A collaborative processing system of private and public domains of AI models based on data security and privacy protection

Abstract

The present invention discloses a private domain and public domain collaborative processing system of an AI model based on data security and privacy protection, the present invention relates to the field of artificial intelligence, data security and privacy protection technology, including a collaborative processing platform, the collaborative processing platform is communicatively connected with a public-private domain strategy selection module, a private domain data processing module, a secure transmission channel module, a public domain model training and optimization module, a privacy protection mechanism module and a model feedback and fusion module, wherein the modules are connected by electrical signals. The private domain and public domain collaborative processing system of an AI model based on data security and privacy protection effectively solves the data security and privacy protection problems faced by the AI ​​model during training and deployment by constructing a secure collaborative processing mechanism between the private domain and the public domain, which not only improves the generalization ability and accuracy of the model, but also ensures the privacy security of user data, providing a solid guarantee for the widespread application of AI technology.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, data security and privacy protection, and specifically to a private domain and public domain collaborative processing system of an AI model based on data security and privacy protection. Background Art

[0002] With the rapid development of artificial intelligence technology, AI models are increasingly used in all walks of life. However, data security and privacy have become important bottlenecks restricting their further application. In actual applications, enterprises often face the problem of data islands, that is, data is scattered in multiple private environments and is difficult to effectively integrate and utilize. At the same time, due to the strict implementation of data privacy protection regulations (such as GDPR, CCPA, etc.), when enterprises use these data to train AI models, they must strictly abide by the privacy protection principles to prevent data leakage.

[0003] In existing technologies, AI models usually need to be trained and deployed in a single data environment. They either rely entirely on private domain data, resulting in insufficient model generalization capabilities; or they upload sensitive data to the public domain, posing a serious risk of privacy leakage. Therefore, how to achieve effective collaborative processing of AI models between private and public domains while ensuring data security and privacy has become an urgent problem to be solved. Summary of the invention

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: a private and public domain collaborative processing system of an AI model based on data security and privacy protection, comprising a collaborative processing platform, wherein the collaborative processing platform is communicatively connected with a public-private strategy selection module, a private data processing module, a secure transmission channel module, a public model training and optimization module, a privacy protection mechanism module, and a model feedback and fusion module, wherein the modules are connected by electrical signals;

[0005] The public-private domain strategy selection module is used to extract information features input by users and match them with AI models for processing, ensuring that different types of data are classified and processed before entering the processing flow, thereby improving processing efficiency and accuracy;

[0006] The private domain data processing module is used to pre-process sensitive data in a local or private cloud environment. The pre-processing operation includes data encryption, desensitization, and anonymization steps to ensure that the data meets security standards before leaving the private domain, effectively protecting the privacy of user data and preventing data leakage risks;

[0007] The secure transmission channel module is used to establish an encrypted and verifiable data transmission channel to securely transmit processed data and model parameters between the private domain and the public domain, prevent the data from being intercepted or tampered with during transmission, and ensure the integrity and security of the data;

[0008] The public domain model training and optimization module is used to train, verify and optimize the model in a public domain environment using data and model parameters securely transmitted from the private domain, thereby improving the generalization ability and accuracy of the model, and applying a privacy protection mechanism to ensure that the privacy information of the private domain data is not leaked during the model training process;

[0009] The privacy protection mechanism module is used to adopt differential privacy and federated learning technologies to protect the privacy information of private domain data in model training in the public domain. Differential privacy protects individual privacy by adding random noise to the data while ensuring the accuracy of statistical results. Federated learning allows multiple participants to train models locally and only share model parameter updates without directly sharing data.

[0010] The model feedback and fusion module is used to feed back the optimized model parameters and knowledge to the private domain, and fuse them with the original model in the private domain to form a new AI model that is more accurate and protects privacy, thereby achieving continuous optimization and improvement of the model and improving the performance and accuracy of the model.

[0011] Preferably, in the public-private domain strategy selection module, the process of extracting information features and matching the AI ​​model includes:

[0012] When users input data, the data is parsed through natural language processing technology to identify and extract relevant information features, including the type, format, content, and sensitivity of the data;

[0013] According to the extracted information features, the decision tree classification algorithm is used to classify the data, calculate the information gain of each feature, analyze the entropy value of the data set, determine the purity of the data set, and classify the data into different categories for subsequent processing and model matching. By calculating the information gain of each feature, the feature that can minimize the entropy is selected for segmentation, and the best decision is made at each node to maximize the distinction of the category;

[0014] Based on the data category and information features, the corresponding AI model is matched for each type of data from the model library. The matching process is based on multiple dimensions such as model performance, scope of application, and data processing capabilities, and the most suitable AI model for the current data category is selected for subsequent processing;

[0015] Before the data enters the matching AI model, it is preprocessed. The preprocessing includes data cleaning, formatting, and normalization operations. According to the sensitivity of the data, the privacy protection requirements during the data processing are evaluated.

