Representation from Transformers) and GPT-3 (Generative Pre-trained Transformer-3), enterprises and organizations hope to leverage these pre-trained models to build their own online intelligent dialogue systems and provide high-quality dialogue interaction experiences to external parties. These large language models possess powerful language understanding and generation capabilities, enabling them to generate human-like responses, making dialogue interactions more natural and fluent. However, to adapt these pre-trained models to specific tasks or domains, enterprises need to fine-tune and process their parameters, utilizing their own data to improve model performance and accuracy. Fine-tuning the model allows it to better adapt to the specific needs of the enterprise, providing more accurate dialogue responses. This parameter fine-tuning and processing process requires the use of the enterprise's own data, which may contain sensitive user information and private content.

[0003] Therefore, protecting user privacy becomes a crucial concern during parameter fine-tuning and processing. Enterprises and organizations want internal users to freely use pre-trained models for dialogue interaction, but they also do not want private data to be accessed or leaked by external users. However, existing privacy protection methods, such as data encryption and data anonymization, are often unsuitable for online intelligent dialogue scenarios with large language models. Traditional methods struggle to balance the high-quality interaction of dialogue systems with the need for user privacy protection. In this context, how to achieve privacy protection in intelligent dialogue, balance dialogue quality and user privacy, and simultaneously improve the flexibility and scalability of the model has become an urgent problem to be solved in this field. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method, apparatus, device, and storage medium for protecting privacy in intelligent dialogue. It utilizes a pre-trained large language model to design an online intelligent dialogue system, and by introducing a context permission vector, it achieves a mapping between user permissions and enterprise data, thereby better protecting enterprise data in the model, enabling authorized access, reducing the risk of model data leakage, and continuously optimizing the model's performance and privacy protection capabilities by combining user evaluation data. The specific solution is as follows:

[0005] Firstly, this application provides a method for protecting privacy in intelligent dialogue, including:

[0006] Determine the user permissions of the target user, and use the target large language model to determine the context permission vector corresponding to the target user based on the user permissions;

[0007] The system acquires the dialogue input by the target user through a preset user interaction interface, and uses the target large language model to respond to the dialogue based on the context permission vector to generate corresponding dialogue responses.

[0008] The evaluation data corresponding to the dialogue response is generated, and the target large language model is optimized based on the target dataset and the evaluation data.

[0009] Optionally, before determining the context permission vector corresponding to the target user using the target big oracle model based on the user permissions, the method further includes:

[0010] An initial large language model is constructed based on a pre-trained large language model, a privacy-adaptive network, a prompting learning module, and a gating unit network. The initial large language model is then trained using the target dataset to obtain the target large oracle model. The gating unit network is used to adjust the parameters of the privacy-adaptive network to control the sensitivity of the target large language model to input and output data.

[0011] Optionally, training the initial large language model using the target dataset includes:

[0012] Obtain preset internal private data of an enterprise, integrate the preset internal private data of the enterprise, classify it according to preset classification rules, and obtain a private dataset corresponding to the user permissions of the target user;

[0013] Obtain a public dataset, and construct the target dataset based on the private dataset and the public dataset, and train the initial large language model using the target dataset.

[0014] Optionally, responding to the dialogue using the target large language model based on the context permission vector includes:

[0015] Determine the data sensitivity level and data access permissions of the preset internal private data in the private dataset;

[0016] A response prompt template is generated based on the data sensitivity level and the data access permissions, and the target large language model is used to respond to the dialogue based on the context permission vector according to the response prompt template.

[0017] Optionally, the process of generating the evaluation data corresponding to the dialogue response further includes:

[0018] The interaction environment of the target large language model is simulated by a preset reinforcement learning simulation environment, and the reward value of the dialogue response is determined by the preset reinforcement learning simulation environment based on the dialogue response according to a preset reward function. The evaluation data of the dialogue response is generated based on the reward value.

[0019] Optionally, determining the reward value of the dialogue response based on the dialogue response according to a preset reward function using the preset reinforcement learning simulation environment includes:

[0020] Obtain the feedback result of the target user in response to the dialogue, construct the preset reward function based on the feedback result using a neural network, and generate the reward value corresponding to the dialogue response using the preset reward function.

[0021] Optionally, optimizing the target large language model based on the target dataset and the evaluation data includes:

[0022] The cumulative reward corresponding to the evaluation data is determined, the cumulative reward is maximized using the PPO reinforcement learning algorithm, and the model parameters of the target large language model are adjusted according to the cumulative reward.

