A system and method for protecting input information privacy based on secure two-party computation

Abstract

The present invention discloses a system and method for protecting the privacy of input information based on secure two-party computing, which relates to the field of privacy security. The system includes: a secure two-party computing protocol module, which provides secret sharing technology. The client includes an information input module, which uses secret sharing technology to convert the received input data into two shares of data in the form of additive secret sharing; and also includes a user corpus module, which receives one of the shares of data in the form of additive secret sharing. The private server includes an intelligent association algorithm module and a first prediction model. The intelligent association algorithm module receives the other share of data in the form of additive secret sharing, and obtains a first prediction result, i.e., a first ciphertext prediction result, by calculating the first prediction model; the third-party public cloud server receives the first ciphertext prediction result, reconstructs the first plaintext prediction result, and then sends it to the client.

Description

Technical Field

[0001] The present invention relates to the field of privacy security, and in particular to a system and method for protecting the privacy of input information based on secure two-party computing. Background Art

[0002] In recent years, with the rapid development of instant messaging technology, people can now communicate freely at any time and anywhere. Text, as the primary carrier of information, requires keyboard or voice input before it can be transmitted to distant locations. Intelligent input method software plays an indispensable role in this, primarily by recognizing user interactions with the keyboard or screen and inputting the user's desired information into smart devices. Its application areas are very broad, such as search, chat, and note-taking.

[0003] With advances in machine learning, current input methods have evolved from simple typewriters to intelligent input methods that integrate advanced features such as smart association, smart image matching, AI-assisted writing, voice input, and smart proofreading (collectively referred to as smart association). Smart association refers to a feature within intelligent input methods where, after a user enters a word or a paragraph, the program extracts the user's keywords, compares them with a corpus, predicts the user's next possible input, and then displays them to the user as alternatives, saving the user time and improving input efficiency. Smart image matching is a technology in which intelligent input methods intelligently recommend images or emoticons based on the user's keywords. Instant messaging apps have become an indispensable part of our lives. During chats, people often use emoticons and images to express themselves more efficiently and vividly. Intelligently recommending emoticons or images that match the user's input has become a very important and popular feature of intelligent input methods. AI writing assistance is the latest feature of the intelligent input method. It can infer the emotions the user wants to express or the scenes the user wants to describe based on the user's typing, and push properly edited sentences based on this information to make the chatter express more perfectly. This feature has high requirements for AI technology and is still in the development stage. At this stage, it mainly recommends sentences in the corpus through similarity comparison.

[0004] The emergence of intelligent input methods has made keyboard input easier and more efficient, and has also made the input more engaging and engaging, earning them widespread favor among smartphone users. To provide these intelligent services, intelligent input methods require uploading user input to the backend. Machine learning algorithms are then used to predict and infer the input, generating intelligent associations. Theoretically, the security of data cannot be guaranteed once it leaves the local machine. Intelligent input methods can predict user behavior based on input, leading to numerous commercial activities. Therefore, technological advancements are needed to provide these services while protecting user privacy.

[0005] Miao Gongxun et al., in their Chinese invention patent application, "Method and System for Implementing a Mobile Input Method with Information Encryption Function" (application number CN202011632602.0), proposed a method for encrypting data before transmission and then decrypting it after transmission, essentially using encryption to ensure the security of communication data. This method's encryption function is limited to the data transmission process, preventing data from being intercepted during transmission. However, this solution cannot guarantee that data will not be leaked after decryption, and does not have the ability to prevent server-side data leakage.

[0006] Fei Teng et al. disclosed a privacy protection method applied to input method clients in their Chinese invention patent application “A method and device for privacy protection” (application number CN201910156868.3). The method includes: when a user is detected operating an electronic device, collecting the user's biometric data; matching the user's biometric data with the preset biometric data in the electronic device; if the match fails, encrypting the privacy data in the electronic device. This invention achieves the technical effect of ensuring that privacy data will not be leaked when the electronic device is borrowed, thereby ensuring data security. However, this invention cannot guarantee the use of data by the input method provider's server, and does not have the ability to prevent server-side data leakage.

[0007] Therefore, technical personnel in this field are committed to developing a new input information privacy protection system and method to solve the defects in existing technical solutions that user privacy data may be leaked when user data is uploaded from the client to the server for data mining and intelligent reasoning, such as privacy leakage may occur after decryption or data may be leaked on the server. Summary of the Invention

[0008] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to avoid the leakage of user privacy data when the user data is uploaded from the client to the server for data mining and intelligent reasoning during the user's use of the intelligent input method, thereby achieving the protection of user privacy data.