[0016] Different classification and processing strategies are developed for different types of data, including data flow, processing order, and AI models used. For highly sensitive data, priority is given to processing in the private domain to avoid the risk of data leakage. For less sensitive data, a wider range of data sets are used in the public domain for model training and optimization;

[0017] Through intelligent AI model matching and classification processing strategies, different types of data are classified and processed, and the formulated classification processing strategies are applied to the actual data flow to ensure that the data has received the most appropriate processing solution before entering the subsequent processing process.

[0018] Preferably, the calculation expression of the information gain is:

[0019] ;in, For the entire dataset The entropy of the dataset is the uncertainty or confusion of the dataset. For the dataset Belong to the A subset of categories, For the The number of samples in each category, is the number of samples in the entire data set, For the The entropy of a subset of categories, is the total number of categories in the dataset;

[0020] The calculation expression of the entropy value of the data set is:

[0021] ;in, For the data set The probability of the categories, is the total number of categories in the dataset.

[0022] Preferably, in the private domain data processing module, the process of preprocessing sensitive data includes:

[0023] Use data discovery tools to scan the entire data set, identify data fields containing sensitive information, and classify them. According to the sensitivity of the data, the data is divided into different sensitivity levels, namely high sensitivity level, medium sensitivity level, and low sensitivity level, and the corresponding encryption algorithm and key length are matched according to the sensitivity of the data;

[0024] Through encryption technology, combined with key management strategies, the identified and classified sensitive data is encrypted and converted into a ciphertext form that cannot be directly read, preventing unauthorized personnel from accessing or leaking data;

[0025] In the data preprocessing stage, the data fields that need to be desensitized are identified, and corresponding desensitization technologies (such as replacement, hashing, generalization, etc.) are used to process sensitive data according to the type and sensitivity of the data, remove or replace sensitive information in the data, reduce the risk of data leakage, and maintain the availability and analytical value of the data;

[0026] Identify data fields containing personal identification information and process the data using anonymization techniques (such as differential privacy, k-anonymity, l-diversity, etc.) to ensure that the data cannot be traced back to specific individuals while retaining its analytical value. Remove or replace personal identification information in the data so that the data cannot be directly associated with a specific individual, thereby protecting personal privacy.

[0027] Preferably, in the secure transmission channel module, the process of establishing the data transmission channel includes:

[0028] The initial data transmission connection is established between the private domain and the public domain. The two parties use the pre-agreed security protocol to handshake and confirm the identity and communication parameters of the other party. During the handshake process, the two parties use digital certificates to confirm the identity of the other party to ensure that the communicating parties are the expected entities and prevent man-in-the-middle attacks.

[0029] Use key exchange protocols to securely exchange keys between communicating parties, and regularly replace and update keys to reduce the risk of key cracking;

[0030] Before data is sent, the negotiated encryption algorithm and key are used to encrypt sensitive data, converting plaintext data into ciphertext. Even if the data is intercepted during transmission, it cannot be read by unauthorized parties. The sender uses an asymmetric encryption algorithm to digitally sign the data so that the receiver can verify the source and integrity of the data by verifying the digital signature.

[0031] The encrypted data is sent through the established transmission channel. During the transmission process, the data is encapsulated in the transmission protocol for transmission to ensure the integrity of the data during transmission and prevent tampering;

[0032] When the public domain end receives the data, it uses the corresponding decryption algorithm and key to decrypt the data, and uses the digital signature to verify the integrity of the data to ensure that the data has not been tampered with during transmission.

[0033] Preferably, in the public domain model training and optimization module, the process of model training, verification and optimization includes:

[0034] The public domain environment receives data and model parameters from the private domain through a secure transmission channel;

[0035] Decrypt the data using the corresponding decryption key, convert the encrypted data into a usable plaintext format for model training, and preprocess the decrypted data, including data cleaning, normalization, and feature selection, to improve data quality and make it more suitable for model training;

[0036] Initialize the AI ​​model in the public domain using the received model parameters, input the preprocessed data into the AI ​​model for training, and during the training process, apply the optimization algorithm to minimize the loss function, update the parameters of the AI ​​model, and through multiple iterations of training, allow the AI ​​model to gradually learn the inherent laws and characteristics of the data;

[0037] Use a validation dataset independent of the training data to evaluate the performance of the model, quantify the performance of the model by calculating the accuracy, recall, and F1 score indicators, adjust and optimize the model based on the validation results to improve its generalization ability, and adjust the hyperparameters of the AI ​​model based on the validation results.

