Sensitive information processing method and related apparatus

Abstract

Disclosed in embodiments of the present application are a sensitive information processing method and a related apparatus. The method comprises: receiving a first request inputted by a user, the first request relating to user sensitive information; inputting the first request into a first preset model for processing, and outputting a second request, wherein the second request comprises the first request obtained by performing masking processing on the user sensitive information according to a target masking scheme, and the target masking scheme is a scheme, among a plurality of preset masking schemes, matching the user sensitive information or a scheme having the highest matching degree; and inputting the second request into a second preset model for processing, and outputting a first result in response to the second request, the first result comprising a solution corresponding to the first request having undergone masking processing. By means of the embodiments of the present application, leakage of sensitive information of personal information of a user during use of a large language model can be prevented, thereby improving the security and privacy of the sensitive information.

Description

Sensitive information processing method and related device

[0001] The present application claims priority to the Chinese patent application No. 202410673920.3, filed on May 27, 2024, with the State Intellectual Property Office of China, and entitled "Sensitive information processing method and related device", the whole content of which is incorporated herein by reference. TECHNICAL FIELD

[0002] The present application relates to the field of artificial intelligence (AI), in particular to a sensitive information processing method and related device. BACKGROUND

[0003] Large language models enable thousands of industries, and are the core of the future artificial intelligence (AI) ecosystem. In the application stage of large language models, a large amount of sensitive personal information and even business secrets may be contained in the multi-round dialogue between users and large language models. The improper management of conversation records and the frequent malicious prompt attacks may lead to frequent dialogue privacy leakage events, which brings great challenges to the reasoning data privacy protection of large language models.

[0004] Therefore, how to provide a sensitive information processing method and related device to prevent the leakage of user personal sensitive information in the process of using a large language model and improve the security and privacy of sensitive information is an urgent problem to be solved. SUMMARY

[0005] The embodiments of the present application provide a sensitive information processing method and related device, which can improve the security and privacy of sensitive information.

[0006] In a first aspect, the embodiments of the present application provide a sensitive information processing method, which comprises: receiving a first request input by a user, the first request involving user sensitive information; inputting the first request into a first preset model for processing, outputting a second request, the second request comprising the first request after being desensitized according to a target desensitization scheme; the target desensitization scheme being a scheme or a scheme with the highest matching degree in a plurality of preset desensitization schemes that matches the user sensitive information; inputting the second request into a second preset model for processing, outputting a first result in response to the second request; the first result comprising a solution corresponding to the desensitized first request.

[0007] In the embodiments of the present application, the first request input by the user is not directly input into the large language model for response, but can be first input into the end-side small language model. The end-side small language model can determine a specific desensitization scheme for the user sensitive information in the first request, that is, by analyzing the dialogue content, the matching degree between the plurality of preset desensitization schemes and the user sensitive information is determined, a scheme that matches or has the highest matching degree with the user sensitive information is selected from the plurality of preset desensitization schemes as a target desensitization scheme, and the result after the first request is desensitized by the target desensitization scheme has a small influence on the processing result of the large language model, avoiding that all requests are desensitized by the same desensitization scheme, resulting in over-desensitization of the request and the large language model cannot accurately reply, thereby ensuring the accuracy of the reply of the large language model. Further, the user sensitive information in the first request can be desensitized according to the target preset desensitization scheme to output a second request after desensitization, and the desensitization of the text data input by the user is completed. Then, the second request is input into the large language model for response, avoiding directly inputting the first request into the large language model for processing, thereby protecting the personal privacy data of the user.

[0008] In some embodiments, the first preset model includes a first sub-model, a second sub-model and a third sub-model, and inputting the first request into the first preset model for processing to output the second request includes: inputting the plurality of preset desensitization schemes, the first request and the user sensitive information into the first sub-model for processing to output the matching degree between each preset desensitization scheme in the plurality of preset desensitization schemes and the user sensitive information; inputting the matching degree corresponding to each preset desensitization scheme into the second sub-model for processing to output the target desensitization scheme, the target desensitization scheme being the scheme with the highest matching degree between the plurality of preset desensitization schemes and the user sensitive information; and inputting the target desensitization scheme, the first request and the user sensitive information into the third sub-model for processing to output the second request.

[0009] In the embodiments of the present application, the function of the first sub-model is to determine the matching degree between each preset desensitization scheme and the user sensitive information based on the plurality of preset desensitization schemes, the first request and the user sensitive information; the function of the second sub-model is to determine the target desensitization scheme based on the matching degree between each preset desensitization scheme and the user sensitive information; and the function of the third sub-model is to desensitize the user sensitive information in the first request based on the target desensitization scheme to obtain the second request after desensitization of the first request. The specific desensitization scheme of the user sensitive information is determined by the end-side small language model, that is, the matching degrees between the plurality of preset desensitization schemes and the user sensitive information are determined by analyzing the dialogue content, and one of the plurality of preset desensitization schemes with the highest matching degree is selected as the target desensitization scheme. It should be noted that the matching degree between each preset desensitization scheme and the user sensitive information can indicate the influence degree of the result after desensitization of the user sensitive information in the first request by the preset desensitization scheme on the reply of the large language model, the matching degree between the preset desensitization scheme and the user sensitive information is the highest, the influence degree of the result after desensitization of the first request by the preset desensitization scheme on the reply of the large language model is the smallest, the accuracy of the reply of the large language model is high, and then it is indicated that the user sensitive information is most suitable for the current preset desensitization scheme. Further, the user sensitive information in the first request can be desensitized according to the target preset desensitization scheme to output the second request after desensitization, and the desensitization of the text data input by the user is completed. Since the small language model can intelligently determine different desensitization schemes for different types of user sensitive information, it is avoided that all requests are desensitized by using the same desensitization scheme, which leads to excessive desensitization of the request and the large language model cannot accurately reply, so that the accuracy of the reply of the large language model is ensured while protecting the user's personal privacy data.

[0010] In some embodiments, the higher the matching degree between the preset desensitization scheme and the user sensitive information, the smaller the difference between the response result output by the second preset model and the original response result corresponding to the first request.

[0011] In the embodiments of the present application, the matching degree between each preset desensitization scheme and the user sensitive information can be understood as the influence degree of the result after desensitization of the user sensitive information in the target question by the preset desensitization scheme on the reply of the large language model (i.e., the second preset model), the higher the matching degree between the preset desensitization scheme and the user sensitive information, the smaller the difference between the response result output by the large language model and the result obtained by directly inputting the first request into the large language model for processing (i.e., the original response result corresponding to the first request), that is, the smaller the influence degree on the reply of the large language model, and then it is indicated that the user sensitive information is more suitable for the current preset desensitization scheme.

[0012] In some embodiments, the first preset model further includes a fourth sub-model, and before inputting the plurality of preset desensitization schemes, the first request, and the user sensitive information into the first sub-model, the method further includes: inputting the first request into the fourth sub-model for processing to output entity information; the entity information includes one or more of a name, an address, a mobile phone number, and an ID number.