[0023] Secondly, this application provides a smart dialogue privacy protection device, comprising:

[0024] The permission determination module is used to determine the user permissions of the target user, and to determine the context permission vector corresponding to the target user based on the user permissions using the target large language model;

[0025] The response generation module is used to acquire the dialogue input by the target user through a preset user interaction interface, and to respond to the dialogue using the target large language model according to the context permission vector to generate a corresponding dialogue response.

[0026] The model optimization module is used to generate evaluation data corresponding to the dialogue response, and optimize the target large language model based on the target dataset and the evaluation data.

[0027] Thirdly, this application provides an electronic device, which includes a processor and a memory; wherein the memory is used to store a computer program, which is loaded and executed by the processor to implement the aforementioned intelligent dialogue privacy protection method.

[0028] Fourthly, this application provides a computer-readable storage medium for storing a computer program that, when executed by a processor, implements the aforementioned intelligent dialogue privacy protection method.

[0029] In this application, the user permissions of the target user are first determined, and the context permission vector corresponding to the target user is determined using a target large language model based on the user permissions. Then, the dialogue input by the target user through a preset user interaction interface is obtained, and the target large language model responds to the dialogue based on the context permission vector to generate corresponding dialogue responses. Evaluation data corresponding to the dialogue responses is generated, and the target large language model is optimized based on the target dataset and the evaluation data. Through the above technical solution, this application can design an online intelligent dialogue system using a pre-trained large language model. Furthermore, by introducing a context permission vector, the mapping between user permissions and enterprise data is realized, thereby better protecting the enterprise data in the model, enabling authorized access, ensuring that only authorized personnel can access and use the data, reducing the risk of model data leakage, and balancing the high-quality interaction of the dialogue system with the user privacy protection requirements. In addition, by combining user evaluation data, the performance and privacy protection capabilities of the model are continuously optimized, providing a more intelligent and privacy-secure dialogue experience. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0031] Figure 1 A flowchart of a method for protecting privacy in intelligent dialogue provided in this application;

[0032] Figure 2 A schematic diagram of the constituent modules of an intelligent dialogue system provided in this application;

[0033] Figure 3 A flowchart for training an intelligent dialogue model is provided in this application;

[0034] Figure 4 A flowchart illustrating a specific method for protecting privacy in intelligent dialogue, as provided in this application;

[0035] Figure 5 A schematic diagram of a smart dialogue privacy protection device provided in this application;

[0036] Figure 6 This application provides a structural diagram of an electronic device. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] When utilizing online intelligent dialogue systems, enterprises want internal users to freely use pre-trained models for dialogue interaction, but they also do not want private data to be accessed or leaked by external users. However, existing privacy protection methods are often unsuitable for online intelligent dialogue scenarios with large language models, making it difficult to balance the high-quality interaction of the dialogue system with the user privacy protection requirements. This application discloses a hierarchical privacy protection method for online intelligent dialogue based on pre-trained large language models, involving the fields of intelligent dialogue, large language models, prompting learning, and privacy protection. By pre-training a large language model and introducing context permission vectors, authorized access is achieved, ensuring that only authorized personnel can access and use the data, reducing the risk of model data leakage, and balancing the high-quality interaction of the dialogue system with the user privacy protection requirements.

[0039] See Figure 1 As shown, this embodiment of the invention discloses a method for protecting privacy in intelligent dialogue, including:

[0040] Step S11: Determine the user permissions of the target user, and determine the context permission vector corresponding to the target user based on the user permissions using the target large language model.

[0041] In this embodiment, it is first necessary to determine the user permissions corresponding to the target user of the online intelligent dialogue system constructed in this embodiment, and then determine the corresponding context permission vector based on the determined user permissions using the target large language model, so as to use the context permission vector to map the user role's permissions to enterprise data.

[0042] It should be noted that the aforementioned online intelligent dialogue system was pre-built using a large language model in this embodiment, specifically as follows: Figure 2As shown, the aforementioned online intelligent dialogue system mainly consists of modules such as data storage and processing, dialogue management and control, user interaction interface / application open API (Application Programming Interface), privacy protection and access control. Its core is built around a large language model and deployed and run on the enterprise's IT (Internet Technology) infrastructure. The system uses the enterprise's private data to complete model training through methods such as model fine-tuning and prompt learning to provide a high-quality business dialogue interaction experience.

