Artificial intelligence platform construction method and system based on privacy protection

Abstract

The invention discloses an artificial intelligence platform construction method and system based on privacy protection, and the platform solves the possible privacy leakage problem in the data processing process through integrating the privacy protection technologies of homomorphic encryption, federated learning, secure multi-party computing and the like. In the privacy protection model training process, privacy technologies such as homomorphic encryption and secure multi-party computing technology are adopted, and efficient training of a privacy protection pre-training model under privacy protection is achieved. In the privacy protection model reasoning process, more accurate cue words can be provided according to the memory module to guide model reasoning so that reasoning is more efficient, in addition, cross-device and cross-scene privacy protection calculation is achieved through the privacy protection-based artificial intelligence platform, and safe circulation and efficient utilization of data are ensured. Omnibearing technical guarantee is provided for privacy protection and data security, and application of the artificial intelligence technology in the privacy protection field is promoted.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and data privacy protection technology, specifically relating to a method and system for constructing an artificial intelligence platform based on privacy protection. Background Technology

[0002] With the continuous development of artificial intelligence theory and technology, AI is being applied more and more widely across various industries. However, in the process of training and using AI models, an increasing amount of sensitive personal data is being collected and processed, leading to a growing risk of privacy breaches. In many cases, users are required to submit sensitive personal data to use AI services, which inevitably brings the potential for privacy leaks.

[0003] To address this issue, how to fully leverage the advantages of artificial intelligence technology while ensuring data privacy has become a critical challenge. While existing privacy protection technologies can safeguard data security to some extent, the efficiency and security of these measures still face challenges in handling large-scale data and complex model training and inference. Therefore, developing an efficient, secure, and transparent privacy-preserving AI platform is of paramount importance. Summary of the Invention

[0004] Based on the above background and the problems existing in the prior art, this invention proposes a method and system for constructing an artificial intelligence platform based on privacy protection.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A privacy-preserving pre-trained model is constructed and trained to obtain a privacy-preserving model. The privacy-preserving pre-trained model includes a logistic regression model and a longitudinal decision tree. Specifically, the modal data of users is acquired and encrypted using a fully homomorphic encryption algorithm. The encrypted data is then used in the logistic regression model. Through a federated learning framework, the privacy data of the participants is acquired and features are extracted to obtain the maximum split value. A longitudinal decision tree model is then constructed. Obtain user question information, construct a memory module, filter and extract from the memory module based on the question information to obtain the first question set; perform inference using a privacy protection model based on the question information and the target prompt words of the first question set, and update the memory module based on the inference results; A privacy-preserving artificial intelligence platform is built based on a privacy-preserving model and an updated memory module, providing privacy-preserving computing and data flow across devices and scenarios.

[0006] As one possible implementation, the construction and training of the privacy-preserving pre-trained model includes the following steps: Fully homomorphic encryption technology is applied to the logistic regression model to perform encrypted calculations on the initial training data, features and biases of the logistic regression model, in order to achieve efficient logistic regression training in a distributed environment. Based on federated learning, a longitudinal decision tree model is constructed by each party independently calculating feature contributions and exchanging encrypted information, which supports training on large-scale datasets and protects the privacy of the model's inference process. Model training is based on a collaborative mechanism of linear layer homomorphism and nonlinear layer OT transformation. Linear layer homomorphism performs linear transformation of the model by encrypting data, while nonlinear layer OT transformation uses the OT protocol to implement nonlinear functions in the model.

[0007] As one possible implementation, applying fully homomorphic encryption technology to a logistic regression model includes the following steps: A public-private key pair is generated using a fully homomorphic encryption algorithm to form an encrypted covariance matrix. The fully homomorphic encryption algorithm includes CKKS. Receive the encrypted covariance matrix, obtain the plaintext data through decryption, calculate the dimension reduction matrix, and encrypt the dimension reduction matrix to obtain the feature ciphertext vector; By iterating until the maximum number of iterations is reached, the final optimized logistic regression model is obtained.

[0008] As one possible implementation method, the construction of the vertical decision tree model includes the following steps: Based on the decision tree model generated in the previous round, the participants calculate the gradient information of each sample and pass the gradient information to other participants. The gradient information includes the first-order gradient and the second-order gradient. The participants calculate multiple split values ​​based on gradient information and select the largest split value to obtain their own maximum split value; By comparing the maximum split values ​​of each participant, the global maximum split value is selected. Based on this maximum split value, the target split point is determined, and the node sample space is divided. The partitioned sample space information is synchronized to all participants to ensure that each participant shares the global structural information of the decision tree, forming a vertical decision tree model.

[0009] As one possible implementation, the model training based on the collaborative mechanism of linear layer homomorphism and nonlinear layer OT has the following shortcomings: The input data and model weights are processed under the protection of homomorphic encryption; In the encrypted state, the weighted summation and matrix multiplication of the linear layer are calculated using homomorphic encryption to ensure that data privacy is not leaked. The nonlinear layer computes the activation function through the OT protocol, which ensures that computation is performed without exposing the data and reduces computational overhead through efficient approximation methods. By converting the encrypted output into plaintext through decryption, the inference process can be completed, allowing for efficient model training while ensuring privacy.