[0038] Preferably, the process of adjusting the AI ​​model hyperparameters includes:

[0039] Before training begins, set the hyperparameter set , initially set the hyperparameter set Each hyperparameter The range of hyperparameters is clear, and each hyperparameter The range is ;

[0040] Select Bayesian optimization as the hyperparameter optimization strategy, and use Gaussian process as the core of Bayesian optimization to calculate the mean and variance functions of Gaussian process;

[0041] For each set of hyperparameters , train the model and calculate the loss function of the model on the validation set to evaluate its performance;

[0042] Use the loss function results on the validation set to update the kernel matrix of the Gaussian process, and use the new observation data to update the mean and variance of the Gaussian process, record the model performance under each set of hyperparameters, and use the Bayesian optimization acquisition function Measure expected improvement and calculate acquisition function Select the next set of hyperparameters. The maximized hyperparameters are used as the next set of hyperparameters;

[0043] Repeat the process of model training evaluation and hyperparameter recording, with a preset number of iterations , for each iteration , update the Gaussian process model and select new hyperparameters until the predetermined number of iterations is reached After all iterations are completed, the loss function on the validation set is selected The minimized hyperparameters are the set of hyperparameters with the best performance.

[0044] Preferably, the expression of the Gaussian process mean function is:

[0045] ;in, Represents a given input The mean prediction at Represents the kernel function of the Gaussian process, measuring the input Similarity to oneself;

[0046] The expression of the Gaussian process variance function is:

[0047] ;in, Represents a given input The prediction variance at represents the kernel function, represents the variance of the noise, Represents the kernel function, which measures the input With the evaluated point The similarity between Represents the diagonal elements of the kernel function, measuring the evaluated points The similarity between itself and represents the number of evaluated points, that is, the number of observed data points used for prediction in the Gaussian process;

[0048] The expression of the acquisition function of the Bayesian optimization is:

[0049] ;in, represents the acquisition function, which is used to measure the given input Expected improvements in represents the mean prediction of the Gaussian process, represents the standard deviation prediction of the Gaussian process, represents the hyperparameters used to control risk and exploration.

[0050] Preferably, in the privacy protection mechanism module, the process of protecting the privacy information of private domain data includes:

[0051] Integrate the pre-processed private domain data, perform sensitivity calculations during data analysis, analyze the maximum difference in function outputs on adjacent data sets, preset privacy budget parameters, quantify the strength of privacy protection, use the Laplace mechanism to calculate the results after adding noise, output the results after adding noise to protect the privacy of the data, and provide statistical information;

[0052] Each participant trains the model using his or her own dataset in his or her own private environment and updates the model parameters.

[0053] Implement federated learning among multiple participants to conduct local model training and parameter update sharing. After training is completed, each participant sends the model parameter update to a central server in the public domain.

[0054] The central server receives model parameter updates from each participant and uses the federated averaging algorithm to aggregate parameters to update the global model.

[0055] Preferably, in the model feedback and fusion module, the process of forming a new AI model includes:

[0056] In the public domain environment, the trained and optimized model parameters are collected, and the model parameters in the public domain are encrypted and transmitted back to the private domain using a secure transmission channel to ensure the security and integrity of the parameters during the transmission process and prevent the data from being intercepted or tampered with.

[0057] Verify and preprocess the received model parameters to make them suitable for private domain use, ensure the compatibility of the parameters with the private domain model, and prepare for model fusion;

[0058] The preprocessed public domain model parameters are merged with the private domain model parameters to form a new model. The new model combines the specific knowledge of private domain data with the general rules of public domain data. If the private domain model has not been trained before, the public domain parameters are used as initialization. If the private domain model has been trained, the public domain parameters are combined with the private domain parameters.

[0059] Evaluate the performance of the new model after fusion on the private domain validation set to verify the effect of model fusion and ensure that the new model has good performance on private domain data. Fine-tune the new model and adjust hyperparameters based on the verification results to further improve the performance and accuracy of the model.

[0060] Deploy verified and adjusted new models to the production environment to improve the intelligence level of business processes, and continuously monitor model performance in actual applications, collect feedback, and regularly update the model based on business needs and model performance to form a closed loop of continuous optimization.

[0061] The present invention provides an AI model private domain and public domain collaborative processing system based on data security and privacy protection. It has the following beneficial effects:

[0062] 1. This AI model private and public domain collaborative processing system based on data security and privacy protection effectively solves the data security and privacy protection problems faced by AI models during training and deployment by building a secure collaborative processing mechanism between private and public domains. It not only improves the generalization ability and accuracy of the model, but also ensures the privacy and security of user data, providing a solid guarantee for the widespread application of AI technology.