[0043] It is understandable that, such as Figure 3As shown, before determining the context permission vector corresponding to the target user using the target big oracle model based on user permissions, an initial big language model needs to be constructed based on a pre-trained big language model, a privacy-adaptive network, a prompting learning module, and a gating unit network. This initial big language model is then trained using the target dataset to obtain the target big oracle model. The aforementioned gating unit network is used to adjust the parameters of the privacy-adaptive network to control the sensitivity of the target big language model to input and output data. The aforementioned big language model is a neural network model constructed based on an existing Transformer architecture pre-trained natural language model. It is the core component of the online intelligent dialogue system in this embodiment, capable of understanding user-input questions and generating natural and accurate dialogue responses. It mainly consists of a pre-trained big language model (LLM), a privacy-adaptive network (Adapter), a prompting learning module (Prompt), and a gating unit network (GUM, specifically referring to the gating unit module constructed by the gating unit network in this embodiment, which can also be a GRU, Gate Recurrent Unit). Specifically, the aforementioned pre-trained large language model (LLM) is a language model pre-trained on large-scale text data, used to provide general language understanding and generation capabilities for various dialogue scenarios; the aforementioned privacy adaptation layer network (Adapter) is used to adapt and adjust the output of the large language model according to the user's permissions and privacy requirements, ensuring that the dialogue content meets the user's privacy requirements; the aforementioned gated unit network (GUM) is used to adjust the behavior and output of the adaptation layer network, controlling the dialogue output content according to the permission and privacy policy by adjusting the parameters of the privacy adaptation layer network; the aforementioned prompt learning module (Prompt) is used to learn and optimize the prompt generation strategy of the dialogue system. It can add permission prompt data to the preset private dataset data, and combine it with the preset external public dataset obtained by the dataset construction module. Using the prompt learning module and the aforementioned gated unit network, the parameters of the pre-trained large language model are fine-tuned based on the internal and external datasets. The focus is on adjusting the network parameters of the aforementioned gated unit network and the aforementioned privacy adaptation network (Adapter), so that by analyzing and learning from user feedback, dialogue history, etc., better dialogue prompts can be provided to improve dialogue effect and user satisfaction, while guiding the model to generate dialogue content that conforms to the user's privacy preferences. In this way, pre-trained large language models can be extended by designing privacy-adaptive networks and gating unit networks. This can fully leverage the generality and adaptability of the pre-trained model while maintaining the stability of the original structure, and at the same time enhance the model's personalization capabilities.

[0044] In this embodiment, the initial large language model needs to be trained using a target dataset. Therefore, it is necessary to acquire preset internal private data of the enterprise and integrate it to classify it according to preset classification rules, thereby obtaining a private dataset corresponding to the user permissions of the target user. Additionally, a public dataset needs to be acquired, and a target dataset needs to be constructed based on the private and public datasets to train the initial large language model. It is understood that the aforementioned internal private data represents the enterprise's internal data resources, including the enterprise's own dialogue data, knowledge base, product documents, etc.; the public dataset is constructed by collecting internet data and includes various public resources, such as online text and open knowledge bases. After acquiring the internal private data, the dataset construction module of the online intelligent dialogue system integrates and classifies the internal private data, and then classifies it according to data type, user permissions, and sensitivity, forming a dataset corresponding to the user's permission role. Then, the dataset construction module integrates and processes the enterprise's private data and external public data to generate a target dataset suitable for training the large language model. In this way, after acquiring the target dataset, the aforementioned online intelligent dialogue system utilizes the data storage and processing modules to store and process the dialogue data generated during the dialogue. This dialogue data includes, but is not limited to, user input data, dialogue history, and dialogue responses. Furthermore, the dialogue management and control module is responsible for managing and controlling the dialogue process, specifically including receiving user input, generating dialogue responses, and processing the dialogue context. It is understood that, to ensure the security of the enterprise's private data within the target dataset, the aforementioned privacy protection and access control module is needed to manage user data privacy protection and access control during the dialogue process. This ensures that only authorized personnel can access and use the data, monitors privacy risks during the dialogue process, promptly identifies and addresses potential privacy issues, and prevents the leakage of sensitive information. It is also understood that when users interact with the system, an interactive interface between the user and the intelligent dialogue system can be provided based on the user interface / application open API. Specifically, this interactive interface can be a graphical user interface (GUI), a mobile application, or a set of open APIs for other applications to call. In this way, users can interact with the system through this interface or API, send questions, receive responses, and control and manage the dialogue. In this way, by combining the aforementioned large language model with prompting learning techniques and pre-trained large language models, and adjusting data processing, training strategies, and access control permissions, hierarchical permission privacy protection for intelligent dialogue can be achieved. User role permissions are mapped to enterprise private data, and parameters are adjusted based on public datasets and enterprise private data to meet the personalized business requirements of enterprise intelligent dialogue. This approach can balance dialogue quality and user privacy, and improve the model's flexibility and scalability.