[0013] In the embodiments of the present application, the first preset model can include a plurality of sub-models, and each sub-model in the plurality of sub-models can correspond to at least one function. Among them, the fourth sub-model included in the first preset model has an entity information extraction function, and the target question in the first request can be input into the fourth sub-model for processing to extract entity information from the target question, and output the entity information in the first request. Since entity information usually involves sensitive information such as names and addresses, the sensitive information in the first request can be identified.

[0014] In some embodiments, the first preset model further includes a fifth sub-model, and the method further includes: inputting the entity information and the user portrait information into the fifth sub-model for processing to output the user sensitive information; the user portrait information includes one or more of user basic attribute information, social attribute information, behavior attribute information, and psychological attribute information.

[0015] In the embodiments of the present application, the user portrait information can be used to identify sensitive information related only to the user from the entity information, that is, to identify sensitive information related only to the user in a personalized manner, and then these sensitive information related only to the user can be determined as sensitive information for desensitization processing. The problem of excessive desensitization of the first request caused by directly determining the entity information as sensitive information for desensitization processing in the prior art, and the problem of performance loss of the large language model are solved, thereby avoiding excessive desensitization of the target question, ensuring the accuracy of the answer of the large language model while protecting the personal privacy data of the user.

[0016] In some embodiments, the method further includes: inputting the first result into the first preset model for processing to output a second result, the second result being a result obtained by desensitizing and restoring the first result based on the target desensitization scheme.

[0017] In the embodiments of the present application, since the large language model responds to the second request after desensitization processing, the response result includes a solution corresponding to the first request after desensitization processing. There are some differences between the response result and the result obtained by directly inputting the first request into the large language model for processing (i.e., the original response result corresponding to the first request), so the first request can be desensitized and restored based on the target desensitization scheme to output the solution corresponding to the first request.

[0018] In some embodiments, the first preset model is a model trained by taking the preset request as a training sample and taking the desensitization request as an output label, and the desensitization request is a request obtained by desensitizing the preset request according to an optimal desensitization scheme, and the optimal desensitization scheme is a scheme matching the sensitive information in the preset request or a scheme with the highest matching degree among the plurality of preset desensitization schemes.

[0019] In the embodiments of the present application, the first preset model can be trained in advance through the preset request, so that subsequent intelligent decision of different desensitization schemes for different types of user sensitive information can be made based on the first preset model, and then the user sensitive information in the first request can be desensitized based on the desensitization scheme most matching the user sensitive information or the scheme with the highest matching degree, and then the result of the desensitization processing is input into the large language model for processing.

[0020] In a second aspect, the embodiments of the present application provide a sensitive information processing apparatus, which comprises: a first receiving module configured to receive a first request input by a user, the first request involving user sensitive information; a first processing module configured to input the first request into a first preset model for processing, and output a second request, the second request comprising the first request desensitized according to a target desensitization scheme; the target desensitization scheme being a scheme matching the user sensitive information or a scheme with the highest matching degree among a plurality of preset desensitization schemes; and a first output module configured to input the second request into a second preset model for processing, and output a first result in response to the second request; the first result comprising a solution corresponding to the desensitized first request.

[0021] In some embodiments, the first preset model comprises a first sub-model, a second sub-model and a third sub-model, and the first processing module is specifically configured to: input the plurality of preset desensitization schemes, the first request and the user sensitive information into the first sub-model for processing, and output a matching degree between each preset desensitization scheme and the user sensitive information among the plurality of preset desensitization schemes; input the matching degree corresponding to each preset desensitization scheme into the second sub-model for processing, and output a target desensitization scheme, the target desensitization scheme being a scheme with the highest matching degree between the user sensitive information among the plurality of preset desensitization schemes; and input the target desensitization scheme, the first request and the user sensitive information into the third sub-model for processing, and output the second request.

[0022] In some embodiments, the higher the matching degree between the preset desensitization scheme and the user sensitive information, the smaller the difference between the response result output by the second preset model and the original response result corresponding to the first request.

[0023] In some embodiments, the first preset model further includes a fourth sub-model, and the device further includes a second processing module configured to: input the first request into the fourth sub-model for processing, and output entity information; and the entity information includes one or more of a name, an address, a mobile phone number, and an ID number.

[0024] In some embodiments, the first preset model further includes a fifth sub-model, and the device further includes a third processing module configured to: input the entity information and user portrait information into the fifth sub-model for processing, and output user sensitive information; and the user portrait information includes one or more of user basic attribute information, social attribute information, behavior attribute information, and psychological attribute information.

[0025] In some embodiments, the device further includes a second output module configured to: input the first result into the first preset model for processing, and output a second result, the second result being a result obtained by performing desensitization and restoration processing on the first result based on a target desensitization scheme.

[0026] In some embodiments, the first preset model is a model trained by taking a preset request as a training sample and taking a desensitized request as an output label; the desensitized request is a request obtained by performing desensitization processing on the preset request according to an optimal desensitization scheme; and the optimal desensitization scheme is a scheme that matches sensitive information in the preset request or a scheme with the highest matching degree among a plurality of preset desensitization schemes.

[0027] In a third aspect, the present application provides a sensitive information processing device, including a processor and a memory; wherein the memory is configured to store program code, and the processor is configured to invoke the program code to execute the method provided in any possible implementation manner of the first aspect.

[0028] In a fourth aspect, the present application provides a computer readable storage medium, characterized in that the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the method provided in any possible implementation manner of the first aspect.

[0029] In a fifth aspect, the present application provides a computer program product, characterized in that when the computer program product runs on a computer, the computer program product makes the computer execute the method provided in any possible implementation manner of the first aspect.

[0030] In a sixth aspect, the present application provides a chip system, which includes a processor configured to support an electronic device to implement the functions involved in the first aspect, for example, to generate or process the information involved in the sensitive information processing method. In a possible design, the chip system further includes a memory configured to save necessary program instructions and data of the electronic device. The chip system can be composed of a chip, or can include a chip and other discrete devices. Attached Figure Description

[0031] Figure 1 is a schematic diagram of a sensitive information processing system applicable to an embodiment of this application.

[0032] Figure 2a is a schematic diagram of a neuron structure provided in an embodiment of this application.

[0033] Figure 2b is a schematic diagram of a neural network provided in an embodiment of this application.

[0034] Figure 2c is a schematic diagram of an AI application framework provided in an embodiment of this application.

[0035] Figure 3 is a flowchart illustrating a sensitive information processing method provided in an embodiment of this application.

[0036] Figure 4 is a schematic diagram of an entity information list provided in an embodiment of this application.

[0037] Figure 5 is a schematic diagram of the workflow of a sub-model of a first preset model provided in an embodiment of this application.

[0038] Figure 6 is a schematic diagram of user profile data management provided in an embodiment of this application.

[0039] Figure 7 is a schematic diagram of a process provided in this application embodiment of a first preset model to de-identify a first request and output a second request.

[0040] Figure 8 is a schematic diagram of a data anonymization intelligent decision-making scheme based on a small language model provided in an embodiment of this application.

[0041] Figure 9 is a schematic diagram of a user interface provided in an embodiment of this application.

[0042] Figure 10 is a schematic diagram of an edge language model training scheme for a sensitive information importance assessment module provided in an embodiment of this application.

[0043] Figure 11 is a schematic diagram of an edge-side language model training scheme provided in an embodiment of this application.