[0063] 2. This AI model private and public domain collaborative processing system based on data security and privacy protection breaks the data silos by building a collaborative processing mechanism between private and public domains, realizes the effective integration and utilization of cross-domain data, and improves the utilization efficiency of data resources. It adopts differential privacy and federated learning technology to realize cross-domain training and optimization of AI models under the premise of ensuring data privacy, avoiding the risk of privacy leakage caused by data sharing.

[0064] 3. This AI model private and public domain collaborative processing system based on data security and privacy protection enhances the generalization ability of the model by using a wider range of data sets for model training in a public domain environment, enabling it to better adapt to different application scenarios and data distribution, and establishes an encrypted and verifiable data transmission channel to ensure the security and integrity of data during transmission between private and public domains. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a module structure diagram of a private domain and public domain collaborative processing system of an AI model based on data security and privacy protection in the present invention;

[0066] Figure 2 A flowchart for extracting information features and matching AI models for the present invention;

[0067] Figure 3 The figure is a flow chart of model training, verification and optimization of the present invention. DETAILED DESCRIPTION

[0068] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention are provided for the purpose of illustration and description, and are not intended to be exhaustive or to limit the present invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention, and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.

[0069] The first embodiment, as Figure 1 , Figure 2As shown, the present invention provides a technical solution: a private domain and public domain collaborative processing system of an AI model based on data security and privacy protection, including a collaborative processing platform, the collaborative processing platform is communicatively connected with a public-private domain strategy selection module, a private domain data processing module, a secure transmission channel module, a public domain model training and optimization module, a privacy protection mechanism module, and a model feedback and fusion module, wherein the modules are connected by electrical signals;

[0070] The public and private domain strategy selection module is used to extract information features based on user input and match the AI ​​model for processing, ensuring that different types of data are classified and processed before entering the processing flow, thereby improving processing efficiency and accuracy. When the user enters data, the data entered by the user is parsed through natural language processing technology to identify and extract relevant information features, including the type, format, content, and sensitivity of the data. According to the extracted information features, the decision tree classification algorithm is used to classify the data, calculate the information gain of each feature, and analyze the entropy value of the data set to determine the purity of the data set. The data is classified into different categories for subsequent processing and model matching. By calculating the information gain of each feature, the feature that can minimize entropy is selected for segmentation, and the best decision is made at each node to maximize the distinction between categories. Based on the category and information features of the data, the corresponding AI model is matched for each type of data from the model library. The matching process is based on multiple dimensions such as model performance, scope of application, and data processing capabilities. The AI ​​model that best suits the current data category is selected for subsequent processing. Before the data enters the matched AI model, it is preprocessed. The preprocessing includes data cleaning, formatting, and normalization operations. According to the sensitivity of the data, the privacy protection needs in the data processing process are evaluated. Different classification processing strategies are formulated for different types of data, including data flow, processing order, and AI models used. For highly sensitive data, it is preferred to process it in the private domain to avoid the risk of data leakage. For less sensitive data, a wider range of data sets are used in the public domain for model training and optimization. Through intelligent AI model matching and classification processing strategies, different types of data are classified and processed, and the formulated classification processing strategies are applied to the actual data flow to ensure that the data has received the most appropriate processing solution before entering the subsequent processing process.

[0071] Furthermore, the calculation expression of information gain is:

[0072] ;in, For the entire dataset The entropy of the dataset is the uncertainty or confusion of the dataset. For the dataset Belong to the A subset of categories, For the The number of samples in each category, is the number of samples in the entire data set, For the The entropy of a subset of categories, is the total number of categories in the data set. The higher the value of information gain, the more information can be obtained by using this feature to segment the data set, that is, the more useful this feature is for classification;

[0073] The calculation expression of the entropy value of the data set is:

[0074] ;in, For the data set The probability of the categories, is the total number of categories in the data set. The higher the entropy value, the greater the uncertainty of the data set. When the entropy is 0, it means that the data set is completely pure, that is, all samples belong to the same category.

[0075] The private domain data processing module is used to pre-process sensitive data in a local or private cloud environment. The pre-processing operations include data encryption, desensitization, and anonymization steps to ensure that the data meets security standards before leaving the private domain, effectively protecting the privacy and security of user data and preventing data leakage risks. The data discovery tool is used to scan the entire data set, identify data fields containing sensitive information, and classify them. According to the sensitivity of the data, the data is divided into different sensitivity levels, namely highly sensitive, medium sensitive, and low sensitive. The corresponding encryption algorithm and key length are matched according to the sensitivity of the data. Sensitive information includes personal identity information, financial information, medical records, passwords, keys, etc. Among them, highly sensitive data uses symmetric encryption algorithms (such as AES-256) or asymmetric encryption algorithms (such as RSA-4096) to ensure data security. Moderately sensitive data chooses a shorter key length, such as AES-128 or RSA-2048, but still needs to maintain sufficient Security: For low-sensitivity data, a weaker encryption algorithm or simple hash processing is selected according to specific needs. Through encryption technology and key management strategies, the identified and classified sensitive data is encrypted and converted into ciphertext that cannot be directly read to prevent unauthorized personnel from accessing or leaking data. In the data preprocessing stage, the data fields that need to be desensitized are identified, and according to the type and sensitivity of the data, corresponding desensitization technologies (such as replacement, hashing, generalization, etc.) are used to process sensitive data, remove or replace sensitive information in the data, reduce the risk of data leakage, and maintain the availability and analytical value of the data. Data fields containing personal identification information are identified and anonymization technologies (such as differential privacy, k-anonymity, l-diversity, etc.) are used to process the data to ensure that the data cannot be traced back to a specific individual while retaining its analytical value. Personal identification information in the data is removed or replaced so that the data cannot be directly associated with a specific individual, thereby protecting personal privacy.