[0010] As one possible implementation method, the inference using a privacy-preserving model includes the following steps: Obtain user question information and build a memory module; Based on the question information, the memory module is filtered and matched for key prompts, and the first question set is formed by extraction. The privacy-preserving model performs inference based on the question information and the first question set, obtains the inference result, and updates the memory module.

[0011] As one possible implementation, the memory module is obtained through the following steps: Obtain the user's historical questions, the corresponding historical inference results, and the model loss value of the corresponding historical inference results; A preset loss threshold is set, and historical inference results that meet the model loss threshold are selected from historical inference results; A memory module is formed based on historical inference results that meet the loss threshold and the corresponding historical questions.

[0012] As one possible implementation, the privacy-preserving artificial intelligence platform includes an engine layer, an application layer, and a product layer, specifically: The engine layer encapsulates various heterogeneous privacy protection technologies through standardized interfaces and incorporates dynamic scheduling strategies, encapsulating them into independent module components to achieve compatibility and integration of heterogeneous multi-source data, ensuring data flow and efficient processing across devices and scenarios; The application layer provides a variety of functional applications. By integrating LLVM compilation optimization technology, it transforms high-level languages ​​into machine language. At the same time, it adopts a data parallel strategy and general GPU hardware acceleration algorithm to build a layered computing framework. The two work together to form a software and hardware integrated privacy protection engine, providing high-performance basic support for the application layer. The product layer includes various application products in different scenarios such as finance and healthcare. By adopting secure multi-party computation that supports encrypted computation, providing privacy enhancement technologies such as differential privacy with provable privacy protection, and performing encrypted gradient aggregation and privacy protection mechanism query mechanisms, the data privacy of all participants is strictly protected.

[0013] A privacy-preserving artificial intelligence platform construction system includes a privacy-preserving training module, a data inference module, and a privacy-preserving computation and data circulation module. The privacy protection training module constructs and trains a privacy protection pre-trained model to obtain a privacy protection model. The privacy protection pre-trained model includes a logistic regression model and a vertical decision tree. Specifically, it acquires the user's modal data and encrypts the data using a fully homomorphic encryption algorithm, and uses the encrypted data for the logistic regression model. Through a federated learning framework, it acquires the privacy data of the participants and extracts features to obtain the maximum split value, and constructs a vertical decision tree model. The data reasoning module acquires user question information, constructs a memory module, filters and extracts information from the memory module based on the question information to obtain a first question set; it then uses a privacy-preserving model to perform reasoning based on the question information and target prompts from the first question set, and updates the memory module based on the reasoning results. The privacy-preserving computation and data circulation module, based on the privacy-preserving model and the updated memory module, constructs a privacy-preserving artificial intelligence platform, providing privacy-preserving computation and data circulation across devices and scenarios.

[0014] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the following method: A privacy-preserving pre-trained model is constructed and trained to obtain a privacy-preserving model. The privacy-preserving pre-trained model includes a logistic regression model and a longitudinal decision tree. Specifically, the modal data of users is acquired and encrypted using a fully homomorphic encryption algorithm. The encrypted data is then used in the logistic regression model. Through a federated learning framework, the privacy data of the participants is acquired and features are extracted to obtain the maximum split value. A longitudinal decision tree model is then constructed. Obtain user question information, construct a memory module, filter and extract from the memory module based on the question information to obtain the first question set; perform inference using a privacy protection model based on the question information and the target prompt words of the first question set, and update the memory module based on the inference results; A privacy-preserving artificial intelligence platform is built based on a privacy-preserving model and an updated memory module, providing privacy-preserving computing and data flow across devices and scenarios.

[0015] An apparatus for building a privacy-preserving artificial intelligence platform includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the following method: A privacy-preserving pre-trained model is constructed and trained to obtain a privacy-preserving model. The privacy-preserving pre-trained model includes a logistic regression model and a longitudinal decision tree. Specifically, the modal data of users is acquired and encrypted using a fully homomorphic encryption algorithm. The encrypted data is then used in the logistic regression model. Through a federated learning framework, the privacy data of the participants is acquired and features are extracted to obtain the maximum split value. A longitudinal decision tree model is then constructed. Obtain user question information, construct a memory module, filter and extract from the memory module based on the question information to obtain the first question set; perform inference using a privacy protection model based on the question information and the target prompt words of the first question set, and update the memory module based on the inference results; A privacy-preserving artificial intelligence platform is built based on a privacy-preserving model and an updated memory module, providing privacy-preserving computing and data flow across devices and scenarios.