[0044] Figure 12 is a flowchart illustrating a sensitive information processing method provided in an embodiment of this application.

[0045] Figure 13 is a schematic diagram of a sensitive information processing device provided in an embodiment of this application. Detailed Implementation

[0046] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0047] The "multiple" referred to in the embodiments of the present application refers to greater than or equal to two.

[0048] The "first", "second" and the like referred to in the embodiments of the present application are only used for the purpose of distinguishing description, and cannot be understood as indicating or implying relative importance, nor can it be understood as indicating or implying an order.

[0049] The system architecture of the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0050] Please refer to FIG. 1, which is a schematic diagram of a sensitive information processing system to which the embodiments of the present application are applicable. The architecture includes a sensitive information processing device 101 and an AI model 102. Among them:

[0051] The sensitive information processing device 101 can be a server, an electronic device, etc. The server is a device with centralized computing capability. For example, the server can be implemented by a server, a virtual machine, a cloud, etc. For example, the electronic device can be implemented by a smart phone, a smart wearable device (such as a smart watch), etc.

[0052] When the server contains a server, the type of the server includes but is not limited to a general-purpose computer, a gateway server, a special-purpose server computer, a blade server, etc. The present application does not strictly limit the number of servers contained by the server, which can be one or more (such as a server cluster, etc.).

[0053] The virtual machine refers to a computing module that is simulated by software and has complete hardware system functions and runs in a completely isolated environment. Of course, in addition to the virtual machine, the server can also be implemented by other computing instances, such as containers, etc.

[0054] The cloud is a software platform using application virtualization technology, which can allow one or more software and applications to be developed and run in an independent virtualization environment. Optionally, when the server is implemented by the cloud, the cloud can be deployed on a public cloud, a private cloud, or a hybrid cloud, etc.

[0055] The AI model 102 can be deployed on the sensitive information processing device 101, or on other devices, which is not limited by the present application.

[0056] For ease of understanding, the following exemplary descriptions of some concepts related to the embodiments of the present application are given for reference. As described below:

[0057] (1) AI model

[0058] An AI model is a specific implementation of an AI technology function, and the AI model represents a mapping relationship between the input and output of the model. The type of AI model can be a neural network, a linear regression model, a decision tree model, a support vector machine (SVM), a Bayesian network, a Q-learning model, or other machine learning (ML) models.

[0059] (2) Neural network

[0060] A neural network is a specific implementation form of an AI or machine learning technology. According to the universal approximation theorem, a neural network can theoretically approximate any continuous function, thereby enabling the neural network to learn any mapping.

[0061] The idea of a neural network originates from the neuron structure of the brain tissue. For example, each neuron performs a weighted sum operation on its input values and outputs the operation result through an activation function. As shown in FIG. 2a, it is a schematic diagram of a neuron structure. Assuming that the input of a neuron is x = [x0, x1, …, x n ], and the weights corresponding to each input are w = [w0, w1, …, w n ], where w i is the weight of x i , used to weight x i . The bias of the weighted sum operation on the input value is, for example, b. The form of the activation function can be multiple, assuming that the activation function of a neuron is y = f(z) = max(0, z), and the output of the neuron is: For another example, the activation function of a neuron is y = f(z) = z, and the output of the neuron is: where b, w i , x i may be various possible values such as a decimal number, an integer (for example, 0, a positive integer, or a negative integer), or a complex number. The activation functions of different neurons in a neural network can be the same or different.

[0062] Generally, a neural network includes multiple layers, and each layer can include one or more neurons. By increasing the depth and / or width of the neural network, the expressive power of the neural network can be improved, providing a more powerful information extraction and abstract modeling capability for complex systems. The depth of the neural network can refer to the number of layers included in the neural network, and the number of neurons included in each layer can be referred to as the width of the layer. In an implementation manner, the neural network includes an input layer and an output layer. The input layer of the neural network processes the received input information through neurons, and transmits the processing result to the output layer, and the output result of the neural network is obtained by the output layer. In another implementation manner, the neural network includes an input layer, a hidden layer and an output layer, which can be referred to as FIG. 2b. The input layer of the neural network processes the received input information through neurons, and transmits the processing result to the intermediate hidden layer. The hidden layer calculates the received processing result to obtain a calculation result, and transmits the calculation result to the output layer or the adjacent hidden layer. Finally, the output result of the neural network is obtained by the output layer. It should be noted that a neural network can include one hidden layer, or include multiple sequentially connected hidden layers, which is not limited.

[0063] Taking the type of the AI model as the neural network as an example, the AI model involved in the present disclosure can be a deep neural network (DNN). According to the construction manner of the network, the DNN can include a Transformer neural network, a feed forward neural network (FNN), a convolutional neural network (CNN) and a recurrent neural network (RNN).

[0064] (3) Design of AI model

[0065] The design of an AI model mainly includes a data collection link (e.g., collecting training data and / or inference data), a model training link, and a model inference link. Further, an inference result application link can also be included. Referring to FIG. 2c, an AI application framework is shown. In the aforementioned data collection link, a data source is configured to provide a training data set and inference data. In the model training link, an AI model is obtained by analyzing or training the training data provided by the data source. The AI model represents the mapping relationship between the input and output of the model. The AI model is learned by the model training node, which is equivalent to learning the mapping relationship between the input and output of the model using the training data. In the model inference link, the AI model trained in the model training link is used to perform inference based on the inference data provided by the data source, and an inference result is obtained. This link can also be understood as follows: the inference data is input into the AI model, and the output obtained by the AI model is the inference result. The inference result can indicate the configuration parameters used (executed) by the execution object and / or the operation performed by the execution object. In the inference result application link, the inference result is published, for example, the inference result can be uniformly planned by an execution entity, for example, the execution entity can send the inference result to one or more execution objects (e.g., core network equipment, access network equipment, or terminal equipment, etc.) for execution. For example, the execution entity can also feed back the performance of the model to the data source, so as to facilitate subsequent implementation of model update training.

[0066] (4) Large Language Model (LLM)

[0067] The Large Language Model (LLM) refers to a deep neural network model that learns and masters general language knowledge and capabilities through unsupervised, semi-supervised or self-supervised methods using a large amount of text data. The LLM can be fine-tuned or transferred according to the requirements of specific fields and tasks, and can complete multiple natural language processing (NLP) tasks such as text classification, named entity recognition, reading comprehension, and text generation even with only a small amount of samples or even without samples.

[0068] The LLM can be Falcon-40B-instruct, Vicuna, Alpaca, LLaMA (Large Language Model Meta AI), Dolly, etc.

[0069] (5) prompt (prompt)

[0070] The prompt is originally a task-specific input form or template designed by natural language processing (NLP) researchers for downstream tasks, and is now commonly used as a substitute for inputting large language models, i.e., the input text or prefix of the input text of the large language model.

[0071] (6) Sensitive information

[0072] The sensitive information may be, for example, user identity information, perception information, enterprise-internal private information, and the like.

[0073] The user identity information may be, for example, at least one of the following: (1) a user identifier; (2) a device identifier; (3) a user physiological feature, such as a heartbeat, respiration, voice feature, portrait, fingerprint, iris, and the like.