[0076] The secure transmission channel module is used to establish an encrypted and verifiable data transmission channel, securely transmit processed data and model parameters between the private domain and the public domain, prevent data from being intercepted or tampered with during transmission, ensure data integrity and security, establish an initial data transmission connection between the private domain and the public domain, and the two parties use a pre-agreed security protocol to shake hands and confirm each other's identity and communication parameters. During the handshake process, the two parties use digital certificates to confirm each other's identity to ensure that the communicating parties are the expected entities and prevent man-in-the-middle attacks. The key exchange protocol is used to securely exchange keys between the communicating parties to ensure that both parties have the same key for encrypting and decrypting transmitted data, and the key is not obtained by a third party during the exchange process, ensuring the security of the private key and the reliability of the key exchange, and The key is replaced and updated regularly to reduce the risk of key cracking. Before sending data, the negotiated encryption algorithm and key are used to encrypt sensitive data and convert plaintext data into ciphertext. Even if the data is intercepted during transmission, it cannot be read by unauthorized parties. The sender uses an asymmetric encryption algorithm to digitally sign the data so that the receiver can confirm the source and integrity of the data by verifying the digital signature. The encrypted data is sent through the established transmission channel. During the transmission process, the data is encapsulated in the transmission protocol for transmission to ensure the integrity of the data during transmission and prevent tampering. When the public domain end receives the data, the corresponding decryption algorithm and key are used to decrypt the data, and the digital signature is used to verify the integrity of the data to ensure that the data has not been tampered with during transmission.

[0077] The public domain model training and optimization module is used to train, verify and optimize the model in the public domain environment using data and model parameters securely transmitted from the private domain, thereby improving the generalization ability and accuracy of the model. At the same time, a privacy protection mechanism is applied to ensure that the privacy information of the private domain data is not leaked during the model training process.

[0078] The privacy protection mechanism module is used to protect the privacy information of private domain data in model training in the public domain by using differential privacy and federated learning technologies. Differential privacy protects individual privacy by adding random noise to the data while ensuring the accuracy of statistical results. Federated learning allows multiple participants to train models locally and only share model parameter updates without directly sharing data.

[0079] The model feedback and fusion module is used to feed back the optimized model parameters and knowledge to the private domain, and integrate them with the original model in the private domain to form a new AI model that is more accurate and protects privacy, thereby achieving continuous optimization and improvement of the model and improving the performance and accuracy of the model.

[0080] The second embodiment is based on the first embodiment. Figure 3As shown in the figure, in the public domain model training and optimization module, the model training, verification and optimization process includes:

[0081] The public domain environment receives data and model parameters from the private domain through a secure transmission channel. The received data includes sensitive data that has been encrypted, desensitized or anonymized, as well as the initial parameters or updated parameters of the model. The corresponding decryption key is used to decrypt the data, and the encrypted data is converted into a usable plaintext format for model training. The decrypted data is preprocessed, including data cleaning, normalization, and feature selection, to improve data quality and make it more suitable for model training. The received model parameters are used to initialize the AI ​​model in the public domain. If the initial parameters of a new model are received, the model is trained from scratch. If the updated parameters of an existing model are received, the original model is fine-tuned. The preprocessed data is input into the AI ​​model for training. During the training process, an optimization algorithm is applied to minimize the loss function, update the parameters of the AI ​​model, and through multiple iterative training, the AI ​​model gradually learns the inherent laws and characteristics of the data. The performance of the model is evaluated using a validation data set independent of the training data. The performance of the model is quantified by calculating the accuracy, recall, and F1 score indicators. The model is adjusted and optimized based on the validation results to improve its generalization ability, and the hyperparameters of the AI ​​model are adjusted based on the validation results.