[0009] To achieve the above objectives, the present invention provides an input information privacy protection system and method based on secure two-party computing, so that users can use the intelligent input method in a smart and safe manner.

[0010] The present invention provides an input information privacy protection system based on secure two-party computation, comprising:

[0011] A secure two-party computing protocol module, wherein the secure two-party computing protocol module provides secret sharing technology;

[0012] A client, comprising an information input module, the information input module using the secret sharing technology to convert received input data into data in the form of an additive secret sharing of two shares; the client further comprising a user corpus module, the user corpus module connected to the information input module and receiving one of the shares of the data in the form of the additive secret sharing;

[0013] a private server end, the private server end being connected to the client end, comprising an intelligent association algorithm module, the intelligent association algorithm module further comprising a first prediction model having a built-in intelligent association algorithm, the intelligent association algorithm module receiving another share of the data in the additive secret sharing form, and obtaining a first prediction result in the secret sharing form, i.e., a first ciphertext prediction result, by calculating the intelligent association algorithm in the first prediction model, wherein half of the first ciphertext prediction result is on the client end, and the other half is on the private server end;

[0014] The third-party public cloud server receives the first ciphertext prediction result, including half on the client and the other half on the private server, reconstructs the first plaintext prediction result, and then sends the first plaintext prediction result to the client.

[0015] Furthermore, the private server side also includes:

[0016] A user preference association module is connected to the user corpus module, receives the output of the user corpus module and trains a second prediction model. When the prediction accuracy of the second prediction model reaches a preset threshold, the second prediction model is sent to the intelligent association algorithm module as the output of the user preference association module; the intelligent association algorithm module calculates the second prediction result of the second prediction model through the secret sharing technology provided by the secure two-party computing protocol module, and the second prediction result is also in the additive secret sharing form, that is, the second ciphertext form prediction result, half of which is located on the client and the other half is located on the private server; the client and the private server send their respective halves to the third-party public cloud server, and the third-party public cloud server reconstructs the second ciphertext form prediction result to obtain a second plaintext form prediction result, and then sends the second plaintext form prediction result to the client.

[0017] Furthermore, the second prediction model is an LSTM network model.

[0018] Furthermore, the additive secret sharing form is 2-out-of-2.

[0019] Furthermore, the algorithm used by the intelligent association algorithm module is the k-means clustering algorithm.

[0020] Furthermore, the user corpus module is divided into an update area, a cache area, a work area, and a corpus area; the update area is a queue-based storage used to record and predict the input data in real time; the cache area is used to cache the input data for training; the work area is used to integrate and deduplicate the input data and store it in the corpus area.

[0021] Furthermore, the user preference association module uses the output of the update area of ​​the user corpus module as the input of the second prediction model, and uses the second prediction model to train the cache area of ​​the user corpus module.

[0022] The present invention also provides a method for protecting the privacy of input information based on secure two-party computing, the method comprising the following steps:

[0023] Step 1: The information input module located at the client side segments, filters, and encodes the received input data, and then converts the input data into an additive secret sharing form of two shares using the secret sharing technology in the secure two-party computing protocol module;

[0024] Step 2: The intelligent associative algorithm module located on the private server receives one share of the input data in the additive secret sharing form, and a first prediction model with an intelligent associative algorithm built into the intelligent associative algorithm module performs prediction to obtain a first prediction result in the secret sharing form, that is, a first ciphertext prediction result, half of the first ciphertext prediction result is on the client side, and the other half is on the private server side;

[0025] Step 3: The user corpus module located on the client receives another share of the input data in the additive secret sharing form, and the input data enters the update area of ​​the user corpus module. In the update area, a check is made to determine whether a second prediction model in the user preference association module located on the private server exists. If the second prediction model does not exist, the input data enters a cache area to provide data for training the second prediction model. If the second prediction model in the user preference association module exists, the second prediction model predicts the input data in the update area of ​​the user corpus module to obtain a second prediction result of the second prediction model. The second prediction result is also in the additive secret sharing form, i.e., a second ciphertext form prediction result, half of which is located on the client and the other half is located on the private server.

[0026] Step 4: Send the first ciphertext prediction result and the second ciphertext prediction result to the third-party public cloud server. The third-party public cloud server extracts a large amount of text expressions stored in the cloud corpus module according to the ciphertext prediction results, including phrases, famous sentences, classic sentences, daily sentences, and currently popular pictures and emoticons, and sends them to the client to obtain the final result of intelligent association.