[0074] The perception information may be, for example, location information or trajectory information, and the like.

[0075] The enterprise-internal private information may be, for example, company decisions, meeting minutes, customer information, unpublished product information, and the like.

[0076] (7) Data desensitization

[0077] The data desensitization refers to data transformation of certain sensitive information through desensitization rules to achieve reliable protection of sensitive private data.

[0078] (8) User / personal portrait

[0079] The user / personal portrait is a tagged and digitized user model abstracted from real information such as user basic attributes, social attributes, behavior attributes, and psychological attributes.

[0080] The above exemplary explanations of the concepts can be applied in the embodiments below.

[0081] The architecture of the embodiments of the present application is described above, and the method of the embodiments of the present application is described in detail below.

[0082] Please refer to FIG. 3, which is a flow diagram of a sensitive information processing method according to an embodiment of the present application. Optionally, the method can be applied to the sensitive information processing device described above, such as the sensitive information processing device shown in FIG. 1. The sensitive information processing method shown in FIG. 3 can include steps S301-S303. It should be understood that, for the convenience of description, the steps S301-S303 are described in this order, and it is not intended to limit the execution of the steps in the above order. The embodiments of the present application do not limit the order of execution, the time of execution, the number of execution, etc. of one or more steps described above. Hereinafter, the execution subject of steps S301-S303 of the sensitive information processing method is taken as a smartphone as an example, and the present application is also applicable to other execution subjects such as servers, other electronic devices, etc. Steps S301-S303 are as follows:

[0083] Step S301: receiving a first request input by a user.

[0084] Specifically, the first request involves user sensitive information. For example, the first request can include a target question, and the target question involves user sensitive information.

[0085] The user sensitive information is sensitive information related to the user, such as user identity information, user perception information, and user private information facing the enterprise internal, etc.

[0086] For example, the user identity information can be at least one of the following: (1) user identification; (2) device identification; (3) user physiological characteristics, such as user heartbeat, breathing, voice characteristics, portrait, fingerprint, iris, etc.

[0087] For example, the user perception information can be user location information or trajectory information, etc.

[0088] The user private information facing the enterprise internal can be company decisions, meeting minutes, customer information, unpublished product information, etc.

[0089] In some embodiments, when the sensitive information processing device is a user device such as a smartphone, the user can input the first request based on the user device (i.e. the smartphone).

[0090] In some embodiments, when the sensitive information processing device is a server such as a gateway server, the user can input the first request based on a user device, etc., and then the sensitive information processing device such as the gateway server obtains the first request.

[0091] Exemplarily, the gateway server can be a server processing the first request. For example, the gateway server can perform desensitization processing on all the first requests of the user, and input the desensitization-processed results into the large language model for processing.

[0092] This example introduces the first request input by the user, and it can be understood that the first request can also come from other devices, etc., and the present scheme does not limit this.

[0093] The first request can be a query request. Exemplarily, the target question included in the first request can be: I am Zhang San, 25 years old, I have A disease, and after treatment in Shanghai B Hospital, Zhang Si doctor gives C scheme, is this scheme reasonable? For example, the target question included in the first request can be: I plan to go to Harbin to see ice sculptures during the New Year holiday from Shanghai, please help me make a travel guide and arrange the itinerary. For example, the target question included in the first request can be: Help me write an invitation letter, please invite Professor Wang Wu of Tsinghua University to serve as the Chief Technology Officer (CTO).

[0094] Correspondingly, the first request can be used to input into the large language model for querying to obtain an answer corresponding to the query request.

[0095] Optionally, the first request can also include files, data, etc. Exemplarily, the user inputs a target file so that the large language model processes the target file to obtain summary information corresponding to the target file, or obtains picture, voice, table, etc. Information.

[0096] Step S302: input the first request into the first preset model for processing, and output the second request.

[0097] Specifically, the second request includes the first request after desensitization processing of the user sensitive information according to the target desensitization scheme; the target desensitization scheme is a scheme matched with the user sensitive information or a scheme with the highest matching degree in multiple preset desensitization schemes. Each preset desensitization scheme can include one or more of the following technologies: replacement, generalization, and masking.

[0098] In some embodiments, the first preset model includes a sub-model, i.e., the fourth sub-model mentioned above, and the first request, such as the target question in the first request, is input into the fourth sub-model for processing to output entity information; the entity information includes one or more of the following: name, address, mobile phone number, and ID number.

[0099] Specifically, the first preset model can include at least one end-side small language model. The end-side small language model refers to a natural language processing model that is smaller and simpler in scale relative to a large language model (LLM). The end-side small language model usually has fewer parameters and a simpler architecture, and thus is more efficient and economical in training, fine-tuning, and deployment. The first preset model can run on the sensitive information processing device. The first preset model can include a plurality of sub-models, each of which can correspond to at least one function. The sensitive information processing device can input the to-be-processed information into the sub-models for processing, and the sub-models can process the to-be-processed information according to the corresponding functions and output the processed information. Among them, a fourth sub-model included in the first preset model has an entity information extraction function. The sensitive information processing device can input the target question in the first request into the fourth sub-model for processing to extract entity information in the target question and output the entity information in the target question. Since entity information usually involves sensitive information such as names and addresses, the sensitive information in the first request can be identified. An entity refers to a thing, person, or place that can exist independently and has an independent identity. Entity information can include, but is not limited to, names, addresses, mobile phone numbers, identification numbers, holidays, ages, medical cases, schools, and the like.

[0100] For example, if the target question 1 input into the fourth sub-model is: My name is Zhang San, I am 25 years old, I have A disease, and after treatment in Shanghai B Hospital, Dr. Zhang Si gave me a C scheme. Is this scheme reasonable?

[0101] The entity information 1 output by the fourth sub-model can be: name: Zhang San; age: 25 years old; disease name: A disease; address: Shanghai; address: B Hospital; name: Zhang Si; scheme: C scheme.

[0102] Optionally, the entity information 1 output by the fourth sub-model can be output in the form of a list. As shown in FIG. 4, FIG. 4 is a schematic diagram of an entity information list provided by an embodiment of the present application. After the fourth sub-model of the first preset model extracts entity information from the target question 1, the fourth sub-model can output a list of entity information 1 corresponding to the target question 1, which can include a plurality of entries, such as name: Zhang San; age: 25 years old; disease name: A disease; address: Shanghai; address: B Hospital; name: Zhang Si; scheme: C scheme.

[0103] For example, if the target question 2 input into the fourth sub-model is: I plan to go to Harbin to see ice sculptures during the New Year holiday from Shanghai. Please help me develop a travel guide and arrange the itinerary.

[0104] The entity information 2 output by the fourth sub-model can be: holiday: New Year's Day; address: Shanghai; address: Harbin.

[0105] For example, if the target question 3 input into the fourth sub-model is: help me write an invitation letter to invite Professor Wang Wu of Tsinghua University to be the chief technology officer.

[0106] The entity information 3 output by the fourth sub-model can be: school: Tsinghua University; name: Wang Wu.

[0107] In some embodiments, the entity information is determined as user sensitive information. Specifically, the entity information is determined as sensitive information for desensitization processing, which can protect user personal privacy data when using a large language model.