[0082] The process of adjusting AI model hyperparameters includes:

[0083] Before training begins, set the hyperparameter set , initially set the hyperparameter set Each hyperparameter The range of hyperparameters is clear, and each hyperparameter The range is ; Select Bayesian optimization method as the hyperparameter optimization strategy, and use Gaussian process as the core of Bayesian optimization to calculate the mean and variance functions of Gaussian process. For each set of hyperparameters , train the model and calculate the loss function of the model on the validation set To evaluate its performance, the kernel matrix of the Gaussian process is updated using the loss function results on the validation set, and the mean and variance of the Gaussian process are updated using the new observation data. The model performance under each set of hyperparameters is recorded, and the acquisition function of Bayesian optimization is used. Measure expected improvement and calculate acquisition function Select the next set of hyperparameters. The maximized hyperparameters are used as the next set of hyperparameters, and the process of model training evaluation and hyperparameter recording is repeated, with a preset number of iterations. , for each iteration , update the Gaussian process model and select new hyperparameters until the predetermined number of iterations is reached After all iterations are completed, the loss function on the validation set is selected The minimized hyperparameters are the set of hyperparameters with the best performance;

[0084] Furthermore, the expression of the Gaussian process mean function is:

[0085] ;in, Represents a given input The mean prediction at Represents the kernel function of the Gaussian process, measuring the input Similarity to oneself;

[0086] The expression of the Gaussian process variance function is:

[0087] ;in, Represents a given input The prediction variance at represents the kernel function, represents the variance of the noise, Represents the kernel function, which measures the input With the evaluated point The similarity between Represents the diagonal elements of the kernel function, measuring the evaluated points The similarity between itself and represents the number of evaluated points, that is, the number of observed data points used for prediction in the Gaussian process;

[0088] The expression of the acquisition function of Bayesian optimization is:

[0089] ;in, represents the acquisition function, which is used to measure the given input Expected improvements in represents the mean prediction of the Gaussian process, represents the standard deviation prediction of the Gaussian process, represents the hyperparameters used to control risk and exploration, The larger the value of , the more the optimization process tends to explore new input areas. The smaller the value of , the more the optimization process tends to use known information to find the best solution;

[0090] In the privacy protection mechanism module, the protection process of private domain data privacy information includes:

[0091] Integrate the preprocessed private domain data, perform sensitivity calculation during data analysis, analyze the maximum difference in function output on adjacent data sets, preset privacy budget parameters, quantify the strength of privacy protection, use the Laplace mechanism to calculate the result after adding noise, output the result after adding noise to protect the privacy of the data, and provide statistical information. The expression for sensitivity calculation is:

[0092] ; Representation function The sensitivity of is the maximum difference in the function output on adjacent data sets. and Represents adjacent data sets, differing by at most one record;

[0093] The expression for adding noise using the Laplace mechanism is:

[0094] ; represents the result after adding noise, Represents the original query result, Indicates is a Laplace distributed random variable with scale parameter, represents the privacy budget parameter, which is used to quantify the strength of privacy protection. The smaller it is, the stronger the privacy protection;

[0095] Each participant uses its own data set to train the model in its own private environment and updates the model parameters. Federated learning is implemented among multiple participants to perform local model training and parameter update sharing. After training is completed, each participant sends the model parameter update to the central server in the public domain. The central server receives the model parameter updates from each participant and uses the federated averaging algorithm to aggregate the parameters to update the global model.

[0096] The calculation expression for parameter aggregation using the federated average algorithm is:

[0097] ; Represents the global model in Wheel parameters, represents the number of participants participating in the training, Indicates Participants in Model parameters after rounds of training;

[0098] In the model feedback and fusion module, the process of forming a new AI model includes:

[0099] In the public domain environment, the trained and optimized model parameters are collected, and the model parameters in the public domain are encrypted and transmitted back to the private domain using a secure transmission channel to ensure the security and integrity of the parameters during transmission and prevent data from being intercepted or tampered with. The received model parameters are verified and preprocessed to make them suitable for use in the private domain, ensuring the compatibility of the parameters with the private domain model, and preparing for model fusion. The preprocessed public domain model parameters are fused with the private domain model parameters to form a new model. The new model combines the specific knowledge of private domain data with the general rules of public domain data. If the private domain model has not been trained before, the public domain model is used. The public domain parameters are used as initialization. If the private domain model has been trained, the public domain parameters are combined with the private domain parameters, and the performance of the fused new model is evaluated on the private domain validation set to verify the effect of model fusion and ensure that the new model has good performance on private domain data. According to the verification results, the new model is fine-tuned and the hyperparameters are adjusted to further improve the performance and accuracy of the model. The verified and adjusted new model is deployed to the production environment to improve the intelligence level of the business process. In actual applications, the model performance is continuously monitored and feedback is collected. According to business needs and model performance, the model is regularly updated to form a closed loop of continuous optimization.