[0016] This invention, by adopting the above technical solutions, has significant technical effects: This invention addresses the privacy breach issue caused by collecting sensitive personal information during the use of artificial intelligence. It employs fully homomorphic encryption to encrypt client-side modal data and uses this encrypted data in a logistic regression model. Through a federated learning framework, each participant locally computes its own privacy data features, calculates the corresponding split values ​​to obtain the maximum split value, and collaboratively constructs a vertical decision tree. Simultaneously, it introduces deep learning by combining "linear layer homomorphism" and "non-linear layer OT" mechanisms, applying the privacy mechanism to the training of a classic AI model jointly constructed by decision trees, logistic regression, and deep learning. This privacy-preserving AI platform, including the collaborative work of the engine layer, application layer, and product layer, provides privacy-preserving computation and data flow across devices and scenarios. This privacy-preserving technology ensures data security at every stage, preventing data leakage during transmission or storage. Employing advanced privacy technologies such as homomorphic encryption and federated learning, this system ensures efficient and secure training of classic AI models in a distributed environment. Simultaneously, for inference using the classic AI model, it acquires user query information, extracts a first query set from a pre-built memory module, and filters it based on target prompts matching the query information. The prompts in the first query set contain multiple keywords, indicating the number of prompts in the set. Based on the selected query information and target prompts, inference is performed using the classic AI model, updating the user memory module and constructing the classic AI model inference method. Regarding the construction of the privacy-preserving AI platform, through the collaborative work of the engine layer, application layer, and product layer, it achieves privacy protection for data encryption, secure computation, and cross-scenario data flow, ensuring maximum protection of AI security and user privacy, and providing users with an efficient, secure, and reliable AI service experience. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the modules of the system of the present invention; Figure 3 This is a schematic diagram of the model data processing flow of the present invention; Figure 4 This is a schematic diagram of the model reasoning process of the present invention; Figure 5 This is a schematic diagram of the artificial intelligence platform architecture of the present invention. Detailed Implementation

[0018] To describe the present invention in more detail, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Example 1: A privacy-preserving machine implementation method for an artificial intelligence platform, such as Figure 1 As shown, it includes the following steps: S100. Construct and train a privacy-preserving pre-trained model to obtain a privacy-preserving model. The privacy-preserving pre-trained model includes a logistic regression model and a vertical decision tree. Specifically, the model involves: acquiring the user's modal data and encrypting the data using a fully homomorphic encryption algorithm, and using the encrypted data in the logistic regression model; acquiring the privacy data of the participants and extracting features through a federated learning framework to obtain the maximum split value, and constructing a vertical decision tree model. S200: Obtain user's question information, construct a memory module, filter and extract from the memory module based on the question information to obtain the first question set; perform reasoning using a privacy protection model based on the question information and the target prompt words of the first question set, and update the memory module based on the reasoning results; The S300, based on a privacy-preserving model and an updated memory module, builds a privacy-preserving artificial intelligence platform that provides privacy-preserving computing and data flow across devices and scenarios.

[0020] The schematic diagram of the model data processing flow in this embodiment is as follows: Figure 3As shown, a general privacy function evaluation is applied to classic artificial intelligence models to achieve privacy-preserving machine learning model training and prediction. This avoids privacy data leakage when AI models collect user information. This includes applying homomorphic encryption to logistic regression, introducing vertical federated learning into decision trees, and applying "linear layer homomorphism" and "non-linear layer OT" mechanisms to deep learning models. Through the synergy of these three approaches, a secure and reliable privacy-preserving AI platform can be formed, ensuring that client and service provider data are no longer subject to privacy leaks and guaranteeing that data interaction between clients and service providers is free from the risk of privacy breaches. This improves data security and user trust. A specific diagram illustrating the secure inference process based on classic AI models is shown below. Figure 5 As shown, it mainly consists of the following three parts: (1) Homomorphic encryption applied to logistic regression The required public-private key pair is generated using the key generation algorithm in the fully homomorphic CKKS cryptographic algorithm.

[0021] The key process in the encryption algorithm is as follows: Step 1: Initialize the safety parameter λ, the upper limit of depth L, set N to a power of 2, and the modulus q to an initial large integer. All parameters set must meet the safety parameter requirements.

[0022] Step 2: Construct a polynomial business cycle .

[0023] Step 3: Sample the private key-related distribution from the distribution χ on the polynomial quotient ring. random distribution Error distribution .

[0024] Step 4, from A polynomial s is randomly selected from the given polynomials to form the private key. A polynomial e is randomly selected as the error distribution polynomial, and a polynomial a is randomly and uniformly selected from R to form the public key.

[0025] Step 5: Set the private key Public key .

[0026] Receive the encrypted covariance matrix, obtain the plaintext data through decryption, and then calculate the dimensionality reduction matrix; Specifically, the receiver receives the ciphertext of the transmitted covariance matrix. The plaintext covariance matrix X is obtained by decryption using the CKKS algorithm. Then calculate the eigenvalues ​​and eigenvectors of the covariance matrix X; Arrange the eigenvectors in descending order according to their corresponding eigenvalues and form a matrix as row vectors. Screen and retain the top k most important eigenvectors (the first k rows), and then a matrix P for dimensionality reduction transformation can be spanned.