[0045] Through the above technical solution, this embodiment addresses the personalized business needs and data privacy protection requirements of enterprises' online intelligent dialogue. It utilizes existing pre-trained large language models to design an online intelligent dialogue system. By hierarchically classifying and distributing enterprise internal private data according to permissions and sensitive data, a core model for training the dialogue system is formed by combining a private dataset with an external public dataset. Furthermore, based on the pre-trained large language model, a privacy-adaptive network, a gating unit network, and a prompting learning module are designed to construct the core large language model infrastructure for the online intelligent dialogue system. This generates a core large language model with personalized and privacy-protected capabilities, enabling user-controllable online dialogue while ensuring the data privacy of the large language model is protected. In this way, enterprise private data can be protected, hierarchical access authorization can be implemented, and only authorized personnel can access and use the data, achieving hierarchical access permission privacy protection for intelligent dialogue.

[0046] Step S12: Obtain the dialogue input by the target user through a preset user interaction interface, and use the target large language model to respond to the dialogue according to the context permission vector to generate a corresponding dialogue response.

[0047] In this embodiment, when a user engages in intelligent dialogue, the dialogue input by the target user through a preset user interaction interface is acquired. The target large language model then responds to the dialogue based on a contextual permission vector to generate corresponding dialogue responses. It should be noted that when responding to the dialogue, the data sensitivity level and data access permissions of preset internal enterprise private data in the private dataset are first determined. Then, a response prompt template is generated based on the data sensitivity level and data access permissions. The target large language model then responds to the dialogue based on the contextual permission vector according to the response prompt template. It can be understood that by designing prompt templates using the dataset construction module for datasets with different sensitivity levels and access permissions, the correlation between the input and output of the target large language model can be indicated. This allows for a better understanding of the user's intent and context, while also serving as a privacy access restriction to comply with data privacy protection requirements. Furthermore, the prompt templates control the natural fluency of the dialogue interaction output, and data access permission prompts are designed to differentiate the privacy of internal and external datasets, thus balancing dialogue quality and user privacy.

[0048] Step S13: Generate evaluation data corresponding to the dialogue response, and optimize the target large language model based on the target dataset and the evaluation data.

[0049] In this embodiment, feedback data submitted by the user after evaluating the dialogue based on satisfaction and privacy can be obtained. Evaluation data corresponding to the dialogue response can then be generated, and the target large language model can be optimized based on the target dataset and the evaluation data. In this way, by continuously collecting internal and external datasets and combining them with user feedback data, the model's performance and privacy protection capabilities can be continuously optimized, providing a more intelligent and privacy-secure dialogue experience.

[0050] Through the above technical solution, this embodiment constructs an initial large language model based on a pre-trained large language model, a privacy-adaptive network, a prompting learning module, and a gating unit network. The initial large language model is trained on a private dataset and a public dataset to obtain a target large language model. The user permissions of the target user are determined, and the context permission vector corresponding to the target user is determined using the target large language model based on the user permissions. The dialogue input by the target user through a preset user interaction interface is obtained, and the data sensitivity level and data access permissions of preset internal private data in the private dataset are determined. A response prompt template is generated based on the data sensitivity level and data access permissions, and the target large language model responds to the dialogue based on the context permission vector according to the response prompt template to generate a corresponding dialogue response. Evaluation data corresponding to the dialogue response is generated, and the target large language model is optimized based on the target dataset and the evaluation data. In this way, prompting learning techniques and pre-trained large language models can be used to adjust data processing, training strategies, and access control permissions to achieve hierarchical permission privacy protection for intelligent dialogue. User role permissions are mapped to enterprise private data, and parameters are adjusted based on public datasets and enterprise private data to meet the personalized business requirements of enterprise intelligent dialogue. This can balance dialogue quality and user privacy, which is beneficial to the data privacy protection of large language models and improves the flexibility and scalability of the models.