[0100] Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field and related fields without creative work should fall within the scope of protection of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention are implemented according to the conventional means in the field unless otherwise specified and limited.

Claims

1. A private domain and public domain collaborative processing system of AI models based on data security and privacy protection, including a collaborative processing platform, characterized in that: The collaborative processing platform is communicatively connected with a public-private domain strategy selection module, a private domain data processing module, a secure transmission channel module, a public domain model training and optimization module, a privacy protection mechanism module, and a model feedback and fusion module, wherein the modules are connected by electrical signals; The public-private domain strategy selection module is used to extract information features input by the user and match the AI ​​model for processing. In the public-private domain strategy selection module, the process of extracting information features and matching the AI ​​model includes: When users input data, the data is parsed through natural language processing technology to identify and extract relevant information features, including the type, format, content, and sensitivity of the data; According to the extracted information features, the decision tree classification algorithm is used to classify the data, calculate the information gain of each feature, and analyze the entropy value of the data set to determine the purity of the data set and classify the data into different categories; Based on the category and information characteristics of the data, the corresponding AI model is matched for each type of data from the model library. The matching process is based on multiple dimensions such as model performance, scope of application, and data processing capabilities. Before the data enters the matching AI model, it is preprocessed. The preprocessing includes data cleaning, formatting, and normalization operations. According to the sensitivity of the data, the privacy protection requirements during the data processing are evaluated. Develop different classification and processing strategies for different types of data, including data flow, processing order, and AI models used; Through intelligent AI model matching and classification processing strategies, different types of data are classified and processed, and the formulated classification processing strategies are applied to actual data flows; The private domain data processing module is used to pre-process sensitive data in a local or private cloud environment, and the pre-processing operation includes data encryption, desensitization, and anonymization processing steps; The secure transmission channel module is used to establish an encrypted and verifiable data transmission channel; The public domain model training and optimization module is used to perform model training, verification and optimization in a public domain environment using data and model parameters securely transmitted from a private domain; The privacy protection mechanism module is used to protect the privacy information of private domain data by using differential privacy and federated learning technology to train models in the public domain; The model feedback and fusion module is used to feed back the optimized model parameters and knowledge to the private domain, and fuse them with the original model in the private domain to form a new AI model.

2. According to claim 1, a private domain and public domain collaborative processing system of AI models based on data security and privacy protection is characterized by: The calculation expression of the information gain is: ;in, For the entire dataset The entropy of For the dataset Belong to the A subset of categories, For the The number of samples in each category, is the number of samples in the entire data set, For the The entropy of a subset of categories, is the total number of categories in the dataset; The calculation expression of the entropy value of the data set is: ;in, For the data set The probability of the categories, is the total number of categories in the dataset.

3. According to claim 2, a private domain and public domain collaborative processing system of AI models based on data security and privacy protection is characterized by: In the private domain data processing module, the process of preprocessing sensitive data includes: Use data discovery tools to scan the entire data set, identify data fields containing sensitive information, and classify them. According to the sensitivity of the data, the data is divided into different sensitivity levels, namely high sensitivity level, medium sensitivity level, and low sensitivity level, and the corresponding encryption algorithm and key length are matched according to the sensitivity of the data; Through encryption technology, combined with key management strategies, the identified and classified sensitive data is encrypted and converted into a ciphertext form that cannot be directly read; In the data preprocessing stage, the data fields that need to be desensitized are identified, and according to the type and sensitivity of the data, corresponding desensitization technology is used to process the sensitive data and remove or replace the sensitive information in the data; Identify data fields that contain personal identification information, process the data using anonymization technology, and remove or replace personal identification information in the data.

4. According to claim 3, a private domain and public domain collaborative processing system of AI models based on data security and privacy protection is characterized by: In the secure transmission channel module, the process of establishing the data transmission channel includes: An initial data transmission connection is established between the private domain and the public domain. Both parties use a pre-agreed security protocol to perform a handshake and confirm each other's identity and communication parameters. Use key exchange protocols to securely exchange keys between communicating parties, and regularly replace and update keys; Before data is sent, the negotiated encryption algorithm and key are used to encrypt sensitive data, converting plaintext data into ciphertext, and the sender uses an asymmetric encryption algorithm to digitally sign the data so that the receiver can verify the source and integrity of the data by verifying the digital signature; The encrypted data is sent through the established transmission channel. During the transmission process, the data is encapsulated in the transmission protocol for transmission; When the public domain end receives the data, it uses the corresponding decryption algorithm and key to decrypt the data, and uses the digital signature to verify the integrity of the data.