[0027] Finally, perform eigenvector dimensionality reduction on the matrix.

[0028] Encrypt the dimensionality reduction matrix to obtain the encrypted eigenvector, and transmit it to the server side; Specifically, encrypt the dimensionality reduction matrix obtained in the previous step using the CKKS cryptographic algorithm to obtain the encrypted vector, and transmit it to the server.

[0029] After receiving the encrypted gradient vector marked as the j-th iteration from the server perform decryption operations to obtain the plaintext gradient vector , use this gradient vector to correct and update the currently held model parameters; immediately re-encrypt the new round of model parameters to generate the encrypted parameters dedicated to the (j + 1)-th iteration , and finally transmit this new ciphertext back to the server.

[0030] Specifically, when receiving the encrypted gradient vector of the j-th iteration sent by the server, decrypt it using the CKKS cryptographic algorithm to obtain the plaintext gradient vector .

[0031] Then, according to the plaintext gradient vector, correct the new model parameters and calculate the model parameters for the next iteration ; where ; is the preset learning rate, i represents the dimension of the gradient vector, and 1 < i < k, j represents the current iteration number, is the initial model parameter. <L

[0032] Finally, encrypt the corrected model parameters to obtain the encrypted parameters , and transmit them to the server, which can facilitate the next iteration.

[0033] Continuously receive the encrypted gradient vector and repeat the above steps until the maximum number of iterations is reached to obtain the final model parameters.

[0034] [[ID=5,0]](2)Federated learning to construct a decision tree According to the decision tree model generated at t - 1, the participating parties calculate the first-order and second-order gradients of each sample, and each participating party transmits this gradient information to other participating parties; In this context, the participants are those who possess label information and feature information. The explanation for round t-1 is as follows: it is assumed that all participants with label information obtain the inference results of all samples through a secure joint training process. The label information owner calculates the first-order gradient and second-order gradient of each sample using the inference results of all samples and the label information.

[0035] Each participant calculates multiple split values ​​locally based on its own characteristics, and selects the largest split value to obtain its own maximum split value; In this embodiment of the invention, the splitting method for a node is as follows: each participant calculates all possible splitting values ​​for all features locally, using the following calculation formula:

[0036] in, For the split value, Let this be the set of the left children of the node after the sample space is split. Let this be the set of the right children of the node after the sample space is split. This is the set of sample spaces for this node. Each participant selects the maximum split value from among its multiple split values. ; By comparing the maximum split values ​​calculated by each participant, the global maximum split value is selected, and the target split point is determined based on this maximum split value. Furthermore, this step can be explained in detail as follows: After each participant determines its own maximum split value, each participant sends its own maximum split value to the other participants; all participants compare multiple maximum split values ​​and select the maximum value as a global maximum split value, which is the optimal split point for the target.

[0037] The splitting information of the decision tree node is determined based on the maximum splitting value. The split sample space information will be synchronized to all participants to ensure that each participant can share the global structure information of the decision tree, thereby completing the training of the decision tree model generated in the t-th round.

[0038] Furthermore, this step can be explained in detail as follows: During the decision tree construction process, the corresponding owner participant is determined based on the global maximum split value, and this owner participant is the participant that provides the global maximum split value; based on the global maximum split value and its corresponding owner participant, the splitting features and splitting thresholds of the decision tree node are determined; finally, the sample space of the current node is partitioned based on the splitting features and splitting thresholds, wherein the combination of the splitting features and splitting thresholds constitutes the splitting point of the node.

[0039] Furthermore, in the case where a lower-level node becomes a leaf node after a node splits, the participant does not send the set information after the split, but instead calculates the weight value of the leaf node and only informs other participants that the branch has reached the leaf node.

[0040] The above process is the decision tree training process. Each participant is aware of the global structure information of each decision tree. During the search for split information at each decision tree node, each participant sends the maximum split value of the local possible split point only once, which is used to compare with the global maximum split value.

[0041] (3) Apply the “linear layer homomorphism” and “nonlinear layer OT” mechanisms to deep learning models The input data and model weights are processed under the protection of homomorphic encryption; To further explain, before the inference process begins, the client uses the public key of the homomorphic CKKS encryption algorithm to encrypt both the original input data to be inferred and the weight parameters of the deep learning model. This step converts the plaintext data... Plaintext Model Convert it into the corresponding ciphertext format.

[0042] In the encrypted state, the weighted summation and matrix multiplication of the linear layer are calculated using homomorphic encryption to ensure that data privacy is not leaked. For linear operation layers in neural networks, including but not limited to fully connected layers and convolutional layers, linear transformations are used.

[0043] The system directly performs homomorphic addition and homomorphic multiplication operations on the ciphertext to generate the resulting ciphertext.

[0044] The entire process is completed within the ciphertext domain, ensuring that the server cannot obtain any plaintext information about the plaintext data or intermediate results throughout the entire computation process.