[0027] Furthermore, the algorithm used by the first prediction model is the k-means clustering algorithm.

[0028] Furthermore, the second prediction model is an LSTM network model.

[0029] The present invention provides a system and method for protecting input information privacy based on secure two-party computation, which has at least the following technical effects:

[0030] 1. The technical solution provided by this invention proposes a secure multi-party computing solution to the privacy leakage problem existing in the intelligent association function of intelligent input methods. It designs a secure two-party computing protocol between users and service providers that supports k-means clustering and long short-term memory networks. This ensures that the user's keyboard input content does not leave the local computer under the security assumption of a semi-honest adversary. At the same time, the intelligent association results are kept confidential by the service provider, greatly enhancing security.

[0031] 2. The technical solution provided by this invention introduces a third-party public cloud server to reconstruct the prediction results and send the corpus, pictures, and emoticons associated with the prediction results to the user, so that the user input content is more protected;

[0032] 3. The technical solution provided by the present invention proposes a user customization module in the intelligent association function of the input method. In response to the issues of user personalization, data privacy security and efficiency, a full-cycle solution combining user data privacy security with data value is designed. By coupling the user customization of the intelligent association algorithm to the intelligent association system, the prediction efficiency and prediction confidence of the input method intelligent association are further improved, thereby further improving the user experience.

[0033] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the working process of a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0035] The following describes several preferred embodiments of the present invention with reference to the accompanying drawings to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0036] Existing technologies fail to consider the privacy of user input. Previously, user typing required only local software support, as intelligent association functions were not yet implemented. However, with technological advancements, service providers are now uploading user input from local machines to servers to implement intelligent association. During this process, user input information, or the content of the conversation, is potentially leaked and could be used by the service provider for commercial purposes without the user's control. As people become more aware of privacy, this "minor issue" will become a key consideration for users when using intelligent input methods. Therefore, intelligent input methods with privacy-preserving machine learning capabilities will increasingly attract attention.

[0037] Privacy computing is a computational theory and methodology for protecting private information throughout its lifecycle. It provides a computable model and axiomatic system for privacy measurement, privacy leakage costs, privacy protection, and the complexity of privacy analysis when ownership, management, and usage rights of private information are separated. Current privacy computing primarily encompasses secure multi-party computation (SMPC), federated learning, trusted execution environments (TEEs), differential privacy, and data desensitization.

[0038] Research on secure multi-party computation (SMC) primarily addresses the problem of securely computing a function without a trusted third party. SMC is the cryptographic foundation for numerous applications, including electronic elections, threshold signatures, and electronic auctions. In a SMC protocol, each party only knows its own input and selectively shares the computation results. The input data of other parties and the intermediate results of the computation remain completely confidential to the other participants.

[0039] Machine learning specializes in studying how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their own performance. Its working method is mainly to use user behavior to train computers, so that computers have certain judgment and prediction capabilities.

[0040] Secret sharing (SS) is a cryptographic technique that divides and stores secrets. Its purpose is to prevent over-centralization, thereby distributing risk and tolerating intrusions. It is a crucial tool in information security and data confidentiality. The principle behind secret sharing is to appropriately divide a secret into shares, with each share managed by a different participant. A single participant cannot recover the secret information; only the collaborative efforts of several participants are necessary. More importantly, if any participant within the corresponding range fails, the secret can still be fully recovered. The key to secret sharing lies in optimally designing secret division and recovery methods.

[0041] This invention uses cryptographic techniques such as secret sharing and secure two-party computation to address the privacy issues associated with input method software that reads private user input while providing features such as intelligent association, intelligent image matching, and AI-assisted writing. The technical solution provided by this invention utilizes secure multi-party computing technology. Throughout the entire data transmission and processing process, data is stored in a secret-sharing format, preventing the plaintext data from being viewed by any party. This effectively addresses the privacy protection issue throughout the entire operation of intelligent input methods.

[0042] Specifically, the present invention provides an input information privacy protection system based on secure two-party computation, including a complete set of privacy protection mechanisms for intelligent input methods. The main functions of the current intelligent input method are divided into two types: the first is a basic typing function that can be executed completely locally, which can still work in an offline state and does not need to upload information locally; the second is an intelligent association function deployed on the private server of the software owner supported by a machine learning algorithm. When the user uses these functions, the software automatically uploads the input content to the private server. The private server uses two machine learning algorithms, k-means clustering and long short-term memory network (LSTM), to achieve the above functions. The technical solution provided by the present invention deploys a secure two-party computation protocol on the client and the private server. When the user uses the intelligent association function, the input data is converted into a secret sharing and uploaded. After that, the private server and the client complete the intelligent association function under the secure two-party computation protocol and send the result under the secret sharing situation to the third-party public cloud server where the cloud corpus is located. Finally, the third-party public cloud server sends the result of the intelligent association directly to the user. During the entire process, the private server can only obtain the secret sharing of the intermediate process and the calculation result, thereby effectively protecting the user's input content.