[0108] In some embodiments, the first preset model further includes other sub-models, i.e., the fifth sub-model mentioned above, which inputs the entity information and the user portrait information into the fifth sub-model for processing, and outputs the user sensitive information.

[0109] Specifically, the fifth sub-model has the function of filtering the user sensitive information related to the user from the entity information based on the user portrait information. The user portrait information includes one or more of user basic attribute information, social attribute information, behavior attribute information, and psychological attribute information, wherein the user basic attribute information can include age, identity card number, mobile phone number, date of birth, and the like. Optionally, the user portrait information can be presented in the form of a user portrait knowledge graph. The sensitive information processing device can first obtain the user portrait information, and input the user portrait information and the entity information into the fifth sub-model for processing, and the fifth sub-model outputs the user sensitive information. It should be understood that the fifth sub-model filters the user sensitive information related only to the user from the entity information identified by the fourth sub-model based on the user portrait information. In the embodiments of the present application, the user sensitive information related only to the user in the dialogue content can be accurately identified based on the user portrait and the context content, i.e., the user sensitive information related only to the user is identified individually, and then the information related only to the user can be determined as sensitive information for desensitization processing, solving the problem in the prior art that the entity information is directly determined as sensitive information, which causes too much information in the first request to be determined as sensitive information, resulting in excessive desensitization of the first request, and further causing performance loss of the large language model, thereby avoiding excessive desensitization of the target question, protecting user personal privacy data, and ensuring the accuracy of the reply of the large language model.

[0110] Optionally, the entity information, the user individualized configuration information, and the user portrait information are input into the fifth sub-model for processing, and the user sensitive information is output. The user individualized configuration information can be sensitive information configured by the user in advance.

[0111] For example, as shown in FIG. 5, which is a workflow diagram of a sub-model of a first preset model provided in an embodiment of the present application, the target question 1 can be: I am Zhang San, 25 years old, I have A disease, after treatment in B hospital in Shanghai, Zhang Si doctor gave C scheme, is this scheme reasonable?

[0112] The entity information 1 can be presented in the form of a list, such as: name: Zhang San; age: 25 years old; disease name: A disease; address: Shanghai; address: B hospital; name: Zhang Si; scheme: C scheme.

[0113] The user portrait information can be presented in the form of a list, and the user portrait information list includes user basic attribute information, social attribute information, behavior attribute information, and psychological attribute information, such as: name: Zhang San; age: 25 years old; ID number: 12345678910123456; mobile phone number: 13466840000; school: Tsinghua University; major: electronic information; history: A disease; hobby: travel; permanent address: Guangzhou.

[0114] The sensitive information processing device can first obtain the above-mentioned user portrait information list, and input the user portrait information list and the above-mentioned entity information 1 list into the fifth sub-model for processing, and the fifth sub-model outputs user sensitive information. The user sensitive information can be output in the form of a list, and the user sensitive information can include: name: Zhang San; age: 25 years old.

[0115] In the embodiment of the present application, the user sensitive information in the conversation content that is only related to the user can be accurately identified based on the user portrait and the context content, that is, the user sensitive information that is only related to the user is identified individually, and then these information that is only related to the user can be determined as sensitive information for desensitization processing, thereby avoiding over-desensitization of the first request, and thus protecting the user's personal privacy data.

[0116] Optionally, the fifth sub-model in the first preset model can be used to save user portrait information, so as to determine user sensitive information in user conversation content.

[0117] For example, as shown in FIG. 6, which is a schematic diagram of user portrait data management provided by an embodiment of the present application, the fifth sub-model can be used to manage user portrait data. The fifth sub-model can filter sensitive information related to the user by comparing the user portrait and the dialogue content. The fifth sub-model can include a personal data writing module and a personal data filtering module. The personal data filtering module mainly uses keyword matching, fuzzy matching, and vector search technologies to query whether the entity data (i.e., the entity information described above) in the input is included in the user portrait, so as to determine whether the entity data is personal sensitive data related to the user. The main input of the module is a list of entity data, and the output is a list of sensitive data related to the user. The main input of the personal data writing module is the dialogue content between the user and the big prophecy model, and the user-related information such as application use cases. The personal data writing module can use relationship extraction, named entity recognition (NER) extraction, and coreference resolution technologies to extract user portrait information from the user-related information.

[0118] In some embodiments, referring to FIG. 7, which is a flowchart of a process in which a first preset model desensitizes a first request to output a second request, the first preset model includes a first sub-model, a second sub-model, and a third sub-model. The first request is input into the first preset model for processing to output the second request, including: inputting a plurality of preset desensitization schemes, the first request, and user sensitive information into the first sub-model for processing to output a matching degree between each preset desensitization scheme and the user sensitive information in the plurality of preset desensitization schemes; inputting the matching degree corresponding to each preset desensitization scheme into the second sub-model for processing to output a target desensitization scheme; and inputting the target desensitization scheme, the first request, and the user sensitive information into the third sub-model for processing to output the second request.

[0119] Specifically, the first sub-model determines the matching degree between each preset anonymization scheme and the user's sensitive information based on multiple preset anonymization schemes, the first request, and the user's sensitive information. The second sub-model determines the target anonymization scheme based on the matching degree of each preset anonymization scheme. The third sub-model anonymizes the user's sensitive information in the first request using the target anonymization scheme to obtain a second request anonymized from the first request. The matching degree between each preset anonymization scheme and the user's sensitive information can be understood as the degree of influence of the result of anonymizing the user's sensitive information in the first request using that preset anonymization scheme on the response of the large language model (i.e., the second preset model). The higher the matching degree between the preset anonymization scheme and the user's sensitive information, the smaller the influence on the response of the large language model, indicating that the user's sensitive information is more suitable for the current preset anonymization scheme. The matching degree between each preset anonymization scheme and the user's sensitive information can also be understood as the importance of the user's sensitive information under the preset anonymization scheme. The lower the importance, the smaller the influence on the response of the large language model, indicating that the user's sensitive information is more suitable for the current preset anonymization scheme. The target de-identification scheme is the one with the highest matching degree between multiple preset de-identification schemes and the user's sensitive information. In other words, the scheme with the highest matching degree among multiple preset de-identification schemes is determined as the scheme matching the first request. In this embodiment, the specific de-identification scheme for the user's sensitive information can be decided by the client-side small language model. That is, by analyzing the dialogue content, the matching degree between multiple preset de-identification schemes and the user's sensitive information is determined. Without affecting the response content of the large language model, the most reasonable de-identification scheme is intelligently selected as the target de-identification scheme from multiple preset de-identification schemes, such as replacement, generalization, masking, etc. Further, the user's sensitive information in the target question can be de-identified according to the target preset de-identification scheme to output the de-identified second request, completing the de-identification processing of the user's input text data. By intelligently deciding different de-identification schemes for different types of sensitive information through the small language model, the accuracy of the response from the large language model can be ensured while protecting the user's personal privacy data.