[0045] The nonlinear layer computes the activation function through the OT protocol, which ensures that computation is performed without exposing the data and reduces computational overhead through efficient approximation methods. To further explain, the ReLU activation function employs the OT protocol. The client and server interact via the OT protocol to collaboratively compute the activation function. One side's input is the shared share corresponding to the ciphertext, while the other side's input is the auxiliary information required to compute the activation function. Taylor expansion is used instead of precise computation to significantly reduce the number of communication rounds and computational overhead required by the OT protocol, achieving a balance between efficiency and security.

[0046] By converting the encrypted output into plaintext through decryption, the inference process can be completed, allowing for efficient training of neural networks while ensuring privacy.

[0047] After the encrypted data undergoes collaborative computation across all network layers (linear and nonlinear layers), the resulting inference is still ciphertext.

[0048] The encrypted result is sent back to the client, which uses its homomorphic encryption private key to decrypt it and obtain the final plaintext reasoning result.

[0049] A flowchart illustrating the inference process of the privacy-preserving model is shown below. Figure 4 As shown, the specific process is as follows: Obtain user question information; The question information is text or voice information entered by the user on the terminal device.

[0050] Build and store a memory module containing the first set of prompts; In the example, the memory module can be constructed by storing the first set of prompts into a preset database, which can be an initialized database or a database that pre-stores other data, wherein at least some of the prompts can refer to preset prompts used to guide the model in reasoning.

[0051] Extract the first set of questions from the memory module and filter them based on N key prompts that match the question information; In this embodiment, N is 20, but it should be understood that N can also be other positive integers; In this embodiment, the memory module can be constructed by storing the first set of prompts in a preset database. The preset database can be an initialized database or a database that has pre-stored other data. At least some of the prompts are preset prompts used to guide classic artificial intelligence models in performing inference tasks.

[0052] The model performs inference based on the filtered information and prompts, obtains inference information, and updates the memory module. In one exemplary implementation, the question information and N target prompts can be directly input into a large model for reasoning to generate the corresponding answer; alternatively, the question information and N target prompts can be preprocessed before being input into a classic artificial intelligence model for reasoning.

[0053] In this example, the construction process of the memory module is as follows: Collect users' historical questions and obtain the historical inference results for the corresponding historical questions and the model loss values ​​used for the corresponding historical inference results; Filter historical inference results to identify those where the model loss value is less than the corresponding set loss threshold; By filtering out the historical reasoning results that meet the conditions and the corresponding historical questions, a memory module storing the first set of prompts is constructed.

[0054] In the example, historical issues are tagged as The corresponding historical reasoning results are marked as The model calculates the loss value for each historical inference result and uses that loss value as the corresponding model loss value.

[0055] In one alternative implementation, the step involves filtering historical results from the historical inference results to identify those with model loss values ​​less than a corresponding set loss threshold.

[0056] Based on the corresponding model loss value, samples consisting of historical inference results and their corresponding historical questions with model loss values ​​lower than the first set loss threshold are rated as high-quality samples. Target prompt words are then matched from these samples to effectively guide the model to perform accurate inference.

[0057] In one optional implementation, historical reasoning results that meet the conditions and their corresponding historical questions are filtered out to construct the memory module storing the first set of prompts, including: The selected historical reasoning results and their corresponding historical questions are used to construct a dictionary in the form of key-value pairs; The dictionary is stored in a preset memory bank, and the information in the memory bank is encoded using an embedding model to obtain the memory module, wherein the first hint set is at least a portion of the information in the encoded dictionary stored in the memory module.

[0058] In this embodiment, the memory module can be dynamically updated based on the subsequently generated answers and their corresponding user questions, thereby enabling the memory module to provide more accurate prompts for the classic artificial intelligence model during subsequent reasoning and effectively guiding the reasoning process.

[0059] Privacy-preserving AI platform architecture diagram Figure 5 As shown, it consists of an engine layer, an application layer, and a product layer. The engine layer serves as the foundational support layer, providing core capabilities for upper-layer services; the application layer acts as a bridge, both calling upon the capabilities of the engine layer and providing general services to the product layer; the product layer, located at the top, provides specific solutions to different business domains by calling upon the capabilities of the application layer.

[0060] The engine layer is the core component of the platform, responsible for implementing its core computing tasks, including data processing, privacy-preserving computation, algorithm optimization, and security assurance. It comprises a privacy-preserving technology library, automated code auditing, horizontal / vertical federated learning, compatibility with diverse and heterogeneous data, and a user-friendly unified user interface. Through compatibility with federated learning, multi-party security protocols, and other privacy-preserving technologies, it supports various application scenarios such as joint modeling / inference, joint querying / statistics, and privacy information retrieval. The platform is compatible with real-time input of multi-source heterogeneous data, constructing cross-device and cross-scenario joint shared computing models, and establishing a hardware-software collaborative privacy-preserving method to achieve efficient joint computation of models. The platform deploys efficient automated code auditing and static analysis technologies to ensure code functionality, memory safety, and improve code security, guaranteeing the security of privacy-preserving code.