[0051] As can be seen from the previous embodiment, this application can utilize a pre-trained large language model to design an online intelligent dialogue system. Next, this embodiment will elaborate on the optimization process of the aforementioned pre-trained large language model. See [link to previous embodiment]. Figure 4 As shown in the figure, this application discloses a specific method for protecting privacy in intelligent dialogue, including:

[0052] Step S21: Determine the user permissions of the target user, and determine the context permission vector corresponding to the target user based on the user permissions using the target large language model.

[0053] Step S22: Obtain the dialogue input by the target user through a preset user interaction interface, and use the target large language model to respond to the dialogue according to the context permission vector to generate a corresponding dialogue response.

[0054] Step S23: Simulate the interactive environment of the target large language model through a preset reinforcement learning simulation environment, and use the preset reinforcement learning simulation environment to determine the reward value of the dialogue response based on the dialogue response according to a preset reward function, and generate the evaluation data of the dialogue response based on the reward value.

[0055] In this embodiment, user-generated evaluation data can be used for model optimization. Alternatively, when the user is not using the system, a preset reinforcement learning simulation environment can be used to simulate the interaction environment of the target large language model. The preset reinforcement learning simulation environment then determines the reward value of the dialogue response based on a preset reward function, and generates evaluation data for the dialogue response based on the reward value. When determining the reward value of the dialogue response, the feedback from the target user is first obtained. Based on the feedback, a preset reward function is constructed using a neural network, and the reward value corresponding to the dialogue response is generated using the preset reward function.

[0056] It should be noted that this embodiment can utilize a pre-defined reinforcement learning (RL) module to optimize the pre-trained large language model. The aforementioned reinforcement learning module is as follows: Figure 3 As shown, the system mainly includes an online dialogue simulation environment, a reward function module (Reward), and a model optimization module (PPO, Proximal Policy Optimization), which are used for reinforcement learning training and improvement of the core online dialogue model. The aforementioned reinforcement learning simulation environment acts as the actual environment for interacting with the pre-trained language model, receiving dialogue responses generated by the pre-trained language model and calculating a reward value based on the reward function as quality feedback for the dialogue responses. The reward function is used to evaluate the quality of the generated dialogue responses. It is built based on a neural network and designed according to user feedback and multiple criteria. The reward function assigns a numerical value to each dialogue response, reflecting its quality and effectiveness. Understandably, the aforementioned user feedback mainly covers two aspects: the fluency and relevance of the dialogue, the satisfaction rating based on meeting user needs (commonly rated from 1 to 5, where 1 represents the worst and 5 represents the best), and the privacy rating (commonly rated levels include "yes," "no," and "maybe"). Therefore, combining the above embodiments, during model training, a combination of pre-trained large language model fine-tuning, cue learning, and reinforcement learning is used to generate a core large language model with personalized and privacy-preserving capabilities. This enables online dialogue with user-controllable permissions and ensures the data privacy of the large language model is protected.

[0057] Step S24: Determine the cumulative reward corresponding to the evaluation data, maximize the cumulative reward using the PPO reinforcement learning algorithm, and adjust the model parameters of the target large language model according to the cumulative reward.

[0058] In this embodiment, the parameters of the pre-trained large-scale language model are optimized through a model optimization module. The PPO reinforcement learning algorithm is employed to adjust and optimize the model's parameters by maximizing the cumulative reward of the dialogue responses generated by the pre-trained large-scale language model, thereby enabling it to generate better dialogue responses. This allows for the simulation of real-world dialogue scenarios, with testers scoring the satisfaction and privacy aspects of multiple rounds of dialogue results to form a reward function. The PPO algorithm is then used to train the large-scale language model for reinforcement learning, achieving hierarchical and isolated online dialogue privacy protection.

[0059] For more detailed processing procedures of steps S21 and S22, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0060] In this embodiment, after generating the corresponding dialogue response, the interactive environment of the target large language model is simulated through a preset reinforcement learning simulation environment. The reward value of the dialogue response is determined based on a preset reward function using this environment. Evaluation data of the dialogue response is generated based on the reward value, and then the cumulative reward corresponding to the evaluation data is determined. The PPO reinforcement learning algorithm is used to maximize the cumulative reward, and the model parameters of the target large language model are adjusted based on the cumulative reward. This approach combines pre-trained large language model fine-tuning, cue learning, and reinforcement learning. By using cue information and context vectors, user role permissions are mapped to enterprise private data. Parameter adjustments are made based on both public and enterprise private data to meet the personalized business requirements of intelligent dialogue for enterprises. Furthermore, the PPO algorithm is used to train the large language model's reinforcement learning, achieving hierarchical isolation of online dialogue privacy.