[0120] Exemplarily, as shown in FIG. 8, FIG. 8 is a schematic diagram of a data desensitization intelligent decision-making scheme based on a small language model provided by an embodiment of the present application. The first preset model can automatically select the most suitable desensitization means for user sensitive information. The first preset model can include a data desensitization intelligent decision-making module. The data desensitization intelligent decision-making module mainly consists of a sensitive information importance evaluation module (which can also be referred to as a first sub-model) and a desensitization scheme privacy judgment module (which can also be referred to as a second sub-model). The sensitive information importance evaluation module uses an end-side small language model to determine the importance of a certain sensitive data under a specific desensitization scheme in the current dialogue content. If the importance is high, the model's reply will be greatly affected, and then the sensitive data is not suitable for the current desensitization scheme. The main inputs of this module include: the original text input by the user, the recognized sensitive information text, and the fine-tunable prompt corresponding to different desensitization schemes. The main outputs include: a desensitization scheme label list suitable for the current sensitive data. The desensitization scheme privacy judgment module records the privacy ranking of all desensitization schemes based on a dictionary structure. This module takes the desensitization scheme label list as input and outputs the desensitization scheme label with the highest privacy in the list as the final decision result. In the embodiment of the present application, the data desensitization intelligent decision-making based on the small language model intelligently decides and calls different privacy protection plug-ins (which can also be referred to as a third sub-model) for different types of sensitive information, thereby improving the accuracy of the model's reply.

[0121] Exemplarily, the target question 1 can be: I am Zhang San, 25 years old, I have A disease, and after treatment in Shanghai B hospital, Dr. Zhang Si gives C scheme, is this scheme reasonable?

[0122] The user sensitive information can be: name: Zhang San; age: 25 years old.

[0123] The plurality of preset desensitization schemes can be: preset desensitization scheme 1: replacement; preset desensitization scheme 2: generalization; and preset desensitization scheme 3: mask.

[0124] The sensitive information processing device can input the target question 1, the user sensitive information, and the plurality of preset desensitization schemes into the first sub-model for processing. The first sub-model can output that the matching degree corresponding to the preset desensitization scheme 1 is 0.9; the matching degree corresponding to the preset desensitization scheme 2 is 0.5; and the matching degree corresponding to the preset desensitization scheme 3 is 0.3.

[0125] The sensitive information processing device can input the matching degree corresponding to the preset desensitization scheme 1, the matching degree corresponding to the preset desensitization scheme 2, and the matching degree corresponding to the preset desensitization scheme 3 into the second sub-model for processing, and output the target desensitization scheme as the preset desensitization scheme 1.

[0126] In some embodiments, the second sub-model determines the target desensitization scheme according to the desensitization scheme privacy levels corresponding to the plurality of preset desensitization schemes, if the matching degrees corresponding to the plurality of preset desensitization schemes are the same, wherein the higher the desensitization scheme privacy level is, the better the desensitization effect is.

[0127] Optionally, when the sensitive information processing device inputs the target question 1, the user sensitive information and the plurality of preset desensitization schemes into the first sub-model for processing, the matching degree corresponding to each preset desensitization scheme output by the first sub-model is 0 or 1, the matching degree of 0 indicates that the user sensitive information is not suitable for the current preset desensitization scheme, and the matching degree of 1 indicates that the user sensitive information is suitable for the current preset desensitization scheme.

[0128] Suppose that the matching degree corresponding to the preset desensitization scheme 1 output by the first sub-model is 1, the matching degree corresponding to the preset desensitization scheme 2 is 1, and the matching degree corresponding to the preset desensitization scheme 3 is 0.

[0129] Further, the sensitive information processing device can input the matching degree of 1 corresponding to the preset desensitization scheme 1, the matching degree of 1 corresponding to the preset desensitization scheme 2, and the matching degree of 0 corresponding to the preset desensitization scheme 3 into the second sub-model for processing, and output the target desensitization scheme. It should be noted that the desensitization scheme privacy levels corresponding to the plurality of preset desensitization schemes can be preset in the second sub-model in advance, for example, the desensitization scheme privacy level of the preset desensitization scheme 1 is greater than the desensitization scheme privacy level of the preset desensitization scheme 2, which is greater than the desensitization scheme privacy level of the preset desensitization scheme 3, and the higher the desensitization scheme privacy level is, the better the desensitization effect is. Since the matching degree corresponding to the preset desensitization scheme 1 is 1, the matching degree corresponding to the preset desensitization scheme 2 is also 1, but the desensitization scheme privacy level of the preset desensitization scheme 1 is greater than the desensitization scheme privacy level of the preset desensitization scheme 2, therefore, the target desensitization scheme output by the second sub-model is the preset desensitization scheme 1.

[0130] The sensitive information processing device can input the preset desensitization scheme 1, the target question 1 and the user sensitive information into the third sub-model for processing, and output a second request, the second request including a target question after the user sensitive information is desensitized according to the preset desensitization scheme 1, for example, replacing “Zhang San” with “Li Si”, and replacing “25 years old” with “26 years old”, and the target question after desensitization can be: My name is Li Si, I am 26 years old, I have A disease, and after treatment in Shanghai B hospital, Dr. Zhang Si gives me C scheme, is this scheme reasonable?

[0131] Optionally, as shown in FIG. 9, the user interface 1 can display the first request input by the user and the second request after desensitization based on the first request.

[0132] In the embodiments of the present application, different desensitization schemes can be intelligently determined for different types of user sensitive information based on the small language model, and then the user sensitive information of the target question in the first request can be desensitized based on the desensitization scheme that best matches the user sensitive information, and the result after desensitization is input into the large language model for processing, thereby protecting the user's personal privacy data while ensuring the accuracy of the large language model's response.

[0133] In some embodiments, the higher the matching degree between the preset desensitization scheme and the target question, the smaller the difference between the response result output by the second preset model and the original response result corresponding to the first request.

[0134] Specifically, the matching degree between each preset desensitization scheme and the user sensitive information can be understood as the degree of influence of the result after desensitization of the user sensitive information in the target question by the preset desensitization scheme on the response of the large language model (i.e., the second preset model). The higher the matching degree between the preset desensitization scheme and the user sensitive information, the smaller the difference between the response result output by the large language model and the result obtained by directly inputting the first request into the large language model for processing (i.e., the original response result corresponding to the first request), i.e., the degree of influence on the response of the large language model is smaller, which means that the user sensitive information is more suitable for the current preset desensitization scheme. By intelligently determining different desensitization schemes for different types of sensitive information through the small language model, the user's personal privacy data can be protected while ensuring the accuracy of the response of the large language model.

[0135] In some embodiments, the first preset model is a model trained with preset requests as training samples and desensitized requests as output labels.

[0136] Specifically, the preset request can be a plurality of query requests; the desensitized request is a request desensitized from the preset request according to the optimal desensitization scheme, and the optimal desensitization scheme is a scheme that matches the sensitive information in the preset request in the preset request or a scheme with the highest matching degree. In the embodiments of the present application, the first preset model can be trained in advance through the preset request, so that different desensitization schemes can be intelligently determined for different types of user sensitive information based on the first preset model, and then the user sensitive information of the target question in the first request can be desensitized based on the desensitization scheme that best matches the user sensitive information, and the result after desensitization is input into the large language model for processing, thereby protecting the user's personal privacy data while ensuring the accuracy of the large language model's response.