[0043] like Figure 1 FIG. 1 is a schematic diagram of the working process of an input information privacy protection system based on secure two-party computation provided by the present invention, wherein the system includes:

[0044] The secure two-party computing protocol module provides secret sharing technology, including: linear operations, nonlinear operations, similarity measurement and distance comparison. The secure two-party computing protocol module is a secure two-party computing protocol on the client and private server side. When the user uses the intelligent association function, the input data is converted into secret sharing and uploaded. After that, the private server side and the client side complete the intelligent association function under the secure two-party computing protocol and send the results under the secret sharing situation to the third-party public cloud server side where the cloud corpus is located. The secure two-party protocol module is the core part of the present invention, which is responsible for implementing encryption services for all computing work. It uses secret sharing and oblivious transmission in cryptography to implement the secure two-party computing sub-protocols of each operator module. The secure two-party computing can ensure that each party participating in the protocol only knows its own input information and obtains the calculation results according to the protocol. The input data of other parties participating in the protocol are completely confidential to other participants. The intermediate results of the calculation are all stored, transmitted and used in the form of secret sharing.

[0045] The client includes an information input module, which is primarily responsible for extracting keywords from text entered by the user on the keyboard and converting them into mathematical expressions. Its working process includes: first, using string matching to segment the input text and filter keywords; second, character encoding the keywords according to a pre-designed dictionary; and finally, using secret sharing technology to convert the keyword character encoding into a secret sharing value. Specifically, the information input module uses 2-out-of-2 secret sharing technology to convert the received input data into data in the form of two-share additive secret sharing. 2-out-of-2 is a type of secret sharing, defined as t-out-of-n, where n refers to dividing the information into n shares and sharing them with n participants, and t refers to the number of participants greater than or equal to t when the original information can be restored.

[0046] The client also includes a user corpus module, which connects to the information input module and receives one share of data in the form of an additive secret share. The user corpus module is customized for the user's personal input preferences. The client integrates the preceding and following corpus content of the user's input, removes duplicates, and finally outputs the user's private encrypted corpus. The user corpus module is divided into an update area, a cache area, a work area, and a corpus area. The update area is a queue-based storage used to record and predict user input in real time. The cache area is used to cache user input for training. The work area is used to integrate and remove duplicates from the corpus, and then store the integrated corpus in the user corpus area.

[0047] The private server side is connected to the client side and includes an intelligent association algorithm module. The intelligent association algorithm module also includes a first prediction model with a built-in intelligent association algorithm. The intelligent association algorithm module receives another share of data in the form of additive secret sharing, and calculates the first prediction result in the form of secret sharing through the intelligent association algorithm in the first prediction model, that is, the first ciphertext prediction result. Half of the first ciphertext prediction result is on the client side, and the other half is on the private server side. The main function of the intelligent association algorithm module is to use the keywords uploaded by the client as the input of the artificial intelligence algorithm, use the machine learning algorithm to intelligently match word groups with high similarity, and then provide three major functions: intelligent association, intelligent image matching, and AI writing assistance. The main machine learning algorithm used in this module is the clustering algorithm. Compared with other machine learning algorithms, the clustering algorithm is more suitable for multi-category problems and is very suitable for the application scenario of intelligent association. In particular, the algorithm used by the intelligent association algorithm module is the k-means clustering algorithm.

[0048] The private server also includes a user preference association module. The user preference association module connects to the user corpus module, receives the output from the user corpus module, and trains a second prediction model. When the prediction accuracy of the second prediction model reaches a pre-set threshold, the second prediction model is sent to the intelligent association algorithm module as the output of the user preference association module. The intelligent association algorithm module calculates the second prediction result of the second prediction model using the secret sharing technology provided by the secure two-party computation protocol module. This second prediction result is also in the form of additive secret sharing, i.e., a second ciphertext prediction result, half of which is stored on the client and the other half on the private server. The second prediction model is an LSTM network model. The user preference association module is a customized module tailored to the user's private input preferences. The private server uses a pre-built LSTM network to train the user's corpus cache for the user corpus module and periodically outputs the model. The output of the user corpus module's update area is used as the input for model prediction. When the prediction accuracy reaches a certain threshold, the model is put into use as the user preference association module.