[0061] Based on the above embodiments, this application discloses a hierarchical privacy protection method for online intelligent dialogue based on a pre-trained large language model, used for training the core large language model of online intelligent dialogue, including:

[0062] Collect external datasets: Collect large-scale text datasets from the Internet, such as web pages, books, articles, papers, etc., through the dataset building module to form the aforementioned public dataset.

[0063] Private Data Classification and Grading: Using the dataset building module, text datasets within the enterprise are collected. Based on different departments, roles, and other dimensions, the data is divided into different sensitivity levels to form a sensitive data grading system within the enterprise. Furthermore, the data is divided according to access permissions to form the aforementioned private dataset.

[0064] Download an existing pre-trained large language model: On the Internet, select a suitable publicly available pre-trained large language model, download its model parameters, and use them as the initial parameter values ​​for the pre-trained large language model.

[0065] Design a privacy-adaptive network structure: Based on the enterprise's personalized needs and the characteristics of the dialogue task, the dataset construction module adjusts the network structure of the pre-trained large language model and adds a privacy-adaptive network adapter to better suit the needs of enterprise users.

[0066] Design prompt templates: For datasets with different levels of sensitivity and access permissions, the dataset construction module designs prompt data templates to indicate the relationship between input and output, better understand the user's intent and context, and also serve as privacy access restrictions to ensure compliance with data privacy protection requirements.

[0067] Large Language Model Parameter Fine-tuning: Permission prompt data is added to the private dataset, and combined with the external public dataset obtained by the dataset construction module, the parameters of the pre-trained large language model are fine-tuned based on the internal and external datasets using the prompt learning module and the gating unit network, with a focus on adjusting the network parameters of the gating unit network and the privacy adaptation network Adapter.

[0068] Test user feedback: Internal personnel participate in the testing process, score the satisfaction level based on the dialogue results, reflect the subjective evaluation of the dialogue quality, and determine whether the dialogue system has leaked data privacy beyond the user's permissions during the dialogue process, providing feedback in the form of "yes", "no" or "may". Based on the collected scores and privacy leakage feedback data, the training dataset construction module will train the reward function.

[0069] Reward function module network training: Based on the collected rating and privacy breach feedback data, the training dataset building module will train the reward function neural network to mimic real user feedback.

[0070] Reinforcement learning training: The large language model is used as the Action module in reinforcement learning. It generates responses based on the context and prompts of the current dialogue, enabling interaction with the reinforcement learning module. Reinforcement learning training is performed through a reward function, and the PPO algorithm is used to adjust the parameters of the large model to further improve the performance and privacy protection capabilities of the dialogue system.

[0071] Adjusting the parameters of the adaptation network: Fix the parameters of the basic large language model, generate prompt data according to the different permission requirements of users, and adjust the parameters of the privacy adaptation network through the gating unit network to dynamically control the sensitivity of the model to different inputs and outputs in order to meet personalized user permission control.

[0072] Generate permission context vector: Based on the user's permission settings, input the relevant sensitive datasets that their permissions allow into the large language model, generate a context vector and save it, which is used as a representation of their permission vector.

[0073] Based on the aforementioned technical solution and the actual privacy protection needs of enterprises in online intelligent dialogue, this system fully utilizes existing pre-trained large language models to design an online intelligent dialogue system. It employs a combination of large model fine-tuning, cue learning, and reinforcement learning, adjusting and optimizing parameters based on public datasets and enterprise private data to meet the personalized business needs and data privacy protection requirements of enterprises in online intelligent dialogue. Furthermore, by adjusting cue messages and gating unit networks, the model parameters are adaptively adjusted for dynamic permissions. Reinforcement learning is used for model training, and internal personnel conduct testing by setting two evaluation indicators: dialogue satisfaction and privacy. Feedback data is used to train a reward function neural network, completing the interaction with the large language model and parameter optimization updates. By inputting enterprise private data into a large language model with fixed parameters according to different permissions, a context vector is formed as the basic data context for user permissions. This context vector is then loaded in a targeted manner based on the actual dialogue, achieving data separation from the model and reducing the risk of data leakage.