[0061] The engine layer encapsulates various heterogeneous privacy protection technologies through standardized interfaces and incorporates dynamic scheduling strategies, encapsulating them into independent module components to achieve compatibility and integration of heterogeneous multi-source data, ensuring data flow and efficient processing across devices and scenarios.

[0062] The engine layer uses a privacy algorithm to protect the library. By combining homomorphic encryption technology, encrypted computation of data is performed to avoid the risk of data leakage. Specifically, input data... via public key After encryption, the platform can perform addition operations in the encrypted state. The encryption process is represented by the following formula:

[0063] The engine layer uses an automated code auditing module ( This system automatically identifies and repairs potential vulnerabilities in encryption operations, employs a formal verification language to ensure the security of computational steps, and automates code auditing by verifying the algorithm logic. The specific security verification process is as follows:

[0064] Simultaneously, a joint horizontal and vertical learning framework is introduced, enabling collaborative training of data from different data sources and devices. The combination of horizontal and vertical learning achieves joint modeling of distributed heterogeneous data, further improving the generalization ability and training efficiency of the learning model. The derivation process of model updates is expressed by the following formula:

[0065] Among them is The model in the current learning cycle, It's the model from the previous cycle. It is an update item obtained by weighted combination of data from various parties.

[0066] The engine layer also supports diverse and heterogeneous data modules. It standardizes data from different formats, converting heterogeneous data such as images, text, and sensor data into a unified format for processing. Data embedding operations... Ensure that data formats are standardized so that all data can be effectively processed and analyzed.

[0067] Finally, the engine layer provides a simple and efficient user experience through an easy-to-use and unified user interface. Users can configure privacy protection algorithms, perform code audits, launch joint learning tasks, and analyze diverse and heterogeneous data through this interface.

[0068] The application layer provides a variety of functional applications. By integrating LLVM compilation optimization technology, it transforms high-level languages ​​into machine language. At the same time, it adopts a data parallel strategy and general GPU hardware acceleration algorithms to build a layered computing framework. The two work together to form a software and hardware integrated privacy protection engine, providing high-performance basic support for the application layer.

[0069] The application layer consists of joint modeling / inference, joint querying / statistics, joint privacy computation, privacy information retrieval, and privacy data authorization. The underlying implementation of the application layer requires the use of LLVM-based compilation optimization techniques and GPU-based privacy-preserving acceleration algorithms. The application layer is the platform's functional implementation layer, providing various privacy-preserving applications and supporting cross-industry data flow, analysis, and decision-making tasks.

[0070] The application layer source code needs to be compiled with LLVM to transform the source code into an intermediate representation (IR) and perform multi-stage optimization, including dead code elimination, constant folding, and loop optimization techniques to improve execution efficiency. LLVM can optimize code for different hardware architectures. LLVM analyzes the parallel tasks of programs for multi-core GPUs, transforms the parallel execution source code into multi-threaded code, and uses the CUDA parallel interface to adapt the parallel code to the GPU.

[0071] GPUs divide the dataset into multiple sub-data blocks and execute the same encryption operation in parallel on each processing core. By using addition and multiplication operations and employing a data parallelization strategy, GPUs can process multiple computing tasks in parallel at the same time.

[0072] Data-parallel training requires the use of a memory pool management approach. This involves pre-allocating a large block of memory and dividing it into multiple smaller blocks to reduce the overhead of frequent memory allocation and deallocation. At program startup, APIs such as cudaMalloc are used for initialization, allocating memory blocks from the pre-allocated memory pool. The memory pool tracks unused memory blocks by maintaining a free list. When the memory in the memory pool is exhausted, additional memory is allocated through a dynamic expansion mechanism. Memory management is achieved by merging free memory blocks and optimizing memory block size allocation to reduce fragmentation.

[0073] Finally, homomorphic encryption is used to encrypt the GPU. The addition and multiplication operations in homomorphic encryption can be broken down into multiple independent computational tasks and distributed across multiple GPU cores for parallel execution. Leveraging the GPU's CUDA programming model, the GPU-accelerated algorithm achieves a significant improvement in computational efficiency by employing efficient modular arithmetic algorithms and optimized numerical computation techniques.

[0074] It supports the acquisition of data from multiple heterogeneous data sources, encrypts key data fields, and ensures the secure flow and application of data elements. The application layer features an open execution operator development framework, allowing partners and business users to customize and integrate operators into the product, thus expanding product functionality. The product can provide customers with data element security solutions to support their digital transformation.

[0075] The product layer includes various application products in different scenarios such as finance and healthcare. By adopting secure multi-party computation that supports encrypted computation, providing privacy enhancement technologies such as differential privacy with provable privacy protection, and performing encrypted gradient aggregation and privacy protection mechanism query mechanisms, the data privacy of all participants is strictly protected.