5. According to claim 4, a private domain and public domain collaborative processing system of AI models based on data security and privacy protection is characterized in that: In the public domain model training and optimization module, the model training, verification and optimization process includes: The public domain environment receives data and model parameters from the private domain through a secure transmission channel; Decrypt the data using the corresponding decryption key, convert the encrypted data into a usable plaintext format, and preprocess the decrypted data, including data cleaning, normalization, and feature selection; Initialize the AI ​​model in the public domain using the received model parameters, input the preprocessed data into the AI ​​model for training, and during the training process, apply the optimization algorithm to minimize the loss function, update the parameters of the AI ​​model, and through multiple iterations of training, allow the AI ​​model to gradually learn the inherent laws and characteristics of the data; Use a validation dataset independent of the training data to evaluate the performance of the model, adjust and optimize the model based on the validation results, and adjust the hyperparameters of the AI ​​model based on the validation results.

6. The AI ​​model private domain and public domain collaborative processing system based on data security and privacy protection according to claim 5 is characterized by: The process of adjusting the AI ​​model hyperparameters includes: Before training begins, set the hyperparameter set , initially set the hyperparameter set Each hyperparameter The range of hyperparameters is clear, and each hyperparameter The range is ; Select Bayesian optimization as the hyperparameter optimization strategy, and use Gaussian process as the core of Bayesian optimization to calculate the mean and variance functions of Gaussian process; For each set of hyperparameters , train the model and calculate the loss function of the model on the validation set to evaluate its performance; Use the loss function results on the validation set to update the kernel matrix of the Gaussian process, and use the new observation data to update the mean and variance of the Gaussian process, record the model performance under each set of hyperparameters, and use the Bayesian optimization acquisition function Measure expected improvement and calculate acquisition function Select the next set of hyperparameters. The maximized hyperparameters are used as the next set of hyperparameters; Repeat the process of model training evaluation and hyperparameter recording, with a preset number of iterations , for each iteration , update the Gaussian process model and select new hyperparameters until the predetermined number of iterations is reached After all iterations are completed, the loss function on the validation set is selected The minimized hyperparameters are the set of hyperparameters with the best performance.

7. The AI ​​model private domain and public domain collaborative processing system based on data security and privacy protection according to claim 6 is characterized by: The expression of the Gaussian process mean function is: ;in, Represents a given input The mean prediction at Represents the kernel function of the Gaussian process, measuring the input Similarity to oneself; The expression of the Gaussian process variance function is: ;in, Represents a given input The prediction variance at represents the kernel function, represents the variance of the noise, Represents the kernel function, which measures the input With the evaluated point The similarity between Represents the diagonal elements of the kernel function, measuring the evaluated points The similarity between itself and represents the number of evaluated points, that is, the number of observed data points used for prediction in the Gaussian process; The expression of the acquisition function of the Bayesian optimization is: ;in, represents the acquisition function, which is used to measure the given input Expected improvements in represents the mean prediction of the Gaussian process, represents the standard deviation prediction of the Gaussian process, represents the hyperparameters used to control risk and exploration.

8. The AI ​​model private domain and public domain collaborative processing system based on data security and privacy protection according to claim 7 is characterized by: In the privacy protection mechanism module, the protection process of private domain data privacy information includes: Integrate the pre-processed private domain data, perform sensitivity calculations during data analysis, analyze the maximum difference in function outputs on adjacent data sets, preset privacy budget parameters, quantify the strength of privacy protection, use the Laplace mechanism to calculate the results after adding noise, output the results after adding noise to protect the privacy of the data, and provide statistical information; Each participant trains the model using his or her own dataset in his or her own private environment and updates the model parameters. Implement federated learning among multiple participants to conduct local model training and parameter update sharing. After training is completed, each participant sends the model parameter update to a central server in the public domain. The central server receives model parameter updates from each participant and uses the federated averaging algorithm to aggregate parameters to update the global model.

9. The AI ​​model private domain and public domain collaborative processing system based on data security and privacy protection according to claim 8, characterized in that: In the model feedback and fusion module, the process of forming a new AI model includes: In the public domain environment, the trained and optimized model parameters are collected, and the model parameters in the public domain are encrypted and transmitted back to the private domain using a secure transmission channel; Verify and preprocess the received model parameters to make them suitable for private domain use; The preprocessed public domain model parameters are merged with the private domain model parameters to form a new model. The new model combines the specific knowledge of private domain data with the general rules of public domain data. If the private domain model has not been trained before, the public domain parameters are used as initialization. If the private domain model has been trained, the public domain parameters are combined with the private domain parameters. Evaluate the performance of the new fused model on the private domain validation set to verify the effect of model fusion, and fine-tune the new model and adjust the hyperparameters based on the verification results; Deploy the verified and adjusted new model to the production environment, and continuously monitor the model performance in actual applications, collect feedback, and regularly update the model based on business needs and model performance to form a closed loop of continuous optimization.

Images