[0137] For example, as shown in FIG. 10, FIG. 10 is a schematic diagram of an end-side language model training scheme of a sensitive information importance evaluation module provided in an embodiment of the present application. Based on the thinking chain technology, the few-shot prompt can be used to show the multi-step reasoning logic of sensitive data importance evaluation, guide the large language model to generate a chain of thinking tracks for different problems, and form the importance evaluation labeled data set. For example, the thinking chain few-shot prompt can include: prompt: given the following question and keyword, judge whether the keyword will affect the answer to the question after desensitization by

replacement and restoration

[0138] Step S303: inputting the second request into the second preset model for processing, and outputting a first result responding to the second request.

[0139] Specifically, the first result includes a solution corresponding to the first request after the desensitization processing.

[0140] In some embodiments, the first result is input into the first preset model for processing, and a second result is output.

[0141] Specifically, the second result is a result after the desensitization and restoration processing of the first result based on the target desensitization scheme. In the embodiments of the present application, since the large language model responds to the second request after the desensitization processing, the response result includes a solution corresponding to the first request after the desensitization processing. The response result may be different from the result obtained by directly inputting the first request into the large language model for processing (i.e., the original response result corresponding to the first request). Therefore, the desensitization and restoration processing of the first result based on the target desensitization scheme can be performed to output the solution corresponding to the first request, so as to protect the user's personal privacy data while ensuring the accuracy of the large language model.

[0142] For example, as shown in FIG. 12, which is a flowchart of a sensitive information processing method provided by an embodiment of the present application, the privacy assistant can be the first preset model mentioned above, the privacy decision engine can include the first sub-model, the second sub-model, the fourth sub-model, and the fifth sub-model mentioned above, and the privacy protection plug-in can be the third sub-model mentioned above. The large language model service is the second preset model mentioned above. When a user accesses the large language model through the privacy assistant, the sensitive information of the user can be automatically protected to prevent sensitive data from being leaked to the large language model. The implementation process of the privacy assistant includes: receiving user input text data from a client application; identifying entity information in the text data based on a dialogue context and a machine learning method, such as a name, an address, a mobile phone number, a credit card number, etc.; in order to further filter out personal data related to the user, the sensitive information related to the user can be filtered out from the identified entity information based on user personalization configuration and a user portrait knowledge graph (which contains personal attribute information of the user, such as age, ID number, mobile phone number, date of birth, etc.); based on the small language model, a specific desensitization scheme for the sensitive information is decided. By analyzing the dialogue content and the importance of the sensitive information, the most reasonable data desensitization method, such as replacement, generalization, masking, etc., is intelligently decided without affecting the reply content of the large language model; according to the decision result, the corresponding privacy protection plug-in is called through the scheduling module to complete the desensitization of the user input text data; the large language model interface is called, the desensitized text data is input, and the desensitized large language model reply text is obtained; based on the user personal portrait and the small language model, the desensitized large language model reply text is optimized to obtain a personalized reply text for the user, and the personalized reply text is sent to the client application.

[0143] In summary, in the present application, the first request input by the user is not directly input to the large language model for response, but can be first input to the end-side small language model. The end-side small language model can decide a specific desensitization scheme for the user sensitive information in the first request, that is, by analyzing the dialogue content, the matching degree between a plurality of preset desensitization schemes and the user sensitive information is determined, a scheme that matches the user sensitive information or has the highest matching degree is selected from the plurality of preset desensitization schemes as a target desensitization scheme, and the result of the first request after desensitization processing by the target desensitization scheme has a small degree of influence on the large language model reply. Avoiding all requests using the same desensitization scheme for desensitization processing, leading to excessive request desensitization, and the large language model cannot accurately reply, thereby ensuring the accuracy of the large language model reply. Further, the user sensitive information in the first request can be desensitized according to the target preset desensitization scheme to output a desensitized second request, and the desensitized processing of the text data input by the user is completed. Then, the second request is input to the large language model for response, avoiding directly inputting the first request to the large language model for processing, thereby protecting the personal privacy data of the user.

[0144] As shown in FIG. 13, FIG. 13 is a schematic diagram of a sensitive information processing apparatus provided by an embodiment of the present application. The sensitive information processing apparatus 40 comprises a first receiving module 401, a first processing module 402, a first output module 403, a second processing module 404, a third processing module 405, and a second output module 406. Details are as follows:

[0145] The first receiving module 401 is configured to receive a first request input by a user, the first request relating to sensitive information of the user.

[0146] The first processing module 402 is configured to input the first request into a first preset model for processing, and output a second request, the second request comprising the first request after being desensitized according to a target desensitization scheme; the target desensitization scheme being a scheme that matches the sensitive information of the user among a plurality of preset desensitization schemes.

[0147] The first output module 403 is configured to input the second request into a second preset model for processing, and output a first result in response to the second request; the first result comprising a solution corresponding to the first request after being desensitized.

[0148] In some embodiments, the first preset model comprises a first sub-model, a second sub-model, and a third sub-model. The first processing module 402 is specifically configured to: input the plurality of preset desensitization schemes, the first request, and the sensitive information of the user into the first sub-model for processing, and output a matching degree between each preset desensitization scheme and the sensitive information of the user among the plurality of preset desensitization schemes; input the matching degree corresponding to each preset desensitization scheme into the second sub-model for processing, and output the target desensitization scheme, the target desensitization scheme being a scheme that has the highest matching degree between the sensitive information of the user among the plurality of preset desensitization schemes; and input the target desensitization scheme, the first request, and the sensitive information of the user into the third sub-model for processing, and output the second request.

[0149] In some embodiments, the higher the matching degree between the preset desensitization scheme and the sensitive information of the user, the smaller the difference between the response result output by the second preset model and an original response result corresponding to the first request.

[0150] In some embodiments, the first preset model further comprises a fourth sub-model. The second processing module 404 is configured to: input the first request into the fourth sub-model for processing, and output entity information; the entity information comprising one or more of a name, an address, a mobile phone number, and an ID number.

[0151] In some embodiments, the first preset model further includes a fifth sub-model, and the third processing module 405 is configured to: input the entity information and the user portrait information into the fifth sub-model for processing, and output the user sensitive information; the user portrait information includes one or more of user basic attribute information, social attribute information, behavior attribute information, and psychological attribute information.

[0152] In some embodiments, the second output module 406 is configured to: input the first result into the first preset model for processing, and output a second result, the second result being a result obtained by performing desensitization and restoration processing on the first result based on a target desensitization scheme.

[0153] In some embodiments, the first preset model is a model trained by taking a preset request as a training sample and taking a desensitization request as an output label; the desensitization request is a request obtained by performing desensitization processing on the preset request according to an optimal desensitization scheme, and the optimal desensitization scheme is a scheme that matches the sensitive information in the preset request or a scheme with the highest matching degree among a plurality of preset desensitization schemes.

[0154] The above modules can refer to the descriptions of the foregoing embodiments, and will not be described again.

[0155] The present application provides a sensitive information processing device, comprising a processor and a memory; wherein the memory is used to store program code, and the processor is used to call the program code to execute the method provided by any possible implementation manner of the above sensitive information processing method.