[0076] As an important component of this platform, the product layer consists of joint detection of financial retail credit risk, cross-institutional financial blacklist model, joint research on drug development experiments, medical data collaborative evaluation system, and security government information statistical verification mechanism.

[0077] The joint detection of financial retail credit risk constructs a cross-institutional joint modeling framework by combining federated learning and homomorphic encryption. The platform performs model training through encrypted gradient aggregation and distributed training, ensuring that the data privacy of all participants is strictly protected. By adopting risk scoring algorithms to encryptedly aggregate data features, it improves the accuracy of financial risk assessment and the scientific nature of risk control decisions.

[0078] The cross-institutional financial blacklist sharing model employs multi-party secure computation and encrypted query protocols. Each participating party exchanges data through encrypted matching algorithms and secure query protocols, ensuring encrypted protection of data during storage and transmission when sharing sensitive information. Furthermore, encrypted queries optimize the real-time matching and sharing efficiency of blacklist data.

[0079] Collaborative drug development research leverages differential privacy and federated learning to support data sharing and joint modeling across multiple research institutions. Data is encrypted and computed locally, and joint training is performed using privacy-preserving gradient aggregation to protect the privacy of experimental data and prevent the leakage of sensitive information. Differential privacy technology further prevents inferences and leaks based on single data points during model training.

[0080] The medical data collaborative evaluation system employs federated learning, homomorphic encryption, and smart contract technologies to support the secure statistical analysis and verification of government data. In the process of cross-departmental data sharing and risk control modeling, encrypted aggregation and secure collaborative analysis mechanisms are used to ensure the privacy of all participants during data computation, while also providing compliance and risk assessment. Smart contract technology is used to automate the verification and execution of data usage agreements, ensuring the compliance and legality of data access.

[0081] Example 2: A privacy-preserving artificial intelligence platform construction system, such as Figure 2 As shown, it includes a privacy-preserving training module 100, a data inference module 200, and a privacy-preserving computation and data circulation module 300: The privacy protection training module 100 constructs and trains a privacy protection pre-trained model to obtain a privacy protection model. The privacy protection pre-trained model includes a logistic regression model and a vertical decision tree. Specifically, it acquires the user's modal data and encrypts the data using a fully homomorphic encryption algorithm, and uses the encrypted data for the logistic regression model. Through a federated learning framework, it acquires the privacy data of the participants and extracts features to obtain the maximum split value, and constructs a vertical decision tree model. The data reasoning module 200 acquires user question information, constructs a memory module, filters and extracts information from the memory module based on the question information to obtain a first question set; it then uses a privacy protection model to perform reasoning based on the question information and the target prompts in the first question set, and updates the memory module based on the reasoning results. The privacy-preserving computing and data circulation module 300, based on the privacy-preserving model and the updated memory module, constructs a privacy-preserving artificial intelligence platform, providing privacy-preserving computing and data circulation across devices and scenarios.

[0082] The above description of the embodiments is provided to enable those skilled in the art to understand and apply the present invention. It will be apparent to those skilled in the art that various modifications can be made to the above embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made to the present invention by those skilled in the art based on the disclosure thereof should be within the scope of protection of the present invention.

Claims

1. A method for constructing a privacy-preserving artificial intelligence platform, characterized in that, Includes the following steps: A privacy-preserving pre-trained model is constructed and trained to obtain a privacy-preserving model. The privacy-preserving pre-trained model includes a logistic regression model and a longitudinal decision tree. Specifically, the modal data of users is acquired and encrypted using a fully homomorphic encryption algorithm. The encrypted data is then used in the logistic regression model. Through a federated learning framework, the privacy data of the participants is acquired and features are extracted to obtain the maximum split value. A longitudinal decision tree model is then constructed. Obtain user question information, construct a memory module, filter and extract from the memory module based on the question information to obtain the first question set; perform inference using a privacy protection model based on the question information and the target prompt words of the first question set, and update the memory module based on the inference results; A privacy-preserving artificial intelligence platform is built based on a privacy-preserving model and an updated memory module, providing privacy-preserving computing and data flow across devices and scenarios.

2. The method for constructing a privacy-preserving artificial intelligence platform according to claim 1, characterized in that, The construction and training of the privacy-preserving pre-trained model includes the following steps: Fully homomorphic encryption technology is applied to the logistic regression model to perform encrypted calculations on the initial training data, features and biases of the logistic regression model, in order to achieve efficient logistic regression training in a distributed environment. Based on federated learning, a longitudinal decision tree model is constructed by each party independently calculating feature contributions and exchanging encrypted information. This supports training on large-scale datasets and protects the privacy of the model's inference process. Model training is based on a collaborative mechanism of linear layer homomorphism and nonlinear layer OT transformation. Linear layer homomorphism performs linear transformation of the model by encrypting data, while nonlinear layer OT transformation uses the OT protocol to implement nonlinear functions in the model.