[0074] See Figure 6 As shown in the embodiments, this application also discloses an intelligent dialogue privacy protection device, including:

[0075] The permission determination module 11 is used to determine the user permissions of the target user, and to determine the context permission vector corresponding to the target user based on the user permissions using the target large language model;

[0076] The response generation module 12 is used to acquire the dialogue input by the target user through a preset user interaction interface, and to respond to the dialogue using the target large language model according to the context permission vector to generate a corresponding dialogue response.

[0077] The model optimization module 13 is used to generate evaluation data corresponding to the dialogue response, and optimize the target large language model based on the target dataset and the evaluation data.

[0078] In this embodiment, the user permissions of the target user are first determined, and the context permission vector corresponding to the target user is determined using a target large language model based on the user permissions. Then, the dialogue input by the target user through a preset user interaction interface is obtained, and the target large language model responds to the dialogue according to the context permission vector to generate corresponding dialogue responses. Evaluation data corresponding to the dialogue responses is generated, and the target large language model is optimized based on the target dataset and the evaluation data. Through the above technical solution, this embodiment can design an online intelligent dialogue system using a pre-trained large language model. By introducing a context permission vector, the mapping between user permissions and enterprise data is realized, thereby better protecting the enterprise data in the model, realizing authorized access, ensuring that only authorized personnel can access and use the data, reducing the risk of model data leakage, and balancing the high-quality interaction of the dialogue system with the user privacy protection requirements. Furthermore, by combining user evaluation data, the performance and privacy protection capabilities of the model are continuously optimized, providing a more intelligent and privacy-secure dialogue experience.

[0079] In some specific embodiments, the intelligent dialogue privacy protection device further includes:

[0080] The model building module is used to construct an initial large language model based on a pre-trained large language model, a privacy adaptation network, a prompting learning module, and a gating unit network, and to train the initial large language model using the target dataset to obtain the target large oracle model; the gating unit network is used to adjust the parameters of the privacy adaptation network to control the sensitivity of the target large language model to input and output data.

[0081] In some specific embodiments, the model building module specifically includes:

[0082] The data acquisition unit is used to acquire preset private internal data of an enterprise, and to integrate the preset private internal data of the enterprise, and to classify it according to preset classification rules to obtain a private dataset corresponding to the user permissions of the target user.

[0083] The model training unit is used to acquire a public dataset, construct the target dataset based on the private dataset and the public dataset, and train the initial large language model using the target dataset.

[0084] In some specific embodiments, the response generation module 12 specifically includes:

[0085] The permission determination unit is used to determine the data sensitivity level and data access permissions of the preset internal private data in the private dataset.

[0086] The dialogue response unit is used to generate a response prompt template based on the data sensitivity level and the data access permissions, and to respond to the dialogue based on the target large language model according to the context permission vector based on the response prompt template.

[0087] In some specific embodiments, the model optimization module 13 further includes:

[0088] The reward value determination submodule is used to simulate the interactive environment of the target large language model through a preset reinforcement learning simulation environment, and use the preset reinforcement learning simulation environment to determine the reward value of the dialogue response based on the dialogue response according to a preset reward function, and generate the evaluation data of the dialogue response based on the reward value.

[0089] In some specific embodiments, the reward value determination submodule specifically includes:

[0090] The function construction unit is used to obtain the feedback result of the target user in response to the dialogue, construct the preset reward function based on the feedback result and the neural network, and generate the reward value corresponding to the dialogue response using the preset reward function.

[0091] In some specific embodiments, the model optimization module 13 specifically includes:

[0092] The parameter adjustment unit is used to determine the cumulative reward corresponding to the evaluation data, maximize the cumulative reward using the PPO reinforcement learning algorithm, and adjust the model parameters of the target large language model according to the cumulative reward.

[0093] Furthermore, embodiments of this application also disclose an electronic device, Figure 6 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0094] Figure 6 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the intelligent dialogue privacy protection method disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0095] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0096] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0097] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the intelligent dialogue privacy protection method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include a computer program capable of performing other specific tasks.