[0049] The client and private server send their respective prediction results to the third-party public cloud server. The third-party public cloud server reconstructs the prediction results into plain text and then sends the results directly to the client. The third-party public cloud server includes a cloud corpus module. Due to its large size, a third-party public cloud server is generally used to provide storage space. This module is responsible for storing a large amount of text expressions of the intelligent input method, including phrases, famous sentences, classic sentences, daily sentences, as well as currently popular pictures and emoticons. In the present invention, the cloud corpus sends the corresponding instant chat intelligent association content to the client based on the input of the client and server.

[0050] It also includes an interactive communication module, which realizes interactive communication between the user client and the server of the intelligent input method provider according to the interactive communication request sent by the secure two-party protocol module.

[0051] The present invention also provides a method for protecting the privacy of input information based on secure two-party computing, comprising the following steps:

[0052] When a user uses the intelligent association function of the intelligent input method on the client side, the machine learning algorithm in the intelligent association algorithm module finds relevant corpus information related to the user input. The secure computation provided by the secure two-party computation protocol ensures the security of the user input during the clustering algorithm prediction process through secret sharing of the input information and secure computation via the secure two-party computation protocol. The parties involved in the secure three-party computation are the user client, the intelligent input method software's private server, and a third-party public cloud server. When generating an intelligent association task instruction, the client converts the user input into a 2-out-of-2 additive secret sharing format and transmits one of the shares to the input method software's private server. The two parties then use a trained secure two-party k-means clustering algorithm and LSTM network to perform intelligent association based on the user input. After the client and private server obtain the secret sharing of the output result, they send their respective secret shares to the third-party public cloud server according to the protocol. The third-party public cloud server reconstructs the prediction result from the two secret shares and then sends the corresponding corpus or image emoticon package directly to the client. The above process enables users to ensure the confidentiality of input content and perform intelligent association of intelligent input methods under the security model of a semi-honest adversary. Even if a third-party public cloud and the software's private server collude, only prediction results can be obtained, and it is difficult to make directional judgments on the input content.

[0053] Specifically, the method includes the following steps:

[0054] Step 1: The information input module located on the client side segments, filters, and encodes the received input data, and then uses the secret sharing technology in the secure two-party computation protocol module to convert the input data into an additive secret sharing form of two shares;

[0055] Step 2: The intelligent associative algorithm module located on the private server receives one of the shares of input data in the form of additive secret sharing, and a first prediction model with an intelligent associative algorithm built into the intelligent associative algorithm module performs prediction to obtain a first prediction result in the form of secret sharing, i.e., a first ciphertext prediction result, half of which is stored on the client and the other half on the private server.

[0056] Step 3: The user corpus module located on the client side receives another share of input data in the form of additive secret sharing. The input data enters the update area of ​​the user corpus module, and the update area checks whether the second prediction model in the user preference association module located on the private server side exists. If the second prediction model does not exist, the input data enters the cache area to provide data for training the second prediction model. If the second prediction model in the user preference association module exists, the second prediction model predicts the input data in the update area of ​​the user corpus module to obtain a second prediction result of the second prediction model. The second prediction result is also in the form of additive secret sharing, that is, a second ciphertext prediction result, half of which is located on the client side and the other half is located on the private server side.

[0057] Step 4: Send the first ciphertext prediction result and the second ciphertext prediction result to a third-party public cloud server. The third-party public cloud server extracts a large amount of text expressions stored in the cloud corpus module according to the ciphertext prediction results, including phrases, famous sentences, classic sentences, daily sentences, and currently popular pictures and emoticons, and sends them to the client to obtain the final result of intelligent association.

[0058] Among them, the algorithm used by the first prediction model is the k-means clustering algorithm.

[0059] Among them, the second prediction model is the LSTM network model.