3. The method for constructing a privacy-preserving artificial intelligence platform according to claim 2, characterized in that, The application of fully homomorphic encryption technology to the logistic regression model includes the following steps: A public-private key pair is generated using a fully homomorphic encryption algorithm to form an encrypted covariance matrix. The fully homomorphic encryption algorithm includes CKKS. Receive the encrypted covariance matrix, obtain the plaintext data through decryption, calculate the dimension reduction matrix, and encrypt the dimension reduction matrix to obtain the feature ciphertext vector; By iterating until the maximum number of iterations is reached, the final optimized logistic regression model is obtained.

4. The method for constructing a privacy-preserving artificial intelligence platform according to claim 1, characterized in that, The construction of the vertical decision tree model includes the following steps: Based on the decision tree model generated in the previous round, the participants calculate the gradient information of each sample and pass the gradient information to other participants. The gradient information includes the first-order gradient and the second-order gradient. The participants calculate multiple split values ​​based on gradient information and select the largest split value to obtain their own maximum split value; By comparing the maximum split values ​​of each participant, the global maximum split value is selected. Based on this maximum split value, the target split point is determined, and the node sample space is divided. The partitioned sample space information is synchronized to all participants to ensure that each participant shares the global structural information of the decision tree, forming a vertical decision tree model.

5. The method for constructing a privacy-preserving artificial intelligence platform according to claim 2, characterized in that, The proposed model training mechanism based on the combined linear layer homomorphism and nonlinear layer OT has the following shortcomings: The input data and model weights are processed under the protection of homomorphic encryption; In the encrypted state, the weighted summation and matrix multiplication of the linear layer are calculated using homomorphic encryption to ensure that data privacy is not leaked. The nonlinear layer computes the activation function through the OT protocol, which ensures that computation is performed without exposing the data and reduces computational overhead through efficient approximation methods. By converting the encrypted output into plaintext through decryption, the inference process can be completed, allowing for efficient model training while ensuring privacy.

6. The method for constructing a privacy-preserving artificial intelligence platform according to claim 1, characterized in that, The reasoning using the privacy-preserving model includes the following steps: Obtain user question information and build a memory module; Based on the question information, the memory module is filtered and matched for key prompts, and the first question set is formed by extraction. The privacy-preserving model performs inference based on the question information and the first question set, obtains the inference result, and updates the memory module.

7. The method for constructing a privacy-preserving artificial intelligence platform according to claim 1, characterized in that, The memory module is obtained through the following steps: Obtain the user's historical questions, the corresponding historical inference results, and the model loss value of the corresponding historical inference results; A preset loss threshold is set, and historical inference results that meet the model loss threshold are selected from historical inference results; A memory module is formed based on historical inference results that meet the loss threshold and the corresponding historical questions.

8. The method for constructing a privacy-preserving artificial intelligence platform according to claim 1, characterized in that, The privacy-preserving AI platform comprises an engine layer, an application layer, and a product layer, specifically: The engine layer encapsulates various heterogeneous privacy protection technologies through standardized interfaces and incorporates dynamic scheduling strategies, encapsulating them into independent module components to achieve compatibility and integration of heterogeneous multi-source data, ensuring data flow and efficient processing across devices and scenarios; The application layer provides a variety of functional applications. By integrating LLVM compilation optimization technology, it transforms high-level languages ​​into machine language. At the same time, it adopts a data parallel strategy and general GPU hardware acceleration algorithm to build a layered computing framework. The two work together to form a software and hardware integrated privacy protection engine, providing high-performance basic support for the application layer. The product layer includes various application products in different scenarios such as finance and healthcare. By adopting secure multi-party computation that supports encrypted computation, providing privacy enhancement technologies such as differential privacy with provable privacy protection, and performing encrypted gradient aggregation and privacy protection mechanism query mechanisms, the data privacy of all participants is strictly protected.

9. A privacy-preserving artificial intelligence platform construction system, characterized in that, It includes a privacy-preserving training module, a data inference module, and a privacy-preserving computation and data circulation module: The privacy protection training module constructs and trains a privacy protection pre-trained model to obtain a privacy protection model. The privacy protection pre-trained model includes a logistic regression model and a vertical decision tree. Specifically, it acquires the user's modal data and encrypts the data using a fully homomorphic encryption algorithm, and uses the encrypted data for the logistic regression model. Through a federated learning framework, it acquires the privacy data of the participants and extracts features to obtain the maximum split value, and constructs a vertical decision tree model. The data reasoning module acquires user question information, constructs a memory module, filters and extracts information from the memory module based on the question information to obtain a first question set; it then uses a privacy-preserving model to perform reasoning based on the question information and target prompts from the first question set, and updates the memory module based on the reasoning results. The privacy-preserving computation and data circulation module, based on the privacy-preserving model and the updated memory module, constructs a privacy-preserving artificial intelligence platform, providing privacy-preserving computation and data circulation across devices and scenarios.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.

11. A privacy-preserving artificial intelligence platform construction apparatus, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 8.

Images