[0098] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned disclosed intelligent dialogue privacy protection method. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0099] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0100] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0101] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0102] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0103] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for intelligent conversation privacy protection, characterized in that, The method comprises the steps of: determining the user authority of the target user, and determining the context authority vector corresponding to the target user by using a target large language model according to the user authority; obtaining the dialogue input by the target user through a preset user interaction interface, and responding to the dialogue by using the target large language model according to the context authority vector to generate a corresponding dialogue response; generating evaluation data corresponding to the dialogue response, and optimizing the target large language model according to a target data set combined with the evaluation data; wherein, before determining the context authority vector corresponding to the target user by using the target large language model according to the user authority, the method further comprises the steps of: constructing an initial large language model based on a pre-trained large language model, a privacy adaptation network, a prompt learning module and a gating unit network, and training the initial large language model by using the target data set to obtain the target large language model; the gating unit network is used to adjust the parameters of the privacy adaptation network to control the sensitivity of the target large language model to input data and output data; and, the training of the initial large language model by using the target data set comprises the steps of: obtaining preset enterprise internal private data, integrating the preset enterprise internal private data, classifying according to a preset classification rule to obtain a private data set corresponding to the user authority of the target user; obtaining a public data set, and constructing the target data set based on the private data set and the public data set, and training the initial large language model by using the target data set; the responding to the dialogue by using the target large language model according to the context authority vector comprises the steps of: determining the data sensitivity level and data access authority of the preset enterprise internal private data in the private data set; generating a response prompt template according to the data sensitivity level and the data access authority, and responding to the dialogue according to the context authority vector by using the target large language model according to the response prompt template.

2. The intelligent conversation privacy protection method of claim 1, wherein, The process of generating evaluation data corresponding to the dialogue response further comprises the steps of: simulating the interaction environment of the target large language model through a preset reinforcement learning simulation environment, and determining the reward value of the dialogue response according to a preset reward function based on the dialogue response by using the preset reinforcement learning simulation environment, and generating the evaluation data of the dialogue response according to the reward value.

3. The intelligent conversation privacy protection method of claim 2, wherein, The determination of the reward value of the dialogue response according to the preset reward function based on the dialogue response by using the preset reinforcement learning simulation environment comprises the steps of: obtaining the feedback result of the target user for the dialogue response, constructing the preset reward function based on a neural network according to the feedback result, and generating the reward value corresponding to the dialogue response by using the preset reward function.

4. The intelligent conversation privacy protection method of claim 3, wherein, The optimization of the target large language model according to the target data set combined with the evaluation data comprises the steps of: determining the cumulative reward corresponding to the evaluation data, maximizing the cumulative reward by using a PPO reinforcement learning algorithm, and adjusting the model parameters of the target large language model according to the cumulative reward.

5. An intelligent conversation privacy protection apparatus, characterized by comprising: The method comprises the steps of: The permission determination module is configured to determine a user permission of the target user, and determine a context permission vector corresponding to the target user by using the target large language model according to the user permission. The response generation module is configured to obtain a dialogue input by the target user through a preset user interaction interface, and generate a corresponding dialogue response by responding to the dialogue according to the context permission vector by using the target large language model. The model optimization module is configured to generate evaluation data corresponding to the dialogue response, and optimize the target large language model according to a target data set combined with the evaluation data. The intelligent dialogue privacy protection device further comprises: The model construction module is configured to construct an initial large language model based on a pre-trained large language model, a privacy adaptation network, a prompt learning module, and a gate unit network, and train the initial large language model by using the target data set to obtain the target large language model. The gate unit network is configured to adjust parameters of the privacy adaptation network to control sensitivity of the target large language model to input data and output data. The model construction module comprises: The data acquisition unit is configured to obtain preset enterprise internal private data, integrate the preset enterprise internal private data, and classify the integrated data according to a preset classification rule to obtain a private data set corresponding to the user permission of the target user. The model training unit is configured to obtain a public data set, construct the target data set based on the private data set and the public data set, and train the initial large language model by using the target data set. The response generation module comprises: The permission determination unit is configured to determine a data sensitivity level and a data access permission of the preset enterprise internal private data in the private data set. The dialogue response unit is configured to generate a response prompt template according to the data sensitivity level and the data access permission, and respond to the dialogue according to the context permission vector by using the target large language model according to the response prompt template.

6. An electronic device, comprising: The electronic device comprises a processor and a memory; wherein the memory is configured to store a computer program, the computer program is loaded and executed by the processor to implement the intelligent dialogue privacy protection method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, A computer program is stored, and the computer program is executed by a processor to implement the intelligent dialogue privacy protection method according to any one of claims 1 to 4.

Images