[0060] The preferred embodiments of the present invention have been described in detail above. It should be understood that numerous modifications and variations based on the concepts of the present invention are possible without inventive effort by those skilled in the art. Therefore, any technical solution that can be derived by one skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A system for protecting input information privacy based on secure two-party computation, characterized in that: The system comprises: A secure two-party computing protocol module, wherein the secure two-party computing protocol module provides secret sharing technology; A client, comprising an information input module, the information input module using the secret sharing technology to convert received input data into data in the form of an additive secret sharing of two shares; the client further comprising a user corpus module, the user corpus module connected to the information input module and receiving one of the shares of the data in the form of the additive secret sharing; a private server end, the private server end being connected to the client end, comprising an intelligent association algorithm module, the intelligent association algorithm module further comprising a first prediction model having a built-in intelligent association algorithm, the intelligent association algorithm module receiving another share of the data in the additive secret sharing form, and obtaining a first prediction result in the secret sharing form, i.e., a first ciphertext prediction result, by calculating the intelligent association algorithm in the first prediction model, wherein half of the first ciphertext prediction result is on the client end, and the other half is on the private server end; The third-party public cloud server receives the first ciphertext prediction result, including half on the client and the other half on the private server, reconstructs the first plaintext prediction result, and then sends the first plaintext prediction result to the client.

2. The input information privacy protection system based on secure two-party computation according to claim 1, characterized in that: The private server also includes: A user preference association module is connected to the user corpus module, receives the output of the user corpus module and trains a second prediction model. When the prediction accuracy of the second prediction model reaches a preset threshold, the second prediction model is sent to the intelligent association algorithm module as the output of the user preference association module; the intelligent association algorithm module calculates the second prediction result of the second prediction model through the secret sharing technology provided by the secure two-party computing protocol module, and the second prediction result is also in the additive secret sharing form, that is, the second ciphertext form prediction result, half of which is located on the client and the other half is located on the private server; the client and the private server send their respective halves to the third-party public cloud server, and the third-party public cloud server reconstructs the second ciphertext form prediction result to obtain a second plaintext form prediction result, and then sends the second plaintext form prediction result to the client.

3. The input information privacy protection system based on secure two-party computation according to claim 2, characterized in that: The second prediction model is an LSTM network model.

4. The input information privacy protection system based on secure two-party computation according to claim 1, characterized in that: The additive secret sharing form is 2-out-of-2.

5. The input information privacy protection system based on secure two-party computation according to claim 1, characterized in that: The algorithm used by the intelligent association algorithm module is the k-means clustering algorithm.

6. The input information privacy protection system based on secure two-party computation according to claim 2, characterized in that: The user corpus module is divided into an update area, a cache area, a work area, and a corpus area; the update area is a queue-based storage used to record and predict the input data in real time; The buffer area is used to cache the input data for training; the work area is used to integrate and remove duplicates from the input data and store them in the corpus area.

7. The input information privacy protection system based on secure two-party computation according to claim 6, characterized in that: The user preference association module uses the output of the update area of ​​the user corpus module as the input of the second prediction model, and uses the second prediction model to train the cache area of ​​the user corpus module.

8. A method for protecting input information privacy based on secure two-party computation, characterized in that: The method comprises the following steps: Step 1: The information input module located at the client side segments, filters, and encodes the received input data, and then converts the input data into an additive secret sharing form of two shares using the secret sharing technology in the secure two-party computing protocol module; Step 2: The intelligent associative algorithm module located on the private server receives one share of the input data in the additive secret sharing form, and a first prediction model with an intelligent associative algorithm built into the intelligent associative algorithm module performs prediction to obtain a first prediction result in the secret sharing form, that is, a first ciphertext prediction result, half of the first ciphertext prediction result is on the client side, and the other half is on the private server side; Step 3: The user corpus module located on the client receives another share of the input data in the form of additive secret sharing. The input data enters the update area of ​​the user corpus module, and the update area checks whether the second prediction model in the user preference association module located on the private server exists. If the second prediction model does not exist, the input data enters the cache area to provide data for training the second prediction model. If the second prediction model in the user preference association module exists, the second prediction model predicts the input data in the update area of ​​the user corpus module to obtain a second prediction result of the second prediction model. The second prediction result is also in the form of additive secret sharing, that is, a second ciphertext prediction result, half of which is located on the client and the other half is located on the private server. Step 4: Send the first ciphertext prediction result and the second ciphertext prediction result to the third-party public cloud server. The third-party public cloud server extracts a large amount of text expressions stored in the cloud corpus module according to the ciphertext prediction results, including phrases, famous sentences, classic sentences, daily sentences, and currently popular pictures and emoticons, and sends them to the client to obtain the final result of intelligent association.

9. The input information privacy protection method based on secure two-party computation according to claim 8, characterized in that: The algorithm used by the first prediction model is the k-means clustering algorithm.

10. The input information privacy protection method based on secure two-party computation according to claim 8, characterized in that: The second prediction model is an LSTM network model.

Images