[0156] The present application provides a computer storage medium, characterized in that the computer storage medium stores a computer program, and the computer program is executed by a processor to implement any one of the above sensitive information processing methods.

[0157] The present application provides a chip system, which includes a processor for supporting an electronic device to implement the functions involved in the above sensitive information processing method, such as generating or processing the information involved in the above sensitive information processing method. In a possible design, the chip system further includes a memory, and the memory is used to save necessary program instructions and data of the electronic device. The chip system can be composed of a chip, or can include a chip and other discrete devices.

[0158] The present application provides a computer program product, characterized in that the computer program includes instructions, and when the computer program is executed by a computer, the computer executes the method described in any one of the above sensitive information processing methods.

[0159] In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0160] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0161] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0162] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0163] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0164] The integrated unit described above, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc., and specifically can be a processor in the computer device) to execute all or part of the steps of the above-mentioned methods of various embodiments of the present application. Among them, the aforementioned storage medium can include: a U disk, a mobile hard disk, a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, abbreviated as: ROM) or a random access memory (Random Access Memory, abbreviated as: RAM) and various program code storage media.

[0165] The above-described embodiments are only used to illustrate the technical solutions of the present application, not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A sensitive information processing method characterized by, The method comprises: receiving a first request input by a user, the first request relating to user sensitive information; inputting the first request into a first preset model for processing, outputting a second request, the second request comprising the first request after being desensitized according to a target desensitization scheme; the target desensitization scheme being a scheme or a scheme with the highest matching degree among a plurality of preset desensitization schemes that matches the user sensitive information; inputting the second request into a second preset model for processing, outputting a first result in response to the second request; the first result comprising a solution corresponding to the desensitized first request.

2. The method of claim 1, wherein, The first preset model comprises a first sub-model, a second sub-model and a third sub-model, and the inputting the first request into the first preset model for processing, outputting a second request comprises: inputting the plurality of preset desensitization schemes, the first request and the user sensitive information into the first sub-model for processing, outputting a matching degree between each preset desensitization scheme and the user sensitive information among the plurality of preset desensitization schemes; inputting the matching degree corresponding to each preset desensitization scheme into the second sub-model for processing, outputting the target desensitization scheme, the target desensitization scheme being a scheme with the highest matching degree among the plurality of preset desensitization schemes and the user sensitive information; inputting the target desensitization scheme, the first request and the user sensitive information into the third sub-model for processing, outputting the second request.

3. The method of claim 1 or 2, wherein, The higher the matching degree between the preset desensitization scheme and the user sensitive information, the smaller the difference between the response result output by the second preset model and the original response result corresponding to the first request.

4. The method of any of claims 2 or 3, wherein, The first preset model further comprises a fourth sub-model, and before inputting the plurality of preset desensitization schemes, the first request and the user sensitive information into the first sub-model, the method further comprises: inputting the first request into the fourth sub-model for processing, outputting entity information; the entity information comprising one or more of name, address, mobile phone number and certificate number.

5. The method of claim 4, wherein, The first preset model further comprises a fifth sub-model, and the method further comprises: inputting the entity information and user portrait information into the fifth sub-model for processing, outputting the user sensitive information; the user portrait information comprising one or more of user basic attribute information, social attribute information, behavior attribute information and psychological attribute information.

6. The method of any one of claims 1-5, wherein, The method further comprises: inputting the first result into the first preset model for processing, outputting a second result, the second result being a result after desensitization and restoration processing of the first result based on the target desensitization scheme.

7. The method according to any one of claims 1 to 6, wherein the method is performed in a single step. The first preset model is a model trained with a preset request as a training sample and a desensitized request as an output label; the desensitized request is a request after desensitization processing of the preset request according to an optimal desensitization scheme, and the optimal desensitization scheme is a scheme or a scheme with the highest matching degree among the plurality of preset desensitization schemes that matches sensitive information in the preset request.

8. A sensitive information processing apparatus characterized by comprising: The device comprises: The first receiving module is configured to receive a first request input by a user, the first request relating to user sensitive information; The first processing module is configured to input the first request into a first preset model for processing, and output a second request, the second request comprising a first request after desensitization processing of the user sensitive information according to a target desensitization scheme; the target desensitization scheme being a scheme or a scheme with the highest matching degree among a plurality of preset desensitization schemes that matches the user sensitive information; The first output module is configured to input the second request into a second preset model for processing, and output a first result in response to the second request; the first result comprising a solution corresponding to the first request after the desensitization processing.

9. The apparatus of claim 8 wherein, The first preset model comprises a first sub-model, a second sub-model and a third sub-model, and the first processing module is specifically configured to: input the plurality of preset desensitization schemes, the first request and the user sensitive information into the first sub-model for processing, and output a matching degree between each preset desensitization scheme among the plurality of preset desensitization schemes and the user sensitive information; input the matching degree corresponding to each preset desensitization scheme into the second sub-model for processing, and output the target desensitization scheme, the target desensitization scheme being a scheme with the highest matching degree among the plurality of preset desensitization schemes and the user sensitive information; input the target desensitization scheme, the first request and the user sensitive information into the third sub-model for processing, and output the second request.

10. The apparatus of claim 8 or 9, wherein, The higher the matching degree between the preset desensitization scheme and the user sensitive information, the smaller the difference between the response result output by the second preset model and an original response result corresponding to the first request.

11. The apparatus as claimed in claim 9 or 10, characterized in that, The first preset model further comprises a fourth sub-model, and before the plurality of preset desensitization schemes, the first request and the user sensitive information are input into the first sub-model, the device further comprises a second processing module configured to: input the first request into the fourth sub-model for processing, and output entity information; the entity information comprising one or more of a name, an address, a mobile phone number and an ID number.

12. The apparatus of claim 11, wherein, The first preset model further comprises a fifth sub-model, and the device further comprises a third processing module configured to: input the entity information and user portrait information into the fifth sub-model for processing, and output the user sensitive information; the user portrait information comprising one or more of user basic attribute information, social attribute information, behavior attribute information and psychological attribute information.

13. The device of any one of claims 8-12, wherein, The device further comprises a second output module configured to: input the first result into the first preset model for processing, and output a second result, the second result being a result after desensitization and restoration processing of the first result based on the target desensitization scheme.

14. The device of any one of claims 8-13, wherein, The first preset model is a model trained by taking a preset request as a training sample and taking a desensitization request as an output label; the desensitization request is a request obtained by desensitizing the preset request according to an optimal desensitization scheme; the optimal desensitization scheme is a scheme or a scheme with the highest matching degree in the plurality of preset desensitization schemes that matches the sensitive information in the preset request.

15. A sensitive information processing device, characterized by comprising: Comprising: a memory and at least one processor; the memory is coupled with the one or more processors, and the memory is configured to store computer program codes, the computer program codes comprising computer instructions, when the one or more processors execute the computer instructions, causing the electronic device to execute the method of any one of claims 1-7.

16. A computer storage medium, comprising, Comprising: computer instructions; when the computer instructions run on an electronic device, causing the electronic device to execute the method of any one of claims 1-7.

17. A computer program product, characterised in that, When the computer program product runs on an electronic device, causing the electronic device to execute the method of any one of claims 1-7